diff --git a/.claude/skills/feature-planning/SKILL.md b/.claude/skills/feature-planning/SKILL.md
new file mode 100644
index 000000000..110ef7256
--- /dev/null
+++ b/.claude/skills/feature-planning/SKILL.md
@@ -0,0 +1,337 @@
+---
+name: feature-planning
+description: >
+  Deep feature research and implementation planning for AI coding agent projects. Use this skill
+  whenever a user asks about a feature they want to implement, improve, add, or design — especially
+  in the context of AI coding agents, CLI tools, terminal agents, or LLM-powered developer tools.
+  Triggers on: "I want to add X feature", "how do I implement X", "can we improve X", "I want to
+  build X into my agent", "feature request for X", "how does X work in these tools", or any phrasing
+  that implies implementing/improving a capability. This skill clones 7 reference repos, spawns
+  sub-agents for deep per-repo research, runs an ultra-QA interview with the user, then produces a
+  comprehensive implementation plan with code, pseudocode, test cases, benchmarks, and direct repo
+  links — so the user can go from idea to working implementation with total confidence.
+---
+
+# Feature Planning Skill
+
+Comprehensive feature research + implementation planning using 9 reference repos as the knowledge base.
+
+## Reference Repositories
+
+| Alias | Repo URL | Stack | What it teaches |
+|-------|----------|-------|-----------------|
+| `oh-my-openagent` | https://github.com/code-yeongyu/oh-my-openagent | TypeScript / OpenCode plugin | Multi-agent orchestration, model routing, tmux sessions, delegate-task patterns |
+| `opencode` | https://github.com/anomalyco/opencode | TypeScript / Bun monorepo | Open-source AI coding agent architecture, provider abstraction, TUI |
+| `oh-my-pi` | https://github.com/can1357/oh-my-pi | TypeScript + Rust / Bun | 40+ providers, 32 tools, LSP+DAP ops, benchmarked edits, IDE wiring |
+| `codebuff` | https://github.com/CodebuffAI/codebuff | TypeScript / multi-agent | File picker + planner + editor + reviewer pipeline, beats Claude Code on evals |
+| `codex` | https://github.com/openai/codex | TypeScript / Node | OpenAI Codex CLI, sandboxed execution, hardened tool use |
+| `claude-code` | https://github.com/claude-code-best/claude-code | TypeScript / Bun | CCB — decompiled Claude Code with Pipe IPC, ACP, remote control, monitoring |
+| `pi-agent-rust` | https://github.com/Dicklesworthstone/pi_agent_rust | Rust 2024 edition | High-perf Rust agent, SQLite sessions, SSE streaming, WASM extension security |
+| `oh-my-claudecode` | https://github.com/Yeachan-Heo/oh-my-claudecode | TypeScript / Claude Code plugin | Claude Code extension with hooks, guards, permission modes, multi-agent tools |
+| `oh-my-codex` | https://github.com/Yeachan-Heo/oh-my-codex | TypeScript / Codex plugin | Codex extension with approval modes, sandbox config, tool gating |
+
+---
+
+## Workflow (follow this order every time)
+
+### Phase 1 — Clone & Sub-agent Research
+
+When the skill is triggered, immediately clone all 9 repos (shallow `--depth=1`) and spawn one research sub-agent per repo. Each sub-agent gets the full repo and the feature request — its job is to autonomously explore **the entire repo** to find everything relevant. The sub-agent decides what to read; nothing is off-limits and nothing is assumed to be the right place to look.
+
+Each sub-agent should:
+
+1. **Map the repo first** — list all files and directories to understand the full shape before diving in. No assumptions about where things live.
+2. **Follow the feature signal** — search for keywords, types, patterns, and concepts related to the requested feature across every file, every directory, every language. If a Rust file has relevant logic, read it. If a config YAML has relevant keys, read it. If a test file shows how a concept is used, read it. If a benchmark shows performance constraints, read it.
+3. **Trace implementations end-to-end** — when a relevant function/type/module is found, follow its call chain in both directions (callers and callees) until the full picture is clear. Don't stop at the first hit.
+4. **Extract everything useful** — architecture patterns, API surfaces, data structures, config hooks, test patterns, benchmark approaches, error handling strategies, extension points, anything that could inform the feature design.
+5. **Return a structured summary** (see **Sub-agent Report Format** below)
+
+The sub-agent must NOT limit itself to any predefined set of files or folders. If it finds something unexpected in an unusual location, it should read it. Thoroughness is the goal.
+
+Run sub-agents in parallel. Collect all 9 reports before continuing.
+
+```bash
+# Clone command template
+for repo in \
+  "https://github.com/code-yeongyu/oh-my-openagent" \
+  "https://github.com/anomalyco/opencode" \
+  "https://github.com/can1357/oh-my-pi" \
+  "https://github.com/CodebuffAI/codebuff" \
+  "https://github.com/openai/codex" \
+  "https://github.com/claude-code-best/claude-code" \
+  "https://github.com/Dicklesworthstone/pi_agent_rust" \
+  "https://github.com/Yeachan-Heo/oh-my-claudecode" \
+  "https://github.com/Yeachan-Heo/oh-my-codex"; do
+  git clone --depth=1 "$repo" /tmp/feature-research/$(basename $repo)
+done
+```
+
+#### Sub-agent Report Format
+
+Each sub-agent returns a structured block:
+
+```
+## [repo-name] Research Report
+
+### Relevance Score: [HIGH / MEDIUM / LOW / NONE]
+### Why relevant: [1-2 sentences]
+
+### Key Files
+- path/to/file.ts — [what it does re: the feature]
+
+### Relevant Code Snippets
+[short excerpts with file:line references]
+
+### Architecture Pattern
+[how this repo approaches the feature domain]
+
+### Direct Links
+- https://github.com/[org]/[repo]/blob/main/[file]#L[line]
+
+### Gaps / What's Missing
+[what this repo doesn't cover that the user might need]
+```
+
+---
+
+### Phase 2 — Present Per-Repo Report to User
+
+After collecting sub-agent reports, present a consolidated **Research Report** to the user with one section per repo. Format:
+
+```
+# Feature Research: [FEATURE NAME]
+
+## Summary
+[2-3 sentence overview of what you found across all repos]
+
+---
+
+## 1. oh-my-openagent
+[sub-agent report content]
+
+## 2. opencode
+...
+
+## 7. pi-agent-rust
+...
+
+---
+
+## Cross-Repo Patterns
+[What approaches are consistent across repos — these are proven patterns]
+
+## Unique Insights
+[Interesting divergences or novel approaches from individual repos]
+```
+
+---
+
+### Phase 3 — Ultra QA Interview
+
+After presenting the research report, enter a deep QA loop with the user. Ask questions in rounds — never dump all questions at once. Use this question bank, picking the most relevant ones for the feature at hand:
+
+**Round 1 — Scope & Goal**
+- What is the exact outcome you want after implementing this? (demo it to me in words)
+- Is this a new feature or improving an existing one? If existing, what's broken/missing?
+- Which repo(s) are you building in / most inspired by?
+- What stack? (TypeScript, Rust, Python, other)
+
+**Round 2 — Constraints & Context**
+- What existing code does this feature touch or depend on?
+- Are there performance requirements? (latency targets, memory limits, throughput)
+- Security constraints? (sandboxing, capability gating, trust levels)
+- Will this need to work across multiple LLM providers or just one?
+
+**Round 3 — Design Preferences**
+- Do you prefer a plugin/extension architecture or embedded implementation?
+- Should this be synchronous, async, or streaming?
+- How should failures be handled? (silent fallback, hard error, user prompt)
+- How will users configure or toggle this feature?
+
+**Round 4 — Testing & Quality**
+- What does a successful implementation look like? How will you verify it?
+- Are there existing tests in the repos we can adapt?
+- Any edge cases you're already worried about?
+
+**Round 5 — Stretch Goals**
+- What would a "10x better" version of this look like?
+- Are there benchmark targets you want to hit?
+- Future integrations you want to leave room for?
+
+Keep asking follow-up questions until you have clear answers to at minimum Round 1 and Round 2. Rounds 3–5 can be inferred from research if the user is in a hurry.
+
+---
+
+### Phase 4 — Comprehensive Implementation Plan
+
+After the QA interview, produce the final plan. This is the deliverable the user keeps. It must include ALL of the following sections:
+
+---
+
+```markdown
+# Implementation Plan: [FEATURE NAME]
+> Generated from research across 9 repos + user interview
+> Goal: [User's stated goal in 1 sentence]
+
+---
+
+## 1. Executive Summary
+[3-5 sentences: what we're building, why this approach, expected outcome]
+
+---
+
+## 2. Architecture Decision
+### Chosen Approach
+[Which pattern from the research repos we're following, and why]
+
+### Alternatives Considered
+| Approach | Source Repo | Pros | Cons | Decision |
+|----------|-------------|------|------|----------|
+
+---
+
+## 3. Data Structures & Types
+
+```typescript  // or Rust, Python, etc.
+// Core types for the feature
+interface FeatureConfig {
+  // ...
+}
+```
+
+---
+
+## 4. Pseudocode — Core Algorithm
+
+```
+FUNCTION implementFeature(input):
+  // Step-by-step logic in plain pseudocode
+  // No language-specific syntax
+  // Shows all branches and edge cases
+```
+
+---
+
+## 5. Implementation Code
+
+### File: [path/to/new-or-modified-file]
+```typescript
+// Full implementation code
+// With inline comments explaining non-obvious choices
+// References to source repos where patterns were borrowed
+```
+
+### File: [path/to/another-file]
+```typescript
+// ...
+```
+
+---
+
+## 6. Configuration & Wiring
+[How to register/hook the feature into the existing system]
+[Config file changes, env vars, flags]
+
+---
+
+## 7. Repo References
+
+Direct links to the most relevant code in each source repo:
+
+| Feature Aspect | Repo | File | Link |
+|----------------|------|------|------|
+| [aspect] | oh-my-openagent | src/agents/... | https://github.com/... |
+| [aspect] | codebuff | packages/... | https://github.com/... |
+| ... | | | |
+
+---
+
+## 8. Test Cases
+
+### Happy Path Tests
+```typescript
+describe('[feature]', () => {
+  it('should [happy case 1]', async () => {
+    // setup
+    // act
+    // assert
+  });
+
+  it('should [happy case 2]', async () => {
+    // ...
+  });
+});
+```
+
+### Edge Cases
+```typescript
+  it('should handle [edge case: empty input]', ...);
+  it('should handle [edge case: provider failure]', ...);
+  it('should handle [edge case: concurrent calls]', ...);
+  it('should handle [edge case: large payload]', ...);
+  it('should handle [edge case: timeout]', ...);
+```
+
+### Integration Tests
+```typescript
+// End-to-end test that exercises the full flow
+```
+
+---
+
+## 9. Benchmarks
+
+### What to Measure
+| Metric | Baseline | Target | How to Measure |
+|--------|----------|--------|----------------|
+| Latency (p50) | - | [Xms] | [method] |
+| Latency (p99) | - | [Xms] | [method] |
+| Memory delta | - | [XMB] | [method] |
+| Throughput | - | [X/s] | [method] |
+
+### Benchmark Code
+```typescript
+// Benchmark harness adapted from oh-my-pi / pi-agent-rust patterns
+```
+
+---
+
+## 10. Migration / Rollout
+[If improving existing feature: how to migrate without breaking changes]
+[Feature flags, gradual rollout, deprecation path]
+
+---
+
+## 11. Known Limitations & Future Work
+- [ ] [Thing not covered in this plan]
+- [ ] [Stretch goal for v2]
+- [ ] [Integration left for later]
+
+---
+
+## 12. Success Criteria Checklist
+- [ ] Core happy path works end-to-end
+- [ ] All edge case tests pass
+- [ ] Performance meets targets from Section 9
+- [ ] No regressions in existing tests
+- [ ] [User's specific success criterion from interview]
+```
+
+---
+
+## Quality Standards
+
+The plan must meet these bars before presenting to the user:
+
+- **No broken links** — all GitHub links must point to real files in the cloned repos
+- **No vague pseudocode** — every step in the pseudocode must be implementable
+- **No placeholder tests** — every test case must have real setup/act/assert
+- **Benchmark section is never empty** — even if targets are TBD, the measurement method must be specified
+- **Every architectural choice has a "why"** referencing a source repo
+- **The user should be able to hand this plan to a junior engineer and get working code back**
+
+---
+
+## References
+
+See `references/repo-summaries.md` for static summaries of all 9 repos (useful when cloning is slow or unavailable).
\ No newline at end of file
diff --git a/.claude/skills/feature-planning/references/repo-summaries.md b/.claude/skills/feature-planning/references/repo-summaries.md
new file mode 100644
index 000000000..79779be1d
--- /dev/null
+++ b/.claude/skills/feature-planning/references/repo-summaries.md
@@ -0,0 +1,177 @@
+# Reference Repo Summaries
+
+Static summaries of all 7 repos for quick lookup without cloning.
+
+---
+
+## 1. oh-my-openagent
+**URL:** https://github.com/code-yeongyu/oh-my-openagent  
+**Stack:** TypeScript, Bun, OpenCode plugin  
+**What it is:** A powerful OpenCode plugin that adds named agents (Prometheus, Atlas, Hephaestus, Sisyphus-Junior) with model-variant routing, tmux session management, and multi-agent delegation via `delegate-task`.
+
+**Key Patterns:**
+- **Agent factory pattern**: `createAtlasAgent()`, `createHephaestusAgent()` — each agent is a factory with model-variant routing
+- **Prompt variants per model**: `default.md` (Claude), `gpt.md`, `gemini.md`, `kimi.md` — same agent, different prompt per provider
+- **Delegate-task orchestration**: Atlas spawns Sisyphus-Junior subagents via `task()` calls; never self-reports, always verifies
+- **Model resolution pipeline**: `resolveModel(input)` → UI override → agent-specific → fallback chain
+- **Tmux integration**: `createTmuxSession()`, `spawnPane()` for multi-pane agent workflows
+- **Session management**: `SessionCursor`, `trackInjectedPath()`, `SessionToolsStore`
+- **Config migration**: `migrateConfigFile()` with `AGENT_NAME_MAP`, `HOOK_NAME_MAP`, `MODEL_VERSION_MAP`
+
+**Key Files:**
+- `src/agents/atlas/agent.ts` — orchestrator agent factory
+- `src/agents/prometheus/system-prompt.ts` — strategic planner prompt loader
+- `src/agents/hephaestus/agent.ts` — autonomous deep worker
+- `src/agents/sisyphus-junior/agent.ts` — category-spawned executor
+- `src/shared/index.ts` — barrel export of 297 utility files
+- `src/shared/model-availability.ts` — `resolveModel()`, `checkModelAvailability()`
+- `packages/prompts-core/` — model-neutral prompt markdown files
+
+---
+
+## 2. opencode
+**URL:** https://github.com/anomalyco/opencode  
+**Stack:** TypeScript, Bun, SST, monorepo (Turbo)  
+**What it is:** The open source AI coding agent. Terminal UI with provider abstraction, extension system, desktop app, and a well-structured monorepo.
+
+**Key Patterns:**
+- **Provider abstraction**: Clean separation between LLM provider and agent logic
+- **Monorepo layout**: `packages/` with `tui/`, `desktop/`, `web/`, `identity/`
+- **SST for infra**: Config-as-code for cloud deployment
+- **Desktop + TUI**: Supports both Electron-style desktop and pure terminal modes
+- **Zed extension**: `packages/extensions/zed/` for IDE integration
+
+**Key Files:**
+- `packages/tui/` — terminal UI implementation
+- `packages/desktop/` — Electron-style desktop wrapper
+- `sst.config.ts` — infrastructure config
+- `turbo.json` — monorepo build pipeline
+
+---
+
+## 3. oh-my-pi
+**URL:** https://github.com/can1357/oh-my-pi  
+**Stack:** TypeScript + Rust, Bun ≥ 1.3.14  
+**What it is:** Fork of Pi by @mariozechner. "A coding agent with the IDE wired in." 40+ providers, 32 built-in tools, 13 LSP ops, 27 DAP ops, ~27k lines of Rust core.
+
+**Key Patterns:**
+- **Benchmarked tool use**: Every tool is tuned for first-attempt success rate; `packages/typescript-edit-benchmark/` has full benchmark harness
+- **LSP integration**: 13 language server protocol operations built in
+- **DAP integration**: 27 debug adapter protocol operations built in
+- **Multi-provider**: 40+ providers with provider-neutral abstraction
+- **Rust core + TS surface**: Performance-critical code in Rust, developer-facing API in TypeScript
+- **Mutation testing**: `src/mutations.ts` for benchmark task generation
+
+**Key Files:**
+- `packages/typescript-edit-benchmark/src/` — full benchmark framework
+- `packages/typescript-edit-benchmark/src/tasks.ts` — benchmark task definitions
+- `packages/typescript-edit-benchmark/src/runner.ts` — benchmark runner
+- `packages/typescript-edit-benchmark/src/prompts/` — benchmark prompt templates
+
+---
+
+## 4. codebuff
+**URL:** https://github.com/CodebuffAI/codebuff  
+**Stack:** TypeScript, multi-agent pipeline  
+**What it is:** AI coding assistant that coordinates specialized agents. Beats Claude Code 61% vs 53% on 175+ coding tasks. Has a `freebuff` free tier.
+
+**Key Patterns:**
+- **4-agent pipeline**: File Picker → Planner → Editor → Reviewer — each is a specialized agent
+- **Tree-sitter code map**: `packages/code-map/` uses tree-sitter for language-aware code parsing across 10+ languages
+- **Agent composition**: Multi-agent as a *strategy*, not just concurrency — each agent has a specific role
+- **Custom agent builder**: `/init` command generates agent scaffolding
+- **Eval-driven development**: `evals/` directory with 175+ tasks across real open-source repos
+
+**Key Files:**
+- `packages/code-map/src/index.ts` — code map entry point
+- `packages/code-map/src/languages.ts` — language detection
+- `packages/code-map/src/tree-sitter-queries/` — per-language AST queries
+- `evals/README.md` — eval methodology
+
+---
+
+## 5. codex (OpenAI Codex CLI)
+**URL:** https://github.com/openai/codex  
+**Stack:** TypeScript, Node.js  
+**What it is:** OpenAI's official local coding agent CLI. Single binary, sandboxed execution, ChatGPT plan integration.
+
+**Key Patterns:**
+- **Sandbox-first execution**: All tool use is sandboxed; firewall init script at `scripts/init_firewall.sh`
+- **Container execution**: `run_in_container.sh` for isolated runs
+- **Hardened tool use**: Security-first design, execution policy, network isolation
+- **Multiple install paths**: npm, Homebrew, binary releases — portable distribution
+- **Bazel build**: `BUILD.bazel`, `MODULE.bazel` for reproducible builds
+
+**Key Files:**
+- `codex-cli/bin/codex.js` — CLI entry point
+- `codex-cli/scripts/init_firewall.sh` — firewall/sandbox setup
+- `codex-cli/scripts/run_in_container.sh` — container execution
+- `codex-cli/package.json` — deps and scripts
+
+---
+
+## 6. claude-code (CCB — Claude Code Best)
+**URL:** https://github.com/claude-code-best/claude-code  
+**Stack:** TypeScript, Bun  
+**What it is:** Decompiled/reconstructed Claude Code (CCB = 踩踩背) with many enterprise features: Pipe IPC multi-instance, ACP protocol (Zed/Cursor IDE), Remote Control Docker deployment, Langfuse monitoring, Web Search, Computer Use, Chrome Use, Voice Mode, Sentry, GrowthBook.
+
+**Key Patterns:**
+- **Pipe IPC**: `main/sub` auto-orchestration + LAN cross-machine zero-config discovery; `/pipes` panel + `Shift+↓` + message broadcast routing
+- **ACP Protocol**: Session resume, Skills, permission bridging for Zed/Cursor
+- **Remote Control**: Docker self-hosted remote UI — watch Claude Code from your phone
+- **Langfuse monitoring**: Every agent loop step is observable and can be converted to datasets
+- **Feature flags**: GrowthBook integration for enterprise feature gating
+- **Memory management**: `/dream` command for memory consolidation
+- **Poor Mode**: Disable memory extraction + typing suggestions to reduce concurrent requests
+
+**Key Files:**
+- `src/types/message.ts` — message types
+- `src/types/tools.ts` — tool type definitions
+- `src/types/plugin.ts` — plugin system types
+- `src/types/hooks.ts` — hook system
+
+---
+
+## 7. pi-agent-rust
+**URL:** https://github.com/Dicklesworthstone/pi_agent_rust  
+**Stack:** Rust 2024 edition, `asupersync` async runtime, `rich_rust` TUI  
+**What it is:** High-performance Rust port of Pi Agent by Jeffrey Emanuel. Single binary, <100ms startup, <50MB idle memory, SQLite sessions, WASM extension security, io_uring fast lane.
+
+**Key Patterns:**
+- **SQLite session store**: `src/session_sqlite.rs` — segmented log + offset index, O(index+tail) reopen on large histories
+- **Hostcall security model**: Capability-gated hostcalls: `tool`/`exec`/`http`/`session`/`ui`/`events`; two-stage exec guard; trust lifecycle `pending→acknowledged→trusted→killed`
+- **io_uring fast lane**: `src/hostcall_io_uring_lane.rs` — deterministic dispatch, typed opcodes, bounded shard queues
+- **WASM extension runtime**: `src/pi_wasm.rs` — startup prewarm, warm isolate reuse, DCG/heredoc AST signals for dangerous shell detection
+- **SSE streaming parser**: Tracks scanned bytes, handles UTF-8 tails, normalizes chunk boundaries, interns event-type strings
+- **Multi-provider**: `src/providers/` — Anthropic, OpenAI, Vertex, Azure, Cohere, GitLab, Copilot
+- **Benchmarks**: `benches/` — tools, semantic context, session save, TUI perf, extensions
+- **Shadow dual execution**: Automatic backoff on divergence; compatibility-lane kill switches
+
+**Key Files:**
+- `src/session_sqlite.rs` — session persistence
+- `src/agent_cx.rs` — agent execution context
+- `src/extension_dispatcher.rs` — WASM extension dispatch
+- `src/hostcall_io_uring_lane.rs` — fast-path hostcall routing
+- `src/providers/anthropic.rs` — Anthropic provider
+- `src/pi_wasm.rs` — WASM runtime
+- `benches/` — full benchmark suite
+- `Cargo.toml` — deps: `asupersync`, `rich_rust`, edition 2024
+
+---
+
+## Cross-Repo Quick Reference
+
+| Feature Domain | Best Source Repo(s) |
+|----------------|---------------------|
+| Multi-agent orchestration | oh-my-openagent (Atlas/delegate-task), codebuff (4-agent pipeline) |
+| Model/provider abstraction | oh-my-openagent (resolveModel), oh-my-pi (40+ providers), pi-agent-rust (src/providers/) |
+| Session persistence | pi-agent-rust (SQLite), claude-code (memory/dream) |
+| Security & sandboxing | codex (firewall), pi-agent-rust (capability gates, trust lifecycle) |
+| Benchmarking | oh-my-pi (typescript-edit-benchmark), pi-agent-rust (benches/) |
+| IDE integration | oh-my-pi (LSP/DAP), claude-code (ACP/Zed/Cursor) |
+| Streaming | pi-agent-rust (SSE parser), opencode (provider abstraction) |
+| Extension/plugin system | pi-agent-rust (WASM), oh-my-openagent (OpenCode plugin), claude-code (ACP) |
+| Monitoring/observability | claude-code (Langfuse, Sentry), pi-agent-rust (runtime risk ledger) |
+| TUI design | opencode (terminal UI), pi-agent-rust (rich_rust), oh-my-pi (IDE-wired) |
+| Code understanding | codebuff (tree-sitter code map), oh-my-openagent (ripgrep-cli) |
+| Prompt engineering | oh-my-openagent (per-model prompt variants), oh-my-pi (benchmark prompts) |
\ No newline at end of file
diff --git a/Cargo.lock b/Cargo.lock
index 85b48b5fb..84caf69c0 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -88,7 +88,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ee91c0c2905bae44f84bfa4e044536541df26b7703fd0888deeb9060fcc44289"
 dependencies = [
  "android-properties",
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "cc",
  "cesu8",
  "jni 0.21.1",
@@ -185,6 +185,15 @@ version = "1.0.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "170433209e817da6aae2c51aa0dd443009a613425dd041ebfb2492d1c4c11a25"
 
+[[package]]
+name = "approx"
+version = "0.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cab112f0a86d568ea0e627cc1d6be74a1e9cd55214684db5561995f6dad897c6"
+dependencies = [
+ "num-traits",
+]
+
 [[package]]
 name = "arboard"
 version = "3.6.1"
@@ -321,9 +330,9 @@ checksum = "f2032f911046de80f0a198e0901378627c33f59ea0ac00e363d481118bd70a53"
 
 [[package]]
 name = "aws-config"
-version = "1.8.17"
+version = "1.8.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "517aa062d8bd9015ee23d6daa5e1c1372328412fdae4e6c4c1be9b69c6ad37a2"
+checksum = "e33f815b73a3899c03b380d543532e5865f230dce9678d108dc10732a8682275"
 dependencies = [
  "aws-credential-types",
  "aws-runtime",
@@ -412,9 +421,9 @@ dependencies = [
 
 [[package]]
 name = "aws-sdk-bedrock"
-version = "1.144.0"
+version = "1.145.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b683a930642668b42b19acb7d26d60400252e950dcd1e13d506748a53f1336b6"
+checksum = "bf0517c31b708b01136276121818c06d9b2d34399641ddda055569c55b03e6ef"
 dependencies = [
  "arc-swap",
  "aws-credential-types",
@@ -437,9 +446,9 @@ dependencies = [
 
 [[package]]
 name = "aws-sdk-bedrockruntime"
-version = "1.132.0"
+version = "1.133.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "41a2940faeb61f4f579a434bc3a546e9ab49a89596e94527d329281ef55fd44d"
+checksum = "76314880945928a4ee3956e92af451ef29ef04b734d008bb94b11158a60a1034"
 dependencies = [
  "arc-swap",
  "aws-credential-types",
@@ -465,10 +474,11 @@ dependencies = [
 
 [[package]]
 name = "aws-sdk-sso"
-version = "1.100.0"
+version = "1.101.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bee2719d4a5e5e147bb9e9b77490df6ece750df1094968aa857b09b618a1881a"
+checksum = "b647baea49ff551960b904f905681e9b4765a6c4ea08631e89dc52d8bd3f5896"
 dependencies = [
+ "arc-swap",
  "aws-credential-types",
  "aws-runtime",
  "aws-smithy-async",
@@ -489,9 +499,9 @@ dependencies = [
 
 [[package]]
 name = "aws-sdk-ssooidc"
-version = "1.102.0"
+version = "1.103.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b30d254992d56ef19f430396e5765b11e0f5bd21a7a557cb12fca1c8c18b9636"
+checksum = "7ae401c65ff288aa7873117fe535cd32b7b1bb0bc43751d28901a1d5f20636b9"
 dependencies = [
  "arc-swap",
  "aws-credential-types",
@@ -514,10 +524,11 @@ dependencies = [
 
 [[package]]
 name = "aws-sdk-sts"
-version = "1.105.0"
+version = "1.106.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "59f4f8065fe615dbed9096458ba98dda6d641553ffd5aedd27e37e65211aca9f"
+checksum = "4c80de7bb7d03e9ca8c9fd7b489f20f3948d3f3be91a7953591347d238115408"
 dependencies = [
+ "arc-swap",
  "aws-credential-types",
  "aws-runtime",
  "aws-smithy-async",
@@ -539,9 +550,9 @@ dependencies = [
 
 [[package]]
 name = "aws-sigv4"
-version = "1.4.4"
+version = "1.4.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7083fb918b38474ac65ffbf8a69fc8792d36879f4ac5f1667b43aec61efe9a5"
+checksum = "bae38512beae0ffee7010fc24e7a8a123c53efdfef42a61e80fda4882418dc71"
 dependencies = [
  "aws-credential-types",
  "aws-smithy-eventstream",
@@ -923,9 +934,9 @@ checksum = "bef38d45163c2f1dde094a7dfd33ccf595c92905c8f8f4fdc18d06fb1037718a"
 
 [[package]]
 name = "bitflags"
-version = "2.11.1"
+version = "2.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c4512299f36f043ab09a583e57bceb5a5aab7a73db1805848e8fef3c9e8c78b3"
+checksum = "b4388bee8683e3d04af747c73422af53102d2bd24d9eadb6cbc100baef4b43f8"
 dependencies = [
  "serde_core",
 ]
@@ -1129,7 +1140,7 @@ version = "0.12.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "fba7adb4dd5aa98e5553510223000e7148f621165ec5f9acd7113f6ca4995298"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "log",
  "polling",
  "rustix 0.38.44",
@@ -1256,9 +1267,9 @@ dependencies = [
 
 [[package]]
 name = "chrono"
-version = "0.4.44"
+version = "0.4.45"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c673075a2e0e5f4a1dde27ce9dee1ea4558c7ffe648f576438a20ca1d2acc4b0"
+checksum = "1aa79e62e7697b8e29b513a68abacf485adcd1fe8284a4316c5ae868e6633327"
 dependencies = [
  "iana-time-zone",
  "js-sys",
@@ -1607,6 +1618,12 @@ dependencies = [
  "cfg-if",
 ]
 
+[[package]]
+name = "critical-section"
+version = "1.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "790eea4361631c5e7d22598ecd5723ff611904e3344ce8720784c93e3d83d40b"
+
 [[package]]
 name = "cross_agent_session_resumer"
 version = "0.2.2"
@@ -1697,7 +1714,7 @@ version = "0.29.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d8b9f2e4c67f833b660cdb0a3523065869fb35570177239812ed4c905aeff87b"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "crossterm_winapi",
  "derive_more",
  "document-features",
@@ -1787,7 +1804,7 @@ version = "0.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3e3d747f100290a1ca24b752186f61f6637e1deffe3bf6320de6fcb29510a307"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "libloading 0.8.9",
  "winapi",
 ]
@@ -1934,7 +1951,7 @@ checksum = "be1e0bca6c3637f992fc1cc7cbc52a78c1ef6db076dbf1059c4323d6a2048376"
 [[package]]
 name = "dcg-core"
 version = "0.6.0-rc.1"
-source = "git+https://github.com/quangdang46/destructive_command_guard?branch=main#6ae9e3f5f9dc93ad8ea8b20f64e7bba24740d29e"
+source = "git+https://github.com/quangdang46/destructive_command_guard?branch=main#6002f69cb2806f918daf90c461272ce8b09442c6"
 dependencies = [
  "aho-corasick",
  "chrono",
@@ -2392,7 +2409,7 @@ version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1e0e367e4e7da84520dedcac1901e4da967309406d1e51017ae1abfb97adbd38"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "objc2 0.6.4",
 ]
 
@@ -3303,7 +3320,7 @@ version = "0.20.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7b88256088d75a56f8ecfa070513a775dd9107f6530ef14919dac831af9cfe2b"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "libc",
  "libgit2-sys",
  "log",
@@ -3462,7 +3479,7 @@ version = "0.16.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2409cffa4fe8b303847d5b6ba8df9da9ba65d302fc5ee474ea0cac5afde79840"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "bstr",
  "gix-path",
  "libc",
@@ -3601,7 +3618,7 @@ version = "0.23.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e8546300aee4c65c5862c22a3e321124a69b654a61a8b60de546a9284812b7e2"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "bstr",
  "gix-features",
  "gix-path",
@@ -3649,7 +3666,7 @@ version = "0.45.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9ea6d3e9e11647ba49f441dea0782494cc6d2875ff43fa4ad9094e6957f42051"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "bstr",
  "filetime",
  "fnv",
@@ -3772,7 +3789,7 @@ version = "0.14.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ed9e0c881933c37a7ef45288d6c5779c4a7b3ad240b4c37657e1d9829eb90085"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "bstr",
  "gix-attributes",
  "gix-config-value",
@@ -3853,7 +3870,7 @@ version = "0.39.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "91898c83b18c635696f7355d171cfa74a52f38022ff89581f567768935ebc4c8"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "bstr",
  "gix-commitgraph",
  "gix-date",
@@ -3886,7 +3903,7 @@ version = "0.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ea9962ed6d9114f7f100efe038752f41283c225bb507a2888903ac593dffa6be"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "gix-path",
  "libc",
  "windows-sys 0.61.2",
@@ -3985,7 +4002,7 @@ version = "0.51.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d052b83d1d1744be95ac6448ac02f95f370a8f6720e466be9ce57146e39f5280"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "gix-commitgraph",
  "gix-date",
  "gix-hash",
@@ -4141,7 +4158,7 @@ version = "0.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "fbcd2dba93594b227a1f57ee09b8b9da8892c34d55aa332e034a228d0fe6a171"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "gpu-alloc-types",
 ]
 
@@ -4151,7 +4168,7 @@ version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "98ff03b468aa837d70984d55f5d3f846f6ec31fe34bbb97c4f85219caeee1ca4"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
 ]
 
 [[package]]
@@ -4173,7 +4190,7 @@ version = "0.2.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "cc11df1ace8e7e564511f53af41f3e42ddc95b56fd07b3f4445d2a6048bc682c"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "gpu-descriptor-types",
  "hashbrown 0.14.5",
 ]
@@ -4184,7 +4201,7 @@ version = "0.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6bf0b36e6f090b7e1d8a4b49c0cb81c1f8376f72198c65dd3ad9ff3556b8b78c"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
 ]
 
 [[package]]
@@ -4294,6 +4311,11 @@ name = "hashbrown"
 version = "0.17.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a"
+dependencies = [
+ "allocator-api2",
+ "equivalent",
+ "foldhash 0.2.0",
+]
 
 [[package]]
 name = "hashline"
@@ -4340,7 +4362,7 @@ version = "0.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "af2a7e73e1f34c48da31fb668a907f250794837e08faa144fd24f0b8b741e890"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "com",
  "libc",
  "libloading 0.8.9",
@@ -4377,7 +4399,7 @@ version = "0.22.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ad82d6598ccf1dac15c8b758a1bd282b755b6776be600429176757190a1b0202"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "byteorder",
  "heed-traits",
  "heed-types",
@@ -4651,7 +4673,7 @@ dependencies = [
  "js-sys",
  "log",
  "wasm-bindgen",
- "windows-core 0.58.0",
+ "windows-core 0.62.2",
 ]
 
 [[package]]
@@ -4801,9 +4823,9 @@ dependencies = [
 
 [[package]]
 name = "ignore"
-version = "0.4.25"
+version = "0.4.26"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d3d782a365a015e0f5c04902246139249abf769125006fbe7649e2ee88169b4a"
+checksum = "b915661dd01db3f05050265b2477bcc6527b3792388e2749b41623cc592be67d"
 dependencies = [
  "crossbeam-deque",
  "globset",
@@ -4920,7 +4942,7 @@ version = "0.11.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "533e68a5842e734946fe159fb03fc9bbbb254f590dd0d8ad321ae5ff7beca2c1"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "inotify-sys",
  "libc",
 ]
@@ -5049,7 +5071,7 @@ checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
 
 [[package]]
 name = "jcode"
-version = "0.22.0"
+version = "0.23.0"
 dependencies = [
  "agentgrep",
  "anyhow",
@@ -6202,7 +6224,7 @@ version = "0.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b750dcadc39a09dbadd74e118f6dd6598df77fa01df0cfcdc52c28dece74528a"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "serde",
  "unicode-segmentation",
 ]
@@ -6240,7 +6262,7 @@ version = "1.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "07293a4e297ac234359b510362495713f75ea345d5307140414f20c69ffeb087"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "libc",
 ]
 
@@ -6361,7 +6383,7 @@ version = "0.1.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f02ab6bace2054fb888a3c16f990117b579d14a3088e472d63c6011fa185c9d3"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "libc",
  "plain",
  "redox_syscall 0.8.1",
@@ -6396,7 +6418,7 @@ version = "0.3.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3f50e8f47623268b5407192d26876c4d7f89d686ca130fdc53bced4814cd29f8"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
 ]
 
 [[package]]
@@ -6502,9 +6524,9 @@ dependencies = [
 
 [[package]]
 name = "log"
-version = "0.4.30"
+version = "0.4.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "616ec5685824bcc94416c6d4a7a446eea774a31efd7062c8480ba6fd06d7a6e5"
+checksum = "953f07c43838f8e6f9758cab68bf5bed85465e7587ebe0b823f1bcd81978ad3a"
 
 [[package]]
 name = "lopdf"
@@ -6542,6 +6564,15 @@ dependencies = [
  "hashbrown 0.16.1",
 ]
 
+[[package]]
+name = "lru"
+version = "0.18.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8a860605968fce16869fd239cf4237a82f3ac470723415db603b0e8b6c8d4fb9"
+dependencies = [
+ "hashbrown 0.17.1",
+]
+
 [[package]]
 name = "lru-slab"
 version = "0.1.2"
@@ -6744,7 +6775,7 @@ version = "0.27.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c43f73953f8cbe511f021b58f18c3ce1c3d1ae13fe953293e13345bf83217f25"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "block",
  "core-graphics-types",
  "foreign-types 0.5.0",
@@ -6838,7 +6869,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "50e3524642f53d9af419ab5e8dd29d3ba155708267667c2f3f06c88c9e130843"
 dependencies = [
  "bit-set 0.5.3",
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "codespan-reporting",
  "hexf-parse",
  "indexmap",
@@ -6905,7 +6936,7 @@ version = "0.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2076a31b7010b17a38c01907c45b945e8f11495ee4dd588309718901b1f7a5b7"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "jni-sys 0.3.1",
  "log",
  "ndk-sys",
@@ -6945,7 +6976,7 @@ version = "0.29.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "71e2746dc3a24dd78b3cfcb7be93368c6de9963d30f43a6a73998a9cf4b17b46"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "cfg-if",
  "cfg_aliases 0.2.1",
  "libc",
@@ -6986,7 +7017,7 @@ version = "6.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6205bd8bb1e454ad2e27422015fb5e4f2bcc7e08fa8f27058670d208324a4d2d"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "crossbeam-channel",
  "filetime",
  "fsevent-sys",
@@ -7005,7 +7036,7 @@ version = "9.0.0-rc.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b44b771d4dd781ef14c84078693e67495da6b47f609f72e8a4da8420a861240e"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "inotify 0.11.2",
  "kqueue",
  "libc",
@@ -7025,7 +7056,7 @@ version = "2.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "42b8cfee0e339a0337359f3c88165702ac6e600dc01c0cc9579a92d62b08477a"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
 ]
 
 [[package]]
@@ -7185,7 +7216,7 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d49e936b501e5c5bf01fda3a9452ff86dc3ea98ad5f283e1455153142d97518c"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "objc2 0.6.4",
  "objc2-core-graphics",
  "objc2-foundation",
@@ -7197,7 +7228,7 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2a180dd8642fa45cdb7dd721cd4c11b1cadd4929ce112ebd8b9f5803cc79d536"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "dispatch2",
  "objc2 0.6.4",
 ]
@@ -7208,7 +7239,7 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e022c9d066895efa1345f8e33e584b9f958da2fd4cd116792e15e07e4720a807"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "dispatch2",
  "objc2 0.6.4",
  "objc2-core-foundation",
@@ -7243,7 +7274,7 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e3e0adef53c21f888deb4fa59fc59f7eb17404926ee8a6f59f5df0fd7f9f3272"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "objc2 0.6.4",
  "objc2-core-foundation",
 ]
@@ -7254,7 +7285,7 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "180788110936d59bab6bd83b6060ffdfffb3b922ba1396b312ae795e1de9d81d"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "objc2 0.6.4",
  "objc2-core-foundation",
 ]
@@ -7286,7 +7317,7 @@ version = "6.5.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0cc3cbf698f9438986c11a880c90a6d04b9de27575afd28bbf45b154b6c709e2"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "libc",
  "once_cell",
  "onig_sys",
@@ -7319,7 +7350,7 @@ version = "0.10.80"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a45fa2aa886c42762255da344f0a0d313e254066c46aad76f300c3d3da62d967"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "cfg-if",
  "foreign-types 0.3.2",
  "libc",
@@ -7473,6 +7504,7 @@ version = "0.7.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4cbf71184cc5ecc2e4e1baccdb21026c20e5fc3dcf63028a086131b3ab00b6e6"
 dependencies = [
+ "approx",
  "bytemuck",
  "fast-srgb8",
  "libm",
@@ -7779,7 +7811,7 @@ version = "0.18.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "60769b8b31b2a9f263dae2776c37b1b28ae246943cf719eb6946a1db05128a61"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "crc32fast",
  "fdeflate",
  "flate2",
@@ -7986,7 +8018,7 @@ version = "0.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f86ba2052aebccc42cbbb3ed234b8b13ce76f75c3551a303cb2bcffcff12bb14"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "getopts",
  "memchr",
  "pulldown-cmark-escape",
@@ -8244,9 +8276,9 @@ checksum = "973443cf09a9c8656b574a866ab68dfa19f0867d0340648c7d2f6a71b8a8ea68"
 
 [[package]]
 name = "ratatui"
-version = "0.30.0"
+version = "0.30.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d1ce67fb8ba4446454d1c8dbaeda0557ff5e94d39d5e5ed7f10a65eb4c8266bc"
+checksum = "1695748e3a735b34968c887ceea5a380b43545903868ae8f5b666593100f6b68"
 dependencies = [
  "instability",
  "ratatui-core",
@@ -8254,22 +8286,26 @@ dependencies = [
  "ratatui-macros",
  "ratatui-termwiz",
  "ratatui-widgets",
+ "serde",
 ]
 
 [[package]]
 name = "ratatui-core"
-version = "0.1.0"
+version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5ef8dea09a92caaf73bff7adb70b76162e5937524058a7e5bff37869cbbec293"
+checksum = "42d3603f354bba8c595fa47860e60142d7372b7210c27044c6a7d0e1a4336b44"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "compact_str",
- "hashbrown 0.16.1",
+ "critical-section",
+ "hashbrown 0.17.1",
  "indoc",
  "itertools 0.14.0",
  "kasuari",
- "lru 0.16.4",
- "strum 0.27.2",
+ "lru 0.18.0",
+ "palette",
+ "serde",
+ "strum 0.28.0",
  "thiserror 2.0.18",
  "unicode-segmentation",
  "unicode-truncate",
@@ -8278,9 +8314,9 @@ dependencies = [
 
 [[package]]
 name = "ratatui-crossterm"
-version = "0.1.0"
+version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "577c9b9f652b4c121fb25c6a391dd06406d3b092ba68827e6d2f09550edc54b3"
+checksum = "2b2867bedcbd6a690ca4f8672a687b730ec07660c79844517b084311b529980c"
 dependencies = [
  "cfg-if",
  "crossterm",
@@ -8306,9 +8342,9 @@ dependencies = [
 
 [[package]]
 name = "ratatui-macros"
-version = "0.7.0"
+version = "0.7.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a7f1342a13e83e4bb9d0b793d0ea762be633f9582048c892ae9041ef39c936f4"
+checksum = "80fac59720679490d89d200df411faa249be728681adcabed3d047ae72c48f1d"
 dependencies = [
  "ratatui-core",
  "ratatui-widgets",
@@ -8316,9 +8352,9 @@ dependencies = [
 
 [[package]]
 name = "ratatui-termwiz"
-version = "0.1.0"
+version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0f76fe0bd0ed4295f0321b1676732e2454024c15a35d01904ddb315afd3d545c"
+checksum = "386b8ff8f74ed749509391c56d549761a2fcdb408e1f42e467286bcb7dac8967"
 dependencies = [
  "ratatui-core",
  "termwiz",
@@ -8326,18 +8362,19 @@ dependencies = [
 
 [[package]]
 name = "ratatui-widgets"
-version = "0.3.0"
+version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d7dbfa023cd4e604c2553483820c5fe8aa9d71a42eea5aa77c6e7f35756612db"
+checksum = "7ef4f17dd7ac3abf5adc2b920a03c61eee4bfe6a88fa5191936895525371d79c"
 dependencies = [
- "bitflags 2.11.1",
- "hashbrown 0.16.1",
+ "bitflags 2.13.0",
+ "hashbrown 0.17.1",
  "indoc",
  "instability",
  "itertools 0.14.0",
  "line-clipping",
  "ratatui-core",
- "strum 0.27.2",
+ "serde",
+ "strum 0.28.0",
  "time",
  "unicode-segmentation",
  "unicode-width 0.2.2",
@@ -8349,7 +8386,7 @@ version = "11.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "498cd0dc59d73224351ee52a95fee0f1a617a2eae0e7d9d720cc622c73a54186"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
 ]
 
 [[package]]
@@ -8440,7 +8477,7 @@ version = "0.5.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ed2bf2547551a7053d6fdfafda3f938979645c44812fbfcda098faae3f1a362d"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
 ]
 
 [[package]]
@@ -8449,7 +8486,7 @@ version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5b44b894f2a6e36457d665d1e08c3866add6ed5e70050c1b4ba8a8ddedb02ce7"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
 ]
 
 [[package]]
@@ -8664,7 +8701,7 @@ version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "28b19f5711867dc33a82cdbfd437c03b4089308f63a7ec3ee6ab34a9d74ff519"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "crossterm",
  "fancy-regex 0.17.0",
  "log",
@@ -8749,7 +8786,7 @@ version = "0.33.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1c6d5e5acb6f6129fe3f7ba0a7fc77bca1942cb568535e18e7bc40262baf3110"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "fallible-iterator",
  "fallible-streaming-iterator",
  "hashlink",
@@ -8808,7 +8845,7 @@ version = "0.38.44"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "fdb5bc1ae2baa591800df16c9ca78619bf65c0488b41b96ccec5d11220d8c154"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "errno",
  "libc",
  "linux-raw-sys 0.4.15",
@@ -8821,7 +8858,7 @@ version = "1.1.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6fe4565b9518b83ef4f91bb47ce29620ca828bd32cb7e408f0062e9930ba190"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "errno",
  "libc",
  "linux-raw-sys 0.12.1",
@@ -9016,7 +9053,7 @@ version = "0.20.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "fd3c7c96f8a08ee34eff8857b11b49b07d71d1c3f4e88f8a88d4c9e9f90b1702"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "bytemuck",
  "core_maths",
  "log",
@@ -9185,7 +9222,7 @@ version = "2.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "897b2245f0b511c87893af39b033e5ca9cce68824c4d7e7630b5a1d339658d02"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "core-foundation 0.9.4",
  "core-foundation-sys",
  "libc",
@@ -9198,7 +9235,7 @@ version = "3.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b7f4bc775c73d9a02cde8bf7b2ec4c9d12743edf609006c7facc23998404cd1d"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "core-foundation 0.10.1",
  "core-foundation-sys",
  "libc",
@@ -9549,7 +9586,7 @@ version = "0.18.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "922fd3eeab3bd820d76537ce8f582b1cf951eceb5475c28500c7457d9d17f53a"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "calloop",
  "calloop-wayland-source",
  "cursor-icon",
@@ -9603,7 +9640,7 @@ version = "0.3.0+sdk-1.3.268.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "eda41003dc44290527a59b13432d4a0379379fa074b70174882adfbdfd917844"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
 ]
 
 [[package]]
@@ -9714,11 +9751,11 @@ dependencies = [
 
 [[package]]
 name = "strum"
-version = "0.27.2"
+version = "0.28.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "af23d6f6c1a224baef9d3f61e287d2761385a5b88fdab4eb4c6f11aeb54c4bcf"
+checksum = "9628de9b8791db39ceda2b119bbe13134770b56c138ec1d3af810d045c04f9bd"
 dependencies = [
- "strum_macros 0.27.2",
+ "strum_macros 0.28.0",
 ]
 
 [[package]]
@@ -9736,9 +9773,9 @@ dependencies = [
 
 [[package]]
 name = "strum_macros"
-version = "0.27.2"
+version = "0.28.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7695ce3845ea4b33927c055a39dc438a45b059f7c1b3d91d38d10355fb8cbca7"
+checksum = "ab85eea0270ee17587ed4156089e10b9e6880ee688791d45a905f5b1ca36f664"
 dependencies = [
  "heck 0.5.0",
  "proc-macro2",
@@ -9869,7 +9906,7 @@ version = "0.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a13f3d0daba03132c0aa9767f98351b3488edc2c100cda2d2ec2b04f3d8d3c8b"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "core-foundation 0.9.4",
  "system-configuration-sys",
 ]
@@ -9952,7 +9989,7 @@ checksum = "4676b37242ccbd1aabf56edb093a4827dc49086c0ffd764a5705899e0f35f8f7"
 dependencies = [
  "anyhow",
  "base64 0.22.1",
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "fancy-regex 0.11.0",
  "filedescriptor",
  "finl_unicode",
@@ -10428,7 +10465,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4cfcf7e2740e6fc6d4d688b4ef00650406bb94adf4731e43c096c3a19fe40840"
 dependencies = [
  "async-compression",
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "bytes",
  "futures-core",
  "futures-util",
@@ -11430,7 +11467,7 @@ version = "0.244.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "47b807c72e1bac69382b3a6fb3dbe8ea4c0ed87ff5629b8685ae6b9a611028fe"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "hashbrown 0.15.5",
  "indexmap",
  "semver",
@@ -11456,7 +11493,7 @@ version = "0.31.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "645c7c96bb74690c3189b5c9cb4ca1627062bb23693a4fad9d8c3de958260144"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "rustix 1.1.4",
  "wayland-backend",
  "wayland-scanner",
@@ -11468,7 +11505,7 @@ version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "625c5029dbd43d25e6aa9615e88b829a5cad13b2819c4ae129fdbb7c31ab4c7e"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "cursor-icon",
  "wayland-backend",
 ]
@@ -11490,7 +11527,7 @@ version = "0.31.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8f81f365b8b4a97f422ac0e8737c438024b5951734506b0e1d775c73030561f4"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "wayland-backend",
  "wayland-client",
  "wayland-scanner",
@@ -11502,7 +11539,7 @@ version = "0.32.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "563a85523cade2429938e790815fd7319062103b9f4a2dc806e9b53b95982d8f"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "wayland-backend",
  "wayland-client",
  "wayland-scanner",
@@ -11514,7 +11551,7 @@ version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "23803551115ff9ea9bce586860c5c5a971e360825a0309264102a9495a5ff479"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "wayland-backend",
  "wayland-client",
  "wayland-protocols 0.31.2",
@@ -11527,7 +11564,7 @@ version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ad1f61b76b6c2d8742e10f9ba5c3737f6530b4c243132c2a2ccc8aa96fe25cd6"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "wayland-backend",
  "wayland-client",
  "wayland-protocols 0.31.2",
@@ -11540,7 +11577,7 @@ version = "0.3.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "eb04e52f7836d7c7976c78ca0250d61e33873c34156a2a1fc9474828ec268234"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "wayland-backend",
  "wayland-client",
  "wayland-protocols 0.32.12",
@@ -11729,7 +11766,7 @@ checksum = "28b94525fc99ba9e5c9a9e24764f2bc29bad0911a7446c12f446a8277369bf3a"
 dependencies = [
  "arrayvec",
  "bit-vec 0.6.3",
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "cfg_aliases 0.1.1",
  "codespan-reporting",
  "indexmap",
@@ -11757,7 +11794,7 @@ dependencies = [
  "arrayvec",
  "ash",
  "bit-set 0.5.3",
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "block",
  "cfg_aliases 0.1.1",
  "core-graphics-types",
@@ -11798,7 +11835,7 @@ version = "0.19.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b671ff9fb03f78b46ff176494ee1ebe7d603393f42664be55b64dc8d53969805"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "js-sys",
  "web-sys",
 ]
@@ -11908,13 +11945,26 @@ version = "0.58.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6ba6d44ec8c2591c134257ce647b7ea6b20335bf6379a27dac5f1641fcf59f99"
 dependencies = [
- "windows-implement",
- "windows-interface",
+ "windows-implement 0.58.0",
+ "windows-interface 0.58.0",
  "windows-result 0.2.0",
  "windows-strings 0.1.0",
  "windows-targets 0.52.6",
 ]
 
+[[package]]
+name = "windows-core"
+version = "0.62.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
+dependencies = [
+ "windows-implement 0.60.2",
+ "windows-interface 0.59.3",
+ "windows-link",
+ "windows-result 0.4.1",
+ "windows-strings 0.5.1",
+]
+
 [[package]]
 name = "windows-implement"
 version = "0.58.0"
@@ -11926,6 +11976,17 @@ dependencies = [
  "syn 2.0.117",
 ]
 
+[[package]]
+name = "windows-implement"
+version = "0.60.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
 [[package]]
 name = "windows-interface"
 version = "0.58.0"
@@ -11937,6 +11998,17 @@ dependencies = [
  "syn 2.0.117",
 ]
 
+[[package]]
+name = "windows-interface"
+version = "0.59.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
 [[package]]
 name = "windows-link"
 version = "0.2.1"
@@ -12297,7 +12369,7 @@ dependencies = [
  "ahash",
  "android-activity",
  "atomic-waker",
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "bytemuck",
  "calloop",
  "cfg_aliases 0.1.1",
@@ -12424,7 +12496,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9d66ea20e9553b30172b5e831994e35fbde2d165325bec84fc43dbf6f4eb9cb2"
 dependencies = [
  "anyhow",
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "indexmap",
  "log",
  "serde",
@@ -12541,7 +12613,7 @@ version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d039de8032a9a8856a6be89cea3e5d12fdd82306ab7c94d74e6deab2460651c5"
 dependencies = [
- "bitflags 2.11.1",
+ "bitflags 2.13.0",
  "dlib",
  "log",
  "once_cell",
@@ -12592,9 +12664,9 @@ checksum = "c94451ac9513335b5e23d7a8a2b61a7102398b8cca5160829d313e84c9d98be1"
 
 [[package]]
 name = "yoke"
-version = "0.8.2"
+version = "0.8.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "abe8c5fda708d9ca3df187cae8bfb9ceda00dd96231bed36e445a1a48e66f9ca"
+checksum = "709fe23a0424b6a435d82152b1bd3fdfb0833487d5fa90d05d42762a9891fef5"
 dependencies = [
  "stable_deref_trait",
  "yoke-derive",
diff --git a/Cargo.toml b/Cargo.toml
index 329524595..2db664c27 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "jcode"
-version = "0.22.0"
+version = "0.23.0"
 description = "Possibly the greatest coding agent ever built — blazing-fast TUI, multi-model, swarm coordination, 30+ tools"
 edition = "2024"
 autobins = false
@@ -344,6 +344,58 @@ opt-level = 3
 [profile.test.package."jcode-tui-anim"]
 opt-level = 3
 
+# Keep the text-shaping stack optimized even in dev/selfdev/test builds.
+#
+# cosmic-text + rustybuzz + ttf-parser + swash + yazi do all of the desktop
+# transcript glyph shaping. At opt-level = 0 they are 15-40x slower, which made
+# scrolling real (emoji/markdown-heavy) transcripts janky in dev/selfdev builds
+# even though the production release build was smooth. These are stable
+# third-party crates that almost never recompile, so pinning them to opt-level =
+# 3 costs a one-time compile and is then reused across every iterative jcode
+# rebuild (same rationale as jcode-tui-anim above). It does NOT slow down
+# recompiles of jcode's own crates.
+[profile.dev.package.cosmic-text]
+opt-level = 3
+[profile.selfdev.package.cosmic-text]
+opt-level = 3
+[profile.test.package.cosmic-text]
+opt-level = 3
+
+[profile.dev.package.rustybuzz]
+opt-level = 3
+[profile.selfdev.package.rustybuzz]
+opt-level = 3
+[profile.test.package.rustybuzz]
+opt-level = 3
+
+[profile.dev.package.ttf-parser]
+opt-level = 3
+[profile.selfdev.package.ttf-parser]
+opt-level = 3
+[profile.test.package.ttf-parser]
+opt-level = 3
+
+[profile.dev.package.swash]
+opt-level = 3
+[profile.selfdev.package.swash]
+opt-level = 3
+[profile.test.package.swash]
+opt-level = 3
+
+[profile.dev.package.yazi]
+opt-level = 3
+[profile.selfdev.package.yazi]
+opt-level = 3
+[profile.test.package.yazi]
+opt-level = 3
+
+[profile.dev.package.fontdb]
+opt-level = 3
+[profile.selfdev.package.fontdb]
+opt-level = 3
+[profile.test.package.fontdb]
+opt-level = 3
+
 [profile.test]
 debug = 0
 incremental = true
diff --git a/crates/jcode-app-core/src/agent/status.rs b/crates/jcode-app-core/src/agent/status.rs
index 747179f30..e011b5b01 100644
--- a/crates/jcode-app-core/src/agent/status.rs
+++ b/crates/jcode-app-core/src/agent/status.rs
@@ -134,7 +134,9 @@ impl Agent {
     }
 
     pub fn provider_name(&self) -> String {
-        crate::provider_catalog::runtime_provider_display_name(self.provider.name())
+        // `display_name()` resolves the active runtime profile (e.g. NVIDIA NIM)
+        // for the OpenRouter slot; for all other providers it equals `name()`.
+        self.provider.display_name()
     }
 
     pub fn provider_model(&self) -> String {
diff --git a/crates/jcode-app-core/src/agent/turn_loops.rs b/crates/jcode-app-core/src/agent/turn_loops.rs
index 96ccdbd15..99641d931 100644
--- a/crates/jcode-app-core/src/agent/turn_loops.rs
+++ b/crates/jcode-app-core/src/agent/turn_loops.rs
@@ -9,6 +9,7 @@ impl Agent {
 
     pub(super) async fn run_turn(&mut self, print_output: bool) -> Result<String> {
         self.set_log_context();
+        crate::session_metrics::record_turn(&self.session.id);
         let mut final_text = String::new();
         let trace = trace_enabled();
         let mut context_limit_retries = 0u32;
diff --git a/crates/jcode-app-core/src/agent/turn_streaming_mpsc.rs b/crates/jcode-app-core/src/agent/turn_streaming_mpsc.rs
index 0f4b0faf5..257b35751 100644
--- a/crates/jcode-app-core/src/agent/turn_streaming_mpsc.rs
+++ b/crates/jcode-app-core/src/agent/turn_streaming_mpsc.rs
@@ -1,5 +1,47 @@
 use super::*;
 
+/// Largest byte index `<= index` that is a UTF-8 char boundary in `text`.
+/// Equivalent to the unstable `str::floor_char_boundary`, reimplemented so the
+/// incremental marker scan can clamp its scan-window start onto a valid
+/// boundary without re-scanning the whole accumulated response.
+fn floor_char_boundary(text: &str, index: usize) -> usize {
+    if index >= text.len() {
+        return text.len();
+    }
+    let mut boundary = index;
+    while boundary > 0 && !text.is_char_boundary(boundary) {
+        boundary -= 1;
+    }
+    boundary
+}
+
+/// The wrapped-tool-call markers emitted by some models inside plain text.
+const WRAP_TOOL_MARKERS: [&str; 2] = ["to=functions.", "+#+#"];
+
+/// Find the first wrapped-tool-call marker in `accumulated`, scanning only the
+/// newly appended `delta` plus a short overlap from the previous tail (so a
+/// marker straddling the append boundary is still found).
+///
+/// This avoids re-scanning the entire accumulated response on every streamed
+/// delta, which was O(response) per token and O(response^2) over a full answer.
+fn find_wrap_marker_incremental(accumulated: &str, appended_len: usize) -> Option<usize> {
+    let max_marker_len = WRAP_TOOL_MARKERS
+        .iter()
+        .map(|marker| marker.len())
+        .max()
+        .unwrap_or(0);
+    let scan_start = accumulated
+        .len()
+        .saturating_sub(appended_len + max_marker_len.saturating_sub(1));
+    let scan_start = floor_char_boundary(accumulated, scan_start);
+    let window = &accumulated[scan_start..];
+    WRAP_TOOL_MARKERS
+        .iter()
+        .filter_map(|marker| window.find(marker))
+        .min()
+        .map(|rel_idx| scan_start + rel_idx)
+}
+
 fn reload_interrupted_tool_result(tc: &ToolCall, elapsed_secs: f64) -> (String, bool) {
     if tc.name == "selfdev" {
         return ("Reload initiated. Process restarting...".to_string(), false);
@@ -402,9 +444,11 @@ impl Agent {
                         }
                         text_content.push_str(&text);
                         if !text_wrapped_detected {
-                            if let Some(marker_idx) = text_content
-                                .find("to=functions.")
-                                .or_else(|| text_content.find("+#+#"))
+                            // Scan only the new delta (plus a short overlap for
+                            // markers straddling the boundary) instead of the
+                            // whole accumulated response on every token.
+                            if let Some(marker_idx) =
+                                find_wrap_marker_incremental(&text_content, text.len())
                             {
                                 text_wrapped_detected = true;
                                 let clean_prefix =
@@ -781,6 +825,14 @@ impl Agent {
                     usage_cache_read,
                     usage_cache_creation,
                 );
+
+                let input = usage_input.unwrap_or(0);
+                let output = usage_output.unwrap_or(0);
+                let total = input
+                    .saturating_add(output)
+                    .saturating_add(usage_cache_read.unwrap_or(0))
+                    .saturating_add(usage_cache_creation.unwrap_or(0));
+                crate::session_metrics::record_token_usage(&self.session.id, total, output);
             }
 
             if usage_input.is_some()
@@ -1347,4 +1399,74 @@ mod tests {
         assert!(is_error);
         assert!(message.contains("interrupted by server reload"));
     }
+
+    /// Reference O(n) full scan, preserving the original precedence: the
+    /// `to=functions.` marker is checked before `+#+#`.
+    fn find_wrap_marker_full(text: &str) -> Option<usize> {
+        text.find("to=functions.").or_else(|| text.find("+#+#"))
+    }
+
+    /// Simulate streaming `full` in arbitrary deltas and assert the incremental
+    /// scan finds the first marker position, matching a full rescan each step.
+    fn assert_incremental_matches(full: &str, chunk: usize) {
+        let mut acc = String::new();
+        let mut incremental_hit: Option<usize> = None;
+        let bytes = full.as_bytes();
+        let mut i = 0;
+        while i < bytes.len() {
+            let mut end = (i + chunk).min(bytes.len());
+            while end < bytes.len() && !full.is_char_boundary(end) {
+                end += 1;
+            }
+            let delta = &full[i..end];
+            acc.push_str(delta);
+            if incremental_hit.is_none() {
+                incremental_hit = find_wrap_marker_incremental(&acc, delta.len());
+            }
+            i = end;
+        }
+        // The earliest of either marker in the full text.
+        let fn_pos = full.find("to=functions.");
+        let plus_pos = full.find("+#+#");
+        let expected = match (fn_pos, plus_pos) {
+            (Some(a), Some(b)) => Some(a.min(b)),
+            (a, b) => a.or(b),
+        };
+        assert_eq!(
+            incremental_hit, expected,
+            "incremental scan mismatch for {full:?} chunk={chunk}"
+        );
+    }
+
+    #[test]
+    fn wrap_marker_incremental_detects_markers_across_chunk_sizes() {
+        let cases = [
+            "plain answer with no marker at all",
+            "answer then to=functions.foo({})",
+            "answer then +#+# wrapped",
+            "prefix +#+# and later to=functions.bar",
+            "unicode 🔄 résumé then to=functions.baz",
+            "",
+            "to=functions.first",
+            "+#+#",
+        ];
+        for case in cases {
+            for chunk in [1usize, 2, 3, 5, 7, 100] {
+                assert_incremental_matches(case, chunk);
+            }
+        }
+    }
+
+    #[test]
+    fn wrap_marker_incremental_finds_marker_straddling_delta_boundary() {
+        // Feed "to=functions." split right in the middle so the marker only
+        // exists once both halves are appended; the overlap window must catch it.
+        let mut acc = String::new();
+        acc.push_str("answer to=fun");
+        assert_eq!(find_wrap_marker_incremental(&acc, "answer to=fun".len()), None);
+        acc.push_str("ctions.tool");
+        let hit = find_wrap_marker_incremental(&acc, "ctions.tool".len());
+        assert_eq!(hit, find_wrap_marker_full(&acc));
+        assert_eq!(hit, Some("answer ".len()));
+    }
 }
diff --git a/crates/jcode-app-core/src/ambient/prompt.rs b/crates/jcode-app-core/src/ambient/prompt.rs
index c5bb677fb..2cca08b28 100644
--- a/crates/jcode-app-core/src/ambient/prompt.rs
+++ b/crates/jcode-app-core/src/ambient/prompt.rs
@@ -181,50 +181,89 @@ pub fn gather_recent_sessions(since: Option<DateTime<Utc>>) -> Vec<RecentSession
 
     let cutoff = since.unwrap_or_else(|| Utc::now() - chrono::Duration::hours(24));
 
-    let mut recent = Vec::new();
+    // Pre-filter candidate session files by filesystem mtime BEFORE loading and
+    // parsing them. The sessions directory can hold tens of thousands of files;
+    // fully parsing every one via Session::load just to drop those older than
+    // the cutoff is O(all_sessions * parse). A session updated after the cutoff
+    // has a recent mtime, so we keep only files whose mtime is at or after the
+    // cutoff (minus a small margin for clock/write skew), then load newest-first
+    // and stop once we have enough recent sessions.
+    const RECENT_SESSION_LIMIT: usize = 20;
+    let mtime_cutoff = cutoff - chrono::Duration::hours(1);
+
+    let mut candidates: Vec<(std::path::PathBuf, std::time::SystemTime)> = Vec::new();
     if let Ok(entries) = std::fs::read_dir(&sessions_dir) {
         for entry in entries.flatten() {
             let path = entry.path();
-            if path.extension().map(|e| e == "json").unwrap_or(false)
-                && let Some(stem) = path.file_stem().and_then(|s| s.to_str())
-                && let Ok(session) = crate::session::Session::load(stem)
-            {
-                // Skip debug sessions
-                if session.is_debug {
-                    continue;
-                }
-                // Only include sessions updated after cutoff
-                if session.updated_at < cutoff {
-                    continue;
-                }
-                let duration = (session.updated_at - session.created_at)
-                    .num_seconds()
-                    .max(0);
-                let extraction = if session.messages.is_empty() {
-                    "no messages"
-                } else {
-                    // Heuristic: if session closed normally, assume extracted
-                    match &session.status {
-                        crate::session::SessionStatus::Closed => "extracted",
-                        crate::session::SessionStatus::Crashed { .. } => "missed",
-                        crate::session::SessionStatus::Active => "in progress",
-                        _ => "unknown",
-                    }
-                };
-                recent.push(RecentSessionInfo {
-                    id: session.id.clone(),
-                    status: session.status.display().to_string(),
-                    topic: session.display_title().map(ToOwned::to_owned),
-                    duration_secs: duration,
-                    extraction_status: extraction.to_string(),
-                });
+            if !path.extension().map(|e| e == "json").unwrap_or(false) {
+                continue;
+            }
+            let Ok(modified) = entry.metadata().and_then(|meta| meta.modified()) else {
+                // If we can't read mtime, keep the file as a candidate so we
+                // don't silently drop a possibly-recent session.
+                candidates.push((path, std::time::SystemTime::UNIX_EPOCH));
+                continue;
+            };
+            let modified_dt: DateTime<Utc> = modified.into();
+            if modified_dt < mtime_cutoff {
+                continue;
+            }
+            candidates.push((path, modified));
+        }
+    }
+    // Newest files first so we can stop early once we have enough.
+    candidates.sort_by(|a, b| b.1.cmp(&a.1));
+
+    let mut recent = Vec::new();
+    // Load somewhat more than the final limit by mtime so the subsequent
+    // id-based sort/truncate picks the true most-recent set even when file
+    // mtime order and id (timestamp) order disagree near the boundary, while
+    // still bounding work far below "load every session file".
+    let load_budget = RECENT_SESSION_LIMIT.saturating_mul(4).max(RECENT_SESSION_LIMIT);
+    let mut loaded = 0usize;
+    for (path, _modified) in candidates {
+        if loaded >= load_budget {
+            break;
+        }
+        if let Some(stem) = path.file_stem().and_then(|s| s.to_str())
+            && let Ok(session) = crate::session::Session::load(stem)
+        {
+            loaded += 1;
+            // Skip debug sessions
+            if session.is_debug {
+                continue;
+            }
+            // Only include sessions updated after cutoff
+            if session.updated_at < cutoff {
+                continue;
             }
+            let duration = (session.updated_at - session.created_at)
+                .num_seconds()
+                .max(0);
+            let extraction = if session.messages.is_empty() {
+                "no messages"
+            } else {
+                // Heuristic: if session closed normally, assume extracted
+                match &session.status {
+                    crate::session::SessionStatus::Closed => "extracted",
+                    crate::session::SessionStatus::Crashed { .. } => "missed",
+                    crate::session::SessionStatus::Active => "in progress",
+                    _ => "unknown",
+                }
+            };
+            recent.push(RecentSessionInfo {
+                id: session.id.clone(),
+                status: session.status.display().to_string(),
+                topic: session.display_title().map(ToOwned::to_owned),
+                duration_secs: duration,
+                extraction_status: extraction.to_string(),
+            });
         }
     }
 
-    // Sort by most recent first (we don't have created_at easily, sort by id which embeds timestamp)
+    // Sort by most recent first (id embeds a timestamp).
     recent.sort_by(|a, b| b.id.cmp(&a.id));
-    recent.truncate(20); // Cap at 20 to keep prompt reasonable
+    recent.truncate(RECENT_SESSION_LIMIT);
     recent
 }
 
diff --git a/crates/jcode-app-core/src/protocol_tests/comm_responses.rs b/crates/jcode-app-core/src/protocol_tests/comm_responses.rs
index 7c886ddf8..13c0186c3 100644
--- a/crates/jcode-app-core/src/protocol_tests/comm_responses.rs
+++ b/crates/jcode-app-core/src/protocol_tests/comm_responses.rs
@@ -158,6 +158,7 @@ fn test_comm_members_roundtrip_includes_status() -> Result<()> {
             latest_completion_report: None,
             live_attachments: Some(0),
             status_age_secs: Some(12),
+            ..Default::default()
         }],
     };
 
diff --git a/crates/jcode-app-core/src/server/client_comm_channels.rs b/crates/jcode-app-core/src/server/client_comm_channels.rs
index e701e1f6e..54ff66cd2 100644
--- a/crates/jcode-app-core/src/server/client_comm_channels.rs
+++ b/crates/jcode-app-core/src/server/client_comm_channels.rs
@@ -93,6 +93,7 @@ pub(super) async fn handle_comm_channel_members(
                     latest_completion_report: member.latest_completion_report.clone(),
                     live_attachments: Some(member.event_txs.len()),
                     status_age_secs: Some(member.last_status_change.elapsed().as_secs()),
+                    ..Default::default()
                 })
             })
             .collect();
diff --git a/crates/jcode-app-core/src/server/client_comm_context.rs b/crates/jcode-app-core/src/server/client_comm_context.rs
index cfe7bc8b6..b99a1b56a 100644
--- a/crates/jcode-app-core/src/server/client_comm_context.rs
+++ b/crates/jcode-app-core/src/server/client_comm_context.rs
@@ -2,6 +2,7 @@ use super::{
     SharedContext, SwarmEvent, SwarmEventType, SwarmMember, fanout_session_event,
     record_swarm_event,
 };
+use super::debug::ClientConnectionInfo;
 use crate::protocol::{AgentInfo, ContextEntry, NotificationType, ServerEvent};
 use std::collections::{HashMap, HashSet};
 use std::path::PathBuf;
@@ -187,6 +188,10 @@ pub(super) async fn handle_comm_read(
     let _ = client_event_tx.send(ServerEvent::CommContext { id, entries });
 }
 
+#[expect(
+    clippy::too_many_arguments,
+    reason = "comm list joins swarm membership, file touches, live sessions, and connection activity"
+)]
 pub(super) async fn handle_comm_list(
     id: u64,
     req_session_id: String,
@@ -194,6 +199,8 @@ pub(super) async fn handle_comm_list(
     swarm_members: &Arc<RwLock<HashMap<String, SwarmMember>>>,
     swarms_by_id: &Arc<RwLock<HashMap<String, HashSet<String>>>>,
     files_touched_by_session: &Arc<RwLock<HashMap<String, HashSet<PathBuf>>>>,
+    sessions: &super::SessionAgents,
+    client_connections: &Arc<RwLock<HashMap<String, ClientConnectionInfo>>>,
 ) {
     let swarm_id = swarm_id_for_session(&req_session_id, swarm_members).await;
 
@@ -206,37 +213,89 @@ pub(super) async fn handle_comm_list(
                 .unwrap_or_default()
         };
 
-        let members = swarm_members.read().await;
-        let touches = files_touched_by_session.read().await;
-
-        let member_list: Vec<AgentInfo> = swarm_session_ids
-            .iter()
-            .filter_map(|sid| {
-                members.get(sid).map(|member| {
-                    let mut files: Vec<String> = touches
-                        .get(sid)
-                        .into_iter()
-                        .flat_map(|paths| paths.iter())
-                        .map(|path| path.display().to_string())
-                        .collect();
-                    files.sort();
-
-                    AgentInfo {
-                        session_id: sid.clone(),
-                        friendly_name: member.friendly_name.clone(),
-                        files_touched: files,
-                        status: Some(member.status.clone()),
-                        detail: member.detail.clone(),
-                        role: Some(member.role.clone()),
-                        is_headless: Some(member.is_headless),
-                        report_back_to_session_id: member.report_back_to_session_id.clone(),
-                        latest_completion_report: member.latest_completion_report.clone(),
-                        live_attachments: Some(member.event_txs.len()),
-                        status_age_secs: Some(member.last_status_change.elapsed().as_secs()),
-                    }
+        // Snapshot the static member fields first, releasing the members lock
+        // before gathering per-session runtime extras (which briefly lock
+        // individual agents and read the connection map).
+        struct MemberStatic {
+            session_id: String,
+            friendly_name: Option<String>,
+            files: Vec<String>,
+            status: String,
+            detail: Option<String>,
+            role: String,
+            is_headless: bool,
+            report_back_to_session_id: Option<String>,
+            latest_completion_report: Option<String>,
+            live_attachments: usize,
+            status_age_secs: u64,
+        }
+
+        let statics: Vec<MemberStatic> = {
+            let members = swarm_members.read().await;
+            let touches = files_touched_by_session.read().await;
+            swarm_session_ids
+                .iter()
+                .filter_map(|sid| {
+                    members.get(sid).map(|member| {
+                        let mut files: Vec<String> = touches
+                            .get(sid)
+                            .into_iter()
+                            .flat_map(|paths| paths.iter())
+                            .map(|path| path.display().to_string())
+                            .collect();
+                        files.sort();
+                        MemberStatic {
+                            session_id: sid.clone(),
+                            friendly_name: member.friendly_name.clone(),
+                            files,
+                            status: member.status.clone(),
+                            detail: member.detail.clone(),
+                            role: member.role.clone(),
+                            is_headless: member.is_headless,
+                            report_back_to_session_id: member.report_back_to_session_id.clone(),
+                            latest_completion_report: member.latest_completion_report.clone(),
+                            live_attachments: member.event_txs.len(),
+                            status_age_secs: member.last_status_change.elapsed().as_secs(),
+                        }
+                    })
                 })
-            })
-            .collect();
+                .collect()
+        };
+
+        let mut member_list: Vec<AgentInfo> = Vec::with_capacity(statics.len());
+        for m in statics {
+            let extras = super::comm_sync::member_runtime_extras(
+                &m.session_id,
+                m.status == "running",
+                sessions,
+                client_connections,
+            )
+            .await;
+
+            member_list.push(AgentInfo {
+                session_id: m.session_id,
+                friendly_name: m.friendly_name,
+                files_touched: m.files,
+                status: Some(m.status),
+                detail: m.detail,
+                role: Some(m.role),
+                is_headless: Some(m.is_headless),
+                report_back_to_session_id: m.report_back_to_session_id,
+                latest_completion_report: m.latest_completion_report,
+                live_attachments: Some(m.live_attachments),
+                status_age_secs: Some(m.status_age_secs),
+                activity: extras.activity,
+                provider_name: extras.provider_name,
+                provider_model: extras.provider_model,
+                turn_count: extras.turn_count,
+                recent_total_tokens: extras.recent_total_tokens,
+                recent_output_tokens: extras.recent_output_tokens,
+                recent_window_secs: extras.recent_window_secs,
+                cumulative_total_tokens: extras.cumulative_total_tokens,
+                todos_completed: extras.todos_completed,
+                todos_total: extras.todos_total,
+            });
+        }
 
         let _ = client_event_tx.send(ServerEvent::CommMembers {
             id,
diff --git a/crates/jcode-app-core/src/server/client_comm_tests.rs b/crates/jcode-app-core/src/server/client_comm_tests.rs
index 70c2354fd..2c1e72475 100644
--- a/crates/jcode-app-core/src/server/client_comm_tests.rs
+++ b/crates/jcode-app-core/src/server/client_comm_tests.rs
@@ -403,6 +403,11 @@ async fn comm_list_includes_member_status_and_detail() {
         HashSet::from([requester_id.clone(), peer_id.clone()]),
     )])));
     let file_touches = Arc::new(RwLock::new(HashMap::new()));
+    let sessions = Arc::new(RwLock::new(HashMap::from([
+        (requester_id.clone(), requester.clone()),
+        (peer_id.clone(), peer.clone()),
+    ])));
+    let client_connections = Arc::new(RwLock::new(HashMap::new()));
 
     handle_comm_list(
         1,
@@ -411,6 +416,8 @@ async fn comm_list_includes_member_status_and_detail() {
         &swarm_members,
         &swarms_by_id,
         &file_touches,
+        &sessions,
+        &client_connections,
     )
     .await;
 
diff --git a/crates/jcode-app-core/src/server/client_disconnect_cleanup.rs b/crates/jcode-app-core/src/server/client_disconnect_cleanup.rs
index 78745d71c..f2d596775 100644
--- a/crates/jcode-app-core/src/server/client_disconnect_cleanup.rs
+++ b/crates/jcode-app-core/src/server/client_disconnect_cleanup.rs
@@ -213,6 +213,7 @@ pub(super) async fn cleanup_client_connection(
                 (None, None)
             }
         };
+        crate::session_metrics::forget(client_session_id);
 
         if let Some(ref swarm_id) = swarm_id {
             record_swarm_event(
diff --git a/crates/jcode-app-core/src/server/client_lifecycle.rs b/crates/jcode-app-core/src/server/client_lifecycle.rs
index a7d91b8fe..3c7aa5b42 100644
--- a/crates/jcode-app-core/src/server/client_lifecycle.rs
+++ b/crates/jcode-app-core/src/server/client_lifecycle.rs
@@ -1929,6 +1929,8 @@ pub(super) async fn handle_client(
                     &swarm_members,
                     &swarms_by_id,
                     &files_touched_by_session,
+                    &sessions,
+                    &client_connections,
                 )
                 .await;
             }
diff --git a/crates/jcode-app-core/src/server/client_lightweight_control.rs b/crates/jcode-app-core/src/server/client_lightweight_control.rs
index dfdd34189..8ab3ac0e8 100644
--- a/crates/jcode-app-core/src/server/client_lightweight_control.rs
+++ b/crates/jcode-app-core/src/server/client_lightweight_control.rs
@@ -204,6 +204,8 @@ pub(super) async fn handle_lightweight_control_request(
                 swarm_members,
                 swarms_by_id,
                 files_touched_by_session,
+                sessions,
+                client_connections,
             )
             .await;
         }
diff --git a/crates/jcode-app-core/src/server/comm_session.rs b/crates/jcode-app-core/src/server/comm_session.rs
index 540c03de1..aa5badc25 100644
--- a/crates/jcode-app-core/src/server/comm_session.rs
+++ b/crates/jcode-app-core/src/server/comm_session.rs
@@ -226,6 +226,43 @@ async fn resolve_coordinator_spawn_identity(
     }
 }
 
+/// Split a configured swarm model that carries an explicit auth-route prefix
+/// (`openai-api:`, `openai-oauth:`, `claude-api:`, `claude-oauth:`) into a
+/// structured selection so spawned sessions pin the exact provider + auth
+/// method instead of guessing from the bare model name.
+///
+/// Example: `agents.swarm_model = "openai-api:gpt-5.5"` resolves to
+/// `model = gpt-5.5`, `provider_key = openai-api-key`,
+/// `route_api_method = openai-api-key`, which makes every spawned agent use
+/// GPT-5.5 on the OpenAI API key route regardless of the coordinator's model.
+///
+/// Returns `None` for models without such a prefix, or for prefixes that carry
+/// no API-vs-OAuth decision (bare provider aliases, OpenRouter, Copilot, ...).
+/// Those keep their prefixed model and route correctly via the existing
+/// session-restore path.
+fn explicit_route_for_configured_model(model: &str) -> Option<SwarmSpawnSelection> {
+    let (_, prefix, bare) = crate::provider::explicit_model_provider_prefix(model)?;
+    let bare = bare.trim();
+    if bare.is_empty() {
+        return None;
+    }
+    // Stable route ids that `ModelRouteApiMethod::parse` round-trips back into
+    // the exact auth method when the spawned session is restored (see
+    // `MultiProvider::model_switch_request_for_session_route`).
+    let route_id = match prefix {
+        "openai-api:" => "openai-api-key",
+        "openai-oauth:" => "openai-oauth",
+        "claude-api:" => "anthropic-api-key",
+        "claude-oauth:" => "claude-oauth",
+        _ => return None,
+    };
+    Some(SwarmSpawnSelection {
+        model: Some(bare.to_string()),
+        provider_key: Some(route_id.to_string()),
+        route_api_method: Some(route_id.to_string()),
+    })
+}
+
 fn resolve_swarm_spawn_selection(
     configured_swarm_model: Option<String>,
     coordinator: &CoordinatorSpawnIdentity,
@@ -244,6 +281,14 @@ fn resolve_swarm_spawn_selection(
 
     match configured_swarm_model {
         Some(model) => {
+            // A configured model may pin an explicit provider + auth route via a
+            // prefix (e.g. "openai-api:gpt-5.5"). Honor it directly so spawned
+            // agents do NOT inherit the coordinator's model/auth and instead use
+            // the requested model on the requested API route.
+            if let Some(selection) = explicit_route_for_configured_model(&model) {
+                return selection;
+            }
+
             // A concrete configured model only inherits the coordinator's
             // provider_key/route when it targets the same model; otherwise the
             // route would point at the wrong provider/auth mode.
diff --git a/crates/jcode-app-core/src/server/comm_session_tests.rs b/crates/jcode-app-core/src/server/comm_session_tests.rs
index eac745636..e541f01a8 100644
--- a/crates/jcode-app-core/src/server/comm_session_tests.rs
+++ b/crates/jcode-app-core/src/server/comm_session_tests.rs
@@ -494,6 +494,54 @@ fn resolve_swarm_spawn_model_keeps_provider_key_when_config_matches_coordinator(
     assert_eq!(selection.route_api_method.as_deref(), Some("custom-route"));
 }
 
+#[test]
+fn resolve_swarm_spawn_model_openai_api_prefix_pins_api_route_over_coordinator() {
+    // `agents.swarm_model = "openai-api:gpt-5.5"` must spawn agents on GPT-5.5
+    // via the OpenAI API key route, regardless of the coordinator's model/auth.
+    let selection = resolve_swarm_spawn_selection(
+        Some("openai-api:gpt-5.5".to_string()),
+        &coordinator_identity(
+            Some("claude-opus-4-8"),
+            Some("claude-oauth"),
+            Some("claude-oauth"),
+        ),
+    );
+
+    assert_eq!(selection.model.as_deref(), Some("gpt-5.5"));
+    assert_eq!(selection.provider_key.as_deref(), Some("openai-api-key"));
+    assert_eq!(selection.route_api_method.as_deref(), Some("openai-api-key"));
+}
+
+#[test]
+fn resolve_swarm_spawn_model_auth_route_prefixes_pin_expected_routes() {
+    for (configured, expected_model, expected_key) in [
+        ("openai-api:gpt-5.5", "gpt-5.5", "openai-api-key"),
+        ("openai-oauth:gpt-5.5", "gpt-5.5", "openai-oauth"),
+        ("claude-api:claude-opus-4-8", "claude-opus-4-8", "anthropic-api-key"),
+        ("claude-oauth:claude-opus-4-8", "claude-opus-4-8", "claude-oauth"),
+    ] {
+        let selection = resolve_swarm_spawn_selection(
+            Some(configured.to_string()),
+            &coordinator_identity(Some("some-other-model"), Some("some-key"), Some("some-route")),
+        );
+        assert_eq!(
+            selection.model.as_deref(),
+            Some(expected_model),
+            "configured {configured:?} model",
+        );
+        assert_eq!(
+            selection.provider_key.as_deref(),
+            Some(expected_key),
+            "configured {configured:?} provider_key",
+        );
+        assert_eq!(
+            selection.route_api_method.as_deref(),
+            Some(expected_key),
+            "configured {configured:?} route_api_method",
+        );
+    }
+}
+
 #[test]
 fn resolve_swarm_spawn_model_inherit_sentinel_uses_coordinator_model() {
     for sentinel in ["inherit", "INHERIT", "coordinator", " inherit ", ""] {
diff --git a/crates/jcode-app-core/src/server/comm_sync.rs b/crates/jcode-app-core/src/server/comm_sync.rs
index fbafacad0..bd18c44a9 100644
--- a/crates/jcode-app-core/src/server/comm_sync.rs
+++ b/crates/jcode-app-core/src/server/comm_sync.rs
@@ -63,6 +63,86 @@ fn live_activity_snapshot(
         })
 }
 
+/// Recent-token lookback window used when reporting per-agent churn in
+/// `swarm list`. Short enough to reflect "what is this agent doing right now".
+pub(super) const SWARM_LIST_TOKEN_WINDOW_SECS: u64 = 10;
+
+/// Runtime extras for a swarm member, gathered without holding the agent lock
+/// for long. Used to enrich the `swarm list` roster with live activity,
+/// provider/model, token churn, turn count, and todo progress.
+#[derive(Default)]
+pub(super) struct MemberRuntimeExtras {
+    pub(super) activity: Option<SessionActivitySnapshot>,
+    pub(super) provider_name: Option<String>,
+    pub(super) provider_model: Option<String>,
+    pub(super) turn_count: Option<u64>,
+    pub(super) recent_total_tokens: Option<u64>,
+    pub(super) recent_output_tokens: Option<u64>,
+    pub(super) recent_window_secs: Option<u64>,
+    pub(super) cumulative_total_tokens: Option<u64>,
+    pub(super) todos_completed: Option<usize>,
+    pub(super) todos_total: Option<usize>,
+}
+
+/// Gather live runtime extras for a single member session.
+///
+/// `member_is_running` is used as a fallback "processing" hint when no live
+/// client connection is reporting activity (e.g. headless sessions).
+pub(super) async fn member_runtime_extras(
+    session_id: &str,
+    member_is_running: bool,
+    sessions: &SessionAgents,
+    client_connections: &Arc<RwLock<HashMap<String, ClientConnectionInfo>>>,
+) -> MemberRuntimeExtras {
+    let activity = {
+        let connections = client_connections.read().await;
+        live_activity_snapshot(&connections, session_id, member_is_running)
+    };
+
+    let (provider_name, provider_model) = {
+        let agent_sessions = sessions.read().await;
+        if let Some(agent) = agent_sessions.get(session_id) {
+            // Never block on a busy agent: token churn and turns come from the
+            // lock-free metrics registry, so a missing provider name here just
+            // means the agent is mid-turn.
+            if let Ok(agent) = agent.try_lock() {
+                (Some(agent.provider_name()), Some(agent.provider_model()))
+            } else {
+                (None, None)
+            }
+        } else {
+            (None, None)
+        }
+    };
+
+    let metrics = crate::session_metrics::snapshot(
+        session_id,
+        std::time::Duration::from_secs(SWARM_LIST_TOKEN_WINDOW_SECS),
+    );
+
+    let (todos_completed, todos_total) = match crate::todo::load_todos(session_id) {
+        Ok(todos) if !todos.is_empty() => {
+            let completed = todos.iter().filter(|t| t.status == "completed").count();
+            (Some(completed), Some(todos.len()))
+        }
+        _ => (None, None),
+    };
+
+    MemberRuntimeExtras {
+        activity,
+        provider_name,
+        provider_model,
+        turn_count: metrics.map(|m| m.turns),
+        recent_total_tokens: metrics.map(|m| m.recent_total_tokens),
+        recent_output_tokens: metrics.map(|m| m.recent_output_tokens),
+        recent_window_secs: metrics.map(|_| SWARM_LIST_TOKEN_WINDOW_SECS),
+        cumulative_total_tokens: metrics.map(|m| m.cumulative_total_tokens),
+        todos_completed,
+        todos_total,
+    }
+}
+
+
 async fn ensure_same_swarm_access(
     id: u64,
     req_session_id: &str,
diff --git a/crates/jcode-app-core/src/tool/communicate_tests.rs b/crates/jcode-app-core/src/tool/communicate_tests.rs
index d6bb47f6c..dfac2f36a 100644
--- a/crates/jcode-app-core/src/tool/communicate_tests.rs
+++ b/crates/jcode-app-core/src/tool/communicate_tests.rs
@@ -126,6 +126,7 @@ fn in_flight_slot_accounting_counts_queued_workers_not_coordinator() {
             latest_completion_report: None,
             live_attachments: None,
             status_age_secs: None,
+            ..Default::default()
         },
         AgentInfo {
             session_id: "worker-queued".to_string(),
@@ -139,6 +140,7 @@ fn in_flight_slot_accounting_counts_queued_workers_not_coordinator() {
             latest_completion_report: None,
             live_attachments: None,
             status_age_secs: None,
+            ..Default::default()
         },
         AgentInfo {
             session_id: "worker-ready".to_string(),
@@ -152,6 +154,7 @@ fn in_flight_slot_accounting_counts_queued_workers_not_coordinator() {
             latest_completion_report: None,
             live_attachments: None,
             status_age_secs: None,
+            ..Default::default()
         },
     ];
 
diff --git a/crates/jcode-app-core/src/tool/communicate_tests/input_format.rs b/crates/jcode-app-core/src/tool/communicate_tests/input_format.rs
index 6e60d85c7..c3606b095 100644
--- a/crates/jcode-app-core/src/tool/communicate_tests/input_format.rs
+++ b/crates/jcode-app-core/src/tool/communicate_tests/input_format.rs
@@ -92,6 +92,7 @@ fn cleanup_candidates_default_to_owned_terminal_workers() {
             latest_completion_report: None,
             live_attachments: None,
             status_age_secs: None,
+            ..Default::default()
         },
         AgentInfo {
             session_id: "owned-done".to_string(),
@@ -105,6 +106,7 @@ fn cleanup_candidates_default_to_owned_terminal_workers() {
             latest_completion_report: None,
             live_attachments: None,
             status_age_secs: None,
+            ..Default::default()
         },
         AgentInfo {
             session_id: "user-created".to_string(),
@@ -118,6 +120,7 @@ fn cleanup_candidates_default_to_owned_terminal_workers() {
             latest_completion_report: None,
             live_attachments: None,
             status_age_secs: None,
+            ..Default::default()
         },
         AgentInfo {
             session_id: "owned-running".to_string(),
@@ -131,6 +134,7 @@ fn cleanup_candidates_default_to_owned_terminal_workers() {
             latest_completion_report: None,
             live_attachments: None,
             status_age_secs: None,
+            ..Default::default()
         },
     ];
     let statuses = default_cleanup_target_statuses();
@@ -198,18 +202,97 @@ fn format_members_includes_status_and_detail() {
             latest_completion_report: None,
             live_attachments: Some(0),
             status_age_secs: Some(12),
+            ..Default::default()
         }],
     );
 
     assert!(output.output.contains("Status: running — working on tests"));
+    assert!(output.output.contains("· 12s"));
     assert!(output.output.contains("Files: src/main.rs"));
     assert!(
         output
             .output
-            .contains("Meta: headless · owned_by_you · attachments=0 · status_age=12s")
+            .contains("Meta: headless · owned_by_you · attachments=0")
     );
 }
 
+#[test]
+fn format_members_renders_activity_progress_churn_and_turns() {
+    let ctx = test_ctx(
+        "session_self_1234567890_deadbeefcafebabe",
+        std::path::Path::new("."),
+    );
+
+    let output = format_members(
+        &ctx,
+        &[AgentInfo {
+            session_id: "session_peer_1234567890_aaaaaaaaaaaa0001".to_string(),
+            friendly_name: Some("otter".to_string()),
+            files_touched: vec![],
+            status: Some("running".to_string()),
+            detail: Some("implementing".to_string()),
+            role: Some("agent".to_string()),
+            is_headless: Some(false),
+            report_back_to_session_id: None,
+            latest_completion_report: None,
+            live_attachments: Some(1),
+            status_age_secs: Some(8),
+            activity: Some(SessionActivitySnapshot {
+                is_processing: true,
+                current_tool_name: Some("edit".to_string()),
+            }),
+            provider_name: Some("anthropic".to_string()),
+            provider_model: Some("claude-sonnet".to_string()),
+            turn_count: Some(7),
+            recent_total_tokens: Some(12_345),
+            recent_output_tokens: Some(2_000),
+            recent_window_secs: Some(10),
+            cumulative_total_tokens: Some(98_765),
+            todos_completed: Some(3),
+            todos_total: Some(7),
+        }],
+    );
+
+    let text = output.output;
+    assert!(text.contains("Activity: working (edit)"), "got: {text}");
+    assert!(text.contains("Progress: 3/7 todos"), "got: {text}");
+    assert!(text.contains("12.3k tok/10s"), "got: {text}");
+    assert!(text.contains("7 turns"), "got: {text}");
+    assert!(text.contains("98.8k tok total"), "got: {text}");
+    assert!(text.contains("Model: anthropic/claude-sonnet"), "got: {text}");
+    // Running agent shows current-turn duration, not an "idle" label.
+    assert!(text.contains("· 8s"), "got: {text}");
+    assert!(!text.contains("idle"), "got: {text}");
+}
+
+#[test]
+fn format_members_labels_idle_ready_agent() {
+    let ctx = test_ctx(
+        "session_self_1234567890_deadbeefcafebabe",
+        std::path::Path::new("."),
+    );
+
+    let output = format_members(
+        &ctx,
+        &[AgentInfo {
+            session_id: "session_peer_1234567890_bbbbbbbbbbbb0002".to_string(),
+            friendly_name: Some("idle-one".to_string()),
+            files_touched: vec![],
+            status: Some("ready".to_string()),
+            detail: None,
+            role: Some("agent".to_string()),
+            is_headless: None,
+            report_back_to_session_id: None,
+            latest_completion_report: None,
+            live_attachments: Some(0),
+            status_age_secs: Some(90),
+            ..Default::default()
+        }],
+    );
+
+    assert!(output.output.contains("idle 1m"), "got: {}", output.output);
+}
+
 #[test]
 fn format_members_disambiguates_duplicate_friendly_names() {
     let ctx = test_ctx(
@@ -231,6 +314,7 @@ fn format_members_disambiguates_duplicate_friendly_names() {
                 latest_completion_report: None,
                 live_attachments: None,
                 status_age_secs: None,
+                ..Default::default()
             },
             AgentInfo {
                 session_id: "session_shark_1234567890_bbbbbbbbbbbb0002".to_string(),
@@ -244,6 +328,7 @@ fn format_members_disambiguates_duplicate_friendly_names() {
                 latest_completion_report: None,
                 live_attachments: None,
                 status_age_secs: None,
+                ..Default::default()
             },
         ],
     );
diff --git a/crates/jcode-app-core/src/tool/gmail.rs b/crates/jcode-app-core/src/tool/gmail.rs
index 9cbc2b2df..7afc720ff 100644
--- a/crates/jcode-app-core/src/tool/gmail.rs
+++ b/crates/jcode-app-core/src/tool/gmail.rs
@@ -4,7 +4,6 @@ use async_trait::async_trait;
 use serde::Deserialize;
 use serde_json::{Value, json};
 
-use crate::auth::google;
 use crate::gmail::{self, GmailClient, MessageFormat};
 
 pub struct GmailTool {
@@ -68,8 +67,8 @@ impl Tool for GmailTool {
                 "intent": super::intent_schema_property(),
                 "action": {
                     "type": "string",
-                    "enum": ["search", "read", "list", "draft", "send", "send_draft", "threads", "thread", "labels", "trash", "modify_labels"],
-                    "description": "Action."
+                    "enum": ["connect", "search", "read", "list", "draft", "send", "send_draft", "threads", "thread", "labels", "trash", "modify_labels"],
+                    "description": "Action. Use 'connect' to set up Gmail access via the Composio managed backend (opens a browser OAuth screen for the user to approve)."
                 },
                 "query": { "type": "string" },
                 "message_id": { "type": "string" },
@@ -92,15 +91,49 @@ impl Tool for GmailTool {
     }
 
     async fn execute(&self, input: Value, _ctx: ToolContext) -> Result<ToolOutput> {
-        if !google::has_tokens() {
+        let params: GmailInput = serde_json::from_value(input)?;
+        let max = params.max_results.unwrap_or(10).min(50);
+
+        // The connect action sets up the Composio managed backend by opening a
+        // browser OAuth screen for the user to approve. It runs before the
+        // is_configured gate so it can establish the very first connection.
+        if params.action == "connect" {
+            if !self.client.supports_connect() {
+                return Ok(ToolOutput::new(
+                    "The 'connect' action is only available with the Composio Gmail backend. \
+                     Set JCODE_GMAIL_BACKEND=composio and COMPOSIO_API_KEY, then retry. \
+                     For the default backend, run `jcode login google` instead.",
+                ));
+            }
+            let no_browser = crate::auth::browser_suppressed(false);
+            match self.client.connect(!no_browser).await {
+                Ok(conn) => {
+                    let who = conn
+                        .email
+                        .clone()
+                        .unwrap_or_else(|| "your Gmail account".to_string());
+                    return Ok(ToolOutput::new(format!(
+                        "Gmail connected via Composio for {}. You can now search, read, draft, and send email.",
+                        who
+                    )));
+                }
+                Err(e) => {
+                    return Ok(ToolOutput::new(format!("Gmail connect failed: {}", e)));
+                }
+            }
+        }
+
+        if !self.client.is_configured() {
+            return Ok(ToolOutput::new(self.client.not_configured_message()));
+        }
+
+        if self.client.needs_connection() {
             return Ok(ToolOutput::new(
-                "Gmail is not configured. Run `jcode login google` to set up Gmail access.",
+                "Gmail (Composio backend) has no connected account yet. Run the gmail tool with \
+                 action 'connect' to authorize your Gmail account, then retry.",
             ));
         }
 
-        let params: GmailInput = serde_json::from_value(input)?;
-        let max = params.max_results.unwrap_or(10).min(50);
-
         match params.action.as_str() {
             "search" | "list" => {
                 let query = params.query.as_deref();
@@ -278,8 +311,7 @@ impl Tool for GmailTool {
             }
 
             "send" => {
-                let tokens = google::load_tokens()?;
-                if !tokens.tier.can_send() {
+                if !self.client.can_send() {
                     return Ok(ToolOutput::new(
                         "Send is not available. Your Gmail access is configured as Read & Draft Only (API-level restriction).\n\
                          The draft has been created - open Gmail to send it manually.\n\
@@ -323,8 +355,7 @@ impl Tool for GmailTool {
             }
 
             "send_draft" => {
-                let tokens = google::load_tokens()?;
-                if !tokens.tier.can_send() {
+                if !self.client.can_send() {
                     return Ok(ToolOutput::new(
                         "Send is not available. Your Gmail access is configured as Read & Draft Only (API-level restriction).\n\
                          Open Gmail to send the draft manually.\n\
@@ -352,8 +383,7 @@ impl Tool for GmailTool {
             }
 
             "trash" => {
-                let tokens = google::load_tokens()?;
-                if !tokens.tier.can_delete() {
+                if !self.client.can_delete() {
                     return Ok(ToolOutput::new(
                         "Trash is not available. Your Gmail access is configured as Read & Draft Only (API-level restriction).\n\
                          To enable delete, rerun `jcode login google --google-access-tier full`.",
diff --git a/crates/jcode-app-core/src/tool/skill.rs b/crates/jcode-app-core/src/tool/skill.rs
index c2211c81c..ac5f238cb 100644
--- a/crates/jcode-app-core/src/tool/skill.rs
+++ b/crates/jcode-app-core/src/tool/skill.rs
@@ -21,7 +21,8 @@ impl SkillTool {
 
 #[derive(Deserialize)]
 struct SkillInput {
-    /// Action to perform: load (default), list, reload, reload_all, read
+    /// Action to perform: load (default), list, reload, reload_all, read.
+    /// `list` shows both loaded skills and the jcode-endorsed catalog.
     #[serde(default = "default_action")]
     action: String,
     /// Skill name (required for load, reload, read)
@@ -119,36 +120,41 @@ impl SkillTool {
 
     async fn list_skills(&self) -> Result<ToolOutput> {
         let registry = self.registry.read().await;
-        let skills = registry.list();
-
-        if skills.is_empty() {
-            return Ok(ToolOutput::new(
-                "No skills available.\n\n\
-                Skills are loaded from:\n\
-                - ~/.claude/skills/<skill-name>/SKILL.md\n\
-                - ./.claude/skills/<skill-name>/SKILL.md\n\n\
-                Create a SKILL.md file with YAML frontmatter:\n\
-                ---\n\
-                name: my-skill\n\
-                description: What this skill does\n\
-                allowed-tools: bash, read, write\n\
-                ---\n\n\
-                # Skill content here",
-            )
-            .with_title("Skills: None available"));
-        }
-
-        let mut output = format!("Available skills: {}\n\n", skills.len());
-
-        for skill in skills {
-            output.push_str(&format!("## /{}\n", skill.name));
-            output.push_str(&format!("  {}\n", skill.description));
-            output.push_str(&format!("  Path: {}\n", skill.path.display()));
-            if let Some(ref tools) = skill.allowed_tools {
-                output.push_str(&format!("  Tools: {}\n", tools.join(", ")));
+        let mut skills = registry.list();
+        skills.sort_by(|a, b| a.name.cmp(&b.name));
+
+        let installed: std::collections::HashSet<&str> =
+            skills.iter().map(|s| s.name.as_str()).collect();
+
+        let mut output = if skills.is_empty() {
+            "No skills loaded.\n\n\
+            Skills are loaded from:\n\
+            - ~/.jcode/skills/<skill-name>/SKILL.md (global)\n\
+            - ./.jcode/skills/<skill-name>/SKILL.md (project-local)\n\
+            - ./.claude/skills/<skill-name>/SKILL.md (compatibility)\n\n\
+            Create a SKILL.md file with YAML frontmatter:\n\
+            ---\n\
+            name: my-skill\n\
+            description: What this skill does\n\
+            allowed-tools: bash, read, write\n\
+            ---\n\n\
+            # Skill content here\n"
+                .to_string()
+        } else {
+            let mut output = format!("Loaded skills: {}\n\n", skills.len());
+            for skill in &skills {
+                output.push_str(&format!("## /{}\n", skill.name));
+                output.push_str(&format!("  {}\n", skill.description));
+                output.push_str(&format!("  Path: {}\n", skill.path.display()));
+                if let Some(ref tools) = skill.allowed_tools {
+                    output.push_str(&format!("  Tools: {}\n", tools.join(", ")));
+                }
+                output.push('\n');
             }
-            output.push('\n');
-        }
+            output
+        };
+
+        append_endorsed_skills(&mut output, &installed);
 
         Ok(ToolOutput::new(output).with_title("Skills: List"))
     }
@@ -243,6 +249,61 @@ impl SkillTool {
     }
 }
 
+/// Append the curated jcode-endorsed skill catalog to `output`, grouped by
+/// category and marked with installed/not-installed status. `installed` is the
+/// set of skill names currently loaded in the registry.
+fn append_endorsed_skills(output: &mut String, installed: &std::collections::HashSet<&str>) {
+    let endorsed = crate::skill::endorsed_skills();
+    if endorsed.is_empty() {
+        return;
+    }
+
+    output.push_str("\nEndorsed skills (recommended by jcode)\n");
+
+    // Group by category, preserving first-seen order.
+    let mut category_order: Vec<&str> = Vec::new();
+    for skill in endorsed {
+        if !category_order.contains(&skill.category) {
+            category_order.push(skill.category);
+        }
+    }
+
+    for category in category_order {
+        let in_category: Vec<_> = endorsed.iter().filter(|e| e.category == category).collect();
+        let installed_count = in_category
+            .iter()
+            .filter(|e| installed.contains(e.name))
+            .count();
+        output.push_str(&format!(
+            "\n  {} ({}/{} installed)\n",
+            category,
+            installed_count,
+            in_category.len()
+        ));
+        for skill in in_category {
+            let is_installed = installed.contains(skill.name);
+            let status = if is_installed {
+                "installed"
+            } else {
+                "not installed"
+            };
+            output.push_str(&format!("  - /{} [{}]\n", skill.name, status));
+            output.push_str(&format!("      {}\n", skill.description));
+            output.push_str(&format!("      source: {}\n", skill.source));
+            if !is_installed && let Some(install) = skill.install {
+                output.push_str(&format!("      install: {}\n", install));
+            }
+        }
+    }
+
+    output.push_str(
+        "\nActivate a loaded skill by loading it with skill_manage (action=load) or typing its slash command.\n",
+    );
+    output.push_str(
+        "NVIDIA CUDA-X skills come from the official catalog at https://github.com/NVIDIA/skills.\n",
+    );
+}
+
 fn normalize_skill_name(name: Option<String>, action: &str) -> Result<String> {
     let name = name.ok_or_else(|| anyhow::anyhow!("'name' is required for {} action", action))?;
     let trimmed = name.trim().trim_start_matches('/').to_string();
@@ -319,7 +380,29 @@ mod tests {
         let input = json!({"action": "list"});
 
         let result = tool.execute(input, ctx).await.unwrap();
-        assert!(result.output.contains("No skills available"));
+        assert!(result.output.contains("No skills loaded"));
+        // Even with no skills loaded, the endorsed catalog should be listed.
+        assert!(result.output.contains("Endorsed skills"));
+    }
+
+    #[tokio::test]
+    async fn test_list_includes_endorsed_skills() {
+        let tool = create_test_tool();
+        let ctx = create_test_context();
+        let input = json!({"action": "list"});
+
+        let result = tool.execute(input, ctx).await.unwrap();
+        // Every endorsed skill should appear with an install-status marker.
+        for endorsed in crate::skill::endorsed_skills() {
+            assert!(
+                result.output.contains(&format!("/{}", endorsed.name)),
+                "expected endorsed skill /{} in:\n{}",
+                endorsed.name,
+                result.output
+            );
+        }
+        // No skills are loaded in this tool, so they should be "not installed".
+        assert!(result.output.contains("[not installed]"));
     }
 
     #[tokio::test]
diff --git a/crates/jcode-app-core/src/tool/task.rs b/crates/jcode-app-core/src/tool/task.rs
index 31546dddf..2c61f127a 100644
--- a/crates/jcode-app-core/src/tool/task.rs
+++ b/crates/jcode-app-core/src/tool/task.rs
@@ -227,6 +227,7 @@ impl Tool for SubagentTool {
         // other's `children` entries. Acceptable for experimental Phase 0;
         // a file-lock or in-memory session cache would fix this properly.
         if let Ok(mut parent_session) = Session::load(&ctx.session_id) {
+            session.route_api_method = parent_session.route_api_method.clone();
             parent_session.add_child(session.id.clone());
             let _ = parent_session.save();
         }
diff --git a/crates/jcode-app-core/src/tool/todo.rs b/crates/jcode-app-core/src/tool/todo.rs
index 56174424a..42fcd87ce 100644
--- a/crates/jcode-app-core/src/tool/todo.rs
+++ b/crates/jcode-app-core/src/tool/todo.rs
@@ -57,6 +57,10 @@ impl Tool for TodoTool {
                                 "type": "string",
                                 "description": "ID."
                             },
+                            "group": {
+                                "type": "string",
+                                "description": "Optional group label. Todos sharing a group render together under one header. Use one group per coherent goal (e.g. 'optimize rendering'). When the user steers into new work, start a new group instead of renaming the existing one. Omit for an ungrouped flat list."
+                            },
                             "confidence": {
                                 "type": "integer",
                                 "minimum": 0,
diff --git a/crates/jcode-base/src/auth/live_provider_probes.rs b/crates/jcode-base/src/auth/live_provider_probes.rs
index ebc467e60..a2bb0da03 100644
--- a/crates/jcode-base/src/auth/live_provider_probes.rs
+++ b/crates/jcode-base/src/auth/live_provider_probes.rs
@@ -258,6 +258,123 @@ mod tests {
             "gpt-5.1"
         );
     }
+
+    fn tool_call_with_signature(signature: Option<&str>) -> NativeClaudeToolCall {
+        NativeClaudeToolCall {
+            id: "call_1".to_string(),
+            name: "read".to_string(),
+            input_json: "{}".to_string(),
+            thought_signature: signature.map(str::to_string),
+        }
+    }
+
+    #[test]
+    fn reasoning_capability_classifies_streamed_when_reasoning_text_present() {
+        let outcome = NativeClaudeStreamOutcome {
+            reasoning_text_len: 42,
+            saw_message_end: true,
+            ..Default::default()
+        };
+        assert_eq!(outcome.reasoning_capability(), "streamed");
+    }
+
+    #[test]
+    fn reasoning_capability_classifies_opaque_from_thinking_signature() {
+        // No reasoning text, but a ThinkingSignatureDelta-style signal: opaque.
+        let outcome = NativeClaudeStreamOutcome {
+            saw_reasoning_signal: true,
+            saw_message_end: true,
+            ..Default::default()
+        };
+        assert_eq!(outcome.reasoning_capability(), "opaque");
+    }
+
+    #[test]
+    fn reasoning_capability_classifies_opaque_from_tool_thought_signature() {
+        // A Gemini-3 tool call carrying a thought_signature is an opaque signal
+        // even when no reasoning text streamed.
+        let outcome = NativeClaudeStreamOutcome {
+            tool_calls: vec![tool_call_with_signature(Some("SIG_ABC"))],
+            saw_message_end: true,
+            ..Default::default()
+        };
+        assert_eq!(outcome.reasoning_capability(), "opaque");
+    }
+
+    #[test]
+    fn reasoning_capability_classifies_none_without_any_signal() {
+        // A tool call with no signature is not a reasoning signal.
+        let outcome = NativeClaudeStreamOutcome {
+            tool_calls: vec![tool_call_with_signature(None)],
+            saw_message_end: true,
+            ..Default::default()
+        };
+        assert_eq!(outcome.reasoning_capability(), "none");
+    }
+
+    #[test]
+    fn reasoning_capability_prefers_streamed_over_opaque() {
+        // Streamed reasoning text wins even when an opaque signal is also present.
+        let outcome = NativeClaudeStreamOutcome {
+            reasoning_text_len: 10,
+            saw_reasoning_signal: true,
+            tool_calls: vec![tool_call_with_signature(Some("SIG"))],
+            saw_message_end: true,
+            ..Default::default()
+        };
+        assert_eq!(outcome.reasoning_capability(), "streamed");
+    }
+
+    #[test]
+    fn parallel_tool_use_replays_every_signature_in_one_assistant_message() {
+        let calls = vec![
+            NativeClaudeToolCall {
+                id: "a".to_string(),
+                name: "read".to_string(),
+                input_json: "{\"file_path\":\"/tmp/a\"}".to_string(),
+                thought_signature: Some("SIG_A".to_string()),
+            },
+            NativeClaudeToolCall {
+                id: "b".to_string(),
+                name: "read".to_string(),
+                input_json: "{\"file_path\":\"/tmp/b\"}".to_string(),
+                thought_signature: Some("SIG_B".to_string()),
+            },
+        ];
+        let assistant = assistant_parallel_tool_uses(&calls);
+        assert!(matches!(assistant.role, Role::Assistant));
+        // One assistant message must carry BOTH tool_use blocks, each with its
+        // own signature preserved.
+        assert_eq!(assistant.content.len(), 2);
+        let sigs: Vec<Option<String>> = assistant
+            .content
+            .iter()
+            .map(|block| match block {
+                ContentBlock::ToolUse {
+                    thought_signature, ..
+                } => thought_signature.clone(),
+                other => panic!("expected ToolUse, got {other:?}"),
+            })
+            .collect();
+        assert_eq!(
+            sigs,
+            vec![Some("SIG_A".to_string()), Some("SIG_B".to_string())]
+        );
+
+        // The results message must answer every call with a matching id.
+        let results = parallel_tool_results(&calls);
+        assert!(matches!(results.role, Role::User));
+        assert_eq!(results.content.len(), 2);
+        let ids: Vec<String> = results
+            .content
+            .iter()
+            .map(|block| match block {
+                ContentBlock::ToolResult { tool_use_id, .. } => tool_use_id.clone(),
+                other => panic!("expected ToolResult, got {other:?}"),
+            })
+            .collect();
+        assert_eq!(ids, vec!["a".to_string(), "b".to_string()]);
+    }
 }
 
 pub async fn run_live_openai_compatible_stream_smoke(
@@ -528,6 +645,15 @@ struct NativeClaudeStreamOutcome {
     /// Number of thinking deltas seen (extended/adaptive thinking). Useful when
     /// a turn is consumed entirely by reasoning and emits no visible text.
     thinking_chunk_count: usize,
+    /// Length of streamed reasoning text (sum of `ThinkingDelta` payloads).
+    /// Distinct from `thinking_chunk_count`: a provider can emit a single empty
+    /// `ThinkingStart`/`ThinkingEnd` pair without ever streaming visible
+    /// reasoning text, which we must classify as `opaque`/`none`, not `streamed`.
+    reasoning_text_len: usize,
+    /// Saw an *opaque* reasoning signal: a `thought_signature` (Gemini-3), a
+    /// `ThinkingSignatureDelta`, or an `OpenAIReasoning` item. This is the
+    /// evidence that the model reasoned even though it never streamed the text.
+    saw_reasoning_signal: bool,
     /// Total stream events observed, for diagnosing empty/odd streams.
     total_events: usize,
     saw_message_end: bool,
@@ -576,6 +702,33 @@ impl NativeClaudeStreamOutcome {
             self.tool_calls.len()
         )
     }
+
+    /// Did any captured tool call carry a Gemini-3 `thought_signature`? This is
+    /// an opaque reasoning signal even when the model streamed no reasoning text.
+    fn any_tool_signature(&self) -> bool {
+        self.tool_calls
+            .iter()
+            .any(|call| call.thought_signature.is_some())
+    }
+
+    /// Classify how this turn exposed the model's reasoning:
+    /// - `streamed`: streamed visible reasoning text (`ThinkingDelta`).
+    /// - `opaque`: no reasoning text, but an opaque reasoning signal was present
+    ///   (a `thought_signature`, a `ThinkingSignatureDelta`, or an
+    ///   `OpenAIReasoning` item). Legitimate and common (Gemini-3, OpenAI).
+    /// - `none`: neither was observed.
+    ///
+    /// All three are valid; the reasoning checkpoint records the classification
+    /// and never fails on `none`.
+    fn reasoning_capability(&self) -> &'static str {
+        if self.reasoning_text_len > 0 {
+            "streamed"
+        } else if self.saw_reasoning_signal || self.any_tool_signature() {
+            "opaque"
+        } else {
+            "none"
+        }
+    }
 }
 
 /// Drive any native [`Provider`] runtime's `complete` and fold the resulting
@@ -610,8 +763,19 @@ async fn consume_native_stream(
                     outcome.chunk_count += 1;
                     outcome.text.push_str(&text);
                 }
-                StreamEvent::ThinkingDelta(_) => {
+                StreamEvent::ThinkingDelta(text) => {
                     outcome.thinking_chunk_count += 1;
+                    outcome.reasoning_text_len += text.len();
+                }
+                // Opaque reasoning signals: the model reasoned but the runtime
+                // surfaces only a signature/encrypted item, not readable text.
+                StreamEvent::ThinkingSignatureDelta(signature) => {
+                    if !signature.is_empty() {
+                        outcome.saw_reasoning_signal = true;
+                    }
+                }
+                StreamEvent::OpenAIReasoning { .. } => {
+                    outcome.saw_reasoning_signal = true;
                 }
                 StreamEvent::ToolUseStart { id, name } => {
                     pending_tool = Some(NativeClaudeToolCall {
@@ -634,10 +798,10 @@ async fn consume_native_stream(
                 // Emitted after the matching `ToolUseEnd`; attach it to the most
                 // recent tool call so probes can replay it on the next turn.
                 StreamEvent::ToolUseSignature(signature) => {
-                    if let Some(tool) = outcome.tool_calls.last_mut()
-                        && !signature.is_empty()
-                    {
-                        tool.thought_signature = Some(signature);
+                    if let Some(tool) = outcome.tool_calls.last_mut() {
+                        if !signature.is_empty() {
+                            tool.thought_signature = Some(signature);
+                        }
                     }
                 }
                 StreamEvent::TokenUsage {
@@ -971,13 +1135,28 @@ pub async fn run_live_claude_native_tool_smoke(
     .with_evidence("model", serde_json::json!(model))
     .with_evidence("tool_name", serde_json::json!(tool_call.name))
     .with_evidence("tool_arguments", parsed_arguments)
-    .with_evidence("followup_consumed_result", serde_json::json!(true));
+    .with_evidence(
+        "followup_consumed_result",
+        serde_json::json!(true),
+    );
     if total_input != 0 || total_output != 0 {
         stage = stage.with_evidence("usage", usage_evidence(total_input, total_output, 0, 0));
     }
     Ok(stage)
 }
 
+/// Stage: reasoning capability (observe-only).
+///
+/// Delegates to the shared [`run_live_native_provider_reasoning_smoke`] so the
+/// native Claude runtime records whether the model streamed reasoning text
+/// (extended thinking) or hid it behind an opaque signal.
+pub async fn run_live_claude_native_reasoning_smoke(
+    model: &str,
+) -> anyhow::Result<crate::live_tests::LiveVerificationStage> {
+    let provider = build_native_claude_provider(model)?;
+    run_live_native_provider_reasoning_smoke(&provider, model, "Claude").await
+}
+
 // === Native Antigravity probes ============================================
 //
 // Antigravity is a Google OAuth login provider whose `generateContent` runtime
@@ -1159,6 +1338,18 @@ pub async fn run_live_antigravity_native_tool_smoke(
     run_live_native_provider_tool_smoke(&provider, model, "Antigravity").await
 }
 
+/// Stage: reasoning capability (observe-only).
+///
+/// Delegates to the shared [`run_live_native_provider_reasoning_smoke`] so
+/// Antigravity records whether the resolved model streams reasoning text or
+/// hides it behind an opaque signal (Gemini-3 thought signatures are opaque).
+pub async fn run_live_antigravity_native_reasoning_smoke(
+    model: &str,
+) -> anyhow::Result<crate::live_tests::LiveVerificationStage> {
+    let provider = build_native_antigravity_provider(model)?;
+    run_live_native_provider_reasoning_smoke(&provider, model, "Antigravity").await
+}
+
 // === Generic native-runtime probes ========================================
 //
 // The native Claude and native Antigravity probes above each build a concrete
@@ -1219,10 +1410,7 @@ pub async fn run_live_native_provider_smoke(
     .with_duration_ms(started.elapsed().as_millis() as u64)
     .with_evidence("model", serde_json::json!(model))
     .with_evidence("matched_expected_content", serde_json::json!(true))
-    .with_evidence(
-        "stop_reason",
-        serde_json::json!(outcome.stop_reason.clone()),
-    );
+    .with_evidence("stop_reason", serde_json::json!(outcome.stop_reason.clone()));
     if let Some(usage) = outcome.usage_evidence() {
         stage = stage.with_evidence("usage", usage);
     }
@@ -1310,10 +1498,108 @@ pub async fn run_live_native_provider_stream_smoke(
     .with_evidence("attempts", serde_json::json!(attempts))
     .with_evidence("total_events", serde_json::json!(outcome.total_events))
     .with_evidence("matched_expected_content", serde_json::json!(true))
-    .with_evidence(
-        "stop_reason",
-        serde_json::json!(outcome.stop_reason.clone()),
+    .with_evidence("stop_reason", serde_json::json!(outcome.stop_reason.clone()));
+    if let Some(usage) = outcome.usage_evidence() {
+        stage = stage.with_evidence("usage", usage);
+    }
+    Ok(stage)
+}
+
+/// Stage: reasoning capability (observe-only).
+///
+/// Sends a small multi-step logic/word problem that forces the model to reason
+/// before answering, consumes the stream, and classifies how the model exposed
+/// its reasoning:
+///
+/// - `streamed`: the runtime streamed visible reasoning text (`ThinkingDelta`).
+/// - `opaque`: no reasoning text, but an opaque reasoning signal was present (a
+///   Gemini-3 `thought_signature`, a `ThinkingSignatureDelta`, or an
+///   `OpenAIReasoning` item). This is legitimate and common (Gemini-3 and
+///   OpenAI hide their reasoning), so it MUST be a pass.
+/// - `none`: neither was observed.
+///
+/// The checkpoint passes as long as the turn completes cleanly (a `MessageEnd`
+/// plus a coherent answer); it never hard-fails just because reasoning was
+/// hidden or absent. The classification is recorded as the `reasoning_capability`
+/// evidence. Expected-to-reason gating (a capability list) can layer on later.
+pub async fn run_live_native_provider_reasoning_smoke(
+    provider: &dyn Provider,
+    model: &str,
+    label: &str,
+) -> anyhow::Result<crate::live_tests::LiveVerificationStage> {
+    let started = std::time::Instant::now();
+    // A small logic word problem with a single unambiguous numeric answer (4
+    // cows: chickens c + cows w give c + w = 7 heads and 2c + 4w = 22 legs, so
+    // w = 4). The `REASON_TEST_ANSWER=<n>` sentinel lets us assert a coherent
+    // result without depending on the model's prose, and the problem requires at
+    // least one elimination/arithmetic step so a reasoning model actually reasons.
+    let messages = vec![Message {
+        role: Role::User,
+        content: vec![ContentBlock::Text {
+            text: "Solve this step by step, then give the final answer. A farmer has chickens \
+                   and cows. Together they have 7 heads and 22 legs. How many cows are there? \
+                   After reasoning, end your reply with exactly REASON_TEST_ANSWER=<number> on \
+                   its own final line."
+                .to_string(),
+            cache_control: None,
+        }],
+        timestamp: None,
+        tool_duration_ms: None,
+    }];
+    let system = "You are a live provider reasoning smoke test. Think through the problem, then \
+                  finish with the required REASON_TEST_ANSWER=<number> line.";
+
+    let outcome = consume_native_stream(
+        provider,
+        &messages,
+        &[],
+        system,
+        std::time::Duration::from_secs(120),
+    )
+    .await?;
+
+    ensure!(
+        outcome.saw_message_end,
+        "native {label} reasoning smoke ended without a message_end event ({})",
+        outcome.diagnostics()
     );
+    // Coherence: the turn must produce a real final answer. We accept either the
+    // exact sentinel or the correct numeric answer (4 cows) appearing in the
+    // text, so a model that ignores the formatting instruction but still answers
+    // correctly is not penalized. The reasoning checkpoint is about completion,
+    // not about reasoning visibility.
+    let answered = outcome.text.contains("REASON_TEST_ANSWER=4")
+        || outcome.text.contains("REASON_TEST_ANSWER= 4")
+        || outcome.text.to_ascii_lowercase().contains("4 cows")
+        || outcome.text.contains("REASON_TEST_ANSWER");
+    ensure!(
+        !outcome.text.trim().is_empty() && answered,
+        "native {label} reasoning smoke produced no coherent answer: {:?} ({})",
+        crate::util::truncate_str(outcome.text.trim(), 200),
+        outcome.diagnostics()
+    );
+
+    let classification = outcome.reasoning_capability();
+    let mut stage = crate::live_tests::LiveVerificationStage::passed(
+        crate::live_tests::checkpoints::REASONING_CAPABILITY,
+    )
+    .with_duration_ms(started.elapsed().as_millis() as u64)
+    .with_evidence("model", serde_json::json!(model))
+    .with_evidence("reasoning_capability", serde_json::json!(classification))
+    .with_evidence(
+        "reasoning_text_chars",
+        serde_json::json!(outcome.reasoning_text_len),
+    )
+    .with_evidence(
+        "thinking_delta_count",
+        serde_json::json!(outcome.thinking_chunk_count),
+    )
+    .with_evidence(
+        "saw_opaque_reasoning_signal",
+        serde_json::json!(outcome.saw_reasoning_signal),
+    )
+    .with_evidence("total_events", serde_json::json!(outcome.total_events))
+    .with_evidence("stop_reason", serde_json::json!(outcome.stop_reason.clone()));
     if let Some(usage) = outcome.usage_evidence() {
         stage = stage.with_evidence("usage", usage);
     }
@@ -1323,7 +1609,7 @@ pub async fn run_live_native_provider_stream_smoke(
 /// Stage: tool-call parse + execution loop + result follow-up against an
 /// arbitrary native provider.
 ///
-/// Two phases:
+/// Three phases:
 ///
 /// 1. **Single round-trip (gating):** ask the model to call a tool (assert a
 ///    parseable tool_use), then feed a synthetic tool_result back (assert the
@@ -1341,6 +1627,14 @@ pub async fn run_live_native_provider_stream_smoke(
 ///    signatures at all), the phase records `multi_tool_replay: "skipped"`
 ///    rather than failing, so it never turns a previously-green provider red
 ///    for a non-signature reason.
+/// 3. **Parallel tool calls in one turn (best-effort):** ask the model to call
+///    the tool TWICE in a single assistant message, then replay BOTH `tool_use`
+///    blocks (each with its own `thought_signature`) inside one assistant turn
+///    and answer both `tool_result`s, asserting the backend accepts a single
+///    assistant message carrying two `functionCall` parts. Distinct from the
+///    sequential loop in phase 2. Records `parallel_tool_calls: "verified"` when
+///    the model emitted >=2 calls in one turn and the follow-up was accepted, or
+///    `"skipped"` when the model only emitted one (best-effort, never a fail).
 pub async fn run_live_native_provider_tool_smoke(
     provider: &dyn Provider,
     model: &str,
@@ -1524,6 +1818,86 @@ pub async fn run_live_native_provider_tool_smoke(
         }
     }
 
+    // Phase 3 (best-effort): ask the model to call the tool TWICE in a single
+    // assistant turn (parallel/batch tool calls), then replay BOTH tool_use
+    // blocks inside ONE assistant message (each carrying its own captured
+    // thought_signature) and answer BOTH tool_results. A backend that accepts a
+    // single assistant message containing two `functionCall` parts completes the
+    // follow-up cleanly; one that rejects parallel calls surfaces here. If the
+    // model only emits a single call (common: many models serialize tool use),
+    // we record `parallel_tool_calls: "skipped"` rather than failing.
+    let mut parallel_tool_calls = "skipped";
+    let mut parallel_call_count = 0usize;
+    let parallel_turn = consume_native_stream(
+        provider,
+        &[Message {
+            role: Role::User,
+            content: vec![ContentBlock::Text {
+                text: "In this single turn, make TWO read tool calls at once (in parallel, in \
+                       one message): read /tmp/auth_tool_probe.txt AND read \
+                       /tmp/auth_tool_probe_2.txt. Emit both tool calls now; do not answer in \
+                       text and do not wait for the first result before making the second call."
+                    .to_string(),
+                cache_control: None,
+            }],
+            timestamp: None,
+            tool_duration_ms: None,
+        }],
+        &tools,
+        system,
+        std::time::Duration::from_secs(120),
+    )
+    .await?;
+    total_input += parallel_turn.input_tokens;
+    total_output += parallel_turn.output_tokens;
+
+    if parallel_turn.tool_calls.len() >= 2 {
+        parallel_call_count = parallel_turn.tool_calls.len();
+        // Build ONE assistant message holding every tool_use block (each with
+        // its own signature), then ONE user message holding every tool_result.
+        let assistant = assistant_parallel_tool_uses(&parallel_turn.tool_calls);
+        let results = parallel_tool_results(&parallel_turn.tool_calls);
+        let convo = vec![
+            Message {
+                role: Role::User,
+                content: vec![ContentBlock::Text {
+                    text: "In this single turn, make TWO read tool calls at once (in parallel, \
+                           in one message): read /tmp/auth_tool_probe.txt AND read \
+                           /tmp/auth_tool_probe_2.txt."
+                        .to_string(),
+                    cache_control: None,
+                }],
+                timestamp: None,
+                tool_duration_ms: None,
+            },
+            assistant,
+            results,
+        ];
+        let parallel_followup = consume_native_stream(
+            provider,
+            &convo,
+            &tools,
+            system,
+            std::time::Duration::from_secs(120),
+        )
+        .await
+        .with_context(|| {
+            format!(
+                "native {label} parallel tool-call replay was rejected (one assistant message \
+                 carried {parallel_call_count} functionCall parts; a backend that does not \
+                 accept parallel tool calls in a single message fails here)"
+            )
+        })?;
+        total_input += parallel_followup.input_tokens;
+        total_output += parallel_followup.output_tokens;
+        ensure!(
+            parallel_followup.saw_message_end,
+            "native {label} parallel tool-call follow-up ended without a message_end event ({})",
+            parallel_followup.diagnostics()
+        );
+        parallel_tool_calls = "verified";
+    }
+
     let mut stage = crate::live_tests::LiveVerificationStage::passed(
         crate::live_tests::checkpoints::TOOL_CALL_PARSE,
     )
@@ -1541,6 +1915,14 @@ pub async fn run_live_native_provider_tool_smoke(
         "tool_call_signatures_present",
         serde_json::json!(signatures_present),
     )
+    .with_evidence(
+        "parallel_tool_calls",
+        serde_json::json!(parallel_tool_calls),
+    )
+    .with_evidence(
+        "parallel_tool_call_count",
+        serde_json::json!(parallel_call_count),
+    )
     .with_evidence("followup_consumed_result", serde_json::json!(true));
     if total_input != 0 || total_output != 0 {
         stage = stage.with_evidence("usage", usage_evidence(total_input, total_output, 0, 0));
@@ -1589,3 +1971,47 @@ fn tool_result_then_text(tool_use_id: &str, result: &str) -> Message {
         tool_duration_ms: None,
     }
 }
+
+/// Build a single assistant message that replays *every* captured tool call as a
+/// parallel batch (multiple `ToolUse` blocks in one message), each preserving
+/// its own `thought_signature`. This is the shape the parallel-tool-call phase
+/// asserts the backend accepts as one assistant turn carrying N `functionCall`
+/// parts.
+fn assistant_parallel_tool_uses(calls: &[NativeClaudeToolCall]) -> Message {
+    let content = calls
+        .iter()
+        .map(|call| ContentBlock::ToolUse {
+            id: call.id.clone(),
+            name: call.name.clone(),
+            input: parse_tool_arguments(&call.input_json),
+            thought_signature: call.thought_signature.clone(),
+        })
+        .collect();
+    Message {
+        role: Role::Assistant,
+        content,
+        timestamp: None,
+        tool_duration_ms: None,
+    }
+}
+
+/// Build a single user message answering *every* parallel tool call with a
+/// synthetic `tool_result`, so a parallel assistant turn is fully resolved in
+/// one follow-up message.
+fn parallel_tool_results(calls: &[NativeClaudeToolCall]) -> Message {
+    let content = calls
+        .iter()
+        .enumerate()
+        .map(|(index, call)| ContentBlock::ToolResult {
+            tool_use_id: call.id.clone(),
+            content: format!("Contents of file {}: token_{index}.", index + 1),
+            is_error: Some(false),
+        })
+        .collect();
+    Message {
+        role: Role::User,
+        content,
+        timestamp: None,
+        tool_duration_ms: None,
+    }
+}
diff --git a/crates/jcode-base/src/auth/provider_e2e.rs b/crates/jcode-base/src/auth/provider_e2e.rs
index 0391cce40..05ef8c40a 100644
--- a/crates/jcode-base/src/auth/provider_e2e.rs
+++ b/crates/jcode-base/src/auth/provider_e2e.rs
@@ -22,13 +22,14 @@ use crate::auth::lifecycle::{
     AuthActivationRequest, activate_auth_change, validate_catalog_invariants,
 };
 use crate::auth::live_provider_probes::{
-    fetch_live_openai_compatible_models, run_live_antigravity_native_smoke,
-    run_live_antigravity_native_stream_smoke, run_live_antigravity_native_tool_smoke,
+    fetch_live_openai_compatible_models, run_live_antigravity_native_reasoning_smoke,
+    run_live_antigravity_native_smoke, run_live_antigravity_native_stream_smoke,
+    run_live_antigravity_native_tool_smoke, run_live_claude_native_reasoning_smoke,
     run_live_claude_native_smoke, run_live_claude_native_stream_smoke,
-    run_live_claude_native_tool_smoke, run_live_native_provider_smoke,
-    run_live_native_provider_stream_smoke, run_live_native_provider_tool_smoke,
-    run_live_openai_compatible_smoke, run_live_openai_compatible_stream_smoke,
-    run_live_openai_compatible_tool_smoke,
+    run_live_claude_native_tool_smoke, run_live_native_provider_reasoning_smoke,
+    run_live_native_provider_smoke, run_live_native_provider_stream_smoke,
+    run_live_native_provider_tool_smoke, run_live_openai_compatible_smoke,
+    run_live_openai_compatible_stream_smoke, run_live_openai_compatible_tool_smoke,
 };
 use crate::live_tests::{
     self, LiveVerificationAuth, LiveVerificationEvent, LiveVerificationResult,
@@ -273,6 +274,7 @@ const FULL_PIPELINE_LABELS: &[(&str, &str)] = &[
     (checkpoints::TOOL_EXECUTION_LOOP, "Tool execution loop"),
     (checkpoints::TOOL_RESULT_FOLLOWUP, "Tool-result followup"),
     (checkpoints::REAL_JCODE_TOOL_SMOKE, "Real Jcode tool smoke"),
+    (checkpoints::REASONING_CAPABILITY, "Reasoning capability"),
 ];
 
 fn label_for(checkpoint: &str) -> &'static str {
@@ -291,17 +293,89 @@ fn label_for(checkpoint: &str) -> &'static str {
 /// declined a second tool call). Surfacing it keeps the coverage observable in
 /// the doctor report instead of collapsing to a generic pass string.
 fn tool_stage_detail(stage: &crate::live_tests::LiveVerificationStage) -> String {
-    match stage
+    let multi = match stage
         .evidence
         .get("multi_tool_replay")
         .and_then(|value| value.as_str())
     {
-        Some("verified") => "tool call parsed and executed; multi-call signature replay verified".to_string(),
-        Some("skipped") => {
-            "tool call parsed and executed; multi-call signature replay skipped (no 2nd tool call)"
-                .to_string()
+        Some("verified") => "multi-call signature replay verified",
+        Some("skipped") => "multi-call signature replay skipped (no 2nd tool call)",
+        _ => "",
+    };
+    let parallel = match stage
+        .evidence
+        .get("parallel_tool_calls")
+        .and_then(|value| value.as_str())
+    {
+        Some("verified") => "parallel tool calls verified",
+        Some("skipped") => "parallel tool calls skipped (single call)",
+        _ => "",
+    };
+    let mut detail = "tool call parsed and executed".to_string();
+    for part in [multi, parallel] {
+        if !part.is_empty() {
+            detail.push_str("; ");
+            detail.push_str(part);
         }
-        _ => "tool call parsed and executed".to_string(),
+    }
+    detail
+}
+
+/// Human-readable detail for a passed reasoning-capability stage. The stage
+/// records `reasoning_capability` as `streamed` (visible reasoning text),
+/// `opaque` (no text but a reasoning signal: thought signature, reasoning item,
+/// or reasoning tokens), or `none` (neither). All three are passes; `opaque` and
+/// `none` are legitimate because providers like Gemini-3 and OpenAI hide their
+/// reasoning. Surfacing the classification keeps the observation visible in the
+/// doctor report.
+fn reasoning_stage_detail(stage: &crate::live_tests::LiveVerificationStage) -> String {
+    match stage
+        .evidence
+        .get("reasoning_capability")
+        .and_then(|value| value.as_str())
+    {
+        Some("streamed") => "reasoning streamed (visible thinking text)".to_string(),
+        Some("opaque") => {
+            "reasoning hidden but signaled (opaque: thought signature / reasoning item)".to_string()
+        }
+        Some("none") => "no reasoning signal observed (model hides or skips reasoning)".to_string(),
+        _ => "reasoning turn completed".to_string(),
+    }
+}
+
+/// Fold a reasoning-capability probe result into a [`DoctorCheck`], honoring the
+/// observe-only contract.
+///
+/// A clean turn records a passed checkpoint carrying the `streamed`/`opaque`/
+/// `none` classification (all three are passes; hiding reasoning is legitimate).
+/// A probe *error* (network, or a turn that did not complete with a coherent
+/// answer) is recorded as **skipped**, never failed: this checkpoint must never
+/// flip a provider to "not user-ready", and it is not part of the strict
+/// coverage ladder, so an observational miss should not fail the tier. The
+/// broader chat/streaming checkpoints already guard turn completion.
+fn push_reasoning_check(
+    result: anyhow::Result<LiveVerificationStage>,
+    checks: &mut Vec<DoctorCheck>,
+    spend: &mut DoctorSpend,
+) {
+    match result {
+        Ok(stage) => {
+            spend.accumulate(stage.evidence.get("usage"), stage.evidence.get("cost"));
+            let detail = reasoning_stage_detail(&stage);
+            checks.push(DoctorCheck::passed(
+                checkpoints::REASONING_CAPABILITY,
+                label_for(checkpoints::REASONING_CAPABILITY),
+                detail,
+            ));
+        }
+        Err(error) => checks.push(DoctorCheck::skipped(
+            checkpoints::REASONING_CAPABILITY,
+            label_for(checkpoints::REASONING_CAPABILITY),
+            format!(
+                "observe-only reasoning probe did not complete: {}",
+                format_error_chain(&error)
+            ),
+        )),
     }
 }
 
@@ -314,6 +388,7 @@ const API_DEPENDENT_CHECKPOINTS: &[&str] = &[
     checkpoints::TOOL_EXECUTION_LOOP,
     checkpoints::TOOL_RESULT_FOLLOWUP,
     checkpoints::REAL_JCODE_TOOL_SMOKE,
+    checkpoints::REASONING_CAPABILITY,
 ];
 
 /// Run the strict provider/model diagnostic.
@@ -532,8 +607,8 @@ pub async fn run_claude_native_e2e(
     use crate::provider::Provider;
     use crate::provider::anthropic::AnthropicProvider;
 
-    let normalized =
-        crate::auth::lifecycle::normalized_auth_provider_id(Some(provider_id)).unwrap_or("claude");
+    let normalized = crate::auth::lifecycle::normalized_auth_provider_id(Some(provider_id))
+        .unwrap_or("claude");
     let provider_label = crate::auth::lifecycle::provider_display_label(Some(normalized))
         .unwrap_or_else(|| "Anthropic/Claude".to_string());
     let provider_id = normalized.to_string();
@@ -573,11 +648,7 @@ pub async fn run_claude_native_e2e(
     let credential_is_oauth = if tier.requires_api_key() {
         match provider_runtime.resolve_access_token_for_doctor().await {
             Ok((token, is_oauth)) if !token.trim().is_empty() => {
-                let kind = if is_oauth {
-                    "OAuth (subscription)"
-                } else {
-                    "API key"
-                };
+                let kind = if is_oauth { "OAuth (subscription)" } else { "API key" };
                 checks.push(DoctorCheck::passed(
                     checkpoints::AUTH_CREDENTIAL_LOADED,
                     label_for(checkpoints::AUTH_CREDENTIAL_LOADED),
@@ -850,6 +921,13 @@ async fn run_native_claude_api_checks(
             }
         }
     }
+
+    // Reasoning capability (observe-only; never gates readiness).
+    push_reasoning_check(
+        run_live_claude_native_reasoning_smoke(selected).await,
+        checks,
+        spend,
+    );
 }
 
 /// The wiring contract for the native Antigravity (Google OAuth Cloud Code)
@@ -890,10 +968,14 @@ fn native_antigravity_auth(account: &str) -> LiveVerificationAuth {
 /// the caller fall back to the runtime default.
 fn cheapest_antigravity_model(catalog_models: &[String]) -> Option<String> {
     let is_alias = |m: &&String| m.trim().is_empty() || m.trim() == "default";
-    if let Some(flash) = catalog_models.iter().filter(|m| !is_alias(m)).find(|m| {
-        let lower = m.to_ascii_lowercase();
-        lower.starts_with("gemini") && lower.contains("flash")
-    }) {
+    if let Some(flash) = catalog_models
+        .iter()
+        .filter(|m| !is_alias(m))
+        .find(|m| {
+            let lower = m.to_ascii_lowercase();
+            lower.starts_with("gemini") && lower.contains("flash")
+        })
+    {
         return Some(flash.clone());
     }
     if let Some(gemini) = catalog_models
@@ -903,7 +985,10 @@ fn cheapest_antigravity_model(catalog_models: &[String]) -> Option<String> {
     {
         return Some(gemini.clone());
     }
-    catalog_models.iter().find(|m| !is_alias(m)).cloned()
+    catalog_models
+        .iter()
+        .find(|m| !is_alias(m))
+        .cloned()
 }
 
 /// Run the strict provider/model diagnostic for the **native Antigravity**
@@ -1187,6 +1272,13 @@ async fn run_native_antigravity_api_checks(
             }
         }
     }
+
+    // Reasoning capability (observe-only; never gates readiness).
+    push_reasoning_check(
+        run_live_antigravity_native_reasoning_smoke(selected).await,
+        checks,
+        spend,
+    );
 }
 
 // === Generic native-runtime doctor =========================================
@@ -1344,8 +1436,8 @@ impl NativeProviderKind {
     /// Returns an error only when the runtime cannot be constructed at all (e.g.
     /// Copilot with no credential file); model selection happens later.
     fn build_runtime(self) -> anyhow::Result<std::sync::Arc<dyn crate::provider::Provider>> {
-        use crate::provider::Provider;
         use anyhow::Context as _;
+        use crate::provider::Provider;
         let runtime: std::sync::Arc<dyn Provider> = match self {
             Self::OpenAi => {
                 let credentials = crate::auth::codex::load_credentials().unwrap_or_else(|_| {
@@ -1360,7 +1452,9 @@ impl NativeProviderKind {
                 std::sync::Arc::new(crate::provider::openai::OpenAIProvider::new(credentials))
             }
             Self::Gemini => std::sync::Arc::new(crate::provider::gemini::GeminiProvider::new()),
-            Self::Cursor => std::sync::Arc::new(crate::provider::cursor::CursorCliProvider::new()),
+            Self::Cursor => {
+                std::sync::Arc::new(crate::provider::cursor::CursorCliProvider::new())
+            }
             Self::Copilot => {
                 // `new()` requires a loadable GitHub token; fall back to an empty
                 // token so the offline tier can still construct the runtime for
@@ -1375,14 +1469,18 @@ impl NativeProviderKind {
                 crate::env::set_var("JCODE_COPILOT_PREFETCH_STARTUP_GRACE_MS", "0");
                 let runtime = match crate::provider::copilot::CopilotApiProvider::new() {
                     Ok(runtime) => runtime,
-                    Err(_) => {
-                        crate::provider::copilot::CopilotApiProvider::new_with_token(String::new())
-                    }
+                    Err(_) => crate::provider::copilot::CopilotApiProvider::new_with_token(
+                        String::new(),
+                    ),
                 };
                 std::sync::Arc::new(runtime)
             }
-            Self::Bedrock => std::sync::Arc::new(crate::provider::bedrock::BedrockProvider::new()),
-            Self::Jcode => std::sync::Arc::new(crate::provider::jcode::JcodeProvider::new()),
+            Self::Bedrock => {
+                std::sync::Arc::new(crate::provider::bedrock::BedrockProvider::new())
+            }
+            Self::Jcode => {
+                std::sync::Arc::new(crate::provider::jcode::JcodeProvider::new())
+            }
             Self::Azure => {
                 // Azure OpenAI is the OpenRouter transport configured via Azure
                 // env; apply that env (endpoint/key/header wiring) before building
@@ -1713,14 +1811,8 @@ pub async fn run_generic_native_e2e(
                 ));
             }
         } else {
-            run_generic_native_api_checks(
-                runtime.as_ref(),
-                &selected,
-                spec.label,
-                &mut checks,
-                &mut spend,
-            )
-            .await;
+            run_generic_native_api_checks(runtime.as_ref(), &selected, spec.label, &mut checks, &mut spend)
+                .await;
         }
     } else {
         for checkpoint in API_DEPENDENT_CHECKPOINTS {
@@ -1822,6 +1914,13 @@ async fn run_generic_native_api_checks(
             }
         }
     }
+
+    // Reasoning capability (observe-only; never gates readiness).
+    push_reasoning_check(
+        run_live_native_provider_reasoning_smoke(provider, selected, label).await,
+        checks,
+        spend,
+    );
 }
 
 /// The jcode-side wiring a given compat profile is expected to activate.
@@ -2417,9 +2516,7 @@ mod tests {
         // OpenAI-compatible profiles are driven by the generic doctor, not the
         // native path.
         assert!(!native_doctor_supports_provider("openrouter"));
-        assert!(!native_doctor_supports_provider(
-            "definitely-not-a-provider"
-        ));
+        assert!(!native_doctor_supports_provider("definitely-not-a-provider"));
     }
 
     #[test]
@@ -2476,4 +2573,71 @@ mod tests {
         let anonymous = native_antigravity_auth("");
         assert!(anonymous.source.contains("Antigravity Google OAuth"));
     }
+
+    #[test]
+    fn tool_stage_detail_surfaces_multi_and_parallel_phases() {
+        let verified = LiveVerificationStage::passed(checkpoints::TOOL_CALL_PARSE)
+            .with_evidence("multi_tool_replay", serde_json::json!("verified"))
+            .with_evidence("parallel_tool_calls", serde_json::json!("verified"));
+        let detail = tool_stage_detail(&verified);
+        assert!(detail.contains("tool call parsed and executed"));
+        assert!(detail.contains("multi-call signature replay verified"));
+        assert!(detail.contains("parallel tool calls verified"));
+
+        let skipped = LiveVerificationStage::passed(checkpoints::TOOL_CALL_PARSE)
+            .with_evidence("multi_tool_replay", serde_json::json!("skipped"))
+            .with_evidence("parallel_tool_calls", serde_json::json!("skipped"));
+        let detail = tool_stage_detail(&skipped);
+        assert!(detail.contains("multi-call signature replay skipped"));
+        assert!(detail.contains("parallel tool calls skipped"));
+
+        // With no evidence the base string is unchanged (back-compat).
+        let bare = LiveVerificationStage::passed(checkpoints::TOOL_CALL_PARSE);
+        assert_eq!(tool_stage_detail(&bare), "tool call parsed and executed");
+    }
+
+    #[test]
+    fn reasoning_stage_detail_describes_each_classification() {
+        for (value, needle) in [
+            ("streamed", "reasoning streamed"),
+            ("opaque", "reasoning hidden but signaled"),
+            ("none", "no reasoning signal observed"),
+        ] {
+            let stage = LiveVerificationStage::passed(checkpoints::REASONING_CAPABILITY)
+                .with_evidence("reasoning_capability", serde_json::json!(value));
+            assert!(
+                reasoning_stage_detail(&stage).contains(needle),
+                "classification {value} should mention {needle}"
+            );
+        }
+    }
+
+    #[test]
+    fn push_reasoning_check_records_pass_for_clean_turn() {
+        let mut checks = Vec::new();
+        let mut spend = DoctorSpend::default();
+        let stage = LiveVerificationStage::passed(checkpoints::REASONING_CAPABILITY)
+            .with_evidence("reasoning_capability", serde_json::json!("opaque"));
+        push_reasoning_check(Ok(stage), &mut checks, &mut spend);
+        assert_eq!(checks.len(), 1);
+        assert_eq!(checks[0].checkpoint, checkpoints::REASONING_CAPABILITY);
+        assert_eq!(checks[0].status, LiveVerificationStageStatus::Passed);
+        assert!(!checks[0].is_failure());
+    }
+
+    #[test]
+    fn push_reasoning_check_skips_never_fails_on_probe_error() {
+        // The observe-only reasoning checkpoint must never produce a failure that
+        // could flip the tier to not-ready; a probe error is recorded as skipped.
+        let mut checks = Vec::new();
+        let mut spend = DoctorSpend::default();
+        push_reasoning_check(
+            Err(anyhow::anyhow!("network blip")),
+            &mut checks,
+            &mut spend,
+        );
+        assert_eq!(checks.len(), 1);
+        assert_eq!(checks[0].status, LiveVerificationStageStatus::Skipped);
+        assert!(!checks[0].is_failure());
+    }
 }
diff --git a/crates/jcode-base/src/config/default_file.rs b/crates/jcode-base/src/config/default_file.rs
index 9d05c4695..c4799c611 100644
--- a/crates/jcode-base/src/config/default_file.rs
+++ b/crates/jcode-base/src/config/default_file.rs
@@ -114,8 +114,8 @@ mouse_capture = true
 # Enable debug socket for external control/testing (default: false)
 debug_socket = false
 
-# Show thinking/reasoning content (default: false)
-show_thinking = false
+# Show thinking/reasoning content (default: true)
+show_thinking = true
 
 # How to display reasoning/thinking content: "off", "full", or "current".
 #   off     - never show reasoning
@@ -123,7 +123,7 @@ show_thinking = false
 #   current - show only the live reasoning; collapse it once the model commits
 #             an assistant message or runs a tool, then show the next one
 # When unset, falls back to show_thinking (true => full, false => off).
-# reasoning_display = "current"
+reasoning_display = "current"
 
 # Markdown spacing style: "compact" (chat/TUI) or "document" (docs-like)
 # markdown_spacing = "compact"
diff --git a/crates/jcode-base/src/gmail.rs b/crates/jcode-base/src/gmail.rs
index 8b4309645..f6fc2d7bd 100644
--- a/crates/jcode-base/src/gmail.rs
+++ b/crates/jcode-base/src/gmail.rs
@@ -1,12 +1,145 @@
 use anyhow::Result;
 use serde::{Deserialize, Serialize};
+use serde_json::{Value, json};
 
 use crate::auth::google;
 
 const GMAIL_API_BASE: &str = "https://gmail.googleapis.com/gmail/v1/users/me";
+const COMPOSIO_DEFAULT_BASE: &str = "https://backend.composio.dev/api/v3.1";
+
+/// Where the Gmail tool gets its credentials and authenticated transport.
+///
+/// `Direct` talks to the Google Gmail REST API using locally stored OAuth
+/// tokens (the original behavior). `Composio` routes the *same* Gmail REST
+/// calls through Composio's managed `proxy-execute` endpoint, so a
+/// Google-verified app brokers auth: no unverified-app warning and no 7-day
+/// testing-mode token expiry.
+#[derive(Debug, Clone)]
+pub enum GmailBackend {
+    Direct,
+    Composio(ComposioConfig),
+}
+
+#[derive(Debug, Clone)]
+pub struct ComposioConfig {
+    pub api_key: String,
+    pub base_url: String,
+    pub connected_account_id: Option<String>,
+    pub user_id: Option<String>,
+    /// Auth config that defines the Gmail OAuth blueprint (scopes + managed
+    /// Composio app). Required to initiate a Connect Link flow. Falls back to
+    /// a persisted value or `COMPOSIO_GMAIL_AUTH_CONFIG_ID`.
+    pub auth_config_id: Option<String>,
+}
+
+impl GmailBackend {
+    /// Resolve the backend from environment configuration.
+    ///
+    /// Defaults to `Direct`. Set `JCODE_GMAIL_BACKEND=composio` (with
+    /// `COMPOSIO_API_KEY` present) to broker Gmail through Composio.
+    pub fn from_env() -> Self {
+        let selection = std::env::var("JCODE_GMAIL_BACKEND")
+            .unwrap_or_default()
+            .trim()
+            .to_lowercase();
+        if selection == "composio" {
+            if let Some(cfg) = ComposioConfig::from_env() {
+                return GmailBackend::Composio(cfg);
+            }
+            eprintln!(
+                "JCODE_GMAIL_BACKEND=composio but COMPOSIO_API_KEY is not set; falling back to direct Gmail backend"
+            );
+        }
+        GmailBackend::Direct
+    }
+
+    pub fn label(&self) -> &'static str {
+        match self {
+            GmailBackend::Direct => "direct",
+            GmailBackend::Composio(_) => "composio",
+        }
+    }
+}
+
+impl ComposioConfig {
+    fn from_env() -> Option<Self> {
+        let api_key = std::env::var("COMPOSIO_API_KEY").ok().filter(|s| !s.is_empty())?;
+        let base_url = std::env::var("COMPOSIO_BASE_URL")
+            .ok()
+            .filter(|s| !s.is_empty())
+            .unwrap_or_else(|| COMPOSIO_DEFAULT_BASE.to_string());
+        // A previously completed Connect Link flow persists the connection so
+        // the user does not have to re-run setup each session.
+        let persisted = ComposioConnection::load().ok().flatten();
+        let connected_account_id = std::env::var("COMPOSIO_GMAIL_CONNECTED_ACCOUNT_ID")
+            .ok()
+            .filter(|s| !s.is_empty())
+            .or_else(|| persisted.as_ref().map(|p| p.connected_account_id.clone()));
+        let user_id = std::env::var("COMPOSIO_GMAIL_USER_ID")
+            .or_else(|_| std::env::var("COMPOSIO_USER_ID"))
+            .ok()
+            .filter(|s| !s.is_empty())
+            .or_else(|| persisted.as_ref().map(|p| p.user_id.clone()));
+        let auth_config_id = std::env::var("COMPOSIO_GMAIL_AUTH_CONFIG_ID")
+            .ok()
+            .filter(|s| !s.is_empty())
+            .or_else(|| persisted.as_ref().and_then(|p| p.auth_config_id.clone()));
+        Some(Self {
+            api_key,
+            base_url,
+            connected_account_id,
+            user_id,
+            auth_config_id,
+        })
+    }
+
+    /// Effective user id, defaulting to "default" so a single-user CLI works
+    /// without any extra configuration.
+    pub fn effective_user_id(&self) -> String {
+        self.user_id.clone().unwrap_or_else(|| "default".to_string())
+    }
+}
+
+/// Persisted record of a completed Composio Gmail connection, stored at
+/// `~/.jcode/composio_gmail.json`.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ComposioConnection {
+    pub connected_account_id: String,
+    pub user_id: String,
+    pub auth_config_id: Option<String>,
+    #[serde(default)]
+    pub email: Option<String>,
+}
+
+impl ComposioConnection {
+    pub fn path() -> Result<std::path::PathBuf> {
+        Ok(crate::storage::jcode_dir()?.join("composio_gmail.json"))
+    }
+
+    pub fn load() -> Result<Option<Self>> {
+        let path = Self::path()?;
+        if !path.exists() {
+            return Ok(None);
+        }
+        crate::storage::harden_secret_file_permissions(&path);
+        Ok(crate::storage::read_json(&path).ok())
+    }
+
+    pub fn save(&self) -> Result<()> {
+        let path = Self::path()?;
+        crate::storage::write_json_secret(&path, self)
+    }
+}
+
+/// Result of initiating a Connect Link OAuth flow.
+pub struct ComposioLink {
+    pub connected_account_id: String,
+    pub redirect_url: String,
+}
 
 pub struct GmailClient {
     http: reqwest::Client,
+    backend: GmailBackend,
 }
 
 impl Default for GmailClient {
@@ -17,13 +150,304 @@ impl Default for GmailClient {
 
 impl GmailClient {
     pub fn new() -> Self {
+        Self::with_backend(GmailBackend::from_env())
+    }
+
+    pub fn with_backend(backend: GmailBackend) -> Self {
         Self {
             http: crate::provider::shared_http_client(),
+            backend,
+        }
+    }
+
+    pub fn backend_label(&self) -> &'static str {
+        self.backend.label()
+    }
+
+    /// Whether this backend has credentials available to talk to Gmail.
+    pub fn is_configured(&self) -> bool {
+        match &self.backend {
+            GmailBackend::Direct => google::has_tokens(),
+            GmailBackend::Composio(cfg) => !cfg.api_key.is_empty(),
+        }
+    }
+
+    /// Whether the current backend is allowed to send mail.
+    ///
+    /// The `Direct` backend honors the locally configured access tier
+    /// (read-only logins cannot send). Composio connections request full
+    /// Gmail scopes, so sending is available.
+    pub fn can_send(&self) -> bool {
+        match &self.backend {
+            GmailBackend::Direct => google::load_tokens()
+                .map(|t| t.tier.can_send())
+                .unwrap_or(false),
+            GmailBackend::Composio(_) => true,
+        }
+    }
+
+    /// Whether the current backend is allowed to delete/trash mail.
+    pub fn can_delete(&self) -> bool {
+        match &self.backend {
+            GmailBackend::Direct => google::load_tokens()
+                .map(|t| t.tier.can_delete())
+                .unwrap_or(false),
+            GmailBackend::Composio(_) => true,
+        }
+    }
+
+    pub fn not_configured_message(&self) -> &'static str {
+        match &self.backend {
+            GmailBackend::Direct => {
+                "Gmail is not configured. Run `jcode login google` to set up Gmail access."
+            }
+            GmailBackend::Composio(_) => {
+                "Gmail (Composio backend) is not configured. Set COMPOSIO_API_KEY and connect your \
+                 Gmail account in Composio, then retry."
+            }
+        }
+    }
+
+    /// True only for the Composio backend when no connected account exists yet.
+    /// In that state, Gmail calls will fail until the user completes the
+    /// Connect Link OAuth flow via [`GmailClient::connect`].
+    pub fn needs_connection(&self) -> bool {
+        matches!(&self.backend, GmailBackend::Composio(cfg) if cfg.connected_account_id.is_none())
+    }
+
+    /// Whether the active backend supports an interactive `connect` action.
+    pub fn supports_connect(&self) -> bool {
+        matches!(&self.backend, GmailBackend::Composio(_))
+    }
+
+    /// Initiate a Composio Connect Link OAuth flow, open the consent screen in
+    /// the user's browser, wait for them to approve, then persist the resulting
+    /// connected account so future sessions are already authenticated.
+    ///
+    /// `open_browser` controls whether we try to launch the system browser
+    /// (set false over SSH/headless; the URL is always returned).
+    pub async fn connect(&self, open_browser: bool) -> Result<ComposioConnection> {
+        let cfg = match &self.backend {
+            GmailBackend::Composio(cfg) => cfg,
+            GmailBackend::Direct => {
+                anyhow::bail!(
+                    "The Composio connect flow is only available when JCODE_GMAIL_BACKEND=composio."
+                )
+            }
+        };
+        let auth_config_id = cfg.auth_config_id.clone().ok_or_else(|| {
+            anyhow::anyhow!(
+                "No Composio Gmail auth config configured. Create a Gmail auth config in the \
+                 Composio dashboard and set COMPOSIO_GMAIL_AUTH_CONFIG_ID."
+            )
+        })?;
+        let user_id = cfg.effective_user_id();
+
+        let link = self.create_link(cfg, &auth_config_id, &user_id).await?;
+        if open_browser {
+            let _ = open::that(&link.redirect_url);
+        }
+        eprintln!(
+            "\nOpening Gmail authorization in your browser. If it did not open, visit:\n{}\n",
+            link.redirect_url
+        );
+
+        let account = self
+            .wait_for_connection(cfg, &link.connected_account_id)
+            .await?;
+
+        let email = account
+            .get("data")
+            .and_then(|d| d.get("email"))
+            .or_else(|| account.get("email"))
+            .and_then(|e| e.as_str())
+            .map(|s| s.to_string());
+
+        let connection = ComposioConnection {
+            connected_account_id: link.connected_account_id,
+            user_id,
+            auth_config_id: Some(auth_config_id),
+            email,
+        };
+        connection.save()?;
+        Ok(connection)
+    }
+
+    /// Create a hosted Connect Link auth session.
+    async fn create_link(
+        &self,
+        cfg: &ComposioConfig,
+        auth_config_id: &str,
+        user_id: &str,
+    ) -> Result<ComposioLink> {
+        let endpoint = format!("{}/connected_accounts/link", cfg.base_url.trim_end_matches('/'));
+        let payload = json!({
+            "auth_config_id": auth_config_id,
+            "user_id": user_id,
+        });
+        let resp = self
+            .http
+            .post(&endpoint)
+            .header("x-api-key", &cfg.api_key)
+            .json(&payload)
+            .send()
+            .await?;
+        let status = resp.status();
+        let text = resp.text().await?;
+        if !status.is_success() {
+            return Err(anyhow::anyhow!(
+                "Composio connect-link error {}: {}",
+                status,
+                truncate_error(&text)
+            ));
+        }
+        let body: Value = serde_json::from_str(&text)?;
+        let redirect_url = body
+            .get("redirect_url")
+            .and_then(|v| v.as_str())
+            .ok_or_else(|| anyhow::anyhow!("Composio did not return a redirect_url"))?
+            .to_string();
+        let connected_account_id = body
+            .get("connected_account_id")
+            .and_then(|v| v.as_str())
+            .ok_or_else(|| anyhow::anyhow!("Composio did not return a connected_account_id"))?
+            .to_string();
+        Ok(ComposioLink {
+            connected_account_id,
+            redirect_url,
+        })
+    }
+
+    /// Poll a connected account until it becomes ACTIVE (or a terminal error).
+    async fn wait_for_connection(
+        &self,
+        cfg: &ComposioConfig,
+        connected_account_id: &str,
+    ) -> Result<Value> {
+        // INITIATED links auto-expire after ~10 minutes; poll up to ~5 minutes.
+        const MAX_ATTEMPTS: u32 = 150;
+        const POLL_INTERVAL: std::time::Duration = std::time::Duration::from_secs(2);
+        let endpoint = format!(
+            "{}/connected_accounts/{}",
+            cfg.base_url.trim_end_matches('/'),
+            connected_account_id
+        );
+        for _ in 0..MAX_ATTEMPTS {
+            let resp = self
+                .http
+                .get(&endpoint)
+                .header("x-api-key", &cfg.api_key)
+                .send()
+                .await?;
+            if resp.status().is_success() {
+                let body: Value = resp.json().await?;
+                let status = body
+                    .get("status")
+                    .or_else(|| body.get("data").and_then(|d| d.get("status")))
+                    .and_then(|s| s.as_str())
+                    .unwrap_or("");
+                match status {
+                    "ACTIVE" => return Ok(body),
+                    "FAILED" | "EXPIRED" => {
+                        let reason = body
+                            .get("status_reason")
+                            .and_then(|r| r.as_str())
+                            .unwrap_or("no reason provided");
+                        anyhow::bail!("Gmail connection {}: {}", status, reason);
+                    }
+                    _ => {}
+                }
+            }
+            tokio::time::sleep(POLL_INTERVAL).await;
+        }
+        anyhow::bail!(
+            "Timed out waiting for Gmail authorization. Re-run the connect action and finish the \
+             browser consent within a few minutes."
+        )
+    }
+
+    /// Send an authenticated Gmail REST request and return the parsed JSON
+    /// response. Both backends produce the identical Gmail API JSON shape, so
+    /// callers can deserialize into the same typed structs.
+    async fn request(
+        &self,
+        method: reqwest::Method,
+        url: &str,
+        body: Option<Value>,
+    ) -> Result<Value> {
+        match &self.backend {
+            GmailBackend::Direct => self.request_direct(method, url, body).await,
+            GmailBackend::Composio(cfg) => self.request_composio(cfg, method, url, body).await,
+        }
+    }
+
+    async fn request_direct(
+        &self,
+        method: reqwest::Method,
+        url: &str,
+        body: Option<Value>,
+    ) -> Result<Value> {
+        let token = google::get_valid_token().await?;
+        let mut req = self.http.request(method, url).bearer_auth(&token);
+        if let Some(ref b) = body {
+            req = req.json(b);
+        }
+        let resp = req.send().await?;
+        let status = resp.status();
+        let text = resp.text().await?;
+        if !status.is_success() {
+            return Err(anyhow::anyhow!(
+                "Gmail API error {}: {}",
+                status,
+                truncate_error(&text)
+            ));
         }
+        if text.trim().is_empty() {
+            return Ok(Value::Null);
+        }
+        Ok(serde_json::from_str(&text)?)
     }
 
-    async fn token(&self) -> Result<String> {
-        google::get_valid_token().await
+    async fn request_composio(
+        &self,
+        cfg: &ComposioConfig,
+        method: reqwest::Method,
+        url: &str,
+        body: Option<Value>,
+    ) -> Result<Value> {
+        let payload = build_composio_proxy_payload(cfg, method.as_str(), url, body);
+        let endpoint = format!("{}/tools/execute/proxy", cfg.base_url.trim_end_matches('/'));
+        let resp = self
+            .http
+            .post(&endpoint)
+            .header("x-api-key", &cfg.api_key)
+            .json(&payload)
+            .send()
+            .await?;
+        let status = resp.status();
+        let text = resp.text().await?;
+        if !status.is_success() {
+            return Err(anyhow::anyhow!(
+                "Composio proxy error {}: {}",
+                status,
+                truncate_error(&text)
+            ));
+        }
+        let envelope: Value = serde_json::from_str(&text)?;
+        // Composio wraps the upstream response as { data, status, headers }.
+        if let Some(inner) = envelope.get("status").and_then(|s| s.as_u64()) {
+            if inner >= 400 {
+                return Err(anyhow::anyhow!(
+                    "Gmail API error {} (via Composio): {}",
+                    inner,
+                    truncate_error(&envelope.get("data").map(|d| d.to_string()).unwrap_or_default())
+                ));
+            }
+        }
+        if let Some(err) = envelope.get("error").filter(|e| !e.is_null()) {
+            return Err(anyhow::anyhow!("Composio error: {}", truncate_error(&err.to_string())));
+        }
+        Ok(envelope.get("data").cloned().unwrap_or(Value::Null))
     }
 
     pub async fn list_messages(
@@ -32,7 +456,6 @@ impl GmailClient {
         label_ids: Option<&[&str]>,
         max_results: u32,
     ) -> Result<MessageList> {
-        let token = self.token().await?;
         let mut url = format!("{}/messages?maxResults={}", GMAIL_API_BASE, max_results);
 
         if let Some(q) = query {
@@ -44,61 +467,47 @@ impl GmailClient {
             }
         }
 
-        let resp = self.http.get(&url).bearer_auth(&token).send().await?;
-        handle_error(&resp).await?;
-        let list: MessageList = resp.json().await?;
-        Ok(list)
+        let value = self.request(reqwest::Method::GET, &url, None).await?;
+        Ok(serde_json::from_value(value)?)
     }
 
     pub async fn get_message(&self, id: &str, format: MessageFormat) -> Result<Message> {
-        let token = self.token().await?;
         let url = format!(
             "{}/messages/{}?format={}",
             GMAIL_API_BASE,
             id,
             format.as_str()
         );
-        let resp = self.http.get(&url).bearer_auth(&token).send().await?;
-        handle_error(&resp).await?;
-        let msg: Message = resp.json().await?;
-        Ok(msg)
+        let value = self.request(reqwest::Method::GET, &url, None).await?;
+        Ok(serde_json::from_value(value)?)
     }
 
     pub async fn list_threads(&self, query: Option<&str>, max_results: u32) -> Result<ThreadList> {
-        let token = self.token().await?;
         let mut url = format!("{}/threads?maxResults={}", GMAIL_API_BASE, max_results);
 
         if let Some(q) = query {
             url.push_str(&format!("&q={}", urlencoding::encode(q)));
         }
 
-        let resp = self.http.get(&url).bearer_auth(&token).send().await?;
-        handle_error(&resp).await?;
-        let list: ThreadList = resp.json().await?;
-        Ok(list)
+        let value = self.request(reqwest::Method::GET, &url, None).await?;
+        Ok(serde_json::from_value(value)?)
     }
 
     pub async fn get_thread(&self, id: &str) -> Result<Thread> {
-        let token = self.token().await?;
         let url = format!("{}/threads/{}?format=metadata", GMAIL_API_BASE, id);
-        let resp = self.http.get(&url).bearer_auth(&token).send().await?;
-        handle_error(&resp).await?;
-        let thread: Thread = resp.json().await?;
-        Ok(thread)
+        let value = self.request(reqwest::Method::GET, &url, None).await?;
+        Ok(serde_json::from_value(value)?)
     }
 
     pub async fn list_labels(&self) -> Result<Vec<Label>> {
-        let token = self.token().await?;
         let url = format!("{}/labels", GMAIL_API_BASE);
-        let resp = self.http.get(&url).bearer_auth(&token).send().await?;
-        handle_error(&resp).await?;
-
         #[derive(Deserialize)]
         struct LabelList {
             labels: Option<Vec<Label>>,
         }
 
-        let list: LabelList = resp.json().await?;
+        let value = self.request(reqwest::Method::GET, &url, None).await?;
+        let list: LabelList = serde_json::from_value(value)?;
         Ok(list.labels.unwrap_or_default())
     }
 
@@ -110,7 +519,6 @@ impl GmailClient {
         in_reply_to: Option<&str>,
         thread_id: Option<&str>,
     ) -> Result<Draft> {
-        let token = self.token().await?;
         let url = format!("{}/drafts", GMAIL_API_BASE);
 
         let mut headers = format!(
@@ -127,40 +535,27 @@ impl GmailClient {
         let raw = format!("{}\r\n{}", headers, body);
         let encoded = base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(raw.as_bytes());
 
-        let mut message = serde_json::json!({ "raw": encoded });
+        let mut message = json!({ "raw": encoded });
         if let Some(tid) = thread_id {
-            message["threadId"] = serde_json::Value::String(tid.to_string());
+            message["threadId"] = Value::String(tid.to_string());
         }
 
-        let payload = serde_json::json!({ "message": message });
+        let payload = json!({ "message": message });
 
-        let resp = self
-            .http
-            .post(&url)
-            .bearer_auth(&token)
-            .json(&payload)
-            .send()
+        let value = self
+            .request(reqwest::Method::POST, &url, Some(payload))
             .await?;
-        handle_error(&resp).await?;
-        let draft: Draft = resp.json().await?;
-        Ok(draft)
+        Ok(serde_json::from_value(value)?)
     }
 
     pub async fn send_draft(&self, draft_id: &str) -> Result<Message> {
-        let token = self.token().await?;
         let url = format!("{}/drafts/send", GMAIL_API_BASE);
-        let payload = serde_json::json!({ "id": draft_id });
+        let payload = json!({ "id": draft_id });
 
-        let resp = self
-            .http
-            .post(&url)
-            .bearer_auth(&token)
-            .json(&payload)
-            .send()
+        let value = self
+            .request(reqwest::Method::POST, &url, Some(payload))
             .await?;
-        handle_error(&resp).await?;
-        let msg: Message = resp.json().await?;
-        Ok(msg)
+        Ok(serde_json::from_value(value)?)
     }
 
     pub async fn send_message(
@@ -171,7 +566,6 @@ impl GmailClient {
         in_reply_to: Option<&str>,
         thread_id: Option<&str>,
     ) -> Result<Message> {
-        let token = self.token().await?;
         let url = format!("{}/messages/send", GMAIL_API_BASE);
 
         let mut headers = format!(
@@ -188,28 +582,20 @@ impl GmailClient {
         let raw = format!("{}\r\n{}", headers, body);
         let encoded = base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(raw.as_bytes());
 
-        let mut message = serde_json::json!({ "raw": encoded });
+        let mut message = json!({ "raw": encoded });
         if let Some(tid) = thread_id {
-            message["threadId"] = serde_json::Value::String(tid.to_string());
+            message["threadId"] = Value::String(tid.to_string());
         }
 
-        let resp = self
-            .http
-            .post(&url)
-            .bearer_auth(&token)
-            .json(&message)
-            .send()
+        let value = self
+            .request(reqwest::Method::POST, &url, Some(message))
             .await?;
-        handle_error(&resp).await?;
-        let msg: Message = resp.json().await?;
-        Ok(msg)
+        Ok(serde_json::from_value(value)?)
     }
 
     pub async fn trash_message(&self, id: &str) -> Result<()> {
-        let token = self.token().await?;
         let url = format!("{}/messages/{}/trash", GMAIL_API_BASE, id);
-        let resp = self.http.post(&url).bearer_auth(&token).send().await?;
-        handle_error(&resp).await?;
+        self.request(reqwest::Method::POST, &url, None).await?;
         Ok(())
     }
 
@@ -219,32 +605,49 @@ impl GmailClient {
         add_labels: &[&str],
         remove_labels: &[&str],
     ) -> Result<()> {
-        let token = self.token().await?;
         let url = format!("{}/messages/{}/modify", GMAIL_API_BASE, id);
-        let payload = serde_json::json!({
+        let payload = json!({
             "addLabelIds": add_labels,
             "removeLabelIds": remove_labels,
         });
-        let resp = self
-            .http
-            .post(&url)
-            .bearer_auth(&token)
-            .json(&payload)
-            .send()
+        self.request(reqwest::Method::POST, &url, Some(payload))
             .await?;
-        handle_error(&resp).await?;
         Ok(())
     }
 }
 
-async fn handle_error(resp: &reqwest::Response) -> Result<()> {
-    if resp.status().is_success() {
-        return Ok(());
+/// Build the request body for Composio's `proxy-execute` endpoint, which makes
+/// an authenticated HTTP call to the connected toolkit (Gmail) on our behalf.
+fn build_composio_proxy_payload(
+    cfg: &ComposioConfig,
+    method: &str,
+    url: &str,
+    body: Option<Value>,
+) -> Value {
+    let mut payload = json!({
+        "endpoint": url,
+        "method": method,
+    });
+    if let Some(b) = body {
+        payload["body"] = b;
+    }
+    if let Some(account) = &cfg.connected_account_id {
+        payload["connected_account_id"] = Value::String(account.clone());
+    }
+    if let Some(user) = &cfg.user_id {
+        payload["user_id"] = Value::String(user.clone());
+    }
+    payload
+}
+
+fn truncate_error(text: &str) -> String {
+    const MAX: usize = 400;
+    let trimmed = text.trim();
+    if trimmed.len() <= MAX {
+        trimmed.to_string()
+    } else {
+        format!("{}…", &trimmed[..MAX])
     }
-    Err(anyhow::anyhow!(
-        "Gmail API error {}: check token permissions",
-        resp.status()
-    ))
 }
 
 use base64::Engine;
@@ -446,3 +849,108 @@ pub fn format_message_full(msg: &Message) -> String {
     }
     out
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn cfg() -> ComposioConfig {
+        ComposioConfig {
+            api_key: "test-key".to_string(),
+            base_url: COMPOSIO_DEFAULT_BASE.to_string(),
+            connected_account_id: Some("ca_123".to_string()),
+            user_id: Some("me".to_string()),
+            auth_config_id: Some("ac_123".to_string()),
+        }
+    }
+
+    #[test]
+    fn composio_proxy_payload_get_has_no_body() {
+        let url = format!("{}/messages?maxResults=10", GMAIL_API_BASE);
+        let payload = build_composio_proxy_payload(&cfg(), "GET", &url, None);
+        assert_eq!(payload["endpoint"], url);
+        assert_eq!(payload["method"], "GET");
+        assert!(payload.get("body").is_none());
+        assert_eq!(payload["connected_account_id"], "ca_123");
+        assert_eq!(payload["user_id"], "me");
+    }
+
+    #[test]
+    fn composio_proxy_payload_post_includes_body() {
+        let url = format!("{}/messages/send", GMAIL_API_BASE);
+        let body = json!({ "raw": "abc" });
+        let payload = build_composio_proxy_payload(&cfg(), "POST", &url, Some(body.clone()));
+        assert_eq!(payload["method"], "POST");
+        assert_eq!(payload["body"], body);
+    }
+
+    #[test]
+    fn composio_proxy_payload_omits_optional_account_fields() {
+        let bare = ComposioConfig {
+            api_key: "k".to_string(),
+            base_url: COMPOSIO_DEFAULT_BASE.to_string(),
+            connected_account_id: None,
+            user_id: None,
+            auth_config_id: None,
+        };
+        let payload = build_composio_proxy_payload(&bare, "GET", "http://x/y", None);
+        assert!(payload.get("connected_account_id").is_none());
+        assert!(payload.get("user_id").is_none());
+    }
+
+    #[test]
+    fn direct_backend_label_and_default() {
+        let backend = GmailBackend::Direct;
+        assert_eq!(backend.label(), "direct");
+        let client = GmailClient::with_backend(GmailBackend::Direct);
+        assert_eq!(client.backend_label(), "direct");
+    }
+
+    #[test]
+    fn composio_backend_is_configured_and_can_send() {
+        let client = GmailClient::with_backend(GmailBackend::Composio(cfg()));
+        assert_eq!(client.backend_label(), "composio");
+        assert!(client.is_configured());
+        // Composio connections request full Gmail scopes.
+        assert!(client.can_send());
+        assert!(client.can_delete());
+    }
+
+    #[test]
+    fn truncate_error_caps_length() {
+        let short = truncate_error("  hi  ");
+        assert_eq!(short, "hi");
+        let long = "x".repeat(1000);
+        let capped = truncate_error(&long);
+        assert!(capped.len() <= 401 + 3); // 400 chars + ellipsis byte
+        assert!(capped.ends_with('…'));
+    }
+
+    #[test]
+    fn needs_connection_reflects_connected_account_presence() {
+        // Composio without a connected account needs an interactive connect.
+        let mut without = cfg();
+        without.connected_account_id = None;
+        let client = GmailClient::with_backend(GmailBackend::Composio(without));
+        assert!(client.supports_connect());
+        assert!(client.needs_connection());
+
+        // With a connected account it is ready to make calls.
+        let client = GmailClient::with_backend(GmailBackend::Composio(cfg()));
+        assert!(!client.needs_connection());
+
+        // Direct backend never needs a Composio connection and cannot connect.
+        let direct = GmailClient::with_backend(GmailBackend::Direct);
+        assert!(!direct.supports_connect());
+        assert!(!direct.needs_connection());
+    }
+
+    #[test]
+    fn effective_user_id_defaults_to_default() {
+        let mut c = cfg();
+        c.user_id = None;
+        assert_eq!(c.effective_user_id(), "default");
+        c.user_id = Some("alice".to_string());
+        assert_eq!(c.effective_user_id(), "alice");
+    }
+}
diff --git a/crates/jcode-base/src/lib.rs b/crates/jcode-base/src/lib.rs
index 5aa2c8a4c..bf705bae2 100644
--- a/crates/jcode-base/src/lib.rs
+++ b/crates/jcode-base/src/lib.rs
@@ -64,6 +64,7 @@ pub mod safety;
 pub mod secret_input;
 pub mod session;
 pub mod session_list_cache;
+pub mod session_metrics;
 pub mod side_panel;
 pub mod sidecar;
 pub mod skill;
diff --git a/crates/jcode-base/src/live_tests.rs b/crates/jcode-base/src/live_tests.rs
index f850c3b33..4d42482a6 100644
--- a/crates/jcode-base/src/live_tests.rs
+++ b/crates/jcode-base/src/live_tests.rs
@@ -13,7 +13,7 @@ const DEFAULT_RETEST_DAYS: i64 = 14;
 const LEDGER_ENV: &str = "JCODE_LIVE_TEST_LEDGER";
 const COVERAGE_ENV: &str = "JCODE_LIVE_TEST_COVERAGE";
 
-pub const CHECKPOINT_TAXONOMY_VERSION: u32 = 2;
+pub const CHECKPOINT_TAXONOMY_VERSION: u32 = 3;
 
 pub mod checkpoints {
     pub const AUTH_UX_KEY_ENTRY: &str = "auth_ux_key_entry";
@@ -30,6 +30,10 @@ pub mod checkpoints {
     pub const TOOL_EXECUTION_LOOP: &str = "tool_execution_loop";
     pub const TOOL_RESULT_FOLLOWUP: &str = "tool_result_followup";
     pub const REAL_JCODE_TOOL_SMOKE: &str = "real_jcode_tool_smoke";
+    /// Observe-only: did the model expose its reasoning (`streamed`), hide it
+    /// behind an opaque signal (`opaque`, e.g. Gemini-3 / OpenAI), or emit none
+    /// (`none`)? Never required for user-readiness; hiding reasoning is a pass.
+    pub const REASONING_CAPABILITY: &str = "reasoning_capability";
     pub const RESTART_PERSISTENCE: &str = "restart_persistence";
     pub const NEGATIVE_ERROR_UX: &str = "negative_error_ux";
     pub const MODEL_CAPABILITY_MATRIX: &str = "model_capability_matrix";
@@ -159,6 +163,16 @@ const END_TO_END_CHECKPOINTS: &[LiveVerificationCheckpointDefinition] = &[
         spends_balance: true,
         description: "A normal Jcode agent turn uses the real streamed parser, advertised tool schema, registry execution, tool-result followup, and transcript validation without malformed tool calls.",
     },
+    LiveVerificationCheckpointDefinition {
+        id: checkpoints::REASONING_CAPABILITY,
+        label: "Reasoning capability",
+        category: "reasoning",
+        // Observe-only: a provider that hides its reasoning (opaque) or emits
+        // none is still fully user-ready, so this must never gate readiness.
+        required_for_user_ready: false,
+        spends_balance: true,
+        description: "Records whether the model streams reasoning text, hides it behind an opaque signal (thought_signature/reasoning item/reasoning tokens), or emits none. Passes as long as the reasoning turn completes cleanly; absence of reasoning is recorded, not failed.",
+    },
     LiveVerificationCheckpointDefinition {
         id: checkpoints::RESTART_PERSISTENCE,
         label: "Restart persistence",
@@ -2506,6 +2520,7 @@ mod tests {
             checkpoints::TOOL_EXECUTION_LOOP,
             checkpoints::TOOL_RESULT_FOLLOWUP,
             checkpoints::REAL_JCODE_TOOL_SMOKE,
+            checkpoints::REASONING_CAPABILITY,
             checkpoints::RESTART_PERSISTENCE,
             checkpoints::NEGATIVE_ERROR_UX,
             checkpoints::MODEL_CAPABILITY_MATRIX,
@@ -2519,6 +2534,24 @@ mod tests {
                 .any(|checkpoint| checkpoint.spends_balance),
             "taxonomy should identify balance-spending checkpoints"
         );
+
+        // The reasoning_capability checkpoint is observe-only: it records what
+        // the model exposed (streamed/opaque/none) but a provider that hides its
+        // reasoning is still fully user-ready, so it must never gate readiness or
+        // strict coverage.
+        let reasoning = end_to_end_checkpoint_definitions()
+            .iter()
+            .find(|checkpoint| checkpoint.id == checkpoints::REASONING_CAPABILITY)
+            .expect("reasoning_capability checkpoint must exist in the taxonomy");
+        assert!(
+            !reasoning.required_for_user_ready,
+            "reasoning_capability must not be required for user-readiness"
+        );
+        assert!(
+            !STRICT_PROVIDER_MODEL_COVERAGE_CHECKPOINTS
+                .contains(&checkpoints::REASONING_CAPABILITY),
+            "reasoning_capability must not be a strict-required checkpoint"
+        );
     }
 
     #[test]
diff --git a/crates/jcode-base/src/provider/antigravity.rs b/crates/jcode-base/src/provider/antigravity.rs
index 87434b00b..e1a14b4ae 100644
--- a/crates/jcode-base/src/provider/antigravity.rs
+++ b/crates/jcode-base/src/provider/antigravity.rs
@@ -676,6 +676,7 @@ impl AntigravityProvider {
         tools: &[ToolDefinition],
         system: &str,
         resume_session_id: Option<&str>,
+        force_function_call: bool,
     ) -> Result<CodeAssistGenerateResponse> {
         let mut tokens = antigravity_auth::load_or_refresh_tokens().await?;
         let project = match tokens
@@ -714,13 +715,23 @@ impl AntigravityProvider {
             user_prompt_id: Uuid::new_v4().to_string(),
             request: VertexGenerateContentRequest {
                 contents: super::gemini::build_contents(messages),
-                system_instruction: super::gemini::build_system_instruction(system),
+                system_instruction: super::gemini::build_system_instruction_with_tool_guard(
+                    system,
+                    !tools_is_empty,
+                ),
                 tools,
                 tool_config: if tools_is_empty {
                     None
                 } else {
+                    // On a transparent retry after a MALFORMED_FUNCTION_CALL, force
+                    // function-calling mode `ANY` so the model must emit a real
+                    // functionCall instead of the Python-style pseudo-code that
+                    // triggered the malformed turn (the proven recovery for this
+                    // failure mode). Normal turns use `AUTO`.
                     Some(GeminiToolConfig {
-                        function_calling_config: GeminiFunctionCallingConfig { mode: "AUTO" },
+                        function_calling_config: GeminiFunctionCallingConfig {
+                            mode: if force_function_call { "ANY" } else { "AUTO" },
+                        },
                     })
                 },
                 session_id: resume_session_id
@@ -805,6 +816,51 @@ fn model_is_claude(model: &str) -> bool {
     model.trim().to_ascii_lowercase().contains("claude")
 }
 
+/// Whether a `generateContent` response is an abnormal turn that produced no
+/// usable output (no text, no function call). This is the shape Gemini-3
+/// "thinking" models intermittently return when they emit Python-style
+/// pseudo-code instead of a clean functionCall: `finish_reason ==
+/// MALFORMED_FUNCTION_CALL` (or another non-terminal reason) with empty content.
+/// Such a turn is worth one transparent retry before surfacing an error.
+///
+/// Normal terminal reasons (`STOP`, `MAX_TOKENS`, unspecified) are never treated
+/// as retryable here, even with empty content, so a legitimately empty answer is
+/// not retried in a loop.
+fn is_retryable_empty_turn(response: &CodeAssistGenerateResponse) -> bool {
+    let Some(candidate) = response
+        .response
+        .as_ref()
+        .and_then(|r| r.candidates.as_ref())
+        .and_then(|c| c.first())
+    else {
+        // No candidate at all is handled separately (hard error), not retried here.
+        return false;
+    };
+    let produced_output = candidate
+        .content
+        .as_ref()
+        .map(|content| {
+            content.parts.iter().any(|part| {
+                part.function_call.is_some()
+                    || part.text.as_deref().is_some_and(|text| !text.is_empty())
+            })
+        })
+        .unwrap_or(false);
+    if produced_output {
+        return false;
+    }
+    candidate
+        .finish_reason
+        .as_deref()
+        .map(|reason| {
+            !matches!(
+                reason.to_ascii_uppercase().as_str(),
+                "STOP" | "MAX_TOKENS" | "FINISH_REASON_UNSPECIFIED" | ""
+            )
+        })
+        .unwrap_or(false)
+}
+
 /// Remap model ids that the Antigravity catalog advertises but the
 /// `generateContent`/`streamGenerateContent` backend cannot actually service,
 /// onto an equivalent id that works.
@@ -994,6 +1050,7 @@ impl Provider for AntigravityProvider {
                     &tools,
                     &system,
                     resume_session_id.as_deref(),
+                    false,
                 )
                 .await
             {
@@ -1003,6 +1060,36 @@ impl Provider for AntigravityProvider {
                     return;
                 }
             };
+            // Gemini-3 thinking models intermittently return an empty
+            // `MALFORMED_FUNCTION_CALL` turn (pseudo-code instead of a clean
+            // functionCall). It is transient, so transparently re-request a few
+            // times before surfacing it; this turns a frequent hard failure into a
+            // near-always-successful turn without the agent loop seeing the blip.
+            // The retries force function-calling mode `ANY` so the model must emit
+            // a real functionCall rather than the pseudo-code that failed.
+            let mut response = response;
+            let mut malformed_retries = 0u8;
+            const MAX_MALFORMED_RETRIES: u8 = 2;
+            while is_retryable_empty_turn(&response) && malformed_retries < MAX_MALFORMED_RETRIES {
+                malformed_retries += 1;
+                match provider
+                    .generate_content(
+                        &model,
+                        &messages,
+                        &tools,
+                        &system,
+                        resume_session_id.as_deref(),
+                        true,
+                    )
+                    .await
+                {
+                    Ok(retried) => response = retried,
+                    Err(err) => {
+                        let _ = tx.send(Err(err)).await;
+                        return;
+                    }
+                }
+            }
             let _ = tx
                 .send(Ok(StreamEvent::ConnectionPhase {
                     phase: ConnectionPhase::Streaming,
@@ -1034,6 +1121,13 @@ impl Provider for AntigravityProvider {
                     .await;
                 return;
             };
+            // Track whether this candidate produced any usable output (text or a
+            // tool call). Gemini-3 thinking models intermittently emit Python-style
+            // pseudo-code instead of a clean functionCall and finish with
+            // `MALFORMED_FUNCTION_CALL` (or a bare `OTHER`) and empty content. If we
+            // silently end the turn the agent loop looks like it stalled with no
+            // answer, so we surface an actionable error below instead.
+            let mut produced_output = false;
             if let Some(content) = candidate.content {
                 // Gemini 3 attaches a `thoughtSignature` to function-call parts
                 // (and occasionally to a standalone preceding part). Emit tool
@@ -1050,9 +1144,11 @@ impl Provider for AntigravityProvider {
                         .filter(|sig| !sig.is_empty())
                         .cloned();
                     if let Some(text) = part.text.filter(|text| !text.is_empty()) {
+                        produced_output = true;
                         let _ = tx.send(Ok(StreamEvent::TextDelta(text))).await;
                     }
                     if let Some(function_call) = part.function_call {
+                        produced_output = true;
                         let signature = part_signature.clone().or_else(|| pending_signature.take());
                         let raw_call_id = function_call
                             .id
@@ -1080,6 +1176,50 @@ impl Provider for AntigravityProvider {
                         pending_signature = Some(signature);
                     }
                 }
+                // A thought signature that was never consumed by a following
+                // function call (e.g. a pure-text reasoning turn) is still an
+                // opaque reasoning signal. Surface it as a ThinkingSignatureDelta
+                // rather than dropping it, so reasoning-aware consumers (and the
+                // provider-doctor reasoning probe) can see the model reasoned.
+                if let Some(signature) = pending_signature.take() {
+                    let _ = tx
+                        .send(Ok(StreamEvent::ThinkingSignatureDelta(signature)))
+                        .await;
+                }
+            }
+
+            // An abnormal finish (typically Gemini-3's intermittent
+            // `MALFORMED_FUNCTION_CALL`, where the model writes pseudo-code rather
+            // than a valid functionCall) that yielded no text and no tool call is a
+            // dead turn: surface it as a retryable error instead of a silent empty
+            // `MessageEnd` that looks like the agent gave up. `STOP`/`MAX_TOKENS`
+            // are normal terminal reasons and are left to flow through as usual.
+            if !produced_output {
+                let abnormal = candidate
+                    .finish_reason
+                    .as_deref()
+                    .map(|reason| {
+                        !matches!(
+                            reason.to_ascii_uppercase().as_str(),
+                            "STOP" | "MAX_TOKENS" | "FINISH_REASON_UNSPECIFIED" | ""
+                        )
+                    })
+                    .unwrap_or(false);
+                if abnormal {
+                    let reason = candidate.finish_reason.as_deref().unwrap_or("unknown");
+                    let detail = candidate
+                        .finish_message
+                        .as_deref()
+                        .filter(|msg| !msg.trim().is_empty())
+                        .map(|msg| format!(": {}", crate::util::truncate_str(msg.trim(), 300)))
+                        .unwrap_or_default();
+                    let _ = tx
+                        .send(Err(anyhow::anyhow!(
+                            "Antigravity returned no usable output (finish_reason={reason}){detail}"
+                        )))
+                        .await;
+                    return;
+                }
             }
 
             let _ = tx
diff --git a/crates/jcode-base/src/provider/antigravity_tests.rs b/crates/jcode-base/src/provider/antigravity_tests.rs
index aacb853d9..f3fa23779 100644
--- a/crates/jcode-base/src/provider/antigravity_tests.rs
+++ b/crates/jcode-base/src/provider/antigravity_tests.rs
@@ -556,3 +556,56 @@ fn antigravity_compatible_schema_strips_bounds_and_combiners_for_gpt_oss() {
         serde_json::json!("array")
     );
 }
+
+#[test]
+fn is_retryable_empty_turn_detects_malformed_function_call() {
+    // Empty content + MALFORMED_FUNCTION_CALL is the transient Gemini-3 failure we
+    // retry transparently.
+    let response: CodeAssistGenerateResponse = serde_json::from_value(serde_json::json!({
+        "response": {
+            "candidates": [{
+                "content": {},
+                "finishReason": "MALFORMED_FUNCTION_CALL",
+                "finishMessage": "Malformed function call: print(default_api.read(...))"
+            }]
+        }
+    }))
+    .expect("decode malformed response");
+    assert!(is_retryable_empty_turn(&response));
+}
+
+#[test]
+fn is_retryable_empty_turn_ignores_normal_and_productive_turns() {
+    // A normal STOP turn with text is never retried.
+    let with_text: CodeAssistGenerateResponse = serde_json::from_value(serde_json::json!({
+        "response": {
+            "candidates": [{
+                "content": {"parts": [{"text": "hello"}]},
+                "finishReason": "STOP"
+            }]
+        }
+    }))
+    .expect("decode text response");
+    assert!(!is_retryable_empty_turn(&with_text));
+
+    // A turn with a function call is productive even with no text.
+    let with_call: CodeAssistGenerateResponse = serde_json::from_value(serde_json::json!({
+        "response": {
+            "candidates": [{
+                "content": {"parts": [{"functionCall": {"name": "read", "args": {}}}]},
+                "finishReason": "STOP"
+            }]
+        }
+    }))
+    .expect("decode function call response");
+    assert!(!is_retryable_empty_turn(&with_call));
+
+    // An empty STOP turn (legitimately empty answer) is not retried in a loop.
+    let empty_stop: CodeAssistGenerateResponse = serde_json::from_value(serde_json::json!({
+        "response": {
+            "candidates": [{ "content": {}, "finishReason": "STOP" }]
+        }
+    }))
+    .expect("decode empty stop response");
+    assert!(!is_retryable_empty_turn(&empty_stop));
+}
diff --git a/crates/jcode-base/src/provider/gemini.rs b/crates/jcode-base/src/provider/gemini.rs
index 485fb0786..48674e373 100644
--- a/crates/jcode-base/src/provider/gemini.rs
+++ b/crates/jcode-base/src/provider/gemini.rs
@@ -512,7 +512,10 @@ impl GeminiProvider {
             user_prompt_id: Uuid::new_v4().to_string(),
             request: VertexGenerateContentRequest {
                 contents: build_contents(messages),
-                system_instruction: build_system_instruction(system),
+                system_instruction: build_system_instruction_with_tool_guard(
+                    system,
+                    !tools.is_empty(),
+                ),
                 tools: build_tools(tools),
                 tool_config: if tools.is_empty() {
                     None
@@ -809,6 +812,12 @@ impl Provider for GeminiProvider {
                         .await;
                     return;
                 }
+                // Track whether this candidate produced any usable output (text or
+                // a tool call). Gemini-3 thinking models intermittently emit
+                // Python-style pseudo-code instead of a clean functionCall and
+                // finish with `MALFORMED_FUNCTION_CALL` and empty content; surface
+                // that as a retryable error below rather than a silent empty turn.
+                let mut produced_output = false;
                 if let Some(content) = candidate.content {
                     // Gemini 3 attaches a `thoughtSignature` to function-call
                     // parts (and occasionally to a standalone preceding part).
@@ -826,9 +835,11 @@ impl Provider for GeminiProvider {
                         if let Some(text) = part.text
                             && !text.is_empty()
                         {
+                            produced_output = true;
                             let _ = tx.send(Ok(StreamEvent::TextDelta(text))).await;
                         }
                         if let Some(function_call) = part.function_call {
+                            produced_output = true;
                             let signature =
                                 part_signature.clone().or_else(|| pending_signature.take());
                             let raw_call_id = function_call
@@ -857,6 +868,47 @@ impl Provider for GeminiProvider {
                             pending_signature = Some(signature);
                         }
                     }
+                    // A thought signature not consumed by a following function
+                    // call (e.g. a pure-text reasoning turn) is still an opaque
+                    // reasoning signal. Surface it as a ThinkingSignatureDelta
+                    // instead of dropping it.
+                    if let Some(signature) = pending_signature.take() {
+                        let _ = tx
+                            .send(Ok(StreamEvent::ThinkingSignatureDelta(signature)))
+                            .await;
+                    }
+                }
+
+                // An abnormal finish (typically Gemini-3's intermittent
+                // `MALFORMED_FUNCTION_CALL`) that yielded no text and no tool call
+                // is a dead turn: surface it as a retryable error instead of a
+                // silent empty `MessageEnd`. `STOP`/`MAX_TOKENS` are normal.
+                if !produced_output {
+                    let abnormal = candidate
+                        .finish_reason
+                        .as_deref()
+                        .map(|reason| {
+                            !matches!(
+                                reason.to_ascii_uppercase().as_str(),
+                                "STOP" | "MAX_TOKENS" | "FINISH_REASON_UNSPECIFIED" | ""
+                            )
+                        })
+                        .unwrap_or(false);
+                    if abnormal {
+                        let reason = candidate.finish_reason.as_deref().unwrap_or("unknown");
+                        let detail = candidate
+                            .finish_message
+                            .as_deref()
+                            .filter(|msg| !msg.trim().is_empty())
+                            .map(|msg| format!(": {}", crate::util::truncate_str(msg.trim(), 300)))
+                            .unwrap_or_default();
+                        let _ = tx
+                            .send(Err(anyhow::anyhow!(
+                                "Gemini returned no usable output (finish_reason={reason}){detail}"
+                            )))
+                            .await;
+                        return;
+                    }
                 }
             }
 
@@ -1016,6 +1068,35 @@ pub(crate) fn build_system_instruction(system: &str) -> Option<GeminiContent> {
     }
 }
 
+/// Prevention guidance appended to the Gemini system prompt when tools are
+/// advertised. Gemini-3 "thinking" models intermittently emit Python-style
+/// pseudo-code (e.g. `print(default_api.read(...))`) instead of a clean
+/// `functionCall`, which the backend rejects with `MALFORMED_FUNCTION_CALL` and
+/// empty content. Explicitly forbidding code/namespaces measurably reduces that
+/// failure mode at no latency cost (see the Gemini function-calling guidance and
+/// field reports of this exact behavior).
+const GEMINI_FUNCTION_CALL_GUARD: &str = "\n\n## Function calling\n\
+     - When you call a tool, emit a native function call, not code. Never write \
+     Python (or any language) that calls the tool, and never wrap a call in \
+     print(...) or a code block.\n\
+     - Use the function name exactly as defined. Do not prepend `default_api.` \
+     or any other namespace to the function name.";
+
+/// Build the Gemini `system_instruction`, appending [`GEMINI_FUNCTION_CALL_GUARD`]
+/// when tools are advertised so the model is steered away from the
+/// `MALFORMED_FUNCTION_CALL` pseudo-code failure mode.
+pub(crate) fn build_system_instruction_with_tool_guard(
+    system: &str,
+    has_tools: bool,
+) -> Option<GeminiContent> {
+    if !has_tools {
+        return build_system_instruction(system);
+    }
+    let mut combined = system.trim().to_string();
+    combined.push_str(GEMINI_FUNCTION_CALL_GUARD);
+    build_system_instruction(&combined)
+}
+
 pub(crate) fn build_contents(messages: &[Message]) -> Vec<GeminiContent> {
     messages
         .iter()
diff --git a/crates/jcode-base/src/provider/gemini_tests.rs b/crates/jcode-base/src/provider/gemini_tests.rs
index 3fc03084a..bd856fbae 100644
--- a/crates/jcode-base/src/provider/gemini_tests.rs
+++ b/crates/jcode-base/src/provider/gemini_tests.rs
@@ -726,3 +726,35 @@ fn developer_api_response_parses_without_code_assist_envelope() {
         .expect("missing text");
     assert_eq!(text, "hello from developer api");
 }
+
+#[test]
+fn system_instruction_tool_guard_only_applies_with_tools() {
+    // Without tools, the system instruction is passed through unchanged.
+    let plain = super::build_system_instruction_with_tool_guard("You are helpful.", false)
+        .expect("system instruction present");
+    let plain_text = plain.parts[0].text.clone().unwrap();
+    assert_eq!(plain_text, "You are helpful.");
+    assert!(!plain_text.contains("Function calling"));
+
+    // With tools, the MALFORMED_FUNCTION_CALL prevention guidance is appended.
+    let guarded = super::build_system_instruction_with_tool_guard("You are helpful.", true)
+        .expect("system instruction present");
+    let guarded_text = guarded.parts[0].text.clone().unwrap();
+    assert!(guarded_text.starts_with("You are helpful."));
+    assert!(guarded_text.contains("Function calling"));
+    assert!(guarded_text.contains("native function call, not code"));
+    assert!(guarded_text.contains("default_api."));
+}
+
+#[test]
+fn system_instruction_tool_guard_with_empty_system_still_emits_guidance() {
+    // An empty base system prompt plus tools must still carry the guard so the
+    // model is steered away from pseudo-code tool calls.
+    let guarded = super::build_system_instruction_with_tool_guard("", true)
+        .expect("guard-only instruction present");
+    let text = guarded.parts[0].text.clone().unwrap();
+    assert!(text.contains("Function calling"));
+
+    // Empty system and no tools yields no instruction at all.
+    assert!(super::build_system_instruction_with_tool_guard("", false).is_none());
+}
diff --git a/crates/jcode-base/src/provider/mod.rs b/crates/jcode-base/src/provider/mod.rs
index bfeccd7f9..570adcf2c 100644
--- a/crates/jcode-base/src/provider/mod.rs
+++ b/crates/jcode-base/src/provider/mod.rs
@@ -1042,6 +1042,19 @@ impl Provider for MultiProvider {
         }
     }
 
+    fn display_name(&self) -> String {
+        // The OpenRouter slot multiplexes the public aggregator and every
+        // direct OpenAI-compatible profile (NVIDIA NIM, DeepSeek, ...). Ask the
+        // active execution runtime for its own label so the UI reflects the
+        // profile selected at runtime rather than the fixed "OpenRouter" name.
+        if matches!(self.active_provider(), ActiveProvider::OpenRouter)
+            && let Some(execution) = self.active_openrouter_execution_provider()
+        {
+            return execution.runtime_display_name();
+        }
+        self.name().to_string()
+    }
+
     fn model(&self) -> String {
         match self.active_provider() {
             ActiveProvider::Claude => {
diff --git a/crates/jcode-base/src/provider/openrouter.rs b/crates/jcode-base/src/provider/openrouter.rs
index 18e1ed56a..3590eee5e 100644
--- a/crates/jcode-base/src/provider/openrouter.rs
+++ b/crates/jcode-base/src/provider/openrouter.rs
@@ -1046,6 +1046,45 @@ impl OpenRouterProvider {
         self.supports_provider_features
     }
 
+    /// Human-facing label for the runtime backing this provider instance.
+    ///
+    /// Unlike the env-var based [`crate::provider_catalog::runtime_provider_display_name`],
+    /// this reads the instance's own `profile_id`/`api_base`, so it stays correct
+    /// after a runtime `/model` switch to a different OpenAI-compatible profile
+    /// (e.g. NVIDIA NIM) even though `name()` is fixed at `"openrouter"`.
+    pub(crate) fn runtime_display_name(&self) -> String {
+        // Direct OpenAI-compatible profile (NVIDIA NIM, DeepSeek, Z.AI, ...).
+        if let Some(profile_id) = self.profile_id.as_deref() {
+            if let Some(profile) = openai_compatible_profile_by_id(profile_id) {
+                return profile.display_name.to_string();
+            }
+            return profile_id.to_string();
+        }
+
+        // Non-aggregator endpoint without a known profile id: classify by base
+        // URL so custom OpenAI-compatible endpoints don't masquerade as the
+        // public OpenRouter aggregator.
+        if !self.supports_provider_features {
+            if let Some(profile_id) =
+                crate::provider_catalog::openai_compatible_profile_id_for_api_base(&self.api_base)
+                && let Some(profile) = openai_compatible_profile_by_id(profile_id)
+            {
+                return profile.display_name.to_string();
+            }
+            if std::env::var("JCODE_RUNTIME_PROVIDER")
+                .ok()
+                .is_some_and(|value| value.trim().eq_ignore_ascii_case("azure-openai"))
+            {
+                return "Azure OpenAI".to_string();
+            }
+            if !self.api_base.contains("openrouter.ai") {
+                return "OpenAI-compatible".to_string();
+            }
+        }
+
+        "OpenRouter".to_string()
+    }
+
     pub(crate) fn direct_openai_compatible_route_parts(&self) -> Option<(String, String, String)> {
         if self.supports_provider_features {
             return None;
diff --git a/crates/jcode-base/src/provider/openrouter_provider_impl.rs b/crates/jcode-base/src/provider/openrouter_provider_impl.rs
index cf4c93a75..f3edd04bb 100644
--- a/crates/jcode-base/src/provider/openrouter_provider_impl.rs
+++ b/crates/jcode-base/src/provider/openrouter_provider_impl.rs
@@ -743,6 +743,10 @@ impl Provider for OpenRouterProvider {
         "openrouter"
     }
 
+    fn display_name(&self) -> String {
+        self.runtime_display_name()
+    }
+
     fn model(&self) -> String {
         self.model
             .try_read()
diff --git a/crates/jcode-base/src/provider/openrouter_sse_stream.rs b/crates/jcode-base/src/provider/openrouter_sse_stream.rs
index b9baa3ed5..d72e84c10 100644
--- a/crates/jcode-base/src/provider/openrouter_sse_stream.rs
+++ b/crates/jcode-base/src/provider/openrouter_sse_stream.rs
@@ -445,8 +445,13 @@ impl OpenRouterStream {
         }
 
         while let Some(pos) = self.buffer.find("\n\n") {
+            // Extract this event and remove it (plus the "\n\n" separator) in
+            // place. Reassigning `self.buffer = self.buffer[pos + 2..].to_string()`
+            // copied and reallocated the entire remaining buffer on every event,
+            // which is O(buffer^2) when one network chunk batches many SSE
+            // events. `drain` removes the consumed prefix without reallocating.
             let event_str = self.buffer[..pos].to_string();
-            self.buffer = self.buffer[pos + 2..].to_string();
+            self.buffer.drain(..pos + 2);
 
             // Parse SSE event
             let mut data = None;
diff --git a/crates/jcode-base/src/provider/openrouter_tests.rs b/crates/jcode-base/src/provider/openrouter_tests.rs
index 323e10961..6170f57eb 100644
--- a/crates/jcode-base/src/provider/openrouter_tests.rs
+++ b/crates/jcode-base/src/provider/openrouter_tests.rs
@@ -2194,3 +2194,71 @@ fn strict_openai_schema_endpoint_allows_other_providers() {
         "https://api.openai.com/v1"
     ));
 }
+
+#[test]
+fn runtime_display_name_tracks_active_openai_compatible_profile() {
+    // Regression for issue #329: switching to a direct OpenAI-compatible
+    // profile (NVIDIA NIM) at runtime must surface that profile's display
+    // name, not the fixed "OpenRouter" aggregator label. The machine-facing
+    // `name()` stays "openrouter" because billing/routing logic keys off it.
+    let _lock = ENV_LOCK.lock();
+    let temp = TempDir::new().expect("create temp home");
+    let jcode_home = temp.path().join("jcode-home");
+    let _jcode_home = EnvVarGuard::set("JCODE_HOME", &jcode_home);
+    let _home = EnvVarGuard::set("HOME", temp.path());
+    let _appdata = EnvVarGuard::set("APPDATA", temp.path().join("AppData").join("Roaming"));
+    let _env = isolate_openrouter_autodetect_env();
+
+    // Configure both the OpenRouter aggregator and NVIDIA NIM credentials so
+    // the slot can host either runtime. Set after the isolate guard, which
+    // clears every profile api-key env var.
+    let _or_key = EnvVarGuard::set("OPENROUTER_API_KEY", "or-test-key");
+    let _nim_key = EnvVarGuard::set("NVIDIA_API_KEY", "nim-test-key");
+    crate::config::invalidate_config_cache();
+
+    let provider =
+        crate::provider::MultiProvider::new_with_auth_status(crate::auth::AuthStatus::default());
+
+    // Switch to a NVIDIA NIM model via the profile-prefixed model request.
+    provider
+        .set_model("nvidia-nim:nvidia/llama-3.1-nemotron-ultra-253b-v1")
+        .expect("switch to nvidia-nim profile");
+
+    assert_eq!(
+        Provider::name(&provider),
+        "OpenRouter",
+        "machine-facing name must stay stable for billing/routing"
+    );
+    assert_eq!(
+        Provider::display_name(&provider),
+        "NVIDIA NIM",
+        "header/UI display name must reflect the active runtime profile"
+    );
+
+    // Switching back to the plain OpenRouter aggregator restores the label.
+    provider
+        .set_model("anthropic/claude-sonnet-4")
+        .expect("switch back to openrouter aggregator");
+    assert_eq!(Provider::display_name(&provider), "OpenRouter");
+}
+
+#[test]
+fn runtime_display_name_for_profile_runtime_instance() {
+    // Direct unit coverage of the per-instance resolver used by
+    // `Provider::display_name`.
+    let _lock = ENV_LOCK.lock();
+    let temp = TempDir::new().expect("create temp home");
+    let jcode_home = temp.path().join("jcode-home");
+    let _jcode_home = EnvVarGuard::set("JCODE_HOME", &jcode_home);
+    let _home = EnvVarGuard::set("HOME", temp.path());
+    let _appdata = EnvVarGuard::set("APPDATA", temp.path().join("AppData").join("Roaming"));
+    let _env = isolate_openrouter_autodetect_env();
+    let _key = EnvVarGuard::set("NVIDIA_API_KEY", "nim-test-key");
+
+    let nim = OpenRouterProvider::new_openai_compatible_profile_runtime(
+        crate::provider_catalog::NVIDIA_NIM_PROFILE,
+    )
+    .expect("build nvidia-nim runtime");
+    assert_eq!(nim.runtime_display_name(), "NVIDIA NIM");
+    assert_eq!(Provider::name(&nim), "openrouter");
+}
diff --git a/crates/jcode-base/src/session/render.rs b/crates/jcode-base/src/session/render.rs
index 4c5421dc9..456fdb110 100644
--- a/crates/jcode-base/src/session/render.rs
+++ b/crates/jcode-base/src/session/render.rs
@@ -1,5 +1,6 @@
 use super::{Session, StoredDisplayRole};
 use crate::message::{ContentBlock, Role, ToolCall};
+use jcode_config_types::ReasoningDisplayMode;
 pub use jcode_session_types::{
     RenderedCompactedHistoryInfo, RenderedImage, RenderedImageSource, RenderedMessage,
 };
@@ -16,10 +17,30 @@ pub const DEFAULT_VISIBLE_COMPACTED_HISTORY_MESSAGES: usize = 64;
 /// by the live streaming path. Each line is wrapped via the shared `reasoning_line_markup` so resumed
 /// sessions render reasoning identically to how it streamed, terminated by a
 /// blank line so following answer text renders as a normal paragraph.
+///
+/// Honors the active `reasoning_display` mode so re-rendered history (reload,
+/// resume, remote sync, compaction-window expand) matches the live behavior:
+/// - `Off`: persisted reasoning is hidden entirely.
+/// - `Current`: the block folds down to a single `▸ thought (N lines)` trace,
+///   matching the live collapse animation's end state rather than replaying the
+///   full reasoning back into the transcript on every reload.
+/// - `Full`: every reasoning line is shown (classic behavior).
 fn format_reasoning_markup(text: &str) -> String {
     if text.trim().is_empty() {
         return String::new();
     }
+    let mode = crate::config::config().display.reasoning_display();
+    match mode {
+        ReasoningDisplayMode::Off => return String::new(),
+        ReasoningDisplayMode::Current => {
+            let line_count = text.lines().filter(|l| !l.trim().is_empty()).count();
+            let mut out = jcode_tui_markdown::reasoning_summary_line_markup(line_count);
+            // Blank line terminates the reasoning block.
+            out.push('\n');
+            return out;
+        }
+        ReasoningDisplayMode::Full => {}
+    }
     let mut out = String::new();
     for line in text.split('\n') {
         out.push_str(&jcode_tui_markdown::reasoning_line_markup(line));
diff --git a/crates/jcode-base/src/session_metrics.rs b/crates/jcode-base/src/session_metrics.rs
new file mode 100644
index 000000000..6cfad046c
--- /dev/null
+++ b/crates/jcode-base/src/session_metrics.rs
@@ -0,0 +1,210 @@
+//! Lock-free per-session runtime metrics.
+//!
+//! These metrics are tracked in a process-global registry rather than on the
+//! `Agent` struct itself. That is deliberate: callers such as `swarm list`
+//! read per-agent stats while the agent may be actively processing a turn and
+//! holding its own `Mutex<Agent>` lock. Anything stored behind that lock is
+//! unavailable (`try_lock` fails) exactly when an agent is busiest, which is
+//! when churn/turn data is most interesting. Keeping these counters in a
+//! separate registry lets us observe live activity without contending on the
+//! agent lock.
+//!
+//! The registry stores a small ring of recent token-usage samples per session
+//! so we can report a "tokens churned over the last N seconds" rate, plus a
+//! cumulative turn counter.
+
+use std::collections::HashMap;
+use std::sync::Mutex;
+use std::time::{Duration, Instant};
+
+/// How long an individual token sample stays in the rolling window.
+const SAMPLE_WINDOW: Duration = Duration::from_secs(60);
+
+/// Maximum samples retained per session to bound memory. At one sample per
+/// provider response this comfortably covers the rolling window.
+const MAX_SAMPLES: usize = 256;
+
+#[derive(Clone, Copy)]
+struct TokenSample {
+    at: Instant,
+    /// Total tokens (input + output + cache) observed in this sample.
+    total: u64,
+    /// Output tokens only, the best proxy for "work produced".
+    output: u64,
+}
+
+#[derive(Default)]
+struct SessionMetrics {
+    samples: Vec<TokenSample>,
+    turns: u64,
+    cumulative_total_tokens: u64,
+    cumulative_output_tokens: u64,
+}
+
+impl SessionMetrics {
+    fn prune(&mut self, now: Instant) {
+        let cutoff = now.checked_sub(SAMPLE_WINDOW);
+        self.samples.retain(|sample| match cutoff {
+            Some(cutoff) => sample.at >= cutoff,
+            None => true,
+        });
+        if self.samples.len() > MAX_SAMPLES {
+            let overflow = self.samples.len() - MAX_SAMPLES;
+            self.samples.drain(0..overflow);
+        }
+    }
+}
+
+static REGISTRY: Mutex<Option<HashMap<String, SessionMetrics>>> = Mutex::new(None);
+
+fn with_registry<R>(f: impl FnOnce(&mut HashMap<String, SessionMetrics>) -> R) -> Option<R> {
+    let mut guard = REGISTRY.lock().ok()?;
+    let map = guard.get_or_insert_with(HashMap::new);
+    Some(f(map))
+}
+
+/// Record a token-usage sample for a session. Called from the streaming turn
+/// loop whenever the provider reports usage.
+pub fn record_token_usage(session_id: &str, total_tokens: u64, output_tokens: u64) {
+    if session_id.is_empty() || (total_tokens == 0 && output_tokens == 0) {
+        return;
+    }
+    let now = Instant::now();
+    with_registry(|map| {
+        let entry = map.entry(session_id.to_string()).or_default();
+        entry.samples.push(TokenSample {
+            at: now,
+            total: total_tokens,
+            output: output_tokens,
+        });
+        entry.cumulative_total_tokens = entry.cumulative_total_tokens.saturating_add(total_tokens);
+        entry.cumulative_output_tokens =
+            entry.cumulative_output_tokens.saturating_add(output_tokens);
+        entry.prune(now);
+    });
+}
+
+/// Record that a session completed (or started) a turn.
+pub fn record_turn(session_id: &str) {
+    if session_id.is_empty() {
+        return;
+    }
+    with_registry(|map| {
+        let entry = map.entry(session_id.to_string()).or_default();
+        entry.turns = entry.turns.saturating_add(1);
+    });
+}
+
+/// Snapshot of a session's recent activity.
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub struct SessionMetricsSnapshot {
+    /// Total tokens observed within the lookback window.
+    pub recent_total_tokens: u64,
+    /// Output tokens observed within the lookback window.
+    pub recent_output_tokens: u64,
+    /// Cumulative total tokens for the session lifetime.
+    pub cumulative_total_tokens: u64,
+    /// Cumulative output tokens for the session lifetime.
+    pub cumulative_output_tokens: u64,
+    /// Number of turns recorded for the session.
+    pub turns: u64,
+}
+
+impl SessionMetricsSnapshot {
+    pub fn has_activity(&self) -> bool {
+        self.recent_total_tokens > 0
+            || self.cumulative_total_tokens > 0
+            || self.turns > 0
+    }
+}
+
+/// Read a snapshot of a session's metrics, summing token samples within the
+/// given lookback window. Returns `None` if the session has no recorded
+/// metrics.
+pub fn snapshot(session_id: &str, lookback: Duration) -> Option<SessionMetricsSnapshot> {
+    let now = Instant::now();
+    with_registry(|map| {
+        let entry = map.get_mut(session_id)?;
+        entry.prune(now);
+        let cutoff = now.checked_sub(lookback);
+        let mut recent_total = 0u64;
+        let mut recent_output = 0u64;
+        for sample in &entry.samples {
+            let in_window = match cutoff {
+                Some(cutoff) => sample.at >= cutoff,
+                None => true,
+            };
+            if in_window {
+                recent_total = recent_total.saturating_add(sample.total);
+                recent_output = recent_output.saturating_add(sample.output);
+            }
+        }
+        Some(SessionMetricsSnapshot {
+            recent_total_tokens: recent_total,
+            recent_output_tokens: recent_output,
+            cumulative_total_tokens: entry.cumulative_total_tokens,
+            cumulative_output_tokens: entry.cumulative_output_tokens,
+            turns: entry.turns,
+        })
+    })
+    .flatten()
+}
+
+/// Remove a session's metrics, called when the session leaves the swarm or
+/// disconnects, to avoid unbounded growth.
+pub fn forget(session_id: &str) {
+    with_registry(|map| {
+        map.remove(session_id);
+    });
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn records_and_snapshots_token_usage() {
+        let sid = "session_metrics_test_basic";
+        forget(sid);
+        record_token_usage(sid, 100, 40);
+        record_token_usage(sid, 50, 20);
+        let snap = snapshot(sid, Duration::from_secs(10)).expect("snapshot");
+        assert_eq!(snap.recent_total_tokens, 150);
+        assert_eq!(snap.recent_output_tokens, 60);
+        assert_eq!(snap.cumulative_total_tokens, 150);
+        assert_eq!(snap.cumulative_output_tokens, 60);
+        forget(sid);
+    }
+
+    #[test]
+    fn counts_turns() {
+        let sid = "session_metrics_test_turns";
+        forget(sid);
+        record_turn(sid);
+        record_turn(sid);
+        record_turn(sid);
+        let snap = snapshot(sid, Duration::from_secs(10)).expect("snapshot");
+        assert_eq!(snap.turns, 3);
+        forget(sid);
+    }
+
+    #[test]
+    fn ignores_empty_and_zero() {
+        let sid = "session_metrics_test_zero";
+        forget(sid);
+        record_token_usage(sid, 0, 0);
+        record_token_usage("", 100, 40);
+        assert!(snapshot(sid, Duration::from_secs(10)).is_none());
+        forget(sid);
+    }
+
+    #[test]
+    fn forget_clears_state() {
+        let sid = "session_metrics_test_forget";
+        forget(sid);
+        record_turn(sid);
+        assert!(snapshot(sid, Duration::from_secs(10)).is_some());
+        forget(sid);
+        assert!(snapshot(sid, Duration::from_secs(10)).is_none());
+    }
+}
diff --git a/crates/jcode-base/src/session_tests/cases.rs b/crates/jcode-base/src/session_tests/cases.rs
index 01582b6b2..beb9e4aaa 100644
--- a/crates/jcode-base/src/session_tests/cases.rs
+++ b/crates/jcode-base/src/session_tests/cases.rs
@@ -1069,6 +1069,10 @@ fn test_render_messages_honors_system_display_role_override() {
 fn test_render_messages_renders_persisted_reasoning() {
     use jcode_tui_markdown::REASONING_SENTINEL;
 
+    let _env_lock = lock_env();
+    let _mode = EnvVarGuard::set("JCODE_REASONING_DISPLAY", "full");
+    crate::config::invalidate_config_cache();
+
     let mut session = Session::create_with_id(
         "session_render_reasoning_test".to_string(),
         None,
@@ -1114,6 +1118,10 @@ fn test_render_messages_renders_persisted_reasoning() {
 fn test_render_messages_renders_legacy_reasoning_variant() {
     use jcode_tui_markdown::REASONING_SENTINEL;
 
+    let _env_lock = lock_env();
+    let _mode = EnvVarGuard::set("JCODE_REASONING_DISPLAY", "full");
+    crate::config::invalidate_config_cache();
+
     let mut session = Session::create_with_id(
         "session_render_legacy_reasoning_test".to_string(),
         None,
@@ -1138,6 +1146,87 @@ fn test_render_messages_renders_legacy_reasoning_variant() {
     );
 }
 
+#[test]
+fn test_render_messages_collapses_persisted_reasoning_in_current_mode() {
+    use jcode_tui_markdown::REASONING_SENTINEL;
+
+    let _env_lock = lock_env();
+    let _mode = EnvVarGuard::set("JCODE_REASONING_DISPLAY", "current");
+    crate::config::invalidate_config_cache();
+
+    let mut session = Session::create_with_id(
+        "session_render_reasoning_current_test".to_string(),
+        None,
+        Some("render reasoning current test".to_string()),
+    );
+
+    session.add_message(
+        Role::Assistant,
+        vec![
+            ContentBlock::ReasoningTrace {
+                text: "step one\nstep two\nstep three".to_string(),
+            },
+            ContentBlock::Text {
+                text: "Here is the answer.".to_string(),
+                cache_control: None,
+            },
+        ],
+    );
+
+    let rendered = render_messages(&session);
+    assert_eq!(rendered.len(), 1);
+    let content = &rendered[0].content;
+    // In `current` mode re-rendered history folds the whole reasoning block down
+    // to a single dim/italic trace line, matching the live collapse end state.
+    assert!(
+        content.contains(&format!("*{0}▸ thought (3 lines){0}*", REASONING_SENTINEL)),
+        "expected collapsed reasoning summary, got: {content:?}"
+    );
+    assert!(
+        !content.contains("step one") && !content.contains("step two"),
+        "individual reasoning lines must not be replayed in current mode: {content:?}"
+    );
+    // The answer text is preserved and follows the collapsed trace.
+    assert!(content.contains("Here is the answer."));
+}
+
+#[test]
+fn test_render_messages_hides_persisted_reasoning_in_off_mode() {
+    use jcode_tui_markdown::REASONING_SENTINEL;
+
+    let _env_lock = lock_env();
+    let _mode = EnvVarGuard::set("JCODE_REASONING_DISPLAY", "off");
+    crate::config::invalidate_config_cache();
+
+    let mut session = Session::create_with_id(
+        "session_render_reasoning_off_test".to_string(),
+        None,
+        Some("render reasoning off test".to_string()),
+    );
+
+    session.add_message(
+        Role::Assistant,
+        vec![
+            ContentBlock::ReasoningTrace {
+                text: "secret thought".to_string(),
+            },
+            ContentBlock::Text {
+                text: "Here is the answer.".to_string(),
+                cache_control: None,
+            },
+        ],
+    );
+
+    let rendered = render_messages(&session);
+    assert_eq!(rendered.len(), 1);
+    let content = &rendered[0].content;
+    assert!(
+        !content.contains(REASONING_SENTINEL) && !content.contains("secret thought"),
+        "reasoning must be hidden entirely in off mode: {content:?}"
+    );
+    assert!(content.contains("Here is the answer."));
+}
+
 #[test]
 fn test_render_messages_honors_background_task_display_role_override() {
     let mut session = Session::create_with_id(
diff --git a/crates/jcode-base/src/skill.rs b/crates/jcode-base/src/skill.rs
index 9f55efae5..e35f0bf89 100644
--- a/crates/jcode-base/src/skill.rs
+++ b/crates/jcode-base/src/skill.rs
@@ -976,6 +976,49 @@ mod tests {
         );
     }
 
+    #[test]
+    fn endorsed_skills_include_nvidia_cuda_x_catalog() {
+        let endorsed = endorsed_skills();
+        // Spot-check representative NVIDIA CUDA-X skills sourced from the
+        // official NVIDIA/skills catalog.
+        for expected in [
+            "cuopt-numerical-optimization-api-python",
+            "cupynumeric-install",
+            "accelerated-computing-cudf",
+            "cudaq-guide",
+            "tilegym-adding-cutile-kernel",
+        ] {
+            let skill = endorsed
+                .iter()
+                .find(|s| s.name == expected)
+                .unwrap_or_else(|| panic!("expected endorsed NVIDIA skill {expected}"));
+            assert_eq!(skill.category, "NVIDIA CUDA-X");
+            assert!(
+                skill
+                    .install
+                    .is_some_and(|cmd| cmd.contains("nvidia/skills")),
+                "NVIDIA skill {expected} should have an nvidia/skills install hint"
+            );
+        }
+    }
+
+    #[test]
+    fn endorsed_skills_include_anthropic_frontend_design() {
+        let skill = endorsed_skills()
+            .iter()
+            .find(|s| s.name == "frontend-design")
+            .expect("expected endorsed Anthropic frontend-design skill");
+        assert_eq!(skill.category, "Anthropic Design");
+        assert!(
+            skill.source.contains("anthropics/skills"),
+            "frontend-design should be sourced from anthropics/skills"
+        );
+        assert!(
+            skill.install.is_some_and(|cmd| cmd.contains("anthropics/skills")),
+            "frontend-design should have an anthropics/skills install hint"
+        );
+    }
+
     #[test]
     fn registry_contains_reports_loaded_skills() {
         let temp = tempfile::tempdir().expect("tempdir");
diff --git a/crates/jcode-build-support/src/lib.rs b/crates/jcode-build-support/src/lib.rs
index f5fc9795f..ad8cc2013 100644
--- a/crates/jcode-build-support/src/lib.rs
+++ b/crates/jcode-build-support/src/lib.rs
@@ -766,6 +766,101 @@ pub fn advance_shared_server_if_tracking_stable(version: &str) -> Result<bool> {
     }
 }
 
+/// Outcome of [`repair_stale_shared_server_channel`].
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum SharedServerRepair {
+    /// The `shared-server` channel was repointed at the installed `stable`
+    /// release because stable was strictly newer on disk.
+    Repaired {
+        previous: Option<String>,
+        repaired_to: String,
+    },
+    /// Nothing to do: shared-server is already at/newer than stable, or there is
+    /// no usable stable target.
+    AlreadyCurrent,
+}
+
+/// Drag a *stale* `shared-server` channel forward to the installed `stable`
+/// release so a long-lived daemon can actually reload into a newer binary.
+///
+/// This is the client-side counterpart to [`advance_shared_server_if_tracking_stable`].
+/// Updates advance `stable` but only advance `shared-server` *during the install
+/// path*; a client that is already on the newest release (so `/update` is a
+/// no-op) never re-runs that install path, leaving a long-lived older daemon
+/// pinned to its old `shared-server` binary forever. A newer client that detects
+/// an older server calls this to repoint `shared-server` -> `stable` before
+/// asking the server to reload, so the forced reload has a strictly-newer target
+/// to exec into instead of re-execing the same old binary (the "current client,
+/// stale server" report).
+///
+/// Safety: we only repair when the `stable` binary is *strictly newer by mtime*
+/// than the current `shared-server` binary. That preserves a deliberately-pinned
+/// self-dev `shared-server` build whenever it is at least as fresh as stable (the
+/// case the pin exists to protect), and never downgrades the channel.
+pub fn repair_stale_shared_server_channel() -> Result<SharedServerRepair> {
+    let stable_version = read_stable_version()?;
+    let Some(stable_version) = stable_version
+        .as_deref()
+        .map(str::trim)
+        .filter(|s| !s.is_empty())
+    else {
+        return Ok(SharedServerRepair::AlreadyCurrent);
+    };
+
+    let stable_binary = stable_binary_path()?;
+    if !stable_binary.exists() {
+        return Ok(SharedServerRepair::AlreadyCurrent);
+    }
+
+    // If shared-server already resolves to the same version marker, there is
+    // nothing to repair.
+    let previous = read_shared_server_version()?;
+    if previous.as_deref().map(str::trim).filter(|s| !s.is_empty()) == Some(stable_version) {
+        return Ok(SharedServerRepair::AlreadyCurrent);
+    }
+
+    // Only repair when stable is strictly newer than the current shared-server
+    // binary on disk. This never downgrades, and it preserves a self-dev pin
+    // that is fresher than stable.
+    let shared_binary = shared_server_binary_path()?;
+    if !shared_server_binary_is_strictly_older_than(&shared_binary, &stable_binary) {
+        return Ok(SharedServerRepair::AlreadyCurrent);
+    }
+
+    update_shared_server_symlink(stable_version)?;
+    Ok(SharedServerRepair::Repaired {
+        previous: previous
+            .as_deref()
+            .map(str::trim)
+            .filter(|s| !s.is_empty())
+            .map(str::to_string),
+        repaired_to: stable_version.to_string(),
+    })
+}
+
+/// True when `shared` exists and is strictly older (by mtime) than `stable`, or
+/// when `shared` is missing entirely (nothing to protect). Any mtime
+/// uncertainty on an existing shared binary is treated as "not older" so we
+/// never repair away an unverifiable (possibly newer) pinned build.
+fn shared_server_binary_is_strictly_older_than(
+    shared: &std::path::Path,
+    stable: &std::path::Path,
+) -> bool {
+    let mtime = |p: &std::path::Path| std::fs::metadata(p).ok().and_then(|m| m.modified().ok());
+    let stable_mtime = match mtime(stable) {
+        Some(m) => m,
+        None => return false,
+    };
+    if !shared.exists() {
+        // No deliberate pin on disk; safe to point the channel at stable.
+        return true;
+    }
+    match mtime(shared) {
+        Some(shared_mtime) => shared_mtime < stable_mtime,
+        None => false,
+    }
+}
+
 /// Install release binary into immutable versions, promote it to stable, and also make it the
 /// active current/launcher build.
 pub fn install_local_release(repo_dir: &std::path::Path) -> Result<PathBuf> {
diff --git a/crates/jcode-build-support/src/tests.rs b/crates/jcode-build-support/src/tests.rs
index 00cda73ed..88af9652f 100644
--- a/crates/jcode-build-support/src/tests.rs
+++ b/crates/jcode-build-support/src/tests.rs
@@ -716,3 +716,116 @@ fn selfdev_reload_target_diverges_from_update_probe_when_shared_server_pinned()
         );
     });
 }
+
+/// Write a distinct, real binary into `versions/<version>/jcode` with an
+/// explicit mtime so channel-repair mtime comparisons are deterministic
+/// (install_binary_at_version hard-links and would share an mtime).
+fn write_versioned_binary(version: &str, mtime: std::time::SystemTime) -> PathBuf {
+    let dir = builds_dir().unwrap().join("versions").join(version);
+    std::fs::create_dir_all(&dir).expect("create version dir");
+    let path = dir.join(binary_name());
+    std::fs::write(&path, format!("bin {version}")).expect("write binary");
+    std::fs::File::open(&path)
+        .expect("open binary")
+        .set_modified(mtime)
+        .expect("set mtime");
+    path
+}
+
+#[test]
+fn repair_repoints_stale_shared_server_to_newer_stable() {
+    use std::time::{Duration, SystemTime};
+    with_temp_jcode_home(|| {
+        let base = SystemTime::UNIX_EPOCH + Duration::from_secs(1_000_000);
+        let old = "0.14.6";
+        let new = "0.22.0";
+        // shared-server pinned to the OLD build; stable advanced to the NEW
+        // release (the "current client, no-op /update, stale server" state).
+        write_versioned_binary(old, base);
+        write_versioned_binary(new, base + Duration::from_secs(60));
+        update_shared_server_symlink(old).expect("pin shared-server old");
+        update_stable_symlink(new).expect("stable new");
+
+        let outcome = repair_stale_shared_server_channel().expect("repair");
+        assert_eq!(
+            outcome,
+            SharedServerRepair::Repaired {
+                previous: Some(old.to_string()),
+                repaired_to: new.to_string(),
+            },
+        );
+        assert_eq!(
+            read_shared_server_version().unwrap().as_deref(),
+            Some(new),
+            "shared-server should be dragged forward to stable"
+        );
+    });
+}
+
+#[test]
+fn repair_is_noop_when_shared_server_already_matches_stable() {
+    use std::time::{Duration, SystemTime};
+    with_temp_jcode_home(|| {
+        let base = SystemTime::UNIX_EPOCH + Duration::from_secs(1_000_000);
+        let v = "0.22.0";
+        write_versioned_binary(v, base);
+        update_shared_server_symlink(v).expect("shared");
+        update_stable_symlink(v).expect("stable");
+
+        assert_eq!(
+            repair_stale_shared_server_channel().expect("repair"),
+            SharedServerRepair::AlreadyCurrent,
+        );
+        assert_eq!(read_shared_server_version().unwrap().as_deref(), Some(v));
+    });
+}
+
+#[test]
+fn repair_preserves_fresher_selfdev_pin() {
+    use std::time::{Duration, SystemTime};
+    with_temp_jcode_home(|| {
+        let base = SystemTime::UNIX_EPOCH + Duration::from_secs(1_000_000);
+        let stable_old = "0.14.3";
+        let selfdev_new = "56f43c3d-dirty-deadbeef";
+        // Deliberately-promoted self-dev build that is NEWER than stable must be
+        // preserved (the whole point of pinning shared-server).
+        write_versioned_binary(stable_old, base);
+        write_versioned_binary(selfdev_new, base + Duration::from_secs(120));
+        update_stable_symlink(stable_old).expect("stable");
+        update_shared_server_symlink(selfdev_new).expect("pin newer self-dev");
+
+        assert_eq!(
+            repair_stale_shared_server_channel().expect("repair"),
+            SharedServerRepair::AlreadyCurrent,
+            "must not downgrade a fresher self-dev pin to an older stable"
+        );
+        assert_eq!(
+            read_shared_server_version().unwrap().as_deref(),
+            Some(selfdev_new),
+        );
+    });
+}
+
+#[test]
+fn repair_never_downgrades_when_stable_is_older() {
+    use std::time::{Duration, SystemTime};
+    with_temp_jcode_home(|| {
+        let base = SystemTime::UNIX_EPOCH + Duration::from_secs(1_000_000);
+        let shared_new = "0.22.0";
+        let stable_old = "0.14.3";
+        write_versioned_binary(stable_old, base);
+        write_versioned_binary(shared_new, base + Duration::from_secs(90));
+        update_shared_server_symlink(shared_new).expect("shared new");
+        update_stable_symlink(stable_old).expect("stable old");
+
+        assert_eq!(
+            repair_stale_shared_server_channel().expect("repair"),
+            SharedServerRepair::AlreadyCurrent,
+            "repair must never move shared-server backward to an older stable"
+        );
+        assert_eq!(
+            read_shared_server_version().unwrap().as_deref(),
+            Some(shared_new),
+        );
+    });
+}
diff --git a/crates/jcode-config-types/src/lib.rs b/crates/jcode-config-types/src/lib.rs
index 4a5e7aa86..7391ca91c 100644
--- a/crates/jcode-config-types/src/lib.rs
+++ b/crates/jcode-config-types/src/lib.rs
@@ -643,7 +643,7 @@ pub struct DisplayConfig {
     pub debug_socket: bool,
     /// Center all content (default: false)
     pub centered: bool,
-    /// Show thinking/reasoning content by default (default: false)
+    /// Show thinking/reasoning content by default (default: true)
     pub show_thinking: bool,
     /// How to display reasoning/thinking content (off/full/current).
     /// When unset, falls back to `show_thinking` (true => full, false => off).
@@ -689,8 +689,8 @@ impl Default for DisplayConfig {
             mouse_capture: true,
             debug_socket: false,
             centered: false,
-            show_thinking: false,
-            reasoning_display: None,
+            show_thinking: true,
+            reasoning_display: Some(ReasoningDisplayMode::Current),
             diagram_mode: DiagramDisplayMode::default(),
             markdown_spacing: MarkdownSpacingMode::default(),
             idle_animation: true,
diff --git a/crates/jcode-desktop/src/desktop_benchmark.rs b/crates/jcode-desktop/src/desktop_benchmark.rs
index 6b8e7a021..a85b565be 100644
--- a/crates/jcode-desktop/src/desktop_benchmark.rs
+++ b/crates/jcode-desktop/src/desktop_benchmark.rs
@@ -43,6 +43,38 @@ pub(super) fn resize_render_benchmark_frames(args: &[String]) -> Option<usize> {
     })
 }
 
+/// Parse `--real-transcript-scroll-benchmark[=N]`, the number of scroll frames
+/// to profile against each of the user's largest real on-disk transcripts.
+pub(super) fn real_transcript_scroll_benchmark_frames(args: &[String]) -> Option<usize> {
+    args.iter().enumerate().find_map(|(index, arg)| {
+        arg.strip_prefix("--real-transcript-scroll-benchmark=")
+            .and_then(|value| value.parse::<usize>().ok())
+            .or_else(|| {
+                (arg == "--real-transcript-scroll-benchmark").then(|| {
+                    args.get(index + 1)
+                        .and_then(|value| value.parse::<usize>().ok())
+                        .unwrap_or(600)
+                })
+            })
+    })
+}
+
+/// Parse `--real-transcript-action-benchmark[=N]`, the per-phase frame count for
+/// the multi-action interaction benchmark run against real on-disk transcripts.
+pub(super) fn real_transcript_action_benchmark_frames(args: &[String]) -> Option<usize> {
+    args.iter().enumerate().find_map(|(index, arg)| {
+        arg.strip_prefix("--real-transcript-action-benchmark=")
+            .and_then(|value| value.parse::<usize>().ok())
+            .or_else(|| {
+                (arg == "--real-transcript-action-benchmark").then(|| {
+                    args.get(index + 1)
+                        .and_then(|value| value.parse::<usize>().ok())
+                        .unwrap_or(400)
+                })
+            })
+    })
+}
+
 pub(super) fn benchmark_phase(
     mut frames: usize,
     mut run_frame: impl FnMut(usize) -> usize,
diff --git a/crates/jcode-desktop/src/main.rs b/crates/jcode-desktop/src/main.rs
index 31eca67ad..947f9b096 100644
--- a/crates/jcode-desktop/src/main.rs
+++ b/crates/jcode-desktop/src/main.rs
@@ -145,6 +145,16 @@ const SINGLE_SESSION_CARET_COLOR: [f32; 4] = [0.130, 0.150, 0.190, 0.92];
 const SESSION_SPAWN_REFRESH_DELAY: Duration = Duration::from_millis(350);
 const BACKGROUND_POLL_INTERVAL: Duration = Duration::from_millis(33);
 const BACKEND_REDRAW_FRAME_INTERVAL: Duration = Duration::from_millis(16);
+/// Minimum spacing between animation-driven redraws.
+///
+/// Without this, the desktop render loop re-requests a redraw immediately after
+/// every animated frame (welcome-hero reveal, focus pulse, spinners, smooth
+/// scroll, etc.). Because the surface uses non-blocking `Mailbox` presentation,
+/// `present()` returns instantly, so the unthrottled loop renders at hundreds of
+/// fps and pins the main thread near 100% CPU, starving input handling and the
+/// compositor (the root cause of desktop lag/jank). ~16ms paces continuous
+/// animations to about 60fps, matching typical display refresh.
+const DESKTOP_ANIMATION_FRAME_INTERVAL: Duration = Duration::from_millis(16);
 const SURFACE_TIMEOUT_BACKOFF_MIN: Duration = Duration::from_millis(16);
 const SURFACE_TIMEOUT_BACKOFF_MAX: Duration = Duration::from_millis(250);
 const HEADLESS_CHAT_SMOKE_TIMEOUT: Duration = Duration::from_secs(90);
@@ -383,6 +393,17 @@ fn desktop_background_wake(
     }
 }
 
+/// Compute the next paced animation redraw time.
+///
+/// Returns `Some(now + DESKTOP_ANIMATION_FRAME_INTERVAL)` while an animation is
+/// active and `None` once it settles. Callers schedule this instead of calling
+/// `request_redraw()` immediately, which would render as fast as the CPU allows
+/// (the surface presents without blocking) and pin the main thread near 100%
+/// CPU, starving input handling and the compositor.
+fn next_animation_redraw_at(now: Instant, animation_active: bool) -> Option<Instant> {
+    animation_active.then(|| now + DESKTOP_ANIMATION_FRAME_INTERVAL)
+}
+
 #[derive(Clone, Copy, Debug, PartialEq)]
 struct StreamingTextArrivalStyle {
     opacity: f32,
@@ -695,6 +716,12 @@ async fn run() -> Result<()> {
     if let Some(frames) = scroll_render_benchmark_frames(&args) {
         return run_scroll_render_benchmark(frames);
     }
+    if let Some(frames) = real_transcript_scroll_benchmark_frames(&args) {
+        return run_real_transcript_scroll_benchmark(frames);
+    }
+    if let Some(frames) = real_transcript_action_benchmark_frames(&args) {
+        return run_real_transcript_action_benchmark(frames);
+    }
     if let Some(output_dir) = hero_screenshot_capture_dir(&args) {
         return run_hero_screenshot_capture(&output_dir).await;
     }
@@ -799,6 +826,10 @@ async fn run() -> Result<()> {
     let mut pending_backend_redraw_since: Option<Instant> = None;
     let mut surface_timeout_backoff = SurfaceTimeoutBackoff::default();
     let mut surface_timeout_redraw_at: Option<Instant> = None;
+    // Scheduled time for the next animation-driven redraw. Continuous animations
+    // re-arm this each presented frame so the loop paces itself to roughly the
+    // display refresh rate instead of busy-spinning the main thread.
+    let mut animation_redraw_at: Option<Instant> = None;
     let mut pending_resize: Option<PhysicalSize<u32>> = None;
     let mut space_hold_started_at: Option<Instant> = None;
     let mut space_hold_consumed = false;
@@ -845,6 +876,7 @@ async fn run() -> Result<()> {
             hot_reload_wake,
             space_hold_wake,
             surface_timeout_redraw_at,
+            animation_redraw_at,
         ]
             .into_iter()
             .flatten()
@@ -1608,9 +1640,14 @@ async fn run() -> Result<()> {
                             target.exit();
                             return;
                         }
-                        if frame.animation_active {
-                            window.request_redraw();
-                        }
+                        // Pace continuous animations instead of immediately
+                        // re-requesting a redraw. An immediate request makes the
+                        // event loop render as fast as the CPU allows (the surface
+                        // presents without blocking), pinning the main thread near
+                        // 100% CPU and starving input/compositor scheduling. The
+                        // scheduled wake is serviced in AboutToWait.
+                        animation_redraw_at =
+                            next_animation_redraw_at(Instant::now(), frame.animation_active);
                     }
                     Err(SurfaceError::Lost | SurfaceError::Outdated) => {
                         surface_timeout_backoff.reset();
@@ -1841,6 +1878,18 @@ async fn run() -> Result<()> {
                         }
                     }
                 }
+                // Service the paced animation redraw scheduled by RedrawRequested.
+                // This keeps continuous animations advancing at ~display refresh
+                // without busy-spinning the loop between frames.
+                if let Some(redraw_at) = animation_redraw_at {
+                    let now = Instant::now();
+                    if now >= redraw_at {
+                        animation_redraw_at = None;
+                        if surface_renderable {
+                            window.request_redraw();
+                        }
+                    }
+                }
                 if surface_renderable && app.is_single_session() {
                     let about_to_wait_started = Instant::now();
                     let size = window.inner_size();
@@ -1909,8 +1958,15 @@ async fn run() -> Result<()> {
                 {
                     canvas.needs_initial_frame = false;
                     window.request_redraw();
-                } else if surface_renderable && app.has_frame_animation() {
-                    window.request_redraw();
+                } else if surface_renderable
+                    && app.has_frame_animation()
+                    && animation_redraw_at.is_none()
+                {
+                    // An animation is active but no paced redraw is scheduled yet
+                    // (e.g. it just became active). Schedule one instead of
+                    // requesting a redraw on every loop iteration, which would
+                    // busy-spin the main thread at 100% CPU.
+                    animation_redraw_at = next_animation_redraw_at(Instant::now(), true);
                 }
             }
             _ => {}
@@ -2219,6 +2275,8 @@ const DESKTOP_HELP_LINES: &[&str] = &[
     "  --capture-hero-animation DIR Write deterministic hero animation PNG frames and exit",
     "  --resize-render-benchmark[N]  Print CPU resize/render benchmark JSON and exit",
     "  --scroll-render-benchmark[N]  Print CPU scroll/render benchmark JSON and exit",
+    "  --real-transcript-scroll-benchmark[N]  Profile scrolling against your real on-disk transcripts and exit",
+    "  --real-transcript-action-benchmark[N]  Profile mixed user actions (scroll/resize/typing/pickers/selection/streaming) on real transcripts and exit",
     "  --stream-e2e-benchmark[N]     Print stream event-to-paint guardrail JSON and exit",
     "  --headless-chat-smoke <MSG>  Run a hidden backend smoke test and print JSON events",
     "  --headless-chat-smoke=<MSG>  Same as above",
@@ -5111,6 +5169,879 @@ fn run_scroll_render_benchmark(frames: usize) -> Result<()> {
     Ok(())
 }
 
+/// Profile scrolling against the user's real on-disk transcripts.
+///
+/// This loads the largest real session files (full, untruncated message lists)
+/// and drives the exact production windowed-scroll render path: cached body
+/// wrap, a sliding text-buffer window, viewport extraction, glyph shaping for
+/// the visible window, text areas, and primitive geometry. Per-frame work is
+/// reported per session and aggregated so we can attribute any scroll jank to a
+/// specific stage on real content rather than synthetic fixtures.
+fn run_real_transcript_scroll_benchmark(frames: usize) -> Result<()> {
+    let frames = frames.max(1);
+    let size = PhysicalSize::new(1200, 760);
+    let transcripts = session_data::load_largest_real_transcripts(8, 24)
+        .context("failed to load real transcripts for scroll benchmark")?;
+
+    if transcripts.is_empty() {
+        println!(
+            "{}",
+            serde_json::to_string_pretty(&serde_json::json!({
+                "frames": frames,
+                "sessions": [],
+                "note": "no real transcripts with >=24 messages found under ~/.jcode/sessions",
+            }))?
+        );
+        return Ok(());
+    }
+
+    let mut session_reports = Vec::new();
+    let mut all_frame_samples: Vec<f64> = Vec::new();
+    let mut worst_stage_us = 0.0_f64;
+    let mut worst_stage_name = String::new();
+
+    for transcript in &transcripts {
+        let report = benchmark_real_transcript_scroll(transcript, size, frames);
+        if report.worst_stage_us > worst_stage_us {
+            worst_stage_us = report.worst_stage_us;
+            worst_stage_name = report.worst_stage_name.clone();
+        }
+        all_frame_samples.extend_from_slice(&report.frame_samples);
+        session_reports.push(report);
+    }
+
+    let budget_ms = duration_ms(DESKTOP_120FPS_FRAME_BUDGET);
+    let aggregate_p50 = percentile_ms(&all_frame_samples, 0.50);
+    let aggregate_p95 = percentile_ms(&all_frame_samples, 0.95);
+    let aggregate_p99 = percentile_ms(&all_frame_samples, 0.99);
+    let aggregate_max = max_sample_ms(&all_frame_samples);
+    let passes_budget = aggregate_p99 <= budget_ms;
+
+    let sessions_json = session_reports
+        .iter()
+        .map(RealTranscriptScrollReport::to_json)
+        .collect::<Vec<_>>();
+
+    println!(
+        "{}",
+        serde_json::to_string_pretty(&serde_json::json!({
+            "frames": frames,
+            "size": { "width": size.width, "height": size.height },
+            "target_frame_budget_ms": budget_ms,
+            "sessions_profiled": session_reports.len(),
+            "aggregate_full_scroll_frame": {
+                "frames": all_frame_samples.len(),
+                "p50_ms": aggregate_p50,
+                "p95_ms": aggregate_p95,
+                "p99_ms": aggregate_p99,
+                "max_ms": aggregate_max,
+            },
+            "worst_stage": { "name": worst_stage_name, "max_us_per_frame": worst_stage_us },
+            "passes_120fps_scroll_cpu_budget": passes_budget,
+            "sessions": sessions_json,
+        }))?
+    );
+    Ok(())
+}
+
+struct RealTranscriptScrollReport {
+    session_id: String,
+    title: String,
+    file_bytes: u64,
+    message_count: usize,
+    total_body_lines: usize,
+    max_scroll_lines: usize,
+    body_buffer_rebuilds: usize,
+    frame_samples: Vec<f64>,
+    stage_totals_us: Vec<(&'static str, f64)>,
+    setup_full_relayout_ms: f64,
+    worst_stage_name: String,
+    worst_stage_us: f64,
+    worst_rebuild_us: f64,
+    worst_rebuild_window_lines: usize,
+    worst_rebuild_max_line_chars: usize,
+    worst_rebuild_advanced_lines: usize,
+    worst_rebuild_segments: usize,
+}
+
+impl RealTranscriptScrollReport {
+    fn to_json(&self) -> serde_json::Value {
+        let frames = self.frame_samples.len().max(1);
+        let total_ms = self.frame_samples.iter().sum::<f64>();
+        let stages = self
+            .stage_totals_us
+            .iter()
+            .map(|(name, total_us)| {
+                serde_json::json!({
+                    "name": name,
+                    "mean_us_per_frame": total_us / frames as f64,
+                    "total_ms": total_us / 1000.0,
+                })
+            })
+            .collect::<Vec<_>>();
+        serde_json::json!({
+            "session_id": self.session_id,
+            "title": self.title,
+            "file_bytes": self.file_bytes,
+            "message_count": self.message_count,
+            "total_body_lines": self.total_body_lines,
+            "max_scroll_lines": self.max_scroll_lines,
+            "body_buffer_rebuilds": self.body_buffer_rebuilds,
+            "setup_full_body_relayout_ms": self.setup_full_relayout_ms,
+            "worst_window_rebuild": {
+                "us": self.worst_rebuild_us,
+                "window_lines": self.worst_rebuild_window_lines,
+                "max_line_chars": self.worst_rebuild_max_line_chars,
+                "advanced_shaping_lines": self.worst_rebuild_advanced_lines,
+                "segments": self.worst_rebuild_segments,
+            },
+            "full_scroll_frame": {
+                "frames": self.frame_samples.len(),
+                "mean_ms_per_frame": total_ms / frames as f64,
+                "p50_ms": percentile_ms(&self.frame_samples, 0.50),
+                "p95_ms": percentile_ms(&self.frame_samples, 0.95),
+                "p99_ms": percentile_ms(&self.frame_samples, 0.99),
+                "max_ms": max_sample_ms(&self.frame_samples),
+            },
+            "subphases": stages,
+        })
+    }
+}
+
+/// Build a `SingleSessionApp` backed by a full real transcript, exactly the way
+/// the production resume path hydrates one from disk.
+fn real_transcript_scroll_app(transcript: &session_data::BenchmarkTranscript) -> SingleSessionApp {
+    let mut app = SingleSessionApp::new(None);
+    app.apply_resumed_session_transcript(transcript.messages.clone());
+    app.set_status_label(format!("real transcript: {}", transcript.title));
+    app
+}
+
+fn benchmark_real_transcript_scroll(
+    transcript: &session_data::BenchmarkTranscript,
+    size: PhysicalSize<u32>,
+    frames: usize,
+) -> RealTranscriptScrollReport {
+    let mut app = real_transcript_scroll_app(transcript);
+    let mut font_system = benchmark_font_system();
+
+    // One-time full body wrap (the cost paid when a transcript is first loaded
+    // or the window is resized). After this, scrolling must stay windowed.
+    let setup_started = Instant::now();
+    let body_lines = single_session_rendered_body_lines_for_tick(&app, size, 0);
+    let setup_full_relayout_ms = setup_started.elapsed().as_secs_f64() * 1000.0;
+    let total_body_lines = body_lines.len();
+
+    let max_scroll_lines = single_session_body_scroll_metrics_for_total_lines(
+        &app,
+        size,
+        total_body_lines,
+    )
+    .map(|metrics| metrics.max_scroll_lines)
+    .unwrap_or(0);
+
+    // Prime the sliding text-buffer window at the bottom of the transcript, the
+    // way the app does after hydrating a resumed session.
+    app.scroll_body_to_bottom();
+    let initial_viewport = single_session_body_viewport_from_lines(&app, size, 0.0, &body_lines);
+    let initial_key =
+        single_session_text_key_for_tick_with_rendered_body(&app, size, 0, 0.0, &body_lines);
+    let mut buffers = single_session_text_buffers_from_key(&initial_key, size, &mut font_system);
+    let (mut window_start, mut window_end) =
+        single_session_body_text_window_bounds(&initial_viewport);
+    if let Some(body_buffer) = buffers.get_mut(1) {
+        *body_buffer = single_session_body_text_buffer_from_lines(
+            &mut font_system,
+            &body_lines[window_start..window_end],
+            size,
+            app.text_scale(),
+        );
+        body_buffer.set_scroll(
+            initial_viewport
+                .start_line
+                .saturating_sub(window_start)
+                .min(i32::MAX as usize) as i32,
+        );
+    }
+    let mut last_scroll_start = initial_viewport.start_line;
+
+    // Drive a long scroll sweep from bottom to top and back, one whole line per
+    // frame, so every frame crosses a new line boundary (the worst realistic
+    // continuous-scroll case).
+    let span = max_scroll_lines.max(1);
+    let mut viewport_us = 0.0;
+    let mut window_rebuild_us = 0.0;
+    let mut scroll_us = 0.0;
+    let mut glyph_us = 0.0;
+    let mut areas_us = 0.0;
+    let mut vertices_us = 0.0;
+    let mut body_buffer_rebuilds = 0usize;
+
+    // Optional diagnostic: capture the single slowest window rebuild and describe
+    // the window content so we can attribute the cost (line count, advanced
+    // shaping triggers, longest line) rather than guessing.
+    let diagnose = std::env::var_os("JCODE_DESKTOP_SCROLL_DIAG").is_some();
+    let mut worst_rebuild_us = 0.0_f64;
+    let mut worst_rebuild_window_lines = 0usize;
+    let mut worst_rebuild_max_line_chars = 0usize;
+    let mut worst_rebuild_advanced_lines = 0usize;
+    let mut worst_rebuild_segments = 0usize;
+
+    let (frame_samples, _checksum) = benchmark_frame_samples(frames, |frame| {
+        // Triangle-wave scroll position covering the full transcript height.
+        let phase = frame % (span * 2);
+        let target = if phase <= span { phase } else { span * 2 - phase };
+        app.body_scroll_lines = target as f32;
+        let tick = frame as u64;
+
+        let phase_started = Instant::now();
+        let viewport = single_session_body_viewport_from_lines(&app, size, 0.0, &body_lines);
+        viewport_us += phase_started.elapsed().as_secs_f64() * 1_000_000.0;
+
+        let phase_started = Instant::now();
+        if !single_session_body_text_window_contains(window_start, window_end, &viewport) {
+            (window_start, window_end) = single_session_body_text_window_bounds(&viewport);
+            let rebuild_started = Instant::now();
+            if let Some(body_buffer) = buffers.get_mut(1) {
+                *body_buffer = single_session_body_text_buffer_from_lines(
+                    &mut font_system,
+                    &body_lines[window_start..window_end],
+                    size,
+                    app.text_scale(),
+                );
+            }
+            if diagnose {
+                let rebuild_us = rebuild_started.elapsed().as_secs_f64() * 1_000_000.0;
+                if rebuild_us > worst_rebuild_us {
+                    worst_rebuild_us = rebuild_us;
+                    let window = &body_lines[window_start..window_end];
+                    worst_rebuild_window_lines = window.len();
+                    worst_rebuild_max_line_chars =
+                        window.iter().map(|l| l.text.chars().count()).max().unwrap_or(0);
+                    worst_rebuild_advanced_lines = window
+                        .iter()
+                        .filter(|l| !l.text.is_ascii())
+                        .count();
+                    worst_rebuild_segments =
+                        window.iter().map(|l| l.inline_spans.len() + 1).sum();
+                    if let Ok(path) = std::env::var("JCODE_DESKTOP_SCROLL_DIAG_DUMP") {
+                        let text = window
+                            .iter()
+                            .map(|l| l.text.as_str())
+                            .collect::<Vec<_>>()
+                            .join("\n");
+                        let _ = std::fs::write(
+                            format!("{path}.{}", transcript.session_id),
+                            text,
+                        );
+                    }
+                }
+            }
+            body_buffer_rebuilds += 1;
+            last_scroll_start = usize::MAX;
+        }
+        window_rebuild_us += phase_started.elapsed().as_secs_f64() * 1_000_000.0;
+
+        let phase_started = Instant::now();
+        if viewport.start_line != last_scroll_start {
+            if let Some(body_buffer) = buffers.get_mut(1) {
+                body_buffer.set_scroll(
+                    viewport
+                        .start_line
+                        .saturating_sub(window_start)
+                        .min(i32::MAX as usize) as i32,
+                );
+            }
+            last_scroll_start = viewport.start_line;
+        }
+        scroll_us += phase_started.elapsed().as_secs_f64() * 1_000_000.0;
+
+        let phase_started = Instant::now();
+        let glyph_checksum = buffers
+            .get(1)
+            .map(|body_buffer| {
+                body_buffer
+                    .layout_runs()
+                    .map(|run| run.glyphs.len())
+                    .sum::<usize>()
+            })
+            .unwrap_or_default();
+        glyph_us += phase_started.elapsed().as_secs_f64() * 1_000_000.0;
+
+        let phase_started = Instant::now();
+        let areas = single_session_text_areas_for_app_with_cached_body_viewport(
+            &app, &buffers, size, 0.0, viewport,
+        );
+        areas_us += phase_started.elapsed().as_secs_f64() * 1_000_000.0;
+
+        let phase_started = Instant::now();
+        let vertices = build_single_session_vertices_with_cached_body(
+            &app, size, 0.0, tick, 0.0, 1.0, &body_lines,
+        );
+        vertices_us += phase_started.elapsed().as_secs_f64() * 1_000_000.0;
+
+        buffers.len() ^ areas.len() ^ vertices.len() ^ glyph_checksum
+    });
+
+    let stage_totals_us = vec![
+        ("viewport_extract", viewport_us),
+        ("body_window_rebuild", window_rebuild_us),
+        ("body_scroll_set", scroll_us),
+        ("glyph_layout_count", glyph_us),
+        ("text_areas", areas_us),
+        ("primitive_vertices", vertices_us),
+    ];
+    let frames_f = frames.max(1) as f64;
+    let (worst_stage_name, worst_stage_us) = stage_totals_us
+        .iter()
+        .map(|(name, total)| (name.to_string(), total / frames_f))
+        .fold((String::new(), 0.0_f64), |acc, candidate| {
+            if candidate.1 > acc.1 { candidate } else { acc }
+        });
+
+    RealTranscriptScrollReport {
+        session_id: transcript.session_id.clone(),
+        title: transcript.title.clone(),
+        file_bytes: transcript.file_bytes,
+        message_count: transcript.messages.len(),
+        total_body_lines,
+        max_scroll_lines,
+        body_buffer_rebuilds,
+        frame_samples,
+        stage_totals_us,
+        setup_full_relayout_ms,
+        worst_stage_name,
+        worst_stage_us,
+        worst_rebuild_us,
+        worst_rebuild_window_lines,
+        worst_rebuild_max_line_chars,
+        worst_rebuild_advanced_lines,
+        worst_rebuild_segments,
+    }
+}
+
+/// Profile a realistic mix of user *actions* (not just scrolling) against the
+/// user's largest real on-disk transcripts. Each action phase is measured
+/// separately as per-frame CPU samples and reported as p50/p95/p99/max, plus a
+/// `passes_120fps_cpu_budget` flag against the existing frame budget. This is the
+/// broad interaction-coverage companion to `--real-transcript-scroll-benchmark`.
+fn run_real_transcript_action_benchmark(frames: usize) -> Result<()> {
+    let frames = frames.max(1);
+    let size = PhysicalSize::new(1200, 760);
+    let transcripts = session_data::load_largest_real_transcripts(8, 24)
+        .context("failed to load real transcripts for action benchmark")?;
+
+    if transcripts.is_empty() {
+        println!(
+            "{}",
+            serde_json::to_string_pretty(&serde_json::json!({
+                "frames": frames,
+                "sessions": [],
+                "note": "no real transcripts with >=24 messages found under ~/.jcode/sessions",
+            }))?
+        );
+        return Ok(());
+    }
+
+    let budget_ms = duration_ms(DESKTOP_120FPS_FRAME_BUDGET);
+    // phase name -> all per-frame samples across every session
+    let mut phase_samples: std::collections::BTreeMap<&'static str, Vec<f64>> =
+        std::collections::BTreeMap::new();
+    let mut session_json = Vec::new();
+
+    for transcript in &transcripts {
+        let phases = benchmark_real_transcript_actions(transcript, size, frames);
+        let phase_json = phases
+            .iter()
+            .map(|(name, samples)| {
+                phase_samples
+                    .entry(name)
+                    .or_default()
+                    .extend_from_slice(samples);
+                action_phase_json(name, samples, budget_ms)
+            })
+            .collect::<Vec<_>>();
+        session_json.push(serde_json::json!({
+            "session_id": transcript.session_id,
+            "title": transcript.title,
+            "message_count": transcript.messages.len(),
+            "phases": phase_json,
+        }));
+    }
+
+    let mut aggregate = Vec::new();
+    let mut slowest_phase = String::new();
+    let mut slowest_p99 = 0.0_f64;
+    let mut all_pass = true;
+    for (name, samples) in &phase_samples {
+        let value = action_phase_json(name, samples, budget_ms);
+        let p99 = percentile_ms(samples, 0.99);
+        if p99 > slowest_p99 {
+            slowest_p99 = p99;
+            slowest_phase = (*name).to_string();
+        }
+        if p99 > budget_ms {
+            all_pass = false;
+        }
+        aggregate.push(value);
+    }
+
+    println!(
+        "{}",
+        serde_json::to_string_pretty(&serde_json::json!({
+            "frames_per_phase": frames,
+            "size": { "width": size.width, "height": size.height },
+            "target_frame_budget_ms": budget_ms,
+            "sessions_profiled": transcripts.len(),
+            "aggregate_phases": aggregate,
+            "slowest_phase": { "name": slowest_phase, "p99_ms": slowest_p99 },
+            "passes_120fps_cpu_budget": all_pass,
+            "sessions": session_json,
+        }))?
+    );
+    Ok(())
+}
+
+fn action_phase_json(name: &str, samples: &[f64], budget_ms: f64) -> serde_json::Value {
+    let frames = samples.len().max(1);
+    let total_ms = samples.iter().sum::<f64>();
+    let p99 = percentile_ms(samples, 0.99);
+    serde_json::json!({
+        "name": name,
+        "frames": samples.len(),
+        "mean_ms": total_ms / frames as f64,
+        "p50_ms": percentile_ms(samples, 0.50),
+        "p95_ms": percentile_ms(samples, 0.95),
+        "p99_ms": p99,
+        "max_ms": max_sample_ms(samples),
+        "passes_budget": p99 <= budget_ms,
+    })
+}
+
+/// Run every simulated action phase for one transcript, returning per-phase
+/// per-frame CPU samples (milliseconds). Each phase reproduces the production
+/// render path: cached/wrapped body lines, viewport extraction, a windowed body
+/// text buffer that is reused across frames, text areas, and primitive geometry.
+fn benchmark_real_transcript_actions(
+    transcript: &session_data::BenchmarkTranscript,
+    size: PhysicalSize<u32>,
+    frames: usize,
+) -> Vec<(&'static str, Vec<f64>)> {
+    let base_app = real_transcript_scroll_app(transcript);
+    let body_lines = single_session_rendered_body_lines_for_tick(&base_app, size, 0);
+    let total_lines = body_lines.len();
+    let max_scroll = single_session_body_scroll_metrics_for_total_lines(&base_app, size, total_lines)
+        .map(|metrics| metrics.max_scroll_lines)
+        .unwrap_or(0)
+        .max(1);
+
+    let mut phases: Vec<(&'static str, Vec<f64>)> = Vec::new();
+
+    // 1. Smooth (fractional) scroll: scroll position advances a whole line per
+    //    frame with a fractional offset, the common trackpad-scroll case.
+    phases.push((
+        "smooth_scroll",
+        action_windowed_render_phase(&base_app, &body_lines, size, frames, |app, frame| {
+            let phase = frame % (max_scroll * 2);
+            let target = if phase <= max_scroll {
+                phase
+            } else {
+                max_scroll * 2 - phase
+            };
+            app.body_scroll_lines = target as f32;
+            benchmark_smooth_scroll_lines(frame)
+        }),
+    ));
+
+    // 2. Whole-line scroll: integer line steps, no fractional offset.
+    phases.push((
+        "whole_line_scroll",
+        action_windowed_render_phase(&base_app, &body_lines, size, frames, |app, frame| {
+            let phase = frame % (max_scroll * 2);
+            let target = if phase <= max_scroll {
+                phase
+            } else {
+                max_scroll * 2 - phase
+            };
+            app.body_scroll_lines = target as f32;
+            0.0
+        }),
+    ));
+
+    // 3. Selection drag across the visible transcript while parked mid-scroll.
+    //    This mirrors the real mouse-handler input path, which calls
+    //    single_session_visible_body (a full transcript wrap, now memoized) and
+    //    hit-tests the cursor on every pointer move, then redraws.
+    {
+        let mut app = base_app.clone();
+        app.body_scroll_lines = (max_scroll / 2) as f32;
+        let initial_visible = single_session_visible_body(&app, size);
+        if let Some(point) = single_session_body_point_at_position(size, 40.0, 80.0, &initial_visible)
+        {
+            app.begin_selection(point);
+        } else {
+            app.begin_selection(SelectionPoint { line: 0, column: 0 });
+        }
+        let mut font_system = benchmark_font_system();
+        let (mut buffers, mut window_start, mut window_end, mut last_start) =
+            action_prime_window(&app, &body_lines, size, &mut font_system);
+        let (samples, _) = benchmark_frame_samples(frames, |frame| {
+            // Real input path: resolve the cursor against the visible body
+            // (full-transcript wrap, memoized) and update the selection.
+            let visible = single_session_visible_body(&app, size);
+            let y = 80.0 + (frame % 600) as f32;
+            let x = 40.0 + (frame % 400) as f32;
+            if let Some(point) = single_session_body_point_at_position(size, x, y, &visible) {
+                app.update_selection(point);
+            }
+            action_render_window(
+                &app,
+                &body_lines,
+                size,
+                frame as u64,
+                0.0,
+                &mut font_system,
+                &mut buffers,
+                &mut window_start,
+                &mut window_end,
+                &mut last_start,
+            )
+        });
+        phases.push(("selection_drag", samples));
+    }
+
+    // 3b. Pure input-side selection hit-test cost (no redraw). This isolates the
+    //     real per-mouse-move work the desktop selection handler does:
+    //     single_session_visible_body (a full-transcript wrap, now memoized) plus
+    //     cursor hit-testing. The redraw it triggers is separately cached, so this
+    //     phase exposes the wrap/memo cost that the combined selection_drag phase
+    //     hides behind geometry building.
+    {
+        let mut app = base_app.clone();
+        app.body_scroll_lines = (max_scroll / 2) as f32;
+        app.begin_selection(SelectionPoint { line: 0, column: 0 });
+        let (samples, _) = benchmark_frame_samples(frames, |frame| {
+            let visible = single_session_visible_body(&app, size);
+            let y = 80.0 + (frame % 600) as f32;
+            let x = 40.0 + (frame % 400) as f32;
+            if let Some(point) = single_session_body_point_at_position(size, x, y, &visible) {
+                app.update_selection(point);
+            }
+            visible.len()
+        });
+        phases.push(("selection_input_hittest", samples));
+    }
+
+    // 4. Typing in the composer while parked at the bottom of the transcript.
+    {
+        let mut app = base_app.clone();
+        app.scroll_body_to_bottom();
+        app.draft.clear();
+        app.draft_cursor = 0;
+        let mut font_system = benchmark_font_system();
+        let (mut buffers, mut window_start, mut window_end, mut last_start) =
+            action_prime_window(&app, &body_lines, size, &mut font_system);
+        let (samples, _) = benchmark_frame_samples(frames, |frame| {
+            app.draft.push(benchmark_typing_char(frame));
+            app.draft_cursor = app.draft.len();
+            action_render_window(
+                &app,
+                &body_lines,
+                size,
+                frame as u64,
+                0.0,
+                &mut font_system,
+                &mut buffers,
+                &mut window_start,
+                &mut window_end,
+                &mut last_start,
+            )
+        });
+        phases.push(("composer_typing", samples));
+    }
+
+    // 5. Model picker open/close toggling over the transcript: every other frame
+    //    opens the inline picker card, invalidating the inline-widget geometry.
+    {
+        let mut app = base_app.clone();
+        app.body_scroll_lines = (max_scroll / 3) as f32;
+        let mut font_system = benchmark_font_system();
+        let (mut buffers, mut window_start, mut window_end, mut last_start) =
+            action_prime_window(&app, &body_lines, size, &mut font_system);
+        let (samples, _) = benchmark_frame_samples(frames, |frame| {
+            app.model_picker.open = frame % 2 == 0;
+            app.model_picker.loading = app.model_picker.open;
+            action_render_window(
+                &app,
+                &body_lines,
+                size,
+                frame as u64,
+                0.0,
+                &mut font_system,
+                &mut buffers,
+                &mut window_start,
+                &mut window_end,
+                &mut last_start,
+            )
+        });
+        app.model_picker.open = false;
+        phases.push(("model_picker_toggle", samples));
+    }
+
+    // 6. Session switcher open/close toggling over the transcript.
+    {
+        let mut app = base_app.clone();
+        app.body_scroll_lines = (max_scroll / 3) as f32;
+        let mut font_system = benchmark_font_system();
+        let (mut buffers, mut window_start, mut window_end, mut last_start) =
+            action_prime_window(&app, &body_lines, size, &mut font_system);
+        let (samples, _) = benchmark_frame_samples(frames, |frame| {
+            app.session_switcher.open = frame % 2 == 0;
+            action_render_window(
+                &app,
+                &body_lines,
+                size,
+                frame as u64,
+                0.0,
+                &mut font_system,
+                &mut buffers,
+                &mut window_start,
+                &mut window_end,
+                &mut last_start,
+            )
+        });
+        app.session_switcher.open = false;
+        phases.push(("session_switcher_toggle", samples));
+    }
+
+    // 7. Window resize sweep: each frame is a different surface size, forcing a
+    //    body re-wrap + window rebuild (the worst non-scroll case).
+    //
+    //    Mirrors production (`cached_single_session_body_lines` non-streaming
+    //    branch): the raw styled lines (markdown parse) are generated ONCE and
+    //    cached across sizes; only the width-dependent wrap re-runs per resize.
+    {
+        let app = base_app.clone();
+        let raw_lines = app.body_styled_lines_for_tick(0);
+        let mut font_system = benchmark_font_system();
+        let (samples, _) = benchmark_frame_samples(frames, |frame| {
+            let resize = benchmark_resize_size(frame);
+            let lines = single_session_rendered_body_lines_from_raw_ref(&app, resize, &raw_lines);
+            let viewport = single_session_body_viewport_from_lines(&app, resize, 0.0, &lines);
+            let key =
+                single_session_text_key_for_tick_with_rendered_body(&app, resize, 0, 0.0, &lines);
+            let mut buffers = single_session_text_buffers_from_key(&key, resize, &mut font_system);
+            let (window_start, window_end) = single_session_body_text_window_bounds(&viewport);
+            if let Some(body_buffer) = buffers.get_mut(1) {
+                *body_buffer = single_session_body_text_buffer_from_lines(
+                    &mut font_system,
+                    &lines[window_start..window_end],
+                    resize,
+                    app.text_scale(),
+                );
+            }
+            let areas = single_session_text_areas_for_app_with_cached_body_viewport(
+                &app, &buffers, resize, 0.0, viewport,
+            );
+            let vertices = build_single_session_vertices_with_cached_body(
+                &app, resize, 0.0, frame as u64, 0.0, 1.0, &lines,
+            );
+            buffers.len() ^ areas.len() ^ vertices.len()
+        });
+        phases.push(("window_resize", samples));
+    }
+
+    // 8. Streaming response growth while scrolled near the bottom: a synthetic
+    //    assistant reply grows by a chunk each frame, the live-streaming case.
+    //
+    //    This mirrors the production renderer's incremental path
+    //    (`cached_single_session_body_lines` for the streaming branch): the
+    //    static transcript body is wrapped ONCE, then each frame only truncates
+    //    back to the static base and appends the wrapped streaming tail, rather
+    //    than re-wrapping the whole transcript every frame.
+    {
+        let mut app = base_app.clone();
+        app.scroll_body_to_bottom();
+        app.streaming_response.push_str("Streaming response starting. ");
+        let mut font_system = benchmark_font_system();
+        let static_base = single_session_rendered_static_body_lines_for_streaming(&app, size, 0)
+            .unwrap_or_else(|| single_session_rendered_body_lines_for_tick(&app, size, 0));
+        let static_len = static_base.len();
+        let mut stream_lines = static_base.clone();
+        let (samples, _) = benchmark_frame_samples(frames, |frame| {
+            app.streaming_response.push_str(
+                "Streaming update chunk with `inline code` and prose that wraps across lines. ",
+            );
+            if frame % 9 == 0 {
+                app.streaming_response.push('\n');
+            }
+            // Incremental: reuse the wrapped static base, only re-wrap the tail.
+            stream_lines.truncate(static_len);
+            append_single_session_streaming_response_rendered_body_lines(
+                &app,
+                size,
+                &mut stream_lines,
+            );
+            let viewport = single_session_body_viewport_from_lines(&app, size, 0.0, &stream_lines);
+            let key = single_session_text_key_for_tick_with_rendered_body(
+                &app,
+                size,
+                0,
+                0.0,
+                &stream_lines,
+            );
+            let mut buffers = single_session_text_buffers_from_key(&key, size, &mut font_system);
+            let (window_start, window_end) = single_session_body_text_window_bounds(&viewport);
+            if let Some(body_buffer) = buffers.get_mut(1) {
+                *body_buffer = single_session_body_text_buffer_from_lines(
+                    &mut font_system,
+                    &stream_lines[window_start..window_end],
+                    size,
+                    app.text_scale(),
+                );
+            }
+            let areas = single_session_text_areas_for_app_with_cached_body_viewport(
+                &app, &buffers, size, 0.0, viewport,
+            );
+            let vertices = build_single_session_vertices_with_cached_body(
+                &app,
+                size,
+                0.0,
+                frame as u64,
+                0.0,
+                1.0,
+                &stream_lines,
+            );
+            buffers.len() ^ areas.len() ^ vertices.len()
+        });
+        phases.push(("streaming_growth", samples));
+    }
+
+    phases
+}
+
+/// Prime a reusable text-buffer set and its windowed body buffer for `app`,
+/// matching how the production renderer seeds the sliding window. Returns the
+/// buffers plus the current (window_start, window_end, last_scroll_start).
+fn action_prime_window(
+    app: &SingleSessionApp,
+    body_lines: &[SingleSessionStyledLine],
+    size: PhysicalSize<u32>,
+    font_system: &mut FontSystem,
+) -> (Vec<Buffer>, usize, usize, usize) {
+    let viewport = single_session_body_viewport_from_lines(app, size, 0.0, body_lines);
+    let key = single_session_text_key_for_tick_with_rendered_body(app, size, 0, 0.0, body_lines);
+    let mut buffers = single_session_text_buffers_from_key(&key, size, font_system);
+    let (window_start, window_end) = single_session_body_text_window_bounds(&viewport);
+    if let Some(body_buffer) = buffers.get_mut(1) {
+        *body_buffer = single_session_body_text_buffer_from_lines(
+            font_system,
+            &body_lines[window_start..window_end],
+            size,
+            app.text_scale(),
+        );
+        body_buffer.set_scroll(
+            viewport
+                .start_line
+                .saturating_sub(window_start)
+                .min(i32::MAX as usize) as i32,
+        );
+    }
+    (buffers, window_start, window_end, viewport.start_line)
+}
+
+/// Render one frame through the production windowed path, reusing the body text
+/// buffer and only rebuilding/rescrolling the window when the viewport leaves it.
+#[allow(clippy::too_many_arguments)]
+fn action_render_window(
+    app: &SingleSessionApp,
+    body_lines: &[SingleSessionStyledLine],
+    size: PhysicalSize<u32>,
+    tick: u64,
+    smooth_scroll_lines: f32,
+    font_system: &mut FontSystem,
+    buffers: &mut Vec<Buffer>,
+    window_start: &mut usize,
+    window_end: &mut usize,
+    last_scroll_start: &mut usize,
+) -> usize {
+    let viewport =
+        single_session_body_viewport_from_lines(app, size, smooth_scroll_lines, body_lines);
+    if !single_session_body_text_window_contains(*window_start, *window_end, &viewport) {
+        let (start, end) = single_session_body_text_window_bounds(&viewport);
+        *window_start = start;
+        *window_end = end;
+        if let Some(body_buffer) = buffers.get_mut(1) {
+            *body_buffer = single_session_body_text_buffer_from_lines(
+                font_system,
+                &body_lines[start..end],
+                size,
+                app.text_scale(),
+            );
+        }
+        *last_scroll_start = usize::MAX;
+    }
+    if viewport.start_line != *last_scroll_start {
+        if let Some(body_buffer) = buffers.get_mut(1) {
+            body_buffer.set_scroll(
+                viewport
+                    .start_line
+                    .saturating_sub(*window_start)
+                    .min(i32::MAX as usize) as i32,
+            );
+        }
+        *last_scroll_start = viewport.start_line;
+    }
+    let areas = single_session_text_areas_for_app_with_cached_body_viewport(
+        app,
+        buffers,
+        size,
+        smooth_scroll_lines,
+        viewport,
+    );
+    let vertices = build_single_session_vertices_with_cached_body(
+        app,
+        size,
+        0.0,
+        tick,
+        smooth_scroll_lines,
+        1.0,
+        body_lines,
+    );
+    buffers.len() ^ areas.len() ^ vertices.len()
+}
+
+/// Drive a windowed-scroll render phase, calling `prepare` each frame to mutate
+/// the app's scroll position (and return any fractional smooth-scroll offset).
+fn action_windowed_render_phase(
+    base_app: &SingleSessionApp,
+    body_lines: &[SingleSessionStyledLine],
+    size: PhysicalSize<u32>,
+    frames: usize,
+    mut prepare: impl FnMut(&mut SingleSessionApp, usize) -> f32,
+) -> Vec<f64> {
+    let mut app = base_app.clone();
+    let mut font_system = benchmark_font_system();
+    let (mut buffers, mut window_start, mut window_end, mut last_start) =
+        action_prime_window(&app, body_lines, size, &mut font_system);
+    let (samples, _) = benchmark_frame_samples(frames, |frame| {
+        let smooth = prepare(&mut app, frame);
+        action_render_window(
+            &app,
+            body_lines,
+            size,
+            frame as u64,
+            smooth,
+            &mut font_system,
+            &mut buffers,
+            &mut window_start,
+            &mut window_end,
+            &mut last_start,
+        )
+    });
+    samples
+}
+
 fn run_stream_e2e_benchmark(raw_events: usize) -> Result<()> {
     let result = run_desktop_stream_end_to_end_benchmark(raw_events);
     println!(
@@ -11856,30 +12787,15 @@ fn build_hero_reveal_texture(
     }
 
     let mut values = vec![1.0_f32; (width * height) as usize];
-    let mut min_value = f32::INFINITY;
-    let mut max_value = 0.0_f32;
     let brush_delay_px = (alpha_bounds.height() * 0.10).max(5.0);
 
-    for y in 0..height {
-        for x in 0..width {
-            let pixel_index = (y * width + x) as usize;
-            let alpha = glyph_rgba[pixel_index * 4];
-            if alpha <= 2 {
-                continue;
-            }
-            let (path_progress, distance) = nearest_hero_stroke_progress(
-                x as f32 + 0.5,
-                y as f32 + 0.5,
-                alpha_bounds,
-                &segments,
-            );
-            let width_delay = (distance / brush_delay_px).min(1.0) * 0.045;
-            let value = (path_progress + width_delay).clamp(0.0, 1.0);
-            values[pixel_index] = value;
-            min_value = min_value.min(value);
-            max_value = max_value.max(value);
-        }
-    }
+    // This per-pixel nearest-stroke search dominates the one-time hero mask
+    // build (hundreds of ms on the UI thread). Each lit pixel is independent
+    // and only reads `glyph_rgba`/`segments`, so split the rows across worker
+    // threads. Output is bit-identical to the serial version; min/max are
+    // reduced afterward from the filled buffer.
+    let (min_value, max_value) =
+        fill_hero_reveal_values(&mut values, width, height, glyph_rgba, alpha_bounds, &segments, brush_delay_px);
 
     if !min_value.is_finite() || max_value <= min_value {
         return None;
@@ -11907,6 +12823,103 @@ fn build_hero_reveal_texture(
     Some(reveal_rgba)
 }
 
+/// Fill `values` with each lit pixel's reveal progress and return the
+/// `(min, max)` of the written values.
+///
+/// The work is split into horizontal row bands processed on separate threads
+/// when the image is large enough to amortize the spawn cost. Pixels are
+/// independent, so the result is identical to a serial fill.
+fn fill_hero_reveal_values(
+    values: &mut [f32],
+    width: u32,
+    height: u32,
+    glyph_rgba: &[u8],
+    alpha_bounds: HeroMaskPixelBounds,
+    segments: &[WelcomeHeroStrokeSegment],
+    brush_delay_px: f32,
+) -> (f32, f32) {
+    let row_stride = width as usize;
+    let compute_row = |row_index: u32, row_values: &mut [f32]| -> (f32, f32) {
+        let mut min_value = f32::INFINITY;
+        let mut max_value = 0.0_f32;
+        let row_offset = row_index as usize * row_stride;
+        for x in 0..width {
+            let pixel_index = row_offset + x as usize;
+            let alpha = glyph_rgba[pixel_index * 4];
+            if alpha <= 2 {
+                continue;
+            }
+            let (path_progress, distance) = nearest_hero_stroke_progress(
+                x as f32 + 0.5,
+                row_index as f32 + 0.5,
+                alpha_bounds,
+                segments,
+            );
+            let width_delay = (distance / brush_delay_px).min(1.0) * 0.045;
+            let value = (path_progress + width_delay).clamp(0.0, 1.0);
+            row_values[x as usize] = value;
+            min_value = min_value.min(value);
+            max_value = max_value.max(value);
+        }
+        (min_value, max_value)
+    };
+
+    let total_pixels = row_stride.saturating_mul(height as usize);
+    let worker_count = hero_reveal_worker_count(total_pixels);
+    if worker_count <= 1 || height < 2 {
+        let mut min_value = f32::INFINITY;
+        let mut max_value = 0.0_f32;
+        for (row_index, row_values) in values.chunks_mut(row_stride).enumerate() {
+            let (row_min, row_max) = compute_row(row_index as u32, row_values);
+            min_value = min_value.min(row_min);
+            max_value = max_value.max(row_max);
+        }
+        return (min_value, max_value);
+    }
+
+    let rows_per_band = (height as usize).div_ceil(worker_count).max(1);
+    let mut min_value = f32::INFINITY;
+    let mut max_value = 0.0_f32;
+    std::thread::scope(|scope| {
+        let mut handles = Vec::new();
+        for (band_index, band) in values.chunks_mut(rows_per_band * row_stride).enumerate() {
+            let first_row = (band_index * rows_per_band) as u32;
+            let compute_row = &compute_row;
+            handles.push(scope.spawn(move || {
+                let mut band_min = f32::INFINITY;
+                let mut band_max = 0.0_f32;
+                for (offset, row_values) in band.chunks_mut(row_stride).enumerate() {
+                    let (row_min, row_max) = compute_row(first_row + offset as u32, row_values);
+                    band_min = band_min.min(row_min);
+                    band_max = band_max.max(row_max);
+                }
+                (band_min, band_max)
+            }));
+        }
+        for handle in handles {
+            if let Ok((band_min, band_max)) = handle.join() {
+                min_value = min_value.min(band_min);
+                max_value = max_value.max(band_max);
+            }
+        }
+    });
+    (min_value, max_value)
+}
+
+/// Number of worker threads to use for the hero reveal fill. Returns 1 for
+/// small images where threading overhead would dominate.
+fn hero_reveal_worker_count(total_pixels: usize) -> usize {
+    const MIN_PIXELS_PER_WORKER: usize = 32 * 1024;
+    if total_pixels < MIN_PIXELS_PER_WORKER * 2 {
+        return 1;
+    }
+    let available = std::thread::available_parallelism()
+        .map(|value| value.get())
+        .unwrap_or(1);
+    let by_work = total_pixels / MIN_PIXELS_PER_WORKER;
+    available.min(by_work).max(1)
+}
+
 fn nearest_hero_stroke_progress(
     x: f32,
     y: f32,
diff --git a/crates/jcode-desktop/src/main_tests.rs b/crates/jcode-desktop/src/main_tests.rs
index 4925d7311..6dc4abaaa 100644
--- a/crates/jcode-desktop/src/main_tests.rs
+++ b/crates/jcode-desktop/src/main_tests.rs
@@ -701,6 +701,123 @@ fn desktop_background_wake_only_tracks_active_frame_animation() {
     assert_eq!(desktop_background_wake(now, false, true), None);
 }
 
+#[test]
+fn next_animation_redraw_paces_active_animations_and_settles_when_idle() {
+    let now = Instant::now();
+
+    // While an animation is active, the next redraw is scheduled one frame
+    // interval out rather than immediately, so the loop does not busy-spin.
+    assert_eq!(
+        next_animation_redraw_at(now, true),
+        Some(now + DESKTOP_ANIMATION_FRAME_INTERVAL)
+    );
+    // Once the animation settles, no further redraw is scheduled and the loop
+    // can park on ControlFlow::Wait.
+    assert_eq!(next_animation_redraw_at(now, false), None);
+    // The pacing interval must be positive; a zero interval would reintroduce
+    // the busy-spin it exists to prevent.
+    assert!(DESKTOP_ANIMATION_FRAME_INTERVAL > Duration::ZERO);
+}
+
+#[test]
+fn hero_reveal_worker_count_falls_back_to_serial_for_small_images() {
+    // Tiny images should not pay thread-spawn overhead.
+    assert_eq!(hero_reveal_worker_count(0), 1);
+    assert_eq!(hero_reveal_worker_count(1024), 1);
+    // Large images should use more than one worker when parallelism is available.
+    let big = hero_reveal_worker_count(8 * 1024 * 1024);
+    let available = std::thread::available_parallelism()
+        .map(|value| value.get())
+        .unwrap_or(1);
+    assert!(big >= 1);
+    assert!(big <= available.max(1));
+}
+
+#[test]
+fn fill_hero_reveal_values_matches_serial_reference() {
+    let width = 64_u32;
+    let height = 48_u32;
+    let alpha_bounds = HeroMaskPixelBounds {
+        min_x: 4,
+        min_y: 4,
+        max_x: width - 4,
+        max_y: height - 4,
+    };
+    // A handful of normalized stroke segments tracing a rough path.
+    let segments = vec![
+        WelcomeHeroStrokeSegment {
+            start: [0.1, 0.2],
+            end: [0.4, 0.5],
+            start_progress: 0.0,
+            end_progress: 0.4,
+        },
+        WelcomeHeroStrokeSegment {
+            start: [0.4, 0.5],
+            end: [0.8, 0.3],
+            start_progress: 0.4,
+            end_progress: 0.8,
+        },
+        WelcomeHeroStrokeSegment {
+            start: [0.8, 0.3],
+            end: [0.9, 0.9],
+            start_progress: 0.8,
+            end_progress: 1.0,
+        },
+    ];
+    // Mark a checkerboard of lit pixels so both branches exercise lit/unlit.
+    let mut glyph_rgba = vec![0_u8; (width * height * 4) as usize];
+    for y in 0..height {
+        for x in 0..width {
+            if (x + y) % 3 == 0 {
+                let index = ((y * width + x) * 4) as usize;
+                glyph_rgba[index] = 200;
+            }
+        }
+    }
+    let brush_delay_px = (alpha_bounds.height() * 0.10).max(5.0);
+
+    // Serial reference computed directly here.
+    let mut expected = vec![1.0_f32; (width * height) as usize];
+    let mut expected_min = f32::INFINITY;
+    let mut expected_max = 0.0_f32;
+    for y in 0..height {
+        for x in 0..width {
+            let pixel_index = (y * width + x) as usize;
+            if glyph_rgba[pixel_index * 4] <= 2 {
+                continue;
+            }
+            let (path_progress, distance) = nearest_hero_stroke_progress(
+                x as f32 + 0.5,
+                y as f32 + 0.5,
+                alpha_bounds,
+                &segments,
+            );
+            let width_delay = (distance / brush_delay_px).min(1.0) * 0.045;
+            let value = (path_progress + width_delay).clamp(0.0, 1.0);
+            expected[pixel_index] = value;
+            expected_min = expected_min.min(value);
+            expected_max = expected_max.max(value);
+        }
+    }
+
+    // The parallel implementation must produce bit-identical output regardless
+    // of how many worker threads it chose.
+    let mut actual = vec![1.0_f32; (width * height) as usize];
+    let (actual_min, actual_max) = fill_hero_reveal_values(
+        &mut actual,
+        width,
+        height,
+        &glyph_rgba,
+        alpha_bounds,
+        &segments,
+        brush_delay_px,
+    );
+
+    assert_eq!(actual, expected, "parallel hero reveal fill must match serial");
+    assert_eq!(actual_min.to_bits(), expected_min.to_bits());
+    assert_eq!(actual_max.to_bits(), expected_max.to_bits());
+}
+
 #[test]
 fn desktop_async_job_slots_are_bounded_and_released() -> Result<()> {
     let counter = std::sync::atomic::AtomicUsize::new(0);
diff --git a/crates/jcode-desktop/src/session_data.rs b/crates/jcode-desktop/src/session_data.rs
index 0a363dcc6..72df5ec68 100644
--- a/crates/jcode-desktop/src/session_data.rs
+++ b/crates/jcode-desktop/src/session_data.rs
@@ -90,6 +90,78 @@ pub fn load_session_transcript_by_id(
     Ok(None)
 }
 
+/// A full, uncapped transcript loaded straight from disk, used by the
+/// real-transcript scroll benchmark so we profile the production render path
+/// against the user's actual session content rather than synthetic fixtures.
+#[derive(Debug, Clone)]
+pub struct BenchmarkTranscript {
+    pub session_id: String,
+    pub title: String,
+    pub file_bytes: u64,
+    pub messages: Vec<SessionTranscriptMessage>,
+}
+
+/// Load the largest real session transcripts on disk (by file size), returning
+/// the full message list for each (no card-style truncation). Used only by the
+/// scroll benchmark. Sessions with fewer than `min_messages` are skipped so the
+/// benchmark exercises long, scroll-heavy transcripts.
+pub fn load_largest_real_transcripts(
+    max_sessions: usize,
+    min_messages: usize,
+) -> Result<Vec<BenchmarkTranscript>> {
+    let sessions_dir = jcode_sessions_dir()?;
+    if !sessions_dir.exists() {
+        return Ok(Vec::new());
+    }
+
+    let mut candidates = fs::read_dir(&sessions_dir)
+        .with_context(|| format!("failed to read {}", sessions_dir.display()))?
+        .filter_map(|entry| entry.ok())
+        .filter_map(|entry| {
+            let path = entry.path();
+            session_file_candidate(path.clone())?;
+            let bytes = path.metadata().ok()?.len();
+            Some((path, bytes))
+        })
+        .collect::<Vec<_>>();
+    // Largest files first: they hold the longest transcripts and stress the
+    // windowed-scroll path the most.
+    candidates.sort_by_key(|(_, bytes)| std::cmp::Reverse(*bytes));
+
+    let mut transcripts = Vec::new();
+    for (path, bytes) in candidates {
+        if transcripts.len() >= max_sessions {
+            break;
+        }
+        let session = match load_stored_session(&path) {
+            Ok(session) => session,
+            Err(_) => continue,
+        };
+        let messages = session_transcript_messages(&session);
+        if messages.len() < min_messages {
+            continue;
+        }
+        let id = stored_string(session.id.as_deref())
+            .or_else(|| {
+                path.file_stem()
+                    .map(|stem| stem.to_string_lossy().into_owned())
+            })
+            .unwrap_or_else(|| "unknown-session".to_string());
+        let title = stored_string(session.custom_title.as_deref())
+            .or_else(|| stored_string(session.title.as_deref()))
+            .or_else(|| latest_user_preview(&messages))
+            .unwrap_or_else(|| short_session_name(&id));
+        transcripts.push(BenchmarkTranscript {
+            session_id: id,
+            title,
+            file_bytes: bytes,
+            messages,
+        });
+    }
+
+    Ok(transcripts)
+}
+
 fn load_recent_session_cards_with_limit(limit: usize) -> Result<Vec<SessionCard>> {
     let sessions_dir = jcode_sessions_dir()?;
     if !sessions_dir.exists() {
diff --git a/crates/jcode-desktop/src/single_session_render.rs b/crates/jcode-desktop/src/single_session_render.rs
index 7abf41cf6..e3ba6264c 100644
--- a/crates/jcode-desktop/src/single_session_render.rs
+++ b/crates/jcode-desktop/src/single_session_render.rs
@@ -7209,6 +7209,24 @@ fn push_single_session_inline_code_cards(
     );
 }
 
+/// A thread-local, lazily-initialized `FontSystem` used purely for measuring
+/// glyph layout (inline-code/math pill bounds) during geometry building.
+///
+/// Building a `FontSystem` rescans every system font from disk, costing several
+/// milliseconds per call. The inline-code/math card builder runs on every frame
+/// whose visible window contains inline code or math, so constructing a fresh
+/// `FontSystem` there made scrolling over code blocks janky (multi-ms spikes per
+/// frame). Caching one per render thread keeps repeated measurement cheap. The
+/// system is only used for transient measurement buffers, never for the glyphs
+/// actually uploaded to the GPU, so reuse is safe.
+fn with_measurement_font_system<R>(f: impl FnOnce(&mut FontSystem) -> R) -> R {
+    thread_local! {
+        static MEASUREMENT_FONT_SYSTEM: std::cell::RefCell<FontSystem> =
+            std::cell::RefCell::new(FontSystem::new());
+    }
+    MEASUREMENT_FONT_SYSTEM.with(|cell| f(&mut cell.borrow_mut()))
+}
+
 fn push_single_session_inline_code_cards_from_viewport(
     vertices: &mut Vec<Vertex>,
     app: &SingleSessionApp,
@@ -7245,13 +7263,9 @@ fn push_single_session_inline_code_cards_from_viewport(
         horizontal_pad,
         top_offset_pixels: viewport.top_offset_pixels,
     };
-    let mut font_system = FontSystem::new();
-    let body_buffer = single_session_body_text_buffer_from_lines(
-        &mut font_system,
-        &viewport.lines,
-        size,
-        text_scale,
-    );
+    let body_buffer = with_measurement_font_system(|font_system| {
+        single_session_body_text_buffer_from_lines(font_system, &viewport.lines, size, text_scale)
+    });
     let layout_runs = body_buffer.layout_runs().collect::<Vec<_>>();
 
     let mut occurrences = HashMap::new();
@@ -9073,7 +9087,11 @@ pub(crate) fn single_session_body_viewport_for_tick(
     tick: u64,
     smooth_scroll_lines: f32,
 ) -> SingleSessionBodyViewport {
-    let lines = single_session_rendered_body_lines_for_tick(app, size, tick);
+    // Borrow the memoized full body lines and only clone the visible slice via
+    // `single_session_body_viewport_from_lines`, instead of cloning the whole
+    // transcript. This keeps input-side callers (selection hit-testing on every
+    // mouse-move) O(visible) rather than O(transcript).
+    let lines = single_session_rendered_body_lines_for_tick_shared(app, size, tick);
     single_session_body_viewport_from_lines(app, size, smooth_scroll_lines, &lines)
 }
 
@@ -9117,7 +9135,55 @@ pub(crate) fn single_session_rendered_body_lines_for_tick(
     size: PhysicalSize<u32>,
     tick: u64,
 ) -> Vec<SingleSessionStyledLine> {
-    single_session_rendered_body_lines_from_raw(app, size, app.body_styled_lines_for_tick(tick))
+    (*single_session_rendered_body_lines_for_tick_shared(app, size, tick)).clone()
+}
+
+/// Shared, memoized rendered body lines for the current transcript+layout.
+///
+/// This re-parses markdown and re-wraps the ENTIRE transcript (O(transcript)),
+/// and is called from input handling (every selection mouse-move during a
+/// drag), scroll-metric probing, and several geometry builders. Returning a
+/// shared `Rc` lets callers that only need a slice (the viewport) avoid cloning
+/// the whole transcript on every pointer event. The render hot path uses a
+/// separate Canvas-side cache (`cached_single_session_body_lines`); this
+/// thread-local single-entry memo accelerates the remaining callers. The key is
+/// the body cache key, which already captures the message fingerprint, size,
+/// text scale, and welcome/streaming state, so the cache invalidates whenever
+/// any of those change.
+pub(crate) fn single_session_rendered_body_lines_for_tick_shared(
+    app: &SingleSessionApp,
+    size: PhysicalSize<u32>,
+    tick: u64,
+) -> std::rc::Rc<Vec<SingleSessionStyledLine>> {
+    let layout_size = single_session_body_layout_cache_size(app, size);
+    let key = app.rendered_body_cache_key(layout_size);
+    thread_local! {
+        static RENDERED_BODY_LINES_MEMO: std::cell::RefCell<Option<(u64, std::rc::Rc<Vec<SingleSessionStyledLine>>)>> =
+            const { std::cell::RefCell::new(None) };
+    }
+    // Allow disabling the memo for A/B perf measurement in debug builds only;
+    // the production memo can never be turned off by an env var.
+    let memo_disabled = cfg!(debug_assertions)
+        && std::env::var_os("JCODE_DESKTOP_DISABLE_BODY_MEMO").is_some();
+    if !memo_disabled
+        && let Some(cached) = RENDERED_BODY_LINES_MEMO.with(|cell| {
+            cell.borrow()
+                .as_ref()
+                .filter(|(cached_key, _)| *cached_key == key)
+                .map(|(_, lines)| lines.clone())
+        })
+    {
+        return cached;
+    }
+    let lines =
+        single_session_rendered_body_lines_from_raw(app, size, app.body_styled_lines_for_tick(tick));
+    let shared = std::rc::Rc::new(lines);
+    if !memo_disabled {
+        RENDERED_BODY_LINES_MEMO.with(|cell| {
+            *cell.borrow_mut() = Some((key, shared.clone()));
+        });
+    }
+    shared
 }
 
 pub(crate) fn single_session_rendered_body_lines_from_raw(
diff --git a/crates/jcode-desktop/src/single_session_render/text_style.rs b/crates/jcode-desktop/src/single_session_render/text_style.rs
index 564f8b9a8..0a76c9994 100644
--- a/crates/jcode-desktop/src/single_session_render/text_style.rs
+++ b/crates/jcode-desktop/src/single_session_render/text_style.rs
@@ -80,9 +80,17 @@ pub(super) fn single_session_styled_text_buffer_with_opacity(
     buffer.set_size(font_system, width, height);
     buffer.set_wrap(font_system, wrap);
     let segments = single_session_styled_text_segments_with_opacity(lines, opacity);
-    // Inline span geometry uses glyphon cursors with byte offsets. Basic shaping
-    // reports glyph clusters relative to each styled run, so spans after a
-    // multi-byte marker or a style boundary can shift their pills into prose.
+    // Inline span geometry uses glyphon cursors with byte offsets, and the
+    // glyphon `highlight()` API used to position inline-code/math pills only
+    // works on Advanced-shaped buffers. So any line carrying inline spans must be
+    // Advanced-shaped regardless of script. Advanced shaping is also required for
+    // text containing complex scripts, combining marks, or joiner sequences.
+    //
+    // The expensive case on real transcripts was emoji-rich *prose* lines (no
+    // inline spans): standalone pictographic emoji render identically under Basic
+    // and Advanced shaping, so `char_needs_advanced_shaping` no longer escalates
+    // for them. That keeps the visible-window reshape on every scroll frame cheap
+    // while preserving correct pill geometry for code/math spans.
     let shaping = if lines.iter().any(|line| !line.inline_spans.is_empty())
         || segments
             .iter()
@@ -125,9 +133,16 @@ pub(super) fn char_needs_advanced_shaping(ch: char) -> bool {
             | 0x0590..=0x08FF
             | 0x0900..=0x0DFF
             | 0x1780..=0x18AF
-            // Emoji and symbol sequences often depend on variation selectors / ZWJ.
-            | 0x1F000..=0x1FAFF
+            // Regional indicators combine into flag emoji (pairs need shaping).
+            | 0x1F1E6..=0x1F1FF
     )
+    // Note: standalone pictographic emoji and symbols (e.g. 🔄 ⬜ → ✓) render
+    // identically under Basic and Advanced shaping (single fallback glyph each),
+    // so they intentionally do NOT force Advanced shaping here. Advanced shaping
+    // is several times more expensive and is the dominant per-frame cost when
+    // scrolling emoji-rich transcripts. Only sequences that actually depend on
+    // ligature/joiner shaping (variation selectors, ZWJ, regional-indicator flag
+    // pairs) escalate, which the ranges above already cover.
 }
 
 #[cfg_attr(not(test), allow(dead_code))]
diff --git a/crates/jcode-protocol/src/comm_format.rs b/crates/jcode-protocol/src/comm_format.rs
index c25734014..caa60dfca 100644
--- a/crates/jcode-protocol/src/comm_format.rs
+++ b/crates/jcode-protocol/src/comm_format.rs
@@ -152,6 +152,71 @@ pub fn format_comm_members(current_session_id: &str, members: &[AgentInfo]) -> S
             } else {
                 String::new()
             };
+
+            // Status line: lifecycle + detail, then a contextual age label.
+            // For an idle/ready agent the "age" is how long it has been idle;
+            // for a running agent it is how long the current turn has run.
+            let detail_suffix = member
+                .detail
+                .as_deref()
+                .map(|detail| format!(" — {}", detail))
+                .unwrap_or_default();
+            let age_suffix = match member.status_age_secs {
+                Some(age) if status == "ready" || status == "idle" => {
+                    format!(" · idle {}", format_secs(age))
+                }
+                Some(age) if status == "running" => format!(" · {}", format_secs(age)),
+                Some(age) => format!(" · {} ago", format_secs(age)),
+                None => String::new(),
+            };
+
+            // Live activity: what the agent is doing right now.
+            let activity_suffix = match member.activity.as_ref() {
+                Some(activity) if activity.is_processing => {
+                    match activity.current_tool_name.as_deref() {
+                        Some(tool) => format!("\n    Activity: working ({})", tool),
+                        None => "\n    Activity: thinking".to_string(),
+                    }
+                }
+                _ => String::new(),
+            };
+
+            // Progress: todos completed / total.
+            let progress_suffix = match (member.todos_completed, member.todos_total) {
+                (Some(done), Some(total)) if total > 0 => {
+                    format!("\n    Progress: {}/{} todos", done, total)
+                }
+                _ => String::new(),
+            };
+
+            // Live work signal: recent token churn + cumulative + turns.
+            let mut work_meta = Vec::new();
+            if let (Some(recent), Some(window)) =
+                (member.recent_total_tokens, member.recent_window_secs)
+                && recent > 0
+            {
+                work_meta.push(format!("{} tok/{}s", format_count(recent), window));
+            }
+            if let Some(turns) = member.turn_count.filter(|turns| *turns > 0) {
+                work_meta.push(format!("{} turns", turns));
+            }
+            if let Some(total) = member.cumulative_total_tokens.filter(|total| *total > 0) {
+                work_meta.push(format!("{} tok total", format_count(total)));
+            }
+            let work_suffix = if work_meta.is_empty() {
+                String::new()
+            } else {
+                format!("\n    Work: {}", work_meta.join(" · "))
+            };
+
+            // Model line.
+            let model_suffix = match (member.provider_name.as_deref(), member.provider_model.as_deref())
+            {
+                (Some(provider), Some(model)) => format!("\n    Model: {}/{}", provider, model),
+                (None, Some(model)) => format!("\n    Model: {}", model),
+                _ => String::new(),
+            };
+
             let mut extra_meta = Vec::new();
             if member.is_headless == Some(true) {
                 extra_meta.push("headless".to_string());
@@ -166,37 +231,79 @@ pub fn format_comm_members(current_session_id: &str, members: &[AgentInfo]) -> S
             if let Some(attachments) = member.live_attachments {
                 extra_meta.push(format!("attachments={attachments}"));
             }
-            if let Some(age_secs) = member.status_age_secs {
-                extra_meta.push(format!("status_age={}s", age_secs));
-            }
             let meta_suffix = if extra_meta.is_empty() {
                 String::new()
             } else {
                 format!("\n    Meta: {}", extra_meta.join(" · "))
             };
+
+            // Completion report when the agent has finished.
+            let report_suffix = match member.latest_completion_report.as_deref() {
+                Some(report) if !report.trim().is_empty() => {
+                    format!("\n    Report: {}", truncate_report(report))
+                }
+                _ => String::new(),
+            };
+
             output.push_str(&format!(
-                "  {}{} ({})\n    Status: {}{}{}{}\n",
+                "  {}{} ({})\n    Status: {}{}{}{}{}{}{}{}{}{}\n",
                 name,
                 role_label,
                 if is_me { "you" } else { session },
                 status,
-                member
-                    .detail
-                    .as_deref()
-                    .map(|detail| format!(" — {}", detail))
-                    .unwrap_or_default(),
+                detail_suffix,
+                age_suffix,
+                activity_suffix,
+                progress_suffix,
+                work_suffix,
+                model_suffix,
                 if files.is_empty() {
                     String::new()
                 } else {
                     format!("\n    Files: {}", files)
                 },
-                meta_suffix
+                meta_suffix,
+                report_suffix,
             ));
         }
         output
     }
 }
 
+/// Format a duration in seconds into a compact human label (e.g. `45s`, `3m`, `2h`).
+fn format_secs(secs: u64) -> String {
+    if secs < 60 {
+        format!("{}s", secs)
+    } else if secs < 3600 {
+        format!("{}m", secs / 60)
+    } else {
+        format!("{}h", secs / 3600)
+    }
+}
+
+/// Format a token count compactly (e.g. `850`, `12.3k`, `1.2M`).
+fn format_count(count: u64) -> String {
+    if count < 1_000 {
+        count.to_string()
+    } else if count < 1_000_000 {
+        format!("{:.1}k", count as f64 / 1_000.0)
+    } else {
+        format!("{:.1}M", count as f64 / 1_000_000.0)
+    }
+}
+
+/// Truncate a completion report to a single compact line for the roster view.
+fn truncate_report(report: &str) -> String {
+    const MAX: usize = 120;
+    let one_line: String = report.split_whitespace().collect::<Vec<_>>().join(" ");
+    if one_line.chars().count() > MAX {
+        let truncated: String = one_line.chars().take(MAX).collect();
+        format!("{}…", truncated)
+    } else {
+        one_line
+    }
+}
+
 pub fn format_comm_tool_summary(target: &str, calls: &[ToolCallSummary]) -> String {
     if calls.is_empty() {
         format!("No tool calls found for {}", target)
diff --git a/crates/jcode-protocol/src/lib.rs b/crates/jcode-protocol/src/lib.rs
index e10c42756..20582772c 100644
--- a/crates/jcode-protocol/src/lib.rs
+++ b/crates/jcode-protocol/src/lib.rs
@@ -198,7 +198,7 @@ pub struct ContextEntry {
 }
 
 /// Info about an agent
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
 pub struct AgentInfo {
     pub session_id: String,
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -229,6 +229,36 @@ pub struct AgentInfo {
     /// Seconds since the last status change.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub status_age_secs: Option<u64>,
+    /// Live activity (whether processing + current tool name).
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub activity: Option<SessionActivitySnapshot>,
+    /// Provider name (e.g. "anthropic").
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub provider_name: Option<String>,
+    /// Provider model id.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub provider_model: Option<String>,
+    /// Number of turns the agent has run this session.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub turn_count: Option<u64>,
+    /// Tokens churned (total, including cache) within the recent lookback window.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub recent_total_tokens: Option<u64>,
+    /// Output tokens produced within the recent lookback window.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub recent_output_tokens: Option<u64>,
+    /// Width of the recent-token lookback window, in seconds.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub recent_window_secs: Option<u64>,
+    /// Cumulative total tokens observed for the session lifetime.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub cumulative_total_tokens: Option<u64>,
+    /// Number of completed todos for this agent's session.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub todos_completed: Option<usize>,
+    /// Total number of todos for this agent's session.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub todos_total: Option<usize>,
 }
 
 /// Lightweight status snapshot for a swarm member.
diff --git a/crates/jcode-protocol/src/protocol_tests/comm_responses.rs b/crates/jcode-protocol/src/protocol_tests/comm_responses.rs
index 1bdb0067d..01b0f3147 100644
--- a/crates/jcode-protocol/src/protocol_tests/comm_responses.rs
+++ b/crates/jcode-protocol/src/protocol_tests/comm_responses.rs
@@ -158,6 +158,7 @@ fn test_comm_members_roundtrip_includes_status() -> Result<()> {
             latest_completion_report: Some("Done.".to_string()),
             live_attachments: Some(0),
             status_age_secs: Some(12),
+            ..Default::default()
         }],
     };
 
diff --git a/crates/jcode-provider-core/src/lib.rs b/crates/jcode-provider-core/src/lib.rs
index 73433d8ad..b721ac7df 100644
--- a/crates/jcode-provider-core/src/lib.rs
+++ b/crates/jcode-provider-core/src/lib.rs
@@ -74,8 +74,25 @@ pub trait Provider: Send + Sync {
     }
 
     /// Get the provider name.
+    ///
+    /// This is the stable, machine-facing identifier (e.g. `"openrouter"`,
+    /// `"claude"`). Several surfaces key billing and routing decisions off this
+    /// value, so it must stay constant for a given provider class even when the
+    /// underlying runtime is a specific OpenAI-compatible profile. Use
+    /// [`Provider::display_name`] for anything shown to the user.
     fn name(&self) -> &str;
 
+    /// Human-facing provider label for the *current runtime selection*.
+    ///
+    /// Defaults to [`Provider::name`]. Provider orchestrators that multiplex
+    /// several backends behind one `name()` (notably the OpenRouter slot, which
+    /// also serves direct OpenAI-compatible profiles such as NVIDIA NIM or
+    /// DeepSeek) override this so the UI reflects the profile the user actually
+    /// selected at runtime instead of a fixed aggregator label.
+    fn display_name(&self) -> String {
+        self.name().to_string()
+    }
+
     /// Get the model identifier being used.
     fn model(&self) -> String {
         "unknown".to_string()
@@ -823,7 +840,7 @@ impl ModelCatalogSnapshot {
 
     pub fn from_provider(provider: &dyn Provider) -> Self {
         Self::new(
-            Some(provider.name().to_string()),
+            Some(provider.display_name()),
             Some(provider.model()),
             provider.available_models_display(),
             provider.model_routes(),
diff --git a/crates/jcode-task-types/src/lib.rs b/crates/jcode-task-types/src/lib.rs
index bd14c783c..80bb5d091 100644
--- a/crates/jcode-task-types/src/lib.rs
+++ b/crates/jcode-task-types/src/lib.rs
@@ -200,6 +200,11 @@ pub struct TodoItem {
     pub status: String,
     pub priority: String,
     pub id: String,
+    /// Optional group label. Todos that share a group are displayed together
+    /// under a single header. Use one group per coherent goal; when work is
+    /// steered into a new area, start a new group instead of renaming.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub group: Option<String>,
     /// Forward-looking confidence, from 0-100, that this todo can be completed correctly.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub confidence: Option<u8>,
diff --git a/crates/jcode-tui-core/src/stream_buffer.rs b/crates/jcode-tui-core/src/stream_buffer.rs
index 2b267b483..b8a32c05e 100644
--- a/crates/jcode-tui-core/src/stream_buffer.rs
+++ b/crates/jcode-tui-core/src/stream_buffer.rs
@@ -1,20 +1,53 @@
-//! Semantic stream buffer - chunks streaming text at natural boundaries
+//! Semantic stream buffer - paces streaming text reveal at a smooth rate.
+//!
+//! Providers feed text deltas with wildly different cadences. OpenAI emits many
+//! tiny token-level deltas (a few chars every ~10-15ms), which already looks
+//! smooth. Anthropic coalesces `content_block_delta` events into larger chunks
+//! that arrive in bursts with gaps (e.g. 20-40 chars every ~80-100ms). If we
+//! reveal each burst the instant it arrives, the UI stair-steps: a clump of
+//! text pops in, then nothing for several frames, then another clump.
+//!
+//! To make every provider look the same, this buffer decouples *arrival* from
+//! *reveal*. Incoming text accumulates in a backlog, and a time-paced
+//! proportional controller drips it out: the reveal rate rises with the backlog
+//! so we never fall far behind a fast model, yet a lone burst is spread over
+//! several frames instead of dumped in one. The elapsed-time step is clamped so
+//! an idle gap (connect latency, tool pauses) cannot bank budget that would
+//! instantly dump the next burst.
 
 use serde::Serialize;
 use std::time::{Duration, Instant};
 
-/// Buffer that accumulates streaming text and flushes at semantic boundaries
+/// Steady-state reveal rate (chars/sec) when the backlog is empty. This sets the
+/// floor cadence and how the trailing characters of a burst drain out.
+const BASE_REVEAL_CPS: f32 = 180.0;
+
+/// Additional reveal rate per buffered character. The controller speeds up as the
+/// backlog grows so we track fast models with bounded latency: at steady incoming
+/// rate `R`, the backlog settles near `(R - BASE_REVEAL_CPS) / REVEAL_BACKLOG_GAIN`.
+const REVEAL_BACKLOG_GAIN: f32 = 3.0;
+
+/// Maximum elapsed time credited to a single reveal step. Without this, a long
+/// idle gap before the first/next burst would bank a huge budget and dump the
+/// whole burst at once, reintroducing the choppiness we are trying to remove.
+const MAX_REVEAL_STEP: Duration = Duration::from_millis(50);
+
+/// Buffer that accumulates streaming text and reveals it at a smooth, paced rate.
 pub struct StreamBuffer {
     buffer: String,
-    last_flush: Instant,
-    timeout: Duration,
-    smooth_frame_chars: usize,
+    last_reveal: Instant,
+    /// Fractional reveal budget carried between steps so slow rates still make
+    /// progress instead of rounding down to zero forever.
+    carry: f32,
+    base_cps: f32,
+    backlog_gain: f32,
+    max_step: Duration,
 }
 
 #[derive(Debug, Clone, Serialize)]
 pub struct StreamBufferMemoryProfile {
     pub buffered_text_bytes: usize,
-    pub timeout_ms: u64,
+    pub base_reveal_cps: u32,
 }
 
 impl Default for StreamBuffer {
@@ -27,50 +60,37 @@ impl StreamBuffer {
     pub fn new() -> Self {
         Self {
             buffer: String::new(),
-            last_flush: Instant::now(),
-            timeout: Duration::from_millis(150),
-            smooth_frame_chars: 96,
+            last_reveal: Instant::now(),
+            carry: 0.0,
+            base_cps: BASE_REVEAL_CPS,
+            backlog_gain: REVEAL_BACKLOG_GAIN,
+            max_step: MAX_REVEAL_STEP,
         }
     }
 
-    /// Push text into buffer, returns chunk to display if boundary found
+    /// Push text into the buffer, returning any paced chunk ready to display now.
     pub fn push(&mut self, text: &str) -> Option<String> {
         self.buffer.push_str(text);
-
-        // Find semantic boundary
-        if let Some(boundary) = self.find_boundary() {
-            return Some(self.drain_prefix(boundary.min(self.smooth_frame_boundary())));
-        }
-
-        if self.last_flush.elapsed() >= self.timeout {
-            return self.flush_smooth_frame();
-        }
-
-        None
+        self.reveal_now(Instant::now())
     }
 
-    /// Force flush the entire buffer (call on timeout or message end)
+    /// Force flush the entire buffer (call on message end, commit, or interrupt).
     pub fn flush(&mut self) -> Option<String> {
+        self.carry = 0.0;
+        self.last_reveal = Instant::now();
         if self.buffer.is_empty() {
             None
         } else {
-            self.last_flush = Instant::now();
             Some(std::mem::take(&mut self.buffer))
         }
     }
 
-    /// Flush up to one smooth-render frame worth of text. This is used for
-    /// periodic streaming redraws so large provider/SSE bursts are revealed
-    /// over a few quick frames instead of popping into the TUI all at once.
-    /// Finalization paths should still call [`flush`] to avoid leaving text
-    /// buffered at message boundaries.
+    /// Reveal one paced frame worth of buffered text. Called from the periodic
+    /// redraw tick so the backlog drains smoothly even when no new delta arrived
+    /// this frame. Finalization paths should still call [`flush`] to avoid
+    /// leaving text buffered at message boundaries.
     pub fn flush_smooth_frame(&mut self) -> Option<String> {
-        if self.buffer.is_empty() {
-            None
-        } else {
-            let boundary = self.smooth_frame_boundary().min(self.buffer.len());
-            Some(self.drain_prefix(boundary))
-        }
+        self.reveal_now(Instant::now())
     }
 
     /// Check if buffer is empty
@@ -81,138 +101,205 @@ impl StreamBuffer {
     /// Clear the buffer without returning content
     pub fn clear(&mut self) {
         self.buffer.clear();
-        self.last_flush = Instant::now();
+        self.carry = 0.0;
+        self.last_reveal = Instant::now();
     }
 
     pub fn debug_memory_profile(&self) -> StreamBufferMemoryProfile {
         StreamBufferMemoryProfile {
             buffered_text_bytes: self.buffer.len(),
-            timeout_ms: self.timeout.as_millis() as u64,
+            base_reveal_cps: self.base_cps as u32,
         }
     }
 
-    fn smooth_frame_boundary(&self) -> usize {
-        if self.buffer.chars().count() <= self.smooth_frame_chars {
-            return self.buffer.len();
+    /// Proportional, time-paced reveal. Advances the budget by the (clamped)
+    /// elapsed time times a backlog-scaled rate, then drains that many chars.
+    fn reveal_now(&mut self, now: Instant) -> Option<String> {
+        let backlog = self.buffer.chars().count();
+        if backlog == 0 {
+            // No backlog: reset so an idle gap cannot bank reveal budget.
+            self.carry = 0.0;
+            self.last_reveal = now;
+            return None;
         }
-        self.buffer
-            .char_indices()
-            .map(|(idx, _)| idx)
-            .nth(self.smooth_frame_chars)
-            .unwrap_or(self.buffer.len())
-    }
-
-    fn drain_prefix(&mut self, boundary: usize) -> String {
-        let boundary = floor_char_boundary(&self.buffer, boundary);
-        let chunk = self.buffer[..boundary].to_string();
-        self.buffer = self.buffer[boundary..].to_string();
-        self.last_flush = Instant::now();
-        chunk
-    }
 
-    /// Find a boundary in the buffer (newline-based), returns position after boundary
-    fn find_boundary(&self) -> Option<usize> {
-        let buf = &self.buffer;
+        let dt = now
+            .saturating_duration_since(self.last_reveal)
+            .min(self.max_step)
+            .as_secs_f32();
+        self.last_reveal = now;
 
-        // Code block start/end (```language or ```)
-        if let Some(pos) = buf.find("```") {
-            // Find end of the ``` line
-            if let Some(newline) = buf[pos..].find('\n') {
-                return Some(pos + newline + 1);
-            }
-        }
+        let cps = self.base_cps + backlog as f32 * self.backlog_gain;
+        self.carry += dt * cps;
 
-        // Any newline - simple and predictable
-        if let Some(pos) = buf.find('\n') {
-            return Some(pos + 1);
+        let mut reveal = self.carry.floor() as usize;
+        if reveal == 0 {
+            // Budget hasn't reached a whole char yet; keep accumulating.
+            return None;
         }
-
-        None
+        reveal = reveal.min(backlog);
+        self.carry -= reveal as f32;
+        Some(self.drain_chars(reveal))
     }
-}
 
-fn floor_char_boundary(s: &str, mut index: usize) -> usize {
-    index = index.min(s.len());
-    while index > 0 && !s.is_char_boundary(index) {
-        index -= 1;
+    /// Drain `char_count` characters from the front of the buffer on a UTF-8
+    /// boundary.
+    fn drain_chars(&mut self, char_count: usize) -> String {
+        if char_count == 0 {
+            return String::new();
+        }
+        let end = self
+            .buffer
+            .char_indices()
+            .nth(char_count)
+            .map(|(idx, _)| idx)
+            .unwrap_or(self.buffer.len());
+        let chunk = self.buffer[..end].to_string();
+        self.buffer.replace_range(..end, "");
+        chunk
     }
-    index
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
 
+    /// Drain the buffer to empty using fixed-cadence redraw frames, returning the
+    /// per-frame reveal sizes (in chars).
+    fn drain_frames(buf: &mut StreamBuffer, start: Instant, frame: Duration) -> Vec<usize> {
+        let mut sizes = Vec::new();
+        let mut t = start;
+        let mut guard = 0;
+        while !buf.is_empty() {
+            t += frame;
+            if let Some(chunk) = buf.reveal_now(t) {
+                sizes.push(chunk.chars().count());
+            }
+            guard += 1;
+            assert!(guard < 100_000, "drain did not converge");
+        }
+        sizes
+    }
+
     #[test]
-    fn test_newline_boundary() {
+    fn flush_drains_everything() {
         let mut buf = StreamBuffer::new();
-        let result = buf.push("First line\nSecond line");
-        assert_eq!(result, Some("First line\n".to_string()));
-        assert_eq!(buf.buffer, "Second line");
+        buf.buffer.push_str("remaining content");
+        let result = buf.flush();
+        assert_eq!(result, Some("remaining content".to_string()));
+        assert!(buf.is_empty());
     }
 
     #[test]
-    fn test_code_block_boundary() {
+    fn empty_push_reveals_nothing() {
         let mut buf = StreamBuffer::new();
-        // Code block marker ``` causes flush to include the whole line
-        let result = buf.push("```rust\nfn main() {}");
-        assert_eq!(result, Some("```rust\n".to_string()));
+        assert_eq!(buf.push(""), None);
+        assert!(buf.is_empty());
     }
 
     #[test]
-    fn test_no_boundary() {
+    fn paced_reveal_spreads_a_burst_over_multiple_frames() {
+        let start = Instant::now();
         let mut buf = StreamBuffer::new();
-        let result = buf.push("partial text without newline");
-        assert_eq!(result, None);
-        assert_eq!(buf.buffer, "partial text without newline");
+        buf.last_reveal = start;
+        buf.buffer.push_str(&"a".repeat(40));
+
+        let sizes = drain_frames(&mut buf, start, Duration::from_millis(16));
+        let total: usize = sizes.iter().sum();
+        assert_eq!(total, 40);
+        assert!(
+            sizes.len() >= 3,
+            "a 40-char burst should reveal across multiple frames, got {sizes:?}"
+        );
+        // No single 16ms frame should dump the whole burst.
+        assert!(
+            sizes.iter().all(|&n| n < 40),
+            "no frame should reveal the entire burst, got {sizes:?}"
+        );
     }
 
     #[test]
-    fn test_flush() {
+    fn idle_gap_does_not_dump_the_next_burst() {
+        let start = Instant::now();
         let mut buf = StreamBuffer::new();
-        buf.push("remaining content");
-        let result = buf.flush();
-        assert_eq!(result, Some("remaining content".to_string()));
-        assert!(buf.is_empty());
+        buf.last_reveal = start;
+        // Simulate a long connect/tool pause, then a burst arrives.
+        let arrival = start + Duration::from_secs(5);
+        buf.buffer.push_str(&"b".repeat(30));
+        let first = buf
+            .reveal_now(arrival)
+            .map(|c| c.chars().count())
+            .unwrap_or(0);
+        assert!(
+            first < 30,
+            "the idle gap must not bank budget that dumps the burst, revealed {first}"
+        );
+        // The remainder still drains over subsequent frames.
+        let sizes = drain_frames(&mut buf, arrival, Duration::from_millis(16));
+        assert_eq!(first + sizes.iter().sum::<usize>(), 30);
     }
 
     #[test]
-    fn test_multiple_newlines() {
-        let mut buf = StreamBuffer::new();
-        // First push returns first line
-        let result = buf.push("Line one\nLine two\nLine three");
-        assert_eq!(result, Some("Line one\n".to_string()));
-        // Second push returns second line
-        let result = buf.push("");
-        assert_eq!(result, Some("Line two\n".to_string()));
+    fn bursty_and_steady_feeds_reveal_at_similar_smoothness() {
+        // Steady (OpenAI-like): 4 chars every frame.
+        let start = Instant::now();
+        let frame = Duration::from_millis(16);
+        let mut steady = StreamBuffer::new();
+        steady.last_reveal = start;
+        let mut steady_sizes = Vec::new();
+        let mut t = start;
+        for _ in 0..40 {
+            t += frame;
+            steady.buffer.push_str("abcd");
+            if let Some(c) = steady.reveal_now(t) {
+                steady_sizes.push(c.chars().count());
+            }
+        }
+        steady_sizes.extend(drain_frames(&mut steady, t, frame));
+
+        // Bursty (Anthropic-like): 24 chars every 6th frame.
+        let mut bursty = StreamBuffer::new();
+        bursty.last_reveal = start;
+        let mut bursty_sizes = Vec::new();
+        let mut t = start;
+        for i in 0..60 {
+            t += frame;
+            if i % 6 == 0 {
+                bursty.buffer.push_str(&"x".repeat(24));
+            }
+            if let Some(c) = bursty.reveal_now(t) {
+                bursty_sizes.push(c.chars().count());
+            }
+        }
+        bursty_sizes.extend(drain_frames(&mut bursty, t, frame));
+
+        let max_burst = *bursty_sizes.iter().max().unwrap();
+        // The whole 24-char clump must never appear in a single frame; pacing
+        // should break it into smaller per-frame reveals like the steady feed.
+        assert!(
+            max_burst < 24,
+            "bursty feed should be smoothed, max frame reveal was {max_burst} ({bursty_sizes:?})"
+        );
     }
 
     #[test]
-    fn test_smooth_frame_flush_caps_large_chunks() {
+    fn reveal_respects_utf8_boundaries() {
+        let start = Instant::now();
         let mut buf = StreamBuffer::new();
-        let text = "a".repeat(150);
-        assert_eq!(buf.push(&text), None);
-
-        let first = buf.flush_smooth_frame().unwrap();
-        assert_eq!(first.len(), 96);
-        assert_eq!(buf.buffer.len(), 54);
+        buf.last_reveal = start;
+        buf.buffer.push_str(&"é".repeat(40));
 
-        let rest = buf.flush().unwrap();
-        assert_eq!(rest.len(), 54);
-        assert!(buf.is_empty());
+        let sizes = drain_frames(&mut buf, start, Duration::from_millis(16));
+        assert_eq!(sizes.iter().sum::<usize>(), 40);
     }
 
     #[test]
-    fn test_smooth_frame_flush_respects_utf8_boundaries() {
+    fn small_trailing_text_eventually_drains() {
+        let start = Instant::now();
         let mut buf = StreamBuffer::new();
-        let text = "é".repeat(120);
-        assert_eq!(buf.push(&text), None);
-
-        let first = buf.flush_smooth_frame().unwrap();
-        assert_eq!(first.chars().count(), 96);
-        assert!(first.is_char_boundary(first.len()));
-
-        let rest = buf.flush().unwrap();
-        assert_eq!(rest.chars().count(), 24);
+        buf.last_reveal = start;
+        buf.buffer.push_str("hi");
+        let sizes = drain_frames(&mut buf, start, Duration::from_millis(16));
+        assert_eq!(sizes.iter().sum::<usize>(), 2);
     }
 }
diff --git a/crates/jcode-tui-markdown/src/lib.rs b/crates/jcode-tui-markdown/src/lib.rs
index 0d9a71335..0a0ecfb29 100644
--- a/crates/jcode-tui-markdown/src/lib.rs
+++ b/crates/jcode-tui-markdown/src/lib.rs
@@ -185,6 +185,19 @@ pub fn reasoning_partial_markup(line: &str) -> String {
     }
 }
 
+/// One-line collapsed reasoning summary markup (e.g. `▸ thought (3 lines)`),
+/// styled dim+italic like the live reasoning lines. Used to fold a persisted
+/// reasoning block down to a single trace line when the transcript is
+/// re-rendered from history in `current` reasoning-display mode (so reloaded /
+/// resumed sessions match the live collapse instead of replaying every line).
+pub fn reasoning_summary_line_markup(line_count: usize) -> String {
+    let label = match line_count {
+        0 | 1 => "▸ thought".to_string(),
+        n => format!("▸ thought ({} lines)", n),
+    };
+    reasoning_line_markup(&label)
+}
+
 use render_support::{
     highlight_code_cached, line_plain_text, placeholder_code_block, ranges_overlap, render_table,
 };
diff --git a/crates/jcode-tui-markdown/src/markdown_tests/cases/rendering.rs b/crates/jcode-tui-markdown/src/markdown_tests/cases/rendering.rs
index 4c4082843..f9db404ca 100644
--- a/crates/jcode-tui-markdown/src/markdown_tests/cases/rendering.rs
+++ b/crates/jcode-tui-markdown/src/markdown_tests/cases/rendering.rs
@@ -763,3 +763,51 @@ fn test_reasoning_emphasis_does_not_leak_into_following_text() {
         );
     }
 }
+
+#[test]
+fn test_reasoning_summary_line_markup_folds_to_single_dim_italic_trace() {
+    let sentinel = crate::REASONING_SENTINEL;
+
+    // Pluralized count for multi-line blocks.
+    let many = crate::reasoning_summary_line_markup(3);
+    assert!(
+        many.contains(&format!("*{0}▸ thought (3 lines){0}*", sentinel)),
+        "expected pluralized summary markup, got: {many:?}"
+    );
+
+    // Single/zero-line blocks omit the count.
+    let one = crate::reasoning_summary_line_markup(1);
+    assert!(
+        one.contains(&format!("*{0}▸ thought{0}*", sentinel)) && !one.contains("lines"),
+        "expected bare summary markup, got: {one:?}"
+    );
+    let none = crate::reasoning_summary_line_markup(0);
+    assert!(none.contains(&format!("*{0}▸ thought{0}*", sentinel)), "{none:?}");
+
+    // The summary line renders dim + italic with no sentinel leaking into text.
+    let lines = render_markdown(&many);
+    let dim = md_dim_color();
+    let mut saw_marker = false;
+    for rendered in &lines {
+        for span in &rendered.spans {
+            assert!(
+                !span.content.contains(sentinel),
+                "sentinel leaked into visible summary: {:?}",
+                span.content
+            );
+            if span.content.trim().is_empty() {
+                continue;
+            }
+            if span.content.contains('▸') {
+                saw_marker = true;
+            }
+            assert_eq!(span.style.fg, Some(dim), "summary span not dim: {:?}", span.content);
+            assert!(
+                span.style.add_modifier.contains(Modifier::ITALIC),
+                "summary span not italic: {:?}",
+                span.content
+            );
+        }
+    }
+    assert!(saw_marker, "summary marker '▸' must be visible: {lines:?}");
+}
diff --git a/crates/jcode-tui-messages/src/message.rs b/crates/jcode-tui-messages/src/message.rs
index 90710a1f1..d30da77af 100644
--- a/crates/jcode-tui-messages/src/message.rs
+++ b/crates/jcode-tui-messages/src/message.rs
@@ -175,6 +175,20 @@ impl DisplayMessage {
         }
     }
 
+    /// Create a display-only collapsing reasoning trace ("current" mode). The
+    /// content is sentinel-wrapped dim/italic markup; this message height-collapses
+    /// toward a one-line summary and is excluded from provider/model context.
+    pub fn reasoning(content: impl Into<String>) -> Self {
+        Self {
+            role: "reasoning".to_string(),
+            content: content.into(),
+            tool_calls: Vec::new(),
+            duration_secs: None,
+            title: None,
+            tool_data: None,
+        }
+    }
+
     /// Convert the shared session renderer output into the TUI transcript model.
     pub fn from_rendered_message(item: RenderedMessage) -> Self {
         Self {
diff --git a/crates/jcode-tui-session-picker/src/lib.rs b/crates/jcode-tui-session-picker/src/lib.rs
index bb424c258..9edabd957 100644
--- a/crates/jcode-tui-session-picker/src/lib.rs
+++ b/crates/jcode-tui-session-picker/src/lib.rs
@@ -139,6 +139,10 @@ pub enum SessionFilterMode {
     Codex,
     Pi,
     OpenCode,
+    /// External CLI transcripts (Codex and/or Claude Code) shown together.
+    /// Used by the first-run onboarding "continue where you left off" picker so
+    /// it surfaces every external CLI the user is logged into, not just one.
+    ExternalClis,
 }
 
 impl SessionFilterMode {
@@ -151,6 +155,9 @@ impl SessionFilterMode {
             Self::Codex => Self::Pi,
             Self::Pi => Self::OpenCode,
             Self::OpenCode => Self::All,
+            // ExternalClis is an onboarding-only composite filter, not part of
+            // the user-facing cycle; treat it as a no-op anchor.
+            Self::ExternalClis => Self::All,
         }
     }
 
@@ -163,6 +170,7 @@ impl SessionFilterMode {
             Self::Codex => Self::ClaudeCode,
             Self::Pi => Self::Codex,
             Self::OpenCode => Self::Pi,
+            Self::ExternalClis => Self::All,
         }
     }
 
@@ -175,6 +183,7 @@ impl SessionFilterMode {
             Self::Codex => Some("🧠 Codex"),
             Self::Pi => Some("π Pi"),
             Self::OpenCode => Some("◌ OpenCode"),
+            Self::ExternalClis => Some("🧠 Codex + 🧵 Claude Code"),
         }
     }
 }
diff --git a/crates/jcode-tui/src/tui/app.rs b/crates/jcode-tui/src/tui/app.rs
index 5c787c0c7..192cc2493 100644
--- a/crates/jcode-tui/src/tui/app.rs
+++ b/crates/jcode-tui/src/tui/app.rs
@@ -341,6 +341,28 @@ pub enum ProcessingStatus {
     RunningTool(String),
 }
 
+/// Live "collapse the current reasoning" animation state.
+///
+/// In `current` reasoning-display mode the model's reasoning streams live as
+/// dim+italic lines, then must disappear once the answer commits or a tool runs.
+/// Instead of deleting every reasoning line in a single frame (a jarring upward
+/// jump), the closed reasoning block is moved into a dedicated `"reasoning"`
+/// display message that height-collapses toward a one-line summary over a short
+/// ease-out, leaving a `▸ thought for Xs` trace behind.
+#[derive(Clone, Debug)]
+pub(crate) struct ReasoningCollapse {
+    /// Index into `display_messages` of the `"reasoning"` message being collapsed.
+    pub(crate) msg_index: usize,
+    /// One-line dim summary the block collapses down to (markup for
+    /// "▸ thought for Xs"), always shown at the top of the message.
+    pub(crate) summary_markup: String,
+    /// Per-line dim+italic markup for each reasoning line, in order. The block
+    /// shrinks by dropping leading lines until only `summary_markup` remains.
+    pub(crate) line_markups: Vec<String>,
+    /// When the collapse animation started.
+    pub(crate) started_at: Instant,
+}
+
 #[derive(Clone, Debug, PartialEq, Eq)]
 pub(crate) enum RemoteStartupPhase {
     StartingServer,
@@ -726,6 +748,18 @@ pub struct App {
     // `streaming_text` (the rendered tail of `reasoning_pending_line`). Truncated
     // and re-appended on each delta so the in-progress line updates in place.
     reasoning_partial_len: usize,
+    // Byte offset in `streaming_text` where the current reasoning block began
+    // (recorded by `open_reasoning_region`). Used in `current` mode to slice the
+    // closed reasoning block out of the stream and hand it to the collapse
+    // animation while keeping any answer text that preceded it in order.
+    reasoning_block_start: Option<usize>,
+    // Wall-clock instant the current reasoning region opened, used to label the
+    // collapsed summary ("▸ thought for Xs").
+    reasoning_block_started_at: Option<Instant>,
+    // Active "collapse the current reasoning" animation (current mode only). While
+    // set, a `"reasoning"` display message height-collapses toward its one-line
+    // summary; the redraw loop advances it each frame and finalizes on completion.
+    reasoning_collapse: Option<ReasoningCollapse>,
     // Hot-reload: if set, exec into new binary with this session ID (no rebuild)
     reload_requested: Option<String>,
     // Hot-rebuild: if set, do full git pull + cargo build + tests then exec
diff --git a/crates/jcode-tui/src/tui/app/inline_interactive.rs b/crates/jcode-tui/src/tui/app/inline_interactive.rs
index 1d53715ec..440fd56bd 100644
--- a/crates/jcode-tui/src/tui/app/inline_interactive.rs
+++ b/crates/jcode-tui/src/tui/app/inline_interactive.rs
@@ -2732,11 +2732,22 @@ impl App {
     }
 
     pub(super) fn picker_fuzzy_score(pattern: &str, text: &str) -> Option<i32> {
-        let pat: Vec<char> = pattern
+        let pat = Self::picker_fuzzy_pattern(pattern);
+        Self::picker_fuzzy_score_with_pattern(&pat, text)
+    }
+
+    /// Normalize a fuzzy-match pattern (lowercase, drop whitespace) into chars.
+    /// Hoist this out of per-entry scoring so a filter pass over N entries
+    /// normalizes the pattern once instead of N times per keystroke.
+    pub(super) fn picker_fuzzy_pattern(pattern: &str) -> Vec<char> {
+        pattern
             .to_lowercase()
             .chars()
             .filter(|c| !c.is_whitespace())
-            .collect();
+            .collect()
+    }
+
+    pub(super) fn picker_fuzzy_score_with_pattern(pat: &[char], text: &str) -> Option<i32> {
         let txt: Vec<char> = text.to_lowercase().chars().collect();
         if pat.is_empty() {
             return Some(0);
@@ -2782,13 +2793,16 @@ impl App {
         if picker.filter.is_empty() {
             picker.filtered = (0..picker.entries.len()).collect();
         } else {
+            // Normalize the filter pattern once per keystroke instead of once per
+            // entry inside picker_fuzzy_score.
+            let pat = Self::picker_fuzzy_pattern(&picker.filter);
             let mut scored: Vec<(usize, i32)> = picker
                 .entries
                 .iter()
                 .enumerate()
                 .filter_map(|(i, m)| {
                     let filter_text = picker.filter_text(m);
-                    Self::picker_fuzzy_score(&picker.filter, &filter_text).map(|s| {
+                    Self::picker_fuzzy_score_with_pattern(&pat, &filter_text).map(|s| {
                         let usage_bonus = m.usage_score.min(i32::MAX as u32) as i32;
                         let bonus = usage_bonus + if m.recommended { 5 } else { 0 };
                         (i, s + bonus)
diff --git a/crates/jcode-tui/src/tui/app/input.rs b/crates/jcode-tui/src/tui/app/input.rs
index 3ebec0904..7a0705bea 100644
--- a/crates/jcode-tui/src/tui/app/input.rs
+++ b/crates/jcode-tui/src/tui/app/input.rs
@@ -50,6 +50,48 @@ pub(super) fn strip_reasoning_lines(content: &str) -> String {
     result.trim_end().to_string()
 }
 
+/// Total duration of the "current reasoning collapses away" height animation.
+pub(super) const REASONING_COLLAPSE_DURATION: Duration = Duration::from_millis(280);
+
+/// Split a just-closed reasoning block (sentinel-wrapped dim/italic line markup,
+/// as produced by [`jcode_tui_markdown::reasoning_line_markup`]) into one markup
+/// string per visible reasoning line. Blank separator lines are dropped so the
+/// collapse animates over real thought lines only.
+pub(super) fn reasoning_block_line_markups(block: &str) -> Vec<String> {
+    block
+        .split_inclusive('\n')
+        .filter(|segment| segment.contains(jcode_tui_markdown::REASONING_SENTINEL))
+        .map(|segment| segment.to_string())
+        .collect()
+}
+
+/// One-line dim summary the collapsed reasoning folds into. Includes a `▸` marker
+/// and the thinking duration when known (e.g. `▸ thought for 12s`).
+pub(super) fn reasoning_summary_markup(line_count: usize, elapsed: Option<Duration>) -> String {
+    let label = match elapsed {
+        Some(d) if d.as_secs() >= 1 => format!("▸ thought for {}s", d.as_secs()),
+        Some(_) => "▸ thought".to_string(),
+        None if line_count == 1 => "▸ thought (1 line)".to_string(),
+        None => format!("▸ thought ({} lines)", line_count),
+    };
+    jcode_tui_markdown::reasoning_line_markup(&label)
+}
+
+/// Build the transcript content for a collapsing `"reasoning"` message: the last
+/// `remaining` reasoning lines, or just the summary line once fully collapsed.
+pub(super) fn reasoning_message_content(
+    summary_markup: &str,
+    line_markups: &[String],
+    remaining: usize,
+) -> String {
+    if remaining == 0 || line_markups.is_empty() {
+        return summary_markup.to_string();
+    }
+    let remaining = remaining.min(line_markups.len());
+    let start = line_markups.len() - remaining;
+    line_markups[start..].concat()
+}
+
 pub(super) fn edit_input_in_external_editor(app: &mut App) {
     match edit_text_in_external_editor(&app.input) {
         Ok(edited) => {
@@ -3246,6 +3288,11 @@ impl App {
         self.reasoning_streaming = true;
         self.reasoning_pending_line.clear();
         self.reasoning_partial_len = 0;
+        // Remember where this reasoning block starts in the stream so `current`
+        // mode can later slice it out (without disturbing any preceding answer
+        // text) and hand it to the collapse animation.
+        self.reasoning_block_start = Some(self.streaming_text.len());
+        self.reasoning_block_started_at = Some(Instant::now());
     }
 
     /// Remove the live partial-reasoning tail (the rendered, not-yet-committed
@@ -3311,6 +3358,17 @@ impl App {
                 .push_str(&jcode_tui_markdown::reasoning_line_markup(&pending));
         }
         self.reasoning_streaming = false;
+
+        // In `current` mode, animate the block away instead of leaving it in the
+        // stream to be stripped wholesale at commit time.
+        if matches!(
+            crate::config::config().display.reasoning_display(),
+            crate::config::ReasoningDisplayMode::Current
+        ) {
+            self.begin_reasoning_collapse();
+            return;
+        }
+
         // Terminate the reasoning block with a blank line so following output
         // renders as a normal paragraph.
         if !self.streaming_text.ends_with("\n\n") {
@@ -3323,6 +3381,147 @@ impl App {
         self.refresh_split_view_if_needed();
     }
 
+    /// Slice the just-closed reasoning block out of `streaming_text` and move it
+    /// into a dedicated `"reasoning"` display message, then start (or replace) the
+    /// height-collapse animation. Any answer text streamed *before* the reasoning
+    /// block is left untouched so ordering is preserved. With decorative
+    /// animations disabled (reduced motion / low-power tiers) the block is
+    /// finalized straight to its summary line.
+    pub(super) fn begin_reasoning_collapse(&mut self) {
+        let block_start = self.reasoning_block_start.take().unwrap_or(0);
+        let started_at = self.reasoning_block_started_at.take();
+        // Finalize any previous collapse first so its message snaps to its summary
+        // instead of being orphaned mid-animation.
+        self.finalize_reasoning_collapse();
+
+        let block_start = block_start.min(self.streaming_text.len());
+
+        // Everything from the block start onward is reasoning markup (plus the
+        // separators inserted by open/close). Take it out of the live stream.
+        let block: String = self.streaming_text.split_off(block_start);
+        // Drop a trailing separator the answer-side path would otherwise add.
+        while self.streaming_text.ends_with('\n') {
+            self.streaming_text.pop();
+        }
+        self.refresh_split_view_if_needed();
+
+        let line_markups = reasoning_block_line_markups(&block);
+        if line_markups.is_empty() {
+            // Nothing to show (e.g. empty reasoning); just clear state.
+            self.reasoning_collapse = None;
+            return;
+        }
+
+        let elapsed = started_at.map(|t| t.elapsed());
+        let summary_markup = reasoning_summary_markup(line_markups.len(), elapsed);
+
+        // Build the committed message content: every reasoning line, then the
+        // summary as the final line. The renderer reveals a shrinking suffix.
+        let content =
+            reasoning_message_content(&summary_markup, &line_markups, line_markups.len());
+
+        let msg_index = self.display_messages.len();
+        self.push_display_message(DisplayMessage::reasoning(content));
+
+        let decorative = crate::perf::tui_policy().enable_decorative_animations;
+        if !decorative {
+            // Reduced motion: snap straight to the one-line summary.
+            self.replace_display_message_content(
+                msg_index,
+                reasoning_message_content(&summary_markup, &line_markups, 0),
+            );
+            self.reasoning_collapse = None;
+            return;
+        }
+
+        self.reasoning_collapse = Some(super::ReasoningCollapse {
+            msg_index,
+            summary_markup,
+            line_markups,
+            started_at: Instant::now(),
+        });
+    }
+
+    /// Advance the active reasoning-collapse animation. Returns `true` when the
+    /// transcript changed (so the caller should request a redraw). Finalizes to
+    /// the summary line once the animation completes.
+    pub(super) fn advance_reasoning_collapse(&mut self) -> bool {
+        let Some(collapse) = self.reasoning_collapse.as_ref() else {
+            return false;
+        };
+
+        // If the target message moved or was replaced (compaction/rewind), drop the
+        // animation rather than risk mutating an unrelated message.
+        if self
+            .display_messages
+            .get(collapse.msg_index)
+            .map(|m| m.role.as_str())
+            != Some("reasoning")
+        {
+            self.reasoning_collapse = None;
+            return false;
+        }
+
+        let total = collapse.line_markups.len();
+        let elapsed = collapse.started_at.elapsed();
+        let progress =
+            (elapsed.as_secs_f32() / REASONING_COLLAPSE_DURATION.as_secs_f32()).clamp(0.0, 1.0);
+        // Ease-out cubic so the block decelerates as it folds away.
+        let eased = 1.0 - (1.0 - progress).powi(3);
+        // Number of reasoning lines still visible above the summary. Counts down
+        // from `total` to 0 (only the summary remains).
+        let remaining = ((total as f32) * (1.0 - eased)).round() as usize;
+        let remaining = remaining.min(total);
+
+        let msg_index = collapse.msg_index;
+        let content =
+            reasoning_message_content(&collapse.summary_markup, &collapse.line_markups, remaining);
+        let changed = self.replace_display_message_content(msg_index, content);
+
+        if progress >= 1.0 {
+            self.reasoning_collapse = None;
+        }
+        changed
+    }
+
+    /// Whether a reasoning-collapse animation is currently running.
+    pub(super) fn reasoning_collapse_active(&self) -> bool {
+        self.reasoning_collapse.is_some()
+    }
+
+    /// Test hook: backdate the active collapse's start so `advance_*` observes a
+    /// specific elapsed fraction, and return the number of source reasoning lines.
+    #[cfg(test)]
+    pub(super) fn backdate_reasoning_collapse_for_test(
+        &mut self,
+        elapsed: std::time::Duration,
+    ) -> Option<usize> {
+        let collapse = self.reasoning_collapse.as_mut()?;
+        collapse.started_at = Instant::now()
+            .checked_sub(elapsed)
+            .unwrap_or_else(Instant::now);
+        Some(collapse.line_markups.len())
+    }
+
+    /// Finalize any in-flight reasoning collapse immediately (snap to summary).
+    /// Used when the turn ends or state is reset so no animation is left dangling.
+    pub(super) fn finalize_reasoning_collapse(&mut self) {
+        if let Some(collapse) = self.reasoning_collapse.take() {
+            if self
+                .display_messages
+                .get(collapse.msg_index)
+                .map(|m| m.role.as_str())
+                == Some("reasoning")
+            {
+                let content =
+                    reasoning_message_content(&collapse.summary_markup, &collapse.line_markups, 0);
+                self.replace_display_message_content(collapse.msg_index, content);
+            }
+        }
+        self.reasoning_block_start = None;
+        self.reasoning_block_started_at = None;
+    }
+
     pub(super) fn append_streaming_text(&mut self, text: &str) {
         if text.is_empty() {
             return;
@@ -3356,6 +3555,10 @@ impl App {
         self.reasoning_streaming = false;
         self.reasoning_pending_line.clear();
         self.reasoning_partial_len = 0;
+        // The stream (and any block offset into it) is gone; a running collapse
+        // targets a separate display message and is left to finish on its own.
+        self.reasoning_block_start = None;
+        self.reasoning_block_started_at = None;
         self.refresh_split_view_if_needed();
         self.streaming_md_renderer.borrow_mut().reset();
         crate::tui::mermaid::clear_streaming_preview_diagram();
@@ -3367,6 +3570,8 @@ impl App {
         self.reasoning_streaming = false;
         self.reasoning_pending_line.clear();
         self.reasoning_partial_len = 0;
+        self.reasoning_block_start = None;
+        self.reasoning_block_started_at = None;
         self.refresh_split_view_if_needed();
         self.streaming_md_renderer.borrow_mut().reset();
         crate::tui::mermaid::clear_streaming_preview_diagram();
diff --git a/crates/jcode-tui/src/tui/app/local.rs b/crates/jcode-tui/src/tui/app/local.rs
index b98883f7a..204730a75 100644
--- a/crates/jcode-tui/src/tui/app/local.rs
+++ b/crates/jcode-tui/src/tui/app/local.rs
@@ -55,6 +55,7 @@ pub(super) async fn process_turn_with_input(
 
 pub(super) fn handle_tick(app: &mut App) -> bool {
     let mut needs_redraw = crate::tui::periodic_redraw_required(app);
+    needs_redraw |= app.advance_reasoning_collapse();
     app.maybe_capture_runtime_memory_heartbeat();
     needs_redraw |= app.progress_copy_selection_edge_autoscroll();
     app.progress_mouse_scroll_animation();
@@ -472,6 +473,9 @@ pub(super) fn finish_turn(app: &mut App) {
     app.thought_line_inserted = false;
     app.thinking_prefix_emitted = false;
     app.thinking_buffer.clear();
+    // Snap any in-flight reasoning collapse straight to its summary so no
+    // animation is left running once the turn is idle.
+    app.finalize_reasoning_collapse();
     app.note_runtime_memory_event_force("turn_completed", "local_turn_finished");
     if !app.schedule_auto_poke_followup_if_needed()
         && !app.schedule_overnight_poke_followup_if_needed()
diff --git a/crates/jcode-tui/src/tui/app/onboarding_flow_control.rs b/crates/jcode-tui/src/tui/app/onboarding_flow_control.rs
index 93965c092..ec20f5439 100644
--- a/crates/jcode-tui/src/tui/app/onboarding_flow_control.rs
+++ b/crates/jcode-tui/src/tui/app/onboarding_flow_control.rs
@@ -208,36 +208,17 @@ impl App {
     /// straight into the resume picker (with an onboarding banner + a
     /// "Start a new session" option) instead of asking a separate Yes/No
     /// "continue where you left off" question. When both CLIs are present we
-    /// surface whichever one has the most recent transcript.
+    /// show *both* their transcripts together in one combined, recency-sorted
+    /// list rather than hiding one behind the other.
     pub(super) fn onboarding_after_model_select(&mut self) {
         if !matches!(self.onboarding_phase(), Some(OnboardingPhase::ModelSelect)) {
             return;
         }
-        match self.onboarding_most_recent_external_cli() {
-            Some(cli) => self.onboarding_open_transcript_picker(cli),
-            None => self.onboarding_show_suggestions(),
-        }
-    }
-
-    /// Among the external CLIs whose OAuth credentials are present, pick the one
-    /// with the most recent transcript. Ties (or a CLI with no transcripts yet)
-    /// fall back to detection order (Codex first). Returns `None` when no
-    /// external CLI login is present.
-    fn onboarding_most_recent_external_cli(&self) -> Option<ExternalCli> {
         let present = crate::tui::app::onboarding_flow::detect_external_cli_oauths();
-        match present.as_slice() {
-            [] => None,
-            [only] => Some(*only),
-            _ => {
-                // Multiple logins: rank by newest transcript mtime.
-                present
-                    .iter()
-                    .max_by_key(|cli| {
-                        session_picker::latest_external_cli_session_secs(**cli).unwrap_or(0)
-                    })
-                    .copied()
-                    .or_else(|| present.first().copied())
-            }
+        if present.is_empty() {
+            self.onboarding_show_suggestions();
+        } else {
+            self.onboarding_open_transcript_picker(&present);
         }
     }
 
@@ -283,7 +264,7 @@ impl App {
             _ => return,
         };
         if wants_continue {
-            self.onboarding_open_transcript_picker(cli);
+            self.onboarding_open_transcript_picker(std::slice::from_ref(&cli));
         } else {
             self.onboarding_show_suggestions();
         }
@@ -296,7 +277,7 @@ impl App {
     ///   - `TelemetryConsent`: Left/h -> No, Right/l -> Yes, toggle with
     ///     Up/Down/k/j/Tab; y/n commit directly, Enter/Space commit the
     ///     highlighted default.
-    ///     Returns true if the key was consumed.
+    /// Returns true if the key was consumed.
     pub(super) fn handle_onboarding_continue_prompt_key(&mut self, code: KeyCode) -> bool {
         match self.onboarding_phase() {
             Some(OnboardingPhase::Login { import }) => {
@@ -602,51 +583,89 @@ impl App {
         });
     }
 
-    /// Open a single-select resume-style picker filtered to the external CLI's
-    /// transcripts. Falls back to the session-search prompt if none load.
-    pub(super) fn onboarding_open_transcript_picker(&mut self, cli: ExternalCli) {
-        let filter = match cli {
-            ExternalCli::Codex => SessionFilterMode::Codex,
-            ExternalCli::ClaudeCode => SessionFilterMode::ClaudeCode,
+    /// Open a single-select resume-style picker showing the transcripts of every
+    /// detected external CLI together (Codex and/or Claude Code), sorted by
+    /// recency. Falls back to the session-search prompt if none load.
+    ///
+    /// `clis` is the set of external CLIs the user is logged into. When more than
+    /// one is present we still show them in one combined list so the user never
+    /// has a CLI's history hidden behind the other.
+    pub(super) fn onboarding_open_transcript_picker(&mut self, clis: &[ExternalCli]) {
+        // Choose a representative CLI for the banner/mode headline: the one with
+        // the most recent transcript (falling back to detection order).
+        let headline_cli = clis
+            .iter()
+            .copied()
+            .max_by_key(|cli| session_picker::latest_external_cli_session_secs(*cli).unwrap_or(0))
+            .or_else(|| clis.first().copied())
+            .unwrap_or(ExternalCli::Codex);
+
+        let multi = clis.len() > 1;
+        let filter = if multi {
+            SessionFilterMode::ExternalClis
+        } else {
+            match headline_cli {
+                ExternalCli::Codex => SessionFilterMode::Codex,
+                ExternalCli::ClaudeCode => SessionFilterMode::ClaudeCode,
+            }
         };
 
-        // The onboarding picker only ever shows this one external CLI's
-        // transcripts, so load just those instead of paying the full
-        // `load_sessions_grouped` cost (parsing every jcode snapshot, the other
-        // CLIs, and listing servers). This keeps first-run onboarding snappy.
+        // The onboarding picker only shows external CLI transcripts, so load just
+        // those instead of paying the full `load_sessions_grouped` cost (parsing
+        // every jcode snapshot and listing servers). This keeps first-run
+        // onboarding snappy while still surfacing every logged-in CLI.
         let (server_groups, orphan_sessions) =
-            session_picker::load_external_cli_sessions_grouped(cli);
+            session_picker::load_external_cli_sessions_grouped_multi(clis);
 
         let mut picker = SessionPicker::new_grouped(server_groups, orphan_sessions);
         picker.activate_external_cli_filter(filter);
 
         if picker.visible_session_count() == 0 {
-            self.onboarding_fallback_to_session_search(cli);
+            self.onboarding_fallback_to_session_search(headline_cli);
             return;
         }
 
-        picker.activate_onboarding_banner(Self::onboarding_resume_banner_lines(cli));
+        picker.activate_onboarding_banner(Self::onboarding_resume_banner_lines(clis));
 
         self.session_picker_overlay = Some(RefCell::new(picker));
-        self.session_picker_mode = SessionPickerMode::Onboarding { cli };
+        self.session_picker_mode = SessionPickerMode::Onboarding { cli: headline_cli };
         if let Some(flow) = self.onboarding_flow.as_mut() {
             flow.phase = OnboardingPhase::TranscriptPick {
-                cli,
+                cli: headline_cli,
                 shown_at: Instant::now(),
             };
         }
+        let resume_label = if multi {
+            "Resume a Codex or Claude Code session".to_string()
+        } else {
+            format!("Resume a {} session", headline_cli.label())
+        };
         self.set_status_notice(format!(
-            "Resume a {} session (↑↓ to choose, Enter to resume) or pick \"Start a new session\"",
-            cli.label()
+            "{resume_label} (↑↓ to choose, Enter to resume) or pick \"Start a new session\""
         ));
     }
 
     /// Formatted onboarding prompt shown in the reserved top band of the
     /// resume picker on first run.
-    fn onboarding_resume_banner_lines(cli: ExternalCli) -> Vec<ratatui::text::Line<'static>> {
+    fn onboarding_resume_banner_lines(clis: &[ExternalCli]) -> Vec<ratatui::text::Line<'static>> {
         use ratatui::style::{Color, Modifier, Style};
         use ratatui::text::{Line, Span};
         let accent = crate::tui::color_support::rgb(186, 139, 255);
+        // Describe whichever CLIs were detected: "Codex", "Claude Code", or
+        // "Codex and Claude Code" when both are present.
+        let mut labels: Vec<&'static str> = Vec::new();
+        for cli in clis {
+            let label = cli.label();
+            if !labels.contains(&label) {
+                labels.push(label);
+            }
+        }
+        let found = match labels.as_slice() {
+            [] => "external".to_string(),
+            [only] => (*only).to_string(),
+            [first, second] => format!("{first} and {second}"),
+            _ => labels.join(", "),
+        };
         vec![
             Line::from(vec![Span::styled(
                 "Welcome to jcode 🎉",
@@ -654,8 +673,7 @@ impl App {
             )]),
             Line::from(vec![Span::styled(
                 format!(
-                    "We found your {} sessions. Pick one below to pick up right where you left off,",
-                    cli.label()
+                    "We found your {found} sessions. Pick one below to pick up right where you left off,"
                 ),
                 Style::default().fg(Color::White),
             )]),
@@ -729,10 +747,10 @@ impl App {
     /// so prefer the same resolution the header uses; fall back to the session
     /// model and finally the local provider's model.
     fn onboarding_default_model_id(&self) -> String {
-        if self.is_remote
-            && let Some(model) = self.effective_remote_provider_model()
-        {
-            return model;
+        if self.is_remote {
+            if let Some(model) = self.effective_remote_provider_model() {
+                return model;
+            }
         }
         self.session
             .model
diff --git a/crates/jcode-tui/src/tui/app/remote.rs b/crates/jcode-tui/src/tui/app/remote.rs
index ed04ef207..7cbbea054 100644
--- a/crates/jcode-tui/src/tui/app/remote.rs
+++ b/crates/jcode-tui/src/tui/app/remote.rs
@@ -80,13 +80,16 @@ pub(super) async fn handle_tick(app: &mut App, remote: &mut RemoteConnection) ->
             .is_some_and(|state| state.kind == crate::tui::PickerKind::Model),
     });
     let mut needs_redraw = crate::tui::periodic_redraw_required(app);
+    needs_redraw |= app.advance_reasoning_collapse();
     app.maybe_capture_runtime_memory_heartbeat();
     needs_redraw |= app.progress_copy_selection_edge_autoscroll();
     app.progress_mouse_scroll_animation();
     needs_redraw |= app.update_chat_overscroll();
     needs_redraw |= app.update_pinned_images_auto_hide();
     needs_redraw |= dispatch_compacted_history_load(app, remote).await;
-    if let Some(chunk) = app.stream_buffer.flush() {
+    // Reveal buffered streaming text at the smooth paced rate on each tick, the
+    // same as the local turn loop. Finalization paths still call flush().
+    if let Some(chunk) = app.stream_buffer.flush_smooth_frame() {
         app.append_streaming_text(&chunk);
         needs_redraw = true;
     }
diff --git a/crates/jcode-tui/src/tui/app/remote/server_events.rs b/crates/jcode-tui/src/tui/app/remote/server_events.rs
index 477e98cce..6ad13a8e7 100644
--- a/crates/jcode-tui/src/tui/app/remote/server_events.rs
+++ b/crates/jcode-tui/src/tui/app/remote/server_events.rs
@@ -61,9 +61,20 @@ fn server_release_is_older_than_client(server_version: Option<&str>, client_vers
 /// attached to is not running the binary we expect.
 ///
 /// Precedence:
+/// - The client independently measured the server's release version as strictly
+///   older than its own clean release version -> defer. This wins even over the
+///   server's own `server_has_update: Some(false)` self-report, because a stale
+///   long-lived daemon legitimately reports "no newer binary to reload into"
+///   (its `shared-server` channel still points at its own old build) while the
+///   client can plainly see it is an older release. Trusting the server here is
+///   exactly what left "current client, stale server" stuck (the daemon's reload
+///   decision runs old code that can never drag itself forward). The newer
+///   client is authoritative, so it defers and repairs the channel before
+///   reloading.
 /// - `Some(true)`: the server self-reported a newer binary on disk -> defer.
 /// - `Some(false)`: the server is new enough to self-assess and found nothing
-///   newer to reload into -> trust it, do not fight it with a forced reload.
+///   newer to reload into, AND the client could not prove it is older -> trust
+///   it, do not fight it with a forced reload.
 /// - `None`: the server is too old to self-report. Fall back to our own
 ///   client-side release-version comparison, which is the only signal that can
 ///   catch a pre-self-heal daemon.
@@ -75,10 +86,16 @@ fn should_defer_history_for_runtime_identity_with_allow(
     if allow_mismatch {
         return false;
     }
+    // A client-proven-older server always wins: never let an old daemon's
+    // (locally correct but globally wrong) "no update" self-report veto the
+    // client's own release-order comparison.
+    if client_detected_stale {
+        return true;
+    }
     match server_has_update {
         Some(true) => true,
         Some(false) => false,
-        None => client_detected_stale,
+        None => false,
     }
 }
 
@@ -147,21 +164,31 @@ mod runtime_identity_tests {
     }
 
     #[test]
-    fn client_detection_only_applies_when_server_cannot_self_report() {
+    fn client_detected_older_server_always_defers() {
         // Ancient server (server_has_update: None) that the client independently
         // measured as older -> defer. This is the issue #295 macOS case where a
         // pre-self-heal daemon can never set server_has_update itself.
         assert!(should_defer_history_for_runtime_identity_with_allow(
             None, true, false
         ));
-        // A server new enough to self-assess and report "no newer binary" is
-        // trusted, even if a naive version compare disagrees: forcing a reload
-        // would only loop against a server that has nothing newer to exec into.
-        assert!(!should_defer_history_for_runtime_identity_with_allow(
+        // A server that self-reports "no newer binary" (Some(false)) but that the
+        // client can PROVE is an older release -> still defer. The daemon's
+        // self-report is locally correct (its own shared-server channel points at
+        // its old build) but globally wrong; the newer client is authoritative.
+        // This is the "current client, stale server" report: trusting Some(false)
+        // here is exactly what left the server stuck on the old version forever.
+        assert!(should_defer_history_for_runtime_identity_with_allow(
             Some(false),
             true,
             false
         ));
+        // Same-release/newer server (client could not prove it is older) that
+        // self-reports "no newer binary" -> trust it, do not force a reload loop.
+        assert!(!should_defer_history_for_runtime_identity_with_allow(
+            Some(false),
+            false,
+            false
+        ));
     }
 
     #[test]
@@ -950,7 +977,10 @@ pub(in crate::tui::app) fn handle_server_event(
                 server_has_update,
                 server_version.as_deref(),
             ) {
-                let client_detected_stale = server_has_update.is_none();
+                let client_detected_stale = server_release_is_older_than_client(
+                    server_version.as_deref(),
+                    &client_release_version(),
+                );
                 app.remote_server_version = server_version;
                 app.remote_server_short_name = server_name.clone();
                 app.remote_server_icon = server_icon.clone();
@@ -958,11 +988,29 @@ pub(in crate::tui::app) fn handle_server_event(
                 app.pending_server_reload = true;
                 app.clear_remote_startup_phase();
                 if client_detected_stale {
-                    // The server was too old to self-report an update
-                    // (server_has_update: None), but we independently measured
-                    // its release version as older than ours. This is the
-                    // issue #295 case: a pre-self-heal daemon that would
-                    // otherwise reject newer protocol requests (e.g. set_route).
+                    // The client independently measured the server's release as
+                    // older than its own. This covers both a pre-self-heal daemon
+                    // (server_has_update: None) AND a daemon that self-reports
+                    // "no update" because its own shared-server channel still
+                    // points at its old binary (the "current client, stale
+                    // server" report). Repair the channel client-side so the
+                    // forced reload below has a strictly-newer binary to exec
+                    // into instead of re-execing the same old build.
+                    match crate::build::repair_stale_shared_server_channel() {
+                        Ok(crate::build::SharedServerRepair::Repaired { repaired_to, .. }) => {
+                            crate::logging::info(&format!(
+                                "stale-server repair: repointed shared-server channel to {} before reloading older server",
+                                repaired_to
+                            ));
+                        }
+                        Ok(crate::build::SharedServerRepair::AlreadyCurrent) => {}
+                        Err(err) => {
+                            crate::logging::warn(&format!(
+                                "stale-server repair: failed to repoint shared-server channel: {}",
+                                err
+                            ));
+                        }
+                    }
                     app.set_status_notice(
                         "Connected server is an older release; reloading it before attach",
                     );
diff --git a/crates/jcode-tui/src/tui/app/state_ui.rs b/crates/jcode-tui/src/tui/app/state_ui.rs
index caf099516..9d4514c4b 100644
--- a/crates/jcode-tui/src/tui/app/state_ui.rs
+++ b/crates/jcode-tui/src/tui/app/state_ui.rs
@@ -34,16 +34,37 @@ impl App {
         self.display_edit_tool_message_count = self
             .display_messages
             .iter()
-            .filter(|message| {
-                message
-                    .tool_data
-                    .as_ref()
-                    .map(|tool| tools_ui::is_edit_tool_name(&tool.name))
-                    .unwrap_or(false)
-            })
+            .filter(|message| Self::display_message_is_edit_tool(message))
             .count();
     }
 
+    /// Whether a single display message counts as an edit-tool message for the
+    /// incrementally-maintained `display_edit_tool_message_count`.
+    fn display_message_is_edit_tool(message: &DisplayMessage) -> bool {
+        message
+            .tool_data
+            .as_ref()
+            .map(|tool| tools_ui::is_edit_tool_name(&tool.name))
+            .unwrap_or(false)
+    }
+
+    /// Fold a single message into the cached display-message counters with the
+    /// given sign (+1 when added, -1 when removed). This keeps the counters
+    /// O(1) per mutation instead of rescanning the whole transcript via
+    /// `recompute_display_message_stats`, which made appending M messages one at
+    /// a time cumulatively O(M^2).
+    pub(super) fn adjust_display_message_stats(&mut self, message: &DisplayMessage, added: bool) {
+        let delta: isize = if added { 1 } else { -1 };
+        if message.effective_role() == "user" {
+            self.display_user_message_count =
+                (self.display_user_message_count as isize + delta).max(0) as usize;
+        }
+        if Self::display_message_is_edit_tool(message) {
+            self.display_edit_tool_message_count =
+                (self.display_edit_tool_message_count as isize + delta).max(0) as usize;
+        }
+    }
+
     pub(super) fn active_client_session_id(&self) -> Option<&str> {
         if self.is_remote {
             self.remote_session_id.as_deref()
@@ -85,6 +106,13 @@ impl App {
 
     pub(super) fn bump_display_messages_version(&mut self) {
         self.recompute_display_message_stats();
+        self.bump_display_messages_version_no_stats();
+    }
+
+    /// Bump the display-messages version without rescanning the transcript to
+    /// recompute counters. Callers that have already maintained the cached
+    /// counters incrementally (e.g. a single append) use this to stay O(1).
+    pub(super) fn bump_display_messages_version_no_stats(&mut self) {
         self.display_messages_version = self.display_messages_version.wrapping_add(1);
         self.bump_context_revision();
         self.refresh_split_view_if_needed();
diff --git a/crates/jcode-tui/src/tui/app/state_ui_messages.rs b/crates/jcode-tui/src/tui/app/state_ui_messages.rs
index 782715f3e..01305a12a 100644
--- a/crates/jcode-tui/src/tui/app/state_ui_messages.rs
+++ b/crates/jcode-tui/src/tui/app/state_ui_messages.rs
@@ -79,8 +79,13 @@ impl App {
             return;
         }
         let is_tool = message.role == "tool";
+        // Maintain the cached display-message counters incrementally for this
+        // single append, then bump the version without a full O(M) rescan.
+        // Appending is the hot path; rescanning every append was O(M^2) over a
+        // long session.
+        self.adjust_display_message_stats(&message, true);
         self.display_messages.push(message);
-        self.bump_display_messages_version();
+        self.bump_display_messages_version_no_stats();
         if is_tool && self.diff_mode.has_side_pane() && self.diff_pane_auto_scroll {
             self.diff_pane_scroll = usize::MAX;
         }
@@ -88,6 +93,8 @@ impl App {
 
     pub(super) fn replace_display_messages(&mut self, mut messages: Vec<DisplayMessage>) {
         compact_display_messages_for_storage(&mut messages);
+        // Indices the collapse animation targets no longer apply to the new list.
+        self.reasoning_collapse = None;
         self.display_messages = messages;
         self.sync_compacted_history_lazy_from_display_messages();
         self.bump_display_messages_version();
@@ -350,6 +357,12 @@ impl App {
 
     pub(super) fn clear_display_messages(&mut self) {
         self.compacted_history_lazy = CompactedHistoryLazyState::default();
+        // The transcript (and the index the collapse animation targets) is about
+        // to be discarded; drop any in-flight collapse so it can't mutate a stale
+        // or unrelated message.
+        self.reasoning_collapse = None;
+        self.reasoning_block_start = None;
+        self.reasoning_block_started_at = None;
         if !self.display_messages.is_empty() {
             self.display_messages.clear();
             self.bump_display_messages_version();
diff --git a/crates/jcode-tui/src/tui/app/tests.rs b/crates/jcode-tui/src/tui/app/tests.rs
index 91609f928..b07dc8523 100644
--- a/crates/jcode-tui/src/tui/app/tests.rs
+++ b/crates/jcode-tui/src/tui/app/tests.rs
@@ -596,6 +596,7 @@ fn ancient_server_history_is_deferred_via_client_side_release_check() {
     // it is stale. The client must independently compare release versions and
     // defer + reload anyway, instead of attaching to the ancient daemon (which
     // would then reject newer protocol requests like `set_route`).
+    let _env_guard = crate::storage::lock_test_env();
     crate::env::remove_var("JCODE_ALLOW_SERVER_VERSION_MISMATCH");
     // The test binary's own version is dev/dirty (unorderable), so use the
     // test-only override to give the client a clean release version newer than
@@ -677,11 +678,196 @@ fn ancient_server_history_is_deferred_via_client_side_release_check() {
     );
 }
 
+#[test]
+fn older_server_reporting_no_update_is_still_deferred_via_client_check() {
+    // The "current client, stale server" report: the daemon self-reports
+    // `server_has_update: Some(false)` (its own shared-server channel still
+    // points at its old binary, so locally it sees nothing newer), but the
+    // client can PROVE it is an older release. Before this fix, Some(false)
+    // short-circuited and the client trusted the old server forever. Now the
+    // client's release-order check wins: defer + reload (after repairing the
+    // shared-server channel client-side).
+    let _env_guard = crate::storage::lock_test_env();
+    crate::env::remove_var("JCODE_ALLOW_SERVER_VERSION_MISMATCH");
+    crate::env::set_var("JCODE_TEST_CLIENT_VERSION_OVERRIDE", "v0.22.0 (abcd1234)");
+
+    let mut app = create_test_app();
+    let rt = tokio::runtime::Runtime::new().unwrap();
+    let _guard = rt.enter();
+    let mut remote = crate::tui::backend::RemoteConnection::dummy();
+
+    app.is_remote = true;
+    app.remote_session_id = Some("session_existing".to_string());
+
+    let redraw = app.handle_server_event(
+        crate::protocol::ServerEvent::History {
+            id: 1,
+            session_id: "session_from_old_server".to_string(),
+            messages: vec![],
+            images: vec![],
+            provider_name: Some("p".to_string()),
+            provider_model: Some("m".to_string()),
+            subagent_model: None,
+            autoreview_enabled: None,
+            autojudge_enabled: None,
+            available_models: vec!["m".to_string()],
+            available_model_routes: vec![],
+            mcp_servers: vec![],
+            skills: vec![],
+            total_tokens: None,
+            token_usage_totals: None,
+            all_sessions: vec![],
+            client_count: Some(1),
+            is_canary: Some(false),
+            reload_recovery: None,
+            // Older clean release than the client, but the daemon insists it has
+            // no newer binary to reload into.
+            server_version: Some("v0.14.6 (deadbeef)".to_string()),
+            server_name: Some("old-server".to_string()),
+            server_icon: Some("🕰".to_string()),
+            server_has_update: Some(false),
+            was_interrupted: None,
+            connection_type: Some("websocket".to_string()),
+            status_detail: None,
+            upstream_provider: None,
+            resolved_credential: None,
+            reasoning_effort: None,
+            service_tier: None,
+            compaction_mode: crate::config::CompactionMode::Reactive,
+            activity: None,
+            side_panel: crate::side_panel::SidePanelSnapshot::default(),
+        },
+        &mut remote,
+    );
+
+    crate::env::remove_var("JCODE_TEST_CLIENT_VERSION_OVERRIDE");
+
+    assert!(!redraw);
+    assert!(
+        app.pending_server_reload,
+        "client-proven-older server must defer + reload even when it reports Some(false)"
+    );
+    assert_eq!(app.remote_server_has_update, Some(false));
+    // Remote session state must NOT have been applied from the old server.
+    assert_eq!(app.remote_session_id.as_deref(), Some("session_existing"));
+    assert_eq!(remote.session_id(), None);
+    let content = app.display_messages().last().unwrap().content.clone();
+    assert!(
+        content.contains("older release") && content.contains("jcode server stop"),
+        "{content}"
+    );
+}
+
+#[test]
+fn older_server_history_repairs_stale_shared_server_channel_end_to_end() {
+    // Full-path sandbox: a real temp JCODE_HOME set up in the exact field state
+    // (shared-server pinned to an OLD build, stable advanced to a NEW release by
+    // a previous install). When the current client attaches to a server that
+    // self-reports an older release with `server_has_update: Some(false)`, the
+    // production History handler must repair the shared-server channel so the
+    // forced reload it queues has a strictly-newer binary to exec into.
+    use std::time::{Duration, SystemTime};
+    let _env_guard = crate::storage::lock_test_env();
+    crate::env::remove_var("JCODE_ALLOW_SERVER_VERSION_MISMATCH");
+    crate::env::set_var("JCODE_TEST_CLIENT_VERSION_OVERRIDE", "v0.22.0 (abcd1234)");
+    let temp = tempfile::TempDir::new().expect("temp home");
+    let prev_home = std::env::var_os("JCODE_HOME");
+    crate::env::set_var("JCODE_HOME", temp.path());
+
+    // Build the field state: shared-server -> OLD, stable -> NEW (newer mtime).
+    let base = SystemTime::UNIX_EPOCH + Duration::from_secs(1_000_000);
+    let write_version = |version: &str, mtime: SystemTime| {
+        let dir = crate::build::builds_dir()
+            .unwrap()
+            .join("versions")
+            .join(version);
+        std::fs::create_dir_all(&dir).unwrap();
+        let path = dir.join(crate::build::binary_name());
+        std::fs::write(&path, format!("bin {version}")).unwrap();
+        std::fs::File::open(&path)
+            .unwrap()
+            .set_modified(mtime)
+            .unwrap();
+    };
+    let old = "0.14.6";
+    let new = "0.22.0";
+    write_version(old, base);
+    write_version(new, base + Duration::from_secs(60));
+    crate::build::update_shared_server_symlink(old).expect("pin shared-server old");
+    crate::build::update_stable_symlink(new).expect("stable new");
+
+    let mut app = create_test_app();
+    let rt = tokio::runtime::Runtime::new().unwrap();
+    let _guard = rt.enter();
+    let mut remote = crate::tui::backend::RemoteConnection::dummy();
+    app.is_remote = true;
+    app.remote_session_id = Some("session_existing".to_string());
+
+    let _redraw = app.handle_server_event(
+        crate::protocol::ServerEvent::History {
+            id: 1,
+            session_id: "session_from_old_server".to_string(),
+            messages: vec![],
+            images: vec![],
+            provider_name: Some("p".to_string()),
+            provider_model: Some("m".to_string()),
+            subagent_model: None,
+            autoreview_enabled: None,
+            autojudge_enabled: None,
+            available_models: vec!["m".to_string()],
+            available_model_routes: vec![],
+            mcp_servers: vec![],
+            skills: vec![],
+            total_tokens: None,
+            token_usage_totals: None,
+            all_sessions: vec![],
+            client_count: Some(1),
+            is_canary: Some(false),
+            reload_recovery: None,
+            server_version: Some("v0.14.6 (deadbeef)".to_string()),
+            server_name: Some("old-server".to_string()),
+            server_icon: Some("🕰".to_string()),
+            server_has_update: Some(false),
+            was_interrupted: None,
+            connection_type: Some("websocket".to_string()),
+            status_detail: None,
+            upstream_provider: None,
+            resolved_credential: None,
+            reasoning_effort: None,
+            service_tier: None,
+            compaction_mode: crate::config::CompactionMode::Reactive,
+            activity: None,
+            side_panel: crate::side_panel::SidePanelSnapshot::default(),
+        },
+        &mut remote,
+    );
+
+    let repaired = crate::build::read_shared_server_version().ok().flatten();
+    let pending = app.pending_server_reload;
+
+    // Restore env before asserting so a panic cannot leak global state.
+    crate::env::remove_var("JCODE_TEST_CLIENT_VERSION_OVERRIDE");
+    if let Some(prev_home) = prev_home {
+        crate::env::set_var("JCODE_HOME", prev_home);
+    } else {
+        crate::env::remove_var("JCODE_HOME");
+    }
+
+    assert!(pending, "older server must queue a reload");
+    assert_eq!(
+        repaired.as_deref(),
+        Some(new),
+        "the History handler must repair the stale shared-server channel to the newer stable \
+         release so the queued reload upgrades the server instead of re-execing the old binary"
+    );
+}
+
 #[test]
 fn current_release_server_history_is_not_deferred_by_client_check() {
     // A server on the SAME or NEWER clean release as the client, with
     // server_has_update: None, must be trusted and attached normally. This
     // guards against the client-side check over-firing and looping reloads.
+    let _env_guard = crate::storage::lock_test_env();
     crate::env::remove_var("JCODE_ALLOW_SERVER_VERSION_MISMATCH");
     crate::env::set_var("JCODE_TEST_CLIENT_VERSION_OVERRIDE", "v0.17.0 (d741696f)");
 
diff --git a/crates/jcode-tui/src/tui/app/tests/commands_accounts_02/part_01.rs b/crates/jcode-tui/src/tui/app/tests/commands_accounts_02/part_01.rs
index 109d32c9b..fb3ebb0e9 100644
--- a/crates/jcode-tui/src/tui/app/tests/commands_accounts_02/part_01.rs
+++ b/crates/jcode-tui/src/tui/app/tests/commands_accounts_02/part_01.rs
@@ -669,6 +669,7 @@ fn test_improve_status_summarizes_current_todos() {
             &app.session.id,
             &[
                 crate::todo::TodoItem {
+                    group: None,
                     id: "one".to_string(),
                     content: "Profile startup path".to_string(),
                     status: "in_progress".to_string(),
@@ -679,6 +680,7 @@ fn test_improve_status_summarizes_current_todos() {
                     completion_confidence: None,
                 },
                 crate::todo::TodoItem {
+                    group: None,
                     id: "two".to_string(),
                     content: "Add regression test".to_string(),
                     status: "completed".to_string(),
@@ -770,6 +772,7 @@ fn test_improve_resume_uses_saved_mode_and_current_todos() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "resume1".to_string(),
                 content: "Refactor command parsing".to_string(),
                 status: "in_progress".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/tests/commands_accounts_02/part_02.rs b/crates/jcode-tui/src/tui/app/tests/commands_accounts_02/part_02.rs
index 75318fb15..c69427b69 100644
--- a/crates/jcode-tui/src/tui/app/tests/commands_accounts_02/part_02.rs
+++ b/crates/jcode-tui/src/tui/app/tests/commands_accounts_02/part_02.rs
@@ -127,6 +127,7 @@ fn test_refactor_status_summarizes_current_todos() {
             &app.session.id,
             &[
                 crate::todo::TodoItem {
+                    group: None,
                     id: "one".to_string(),
                     content: "Split giant module".to_string(),
                     status: "in_progress".to_string(),
@@ -137,6 +138,7 @@ fn test_refactor_status_summarizes_current_todos() {
                     completion_confidence: None,
                 },
                 crate::todo::TodoItem {
+                    group: None,
                     id: "two".to_string(),
                     content: "Run review subagent".to_string(),
                     status: "completed".to_string(),
@@ -177,6 +179,7 @@ fn test_refactor_resume_uses_saved_mode_and_current_todos() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "resume1".to_string(),
                 content: "Extract review prompt builder".to_string(),
                 status: "in_progress".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/tests/onboarding_flow.rs b/crates/jcode-tui/src/tui/app/tests/onboarding_flow.rs
index e0aa77f69..f93ecfc9c 100644
--- a/crates/jcode-tui/src/tui/app/tests/onboarding_flow.rs
+++ b/crates/jcode-tui/src/tui/app/tests/onboarding_flow.rs
@@ -411,7 +411,7 @@ fn no_external_transcripts_lands_on_suggestions_without_autosubmit() {
         // Temp home has no Codex transcripts, so opening the picker should land
         // the user on the clean new-session suggestion cards rather than
         // auto-submitting a "search for my last session" turn.
-        app.onboarding_open_transcript_picker(ExternalCli::Codex);
+        app.onboarding_open_transcript_picker(&[ExternalCli::Codex]);
         assert!(matches!(
             app.onboarding_phase(),
             Some(OnboardingPhase::Suggestions)
@@ -432,6 +432,72 @@ fn onboarding_picker_mode_carries_cli() {
     assert_ne!(mode, SessionPickerMode::Resume);
 }
 
+#[test]
+fn onboarding_picker_shows_both_codex_and_claude_transcripts() {
+    use std::fs;
+    with_temp_jcode_home(|| {
+        // Seed one Codex transcript and one Claude Code transcript under the
+        // sandbox-aware external home ($JCODE_HOME/external/...), mirroring a
+        // user who is logged into BOTH CLIs.
+        let home = std::env::var_os("JCODE_HOME").expect("JCODE_HOME");
+        let external = std::path::Path::new(&home).join("external");
+
+        let codex_dir = external.join(".codex/sessions/2026/04/05");
+        fs::create_dir_all(&codex_dir).expect("codex dir");
+        fs::write(
+            codex_dir.join("rollout-2026-04-05T19-00-00-codextest.jsonl"),
+            concat!(
+                "{\"timestamp\":\"2026-04-05T19:00:00Z\",\"type\":\"session_meta\",\"payload\":{\"id\":\"019d-codex-both\",\"timestamp\":\"2026-04-05T18:59:00Z\",\"cwd\":\"/tmp/codex-demo\",\"source\":\"cli\"}}\n",
+                "{\"timestamp\":\"2026-04-05T19:00:03Z\",\"type\":\"response_item\",\"payload\":{\"type\":\"message\",\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"CODEX_MARKER fix the widget\"}]}}\n",
+            ),
+        )
+        .expect("write codex transcript");
+
+        let claude_dir = external.join(".claude/projects/demo-project");
+        fs::create_dir_all(&claude_dir).expect("claude dir");
+        fs::write(
+            claude_dir.join("claude-session-both.jsonl"),
+            concat!(
+                "{\"type\":\"user\",\"uuid\":\"u1\",\"message\":{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"CLAUDE_MARKER fix the flaky test\"}]}}\n",
+                "{\"type\":\"assistant\",\"uuid\":\"a1\",\"parentUuid\":\"u1\",\"message\":{\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"done\"}]}}\n"
+            ),
+        )
+        .expect("write claude transcript");
+
+        let mut app = onboarding_test_app();
+        // Open the combined picker for BOTH detected CLIs.
+        app.onboarding_open_transcript_picker(&[ExternalCli::Codex, ExternalCli::ClaudeCode]);
+
+        // The picker overlay should be up with both CLIs' sessions visible
+        // (not just one).
+        let picker_cell = app
+            .session_picker_overlay
+            .as_ref()
+            .expect("picker overlay should be open");
+        let picker = picker_cell.borrow();
+        assert!(
+            picker.visible_session_count() >= 2,
+            "combined picker should list both CLIs' sessions, got {}",
+            picker.visible_session_count()
+        );
+
+        let mut saw_codex = false;
+        let mut saw_claude = false;
+        for session in picker.visible_session_iter_for_test() {
+            match session.source {
+                jcode_tui_session_picker::SessionSource::Codex => saw_codex = true,
+                jcode_tui_session_picker::SessionSource::ClaudeCode => saw_claude = true,
+                _ => {}
+            }
+        }
+        assert!(saw_codex, "Codex session should be present in combined picker");
+        assert!(
+            saw_claude,
+            "Claude Code session should be present in combined picker"
+        );
+    });
+}
+
 #[test]
 fn startup_check_skips_when_session_already_has_activity() {
     with_temp_jcode_home(|| {
diff --git a/crates/jcode-tui/src/tui/app/tests/reasoning_region.rs b/crates/jcode-tui/src/tui/app/tests/reasoning_region.rs
index 10b5a6075..b4d1ecf76 100644
--- a/crates/jcode-tui/src/tui/app/tests/reasoning_region.rs
+++ b/crates/jcode-tui/src/tui/app/tests/reasoning_region.rs
@@ -17,20 +17,47 @@ fn reasoning_region_emits_dim_italic_lines_no_gutter_header_or_footer() {
 
     app.open_reasoning_region();
     app.append_reasoning_text("Let me think.\nSecond thought.");
-    app.close_reasoning_region(None);
-
-    let text = app.streaming_text();
-    assert!(!text.contains("Thinking"), "no header expected: {text:?}");
-    assert!(!text.contains('>'), "no blockquote gutter expected: {text:?}");
-    assert!(!text.contains("Thought for"), "no footer expected: {text:?}");
+    // While streaming, reasoning is dim+italic markup in the live stream buffer.
+    let streaming = app.streaming_text().to_string();
+    assert!(
+        !streaming.contains("Thinking"),
+        "no header expected: {streaming:?}"
+    );
+    assert!(
+        !streaming.contains('>'),
+        "no blockquote gutter expected: {streaming:?}"
+    );
+    assert!(
+        !streaming.contains("Thought for"),
+        "no footer expected: {streaming:?}"
+    );
     let sentinel = jcode_tui_markdown::REASONING_SENTINEL;
     assert!(
-        text.contains(&format!("*{sentinel}Let me think.{sentinel}*")),
-        "first line not dim+italic: {text:?}"
+        streaming.contains(&format!("*{sentinel}Let me think.{sentinel}*")),
+        "first line not dim+italic: {streaming:?}"
     );
     assert!(
-        text.contains(&format!("*{sentinel}Second thought.{sentinel}*")),
-        "second line not dim+italic: {text:?}"
+        streaming.contains(&format!("*{sentinel}Second thought.{sentinel}*")),
+        "second line not dim+italic: {streaming:?}"
+    );
+
+    // In `current` mode (the default), closing moves the block into a dedicated
+    // collapsing `"reasoning"` display message and clears it from the stream.
+    app.close_reasoning_region(None);
+    assert!(
+        app.streaming_text().is_empty(),
+        "reasoning should leave the live stream once collapsed: {:?}",
+        app.streaming_text()
+    );
+    let reasoning_msg = app
+        .display_messages
+        .iter()
+        .find(|m| m.role == "reasoning")
+        .expect("reasoning message present");
+    assert!(
+        reasoning_msg.content.contains(sentinel),
+        "reasoning message keeps dim+italic markup: {:?}",
+        reasoning_msg.content
     );
 }
 
@@ -44,7 +71,12 @@ fn reasoning_region_closes_before_normal_output() {
     app.close_reasoning_region(None);
     app.append_streaming_text("Final answer.");
 
+    // The answer stays in the live stream and must never be styled as reasoning.
     let text = app.streaming_text();
+    assert!(
+        text.contains("Final answer."),
+        "answer present in stream: {text:?}"
+    );
     let answer_line = text
         .lines()
         .find(|l| l.contains("Final answer."))
@@ -53,9 +85,14 @@ fn reasoning_region_closes_before_normal_output() {
         !answer_line.contains(jcode_tui_markdown::REASONING_SENTINEL),
         "final answer must not be styled as reasoning: {answer_line:?}"
     );
+    // The reasoning collapsed into its own message; it is no longer in the stream.
     assert!(
-        text.contains("\n\nFinal answer."),
-        "missing blank-line separator before output: {text:?}"
+        !text.contains(jcode_tui_markdown::REASONING_SENTINEL),
+        "reasoning must not remain in the answer stream: {text:?}"
+    );
+    assert!(
+        app.display_messages.iter().any(|m| m.role == "reasoning"),
+        "a collapsing reasoning message should exist"
     );
 }
 
@@ -94,11 +131,18 @@ fn reasoning_line_split_across_deltas_stays_one_run() {
     app.append_reasoning_text("two\n");
     app.close_reasoning_region(None);
 
-    let text = app.streaming_text();
+    // The split-across-deltas line is committed as a single emphasis run in the
+    // collapsed reasoning message.
+    let content = app
+        .display_messages
+        .iter()
+        .find(|m| m.role == "reasoning")
+        .map(|m| m.content.clone())
+        .expect("reasoning message present");
     let sentinel = jcode_tui_markdown::REASONING_SENTINEL;
     assert!(
-        text.contains(&format!("*{sentinel}one two{sentinel}*")),
-        "split line must be one emphasis run: {text:?}"
+        content.contains(&format!("*{sentinel}one two{sentinel}*")),
+        "split line must be one emphasis run: {content:?}"
     );
 }
 
@@ -112,7 +156,15 @@ fn reasoning_region_renders_dim_italic_text_without_gutter() {
     app.append_reasoning_text("considering options\n");
     app.close_reasoning_region(None);
 
-    let lines = crate::tui::markdown::render_markdown_with_width(app.streaming_text(), Some(80));
+    // In `current` mode the reasoning now lives in a dedicated collapsing message.
+    let reasoning_content = app
+        .display_messages
+        .iter()
+        .find(|m| m.role == "reasoning")
+        .map(|m| m.content.clone())
+        .expect("reasoning message present");
+
+    let lines = crate::tui::markdown::render_markdown_with_width(&reasoning_content, Some(80));
     let body = lines
         .iter()
         .find(|l| {
@@ -248,7 +300,7 @@ fn reasoning_partial_promotes_to_committed_line_on_newline() {
 #[test]
 fn reasoning_close_promotes_pending_partial_line() {
     // Closing the region with an in-progress (no-newline) partial promotes it to a
-    // committed line exactly once.
+    // committed line exactly once, then collapses into the reasoning message.
     let mut app = create_test_app();
     let sentinel = jcode_tui_markdown::REASONING_SENTINEL;
 
@@ -256,15 +308,152 @@ fn reasoning_close_promotes_pending_partial_line() {
     app.append_reasoning_text("final thought");
     app.close_reasoning_region(None);
 
-    let text = app.streaming_text();
+    // The live stream no longer carries the reasoning; it moved into its message.
+    assert!(
+        app.streaming_text().is_empty(),
+        "reasoning should leave the live stream once collapsed: {:?}",
+        app.streaming_text()
+    );
+    let content = app
+        .display_messages
+        .iter()
+        .find(|m| m.role == "reasoning")
+        .map(|m| m.content.clone())
+        .expect("reasoning message present");
     assert_eq!(
-        text.matches(&format!("*{sentinel}final thought{sentinel}*"))
+        content
+            .matches(&format!("*{sentinel}final thought{sentinel}*"))
             .count(),
         1,
-        "pending partial promoted exactly once on close: {text:?}"
-    );
-    assert!(
-        text.ends_with("\n\n"),
-        "region terminated with blank line: {text:?}"
+        "pending partial promoted exactly once on close: {content:?}"
     );
 }
+
+#[test]
+fn reasoning_block_line_markups_keeps_only_sentinel_lines() {
+    use crate::tui::app::input::{reasoning_block_line_markups, reasoning_message_content};
+
+    let mut block = String::new();
+    block.push_str(&jcode_tui_markdown::reasoning_line_markup("alpha"));
+    block.push('\n'); // a blank separator line (no sentinel)
+    block.push_str(&jcode_tui_markdown::reasoning_line_markup("beta"));
+
+    let lines = reasoning_block_line_markups(&block);
+    assert_eq!(lines.len(), 2, "blank separators are dropped: {lines:?}");
+    let sentinel = jcode_tui_markdown::REASONING_SENTINEL;
+    assert!(lines[0].contains(&format!("{sentinel}alpha{sentinel}")));
+    assert!(lines[1].contains(&format!("{sentinel}beta{sentinel}")));
+
+    // Full content shows every line; remaining==0 shows only the summary.
+    let summary = jcode_tui_markdown::reasoning_line_markup("▸ thought");
+    let full = reasoning_message_content(&summary, &lines, lines.len());
+    assert!(full.contains("alpha") && full.contains("beta"));
+    let collapsed = reasoning_message_content(&summary, &lines, 0);
+    assert!(collapsed.contains("▸ thought"));
+    assert!(!collapsed.contains("alpha") && !collapsed.contains("beta"));
+
+    // A partial reveal keeps the *trailing* lines (oldest fold away first).
+    let partial = reasoning_message_content(&summary, &lines, 1);
+    assert!(partial.contains("beta"), "trailing line kept: {partial:?}");
+    assert!(!partial.contains("alpha"), "leading line folded: {partial:?}");
+}
+
+#[test]
+fn reasoning_summary_markup_uses_duration_when_known() {
+    use crate::tui::app::input::reasoning_summary_markup;
+    use std::time::Duration;
+
+    let with_secs = reasoning_summary_markup(3, Some(Duration::from_secs(12)));
+    assert!(with_secs.contains("▸ thought for 12s"), "{with_secs:?}");
+
+    let no_time = reasoning_summary_markup(4, None);
+    assert!(no_time.contains("▸ thought (4 lines)"), "{no_time:?}");
+}
+
+#[test]
+fn reasoning_collapse_finalizes_to_single_summary_line() {
+    let mut app = create_test_app();
+
+    app.open_reasoning_region();
+    app.append_reasoning_text("first\nsecond\nthird\n");
+    app.close_reasoning_region(None);
+
+    assert!(app.reasoning_collapse_active(), "collapse should start");
+
+    // Snapping finalizes the message to just the summary line.
+    app.finalize_reasoning_collapse();
+    assert!(!app.reasoning_collapse_active(), "collapse cleared on finalize");
+
+    let content = app
+        .display_messages
+        .iter()
+        .find(|m| m.role == "reasoning")
+        .map(|m| m.content.clone())
+        .expect("reasoning message present");
+    assert!(content.contains("▸ thought"), "summary present: {content:?}");
+    assert!(!content.contains("first"), "lines folded away: {content:?}");
+    assert!(!content.contains("third"), "lines folded away: {content:?}");
+}
+
+#[test]
+fn reasoning_collapse_drops_when_target_message_replaced() {
+    let mut app = create_test_app();
+
+    app.open_reasoning_region();
+    app.append_reasoning_text("thinking\n");
+    app.close_reasoning_region(None);
+    assert!(app.reasoning_collapse_active());
+
+    // A transcript reset must invalidate the animation target safely.
+    app.clear_display_messages();
+    assert!(!app.reasoning_collapse_active());
+    // Advancing now is a no-op and must not panic.
+    assert!(!app.advance_reasoning_collapse());
+}
+
+#[test]
+fn reasoning_collapse_visible_lines_shrink_monotonically_over_time() {
+    use crate::tui::app::input::REASONING_COLLAPSE_DURATION;
+    use std::time::Duration;
+
+    let mut app = create_test_app();
+    app.open_reasoning_region();
+    app.append_reasoning_text("l1\nl2\nl3\nl4\nl5\nl6\n");
+    app.close_reasoning_region(None);
+    let sentinel = jcode_tui_markdown::REASONING_SENTINEL;
+
+    let count_visible = |app: &App| -> usize {
+        app.display_messages
+            .iter()
+            .find(|m| m.role == "reasoning")
+            .map(|m| {
+                m.content
+                    .split_inclusive('\n')
+                    .filter(|seg| seg.contains(sentinel))
+                    .filter(|seg| !seg.contains('▸'))
+                    .count()
+            })
+            .unwrap_or(0)
+    };
+
+    // Sample the eased timeline; visible reasoning lines must never increase and
+    // must reach a single summary line (0 source lines) at/after the duration.
+    let dur = REASONING_COLLAPSE_DURATION;
+    let mut prev = usize::MAX;
+    for frac in [0.0_f32, 0.25, 0.5, 0.75, 1.0] {
+        let elapsed = Duration::from_secs_f32(dur.as_secs_f32() * frac);
+        app.backdate_reasoning_collapse_for_test(elapsed)
+            .expect("collapse active");
+        app.advance_reasoning_collapse();
+        let visible = count_visible(&app);
+        assert!(
+            visible <= prev,
+            "visible lines must not increase: frac={frac} visible={visible} prev={prev}"
+        );
+        prev = visible;
+    }
+
+    // Past the duration the animation is finalized to the summary only.
+    assert!(!app.reasoning_collapse_active(), "collapse should finish");
+    assert_eq!(count_visible(&app), 0, "only the summary line remains");
+}
diff --git a/crates/jcode-tui/src/tui/app/tests/remote_events_reload_01/part_01.rs b/crates/jcode-tui/src/tui/app/tests/remote_events_reload_01/part_01.rs
index 1c68ff9b9..a398b3962 100644
--- a/crates/jcode-tui/src/tui/app/tests/remote_events_reload_01/part_01.rs
+++ b/crates/jcode-tui/src/tui/app/tests/remote_events_reload_01/part_01.rs
@@ -1022,6 +1022,7 @@ fn test_remote_done_auto_pokes_again_when_todos_remain() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Continue working".to_string(),
                 status: "pending".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/tests/remote_events_reload_01/part_02.rs b/crates/jcode-tui/src/tui/app/tests/remote_events_reload_01/part_02.rs
index f72f6b552..e1d8dbb7f 100644
--- a/crates/jcode-tui/src/tui/app/tests/remote_events_reload_01/part_02.rs
+++ b/crates/jcode-tui/src/tui/app/tests/remote_events_reload_01/part_02.rs
@@ -82,6 +82,7 @@ fn test_remote_auto_poke_followup_preserves_visible_timer_and_stays_hidden() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Continue working".to_string(),
                 status: "pending".to_string(),
@@ -133,6 +134,7 @@ fn test_remote_auto_poke_completion_above_threshold_only_updates_ui() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Finished work".to_string(),
                 status: "completed".to_string(),
@@ -170,6 +172,7 @@ fn test_remote_auto_poke_completion_below_threshold_tells_model_to_keep_working(
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Needs validation".to_string(),
                 status: "completed".to_string(),
@@ -209,6 +212,7 @@ fn test_remote_poke_status_and_off_update_state() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Continue working".to_string(),
                 status: "pending".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/tests/remote_events_reload_02/part_01.rs b/crates/jcode-tui/src/tui/app/tests/remote_events_reload_02/part_01.rs
index 17a05a4d0..b0741706f 100644
--- a/crates/jcode-tui/src/tui/app/tests/remote_events_reload_02/part_01.rs
+++ b/crates/jcode-tui/src/tui/app/tests/remote_events_reload_02/part_01.rs
@@ -9,6 +9,7 @@ fn test_remote_poke_queues_when_turn_is_in_progress() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Continue working".to_string(),
                 status: "pending".to_string(),
@@ -50,6 +51,7 @@ fn test_remote_poke_queues_when_turn_is_in_progress() {
             &app.session.id,
             &[
                 crate::todo::TodoItem {
+                    group: None,
                     id: "todo-1".to_string(),
                     content: "Continue working".to_string(),
                     status: "pending".to_string(),
@@ -60,6 +62,7 @@ fn test_remote_poke_queues_when_turn_is_in_progress() {
                     completion_confidence: None,
                 },
                 crate::todo::TodoItem {
+                    group: None,
                     id: "todo-2".to_string(),
                     content: "Handle the newly discovered follow-up".to_string(),
                     status: "pending".to_string(),
@@ -148,6 +151,7 @@ fn test_remote_interrupted_auto_poke_requeues_after_deferred_poke() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Resume after interrupt".to_string(),
                 status: "pending".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/tests/remote_events_reload_02/part_02.rs b/crates/jcode-tui/src/tui/app/tests/remote_events_reload_02/part_02.rs
index 58caa3c64..965603b5d 100644
--- a/crates/jcode-tui/src/tui/app/tests/remote_events_reload_02/part_02.rs
+++ b/crates/jcode-tui/src/tui/app/tests/remote_events_reload_02/part_02.rs
@@ -169,6 +169,67 @@ fn test_remove_display_message_bumps_version() {
     assert_ne!(app.display_messages_version, before);
 }
 
+#[test]
+fn test_incremental_display_message_counts_match_full_recompute() {
+    let mut app = create_test_app();
+
+    // Interleave user, assistant, and edit-tool messages via the public append
+    // path, which now maintains the counters incrementally instead of
+    // rescanning the whole transcript.
+    for i in 0..50 {
+        app.push_display_message(DisplayMessage::user(format!("prompt {i}")));
+        app.push_display_message(DisplayMessage::assistant(format!("reply {i}")));
+        if i % 3 == 0 {
+            app.push_display_message(DisplayMessage {
+                role: "tool".to_string(),
+                content: format!("edited file {i}"),
+                tool_calls: vec![],
+                duration_secs: None,
+                title: None,
+                tool_data: Some(crate::message::ToolCall {
+                    id: format!("edit-{i}"),
+                    name: "edit".to_string(),
+                    input: serde_json::json!({"file_path": format!("src/file_{i}.rs")}),
+                    intent: None,
+                    thought_signature: None,
+                }),
+            });
+        }
+    }
+
+    // Remove a few messages to exercise the decrement path.
+    app.remove_display_message(0);
+    app.remove_display_message(5);
+
+    let incremental_user = app.display_user_message_count;
+    let incremental_edit = app.display_edit_tool_message_count;
+
+    let expected_user = app
+        .display_messages
+        .iter()
+        .filter(|m| m.effective_role() == "user")
+        .count();
+    let expected_edit = app
+        .display_messages
+        .iter()
+        .filter(|m| {
+            m.tool_data
+                .as_ref()
+                .map(|tool| crate::tui::ui::tools_ui::is_edit_tool_name(&tool.name))
+                .unwrap_or(false)
+        })
+        .count();
+
+    assert_eq!(
+        incremental_user, expected_user,
+        "incrementally-maintained user count should match a full recompute"
+    );
+    assert_eq!(
+        incremental_edit, expected_edit,
+        "incrementally-maintained edit-tool count should match a full recompute"
+    );
+}
+
 #[test]
 fn test_handle_remote_disconnect_retryable_pending_schedules_retry() {
     let mut app = create_test_app();
diff --git a/crates/jcode-tui/src/tui/app/tests/remote_startup_input_01/part_01.rs b/crates/jcode-tui/src/tui/app/tests/remote_startup_input_01/part_01.rs
index 7e5aa9b50..02eb99f1c 100644
--- a/crates/jcode-tui/src/tui/app/tests/remote_startup_input_01/part_01.rs
+++ b/crates/jcode-tui/src/tui/app/tests/remote_startup_input_01/part_01.rs
@@ -16,6 +16,7 @@ fn test_finish_turn_does_not_duplicate_existing_poke_followup() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Keep going".to_string(),
                 status: "pending".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/tests/remote_startup_input_02/part_01.rs b/crates/jcode-tui/src/tui/app/tests/remote_startup_input_02/part_01.rs
index 8d4113ce7..edee12476 100644
--- a/crates/jcode-tui/src/tui/app/tests/remote_startup_input_02/part_01.rs
+++ b/crates/jcode-tui/src/tui/app/tests/remote_startup_input_02/part_01.rs
@@ -591,7 +591,12 @@ fn test_submit_input_commits_pending_streaming_assistant_text_before_user_messag
     ));
     app.bump_display_messages_version();
     app.streaming_text = "Here is the final paragraph".to_string();
-    assert_eq!(app.stream_buffer.push(" that was still buffered."), None);
+    // Mirror the real streaming caller: append any paced chunk the buffer reveals.
+    // The paced StreamBuffer may reveal part of the text immediately, so commit
+    // (below) must still flush the remainder.
+    if let Some(chunk) = app.stream_buffer.push(" that was still buffered.") {
+        app.append_streaming_text(&chunk);
+    }
 
     app.input = "follow up".to_string();
     app.cursor_pos = app.input.len();
@@ -731,6 +736,7 @@ fn test_create_transfer_session_from_parent_copies_todos_and_uses_compacted_cont
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Carry this forward".to_string(),
                 status: "pending".to_string(),
@@ -866,6 +872,7 @@ fn test_escape_interrupt_disables_auto_poke_while_processing() {
     app.queued_messages
         .push(super::commands::build_poke_message(&[
             crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "keep going".to_string(),
                 status: "pending".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/tests/scroll_copy_01/part_02.rs b/crates/jcode-tui/src/tui/app/tests/scroll_copy_01/part_02.rs
index 370e43373..b50da2af1 100644
--- a/crates/jcode-tui/src/tui/app/tests/scroll_copy_01/part_02.rs
+++ b/crates/jcode-tui/src/tui/app/tests/scroll_copy_01/part_02.rs
@@ -261,6 +261,7 @@ fn test_remote_escape_interrupt_disables_auto_poke_while_processing() {
     app.queued_messages
         .push(super::commands::build_poke_message(&[
             crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "keep going".to_string(),
                 status: "pending".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/tests/scroll_copy_02/part_01.rs b/crates/jcode-tui/src/tui/app/tests/scroll_copy_02/part_01.rs
index bfd0cd27a..797c8e0c6 100644
--- a/crates/jcode-tui/src/tui/app/tests/scroll_copy_02/part_01.rs
+++ b/crates/jcode-tui/src/tui/app/tests/scroll_copy_02/part_01.rs
@@ -115,6 +115,39 @@ fn test_copy_selection_select_all_uses_rendered_chat_text_without_copy_badges()
     );
 }
 
+#[test]
+fn test_copy_selection_metrics_match_built_selection_text() {
+    let _render_lock = scroll_render_test_lock();
+    let (mut app, mut terminal) = create_copy_test_app();
+
+    render_and_snap(&app, &mut terminal);
+    app.handle_key(KeyCode::Char('y'), KeyModifiers::ALT)
+        .unwrap();
+    assert!(app.select_all_in_copy_mode());
+
+    // The allocation-free metrics path used by the status line must agree with
+    // the char/line counts of the actually-built selection text.
+    let range = app
+        .normalized_copy_selection()
+        .expect("normalized selection range");
+    let text = app
+        .current_copy_selection_text()
+        .expect("selection text for full transcript");
+    let (chars, lines) =
+        crate::tui::ui::copy_selection_metrics(range).expect("selection metrics");
+
+    assert_eq!(
+        chars,
+        text.chars().count(),
+        "metrics char count should match built selection text"
+    );
+    assert_eq!(
+        lines,
+        text.lines().count().max(1),
+        "metrics line count should match built selection text"
+    );
+}
+
 #[test]
 fn test_copy_selection_full_user_prompt_line_skips_prompt_chrome() {
     let _render_lock = scroll_render_test_lock();
@@ -978,6 +1011,242 @@ fn test_copy_selection_drag_near_top_edge_keeps_auto_scrolling() {
     ));
 }
 
+#[test]
+fn test_copy_selection_drag_to_bottom_edge_when_pinned_does_not_snap_or_autoscroll() {
+    // Regression: when the transcript is already pinned to the bottom (the common
+    // case), dragging a selection into the bottom edge "hot zone" used to always
+    // snap the cursor to the very last visible line and arm a downward autoscroll,
+    // even though there is nothing more below to scroll into. That made it
+    // impossible to precisely highlight the bottom rows: the selection kept
+    // jumping to the end. With nothing to scroll, the edge band must stay inert so
+    // the selection lands on the exact line under the cursor.
+    let _render_lock = scroll_render_test_lock();
+    let mut app = create_test_app();
+
+    // Tall transcript pinned to the bottom: the bottom rows of the pane are
+    // filled with real content, and there is nothing below to scroll into.
+    let lines = (1..=200)
+        .map(|idx| format!("line {idx:03}"))
+        .collect::<Vec<_>>()
+        .join("\n");
+    app.display_messages = vec![DisplayMessage {
+        role: "assistant".to_string(),
+        content: lines,
+        tool_calls: vec![],
+        duration_secs: None,
+        title: None,
+        tool_data: None,
+    }];
+    app.bump_display_messages_version();
+    app.scroll_offset = 0;
+    app.auto_scroll_paused = false;
+    app.is_processing = false;
+    app.streaming_text.clear();
+    app.status = ProcessingStatus::Idle;
+
+    let backend = ratatui::backend::TestBackend::new(60, 16);
+    let mut terminal = ratatui::Terminal::new(backend).expect("failed to create test terminal");
+    render_and_snap(&app, &mut terminal);
+
+    app.handle_key(KeyCode::Char('y'), KeyModifiers::ALT)
+        .unwrap();
+
+    let (visible_start, visible_end) =
+        crate::tui::ui::copy_viewport_visible_range().expect("visible copy range");
+    let line_count = crate::tui::ui::copy_viewport_line_count().expect("line count");
+    assert_eq!(
+        visible_end, line_count,
+        "test precondition: view must be pinned to the bottom with no content below"
+    );
+    assert!(
+        visible_start > 0,
+        "test precondition: tall transcript must have content scrolled above the view"
+    );
+
+    let layout = crate::tui::ui::last_layout_snapshot().expect("layout snapshot");
+    let area = layout.messages_area;
+    let col = area.x + 1;
+
+    // Pick a real content line near (but not at) the bottom to target.
+    let target_line = visible_end.saturating_sub(2);
+    assert!(target_line >= visible_start, "need a visible target line");
+    let target_row = area.y + (target_line - visible_start) as u16;
+    // The bottom edge band covers the last few rows; target_row must sit inside
+    // it for this regression to be meaningful.
+    let last_row = area.y + area.height - 1;
+    assert!(
+        target_row >= last_row.saturating_sub(2),
+        "target line must fall within the bottom edge hot zone"
+    );
+
+    // Anchor higher up in the viewport.
+    let anchor_row = area.y + 1;
+    app.handle_mouse_event(MouseEvent {
+        kind: MouseEventKind::Down(MouseButton::Left),
+        column: col,
+        row: anchor_row,
+        modifiers: KeyModifiers::empty(),
+    });
+    let before_scroll = app.scroll_offset();
+
+    app.handle_mouse_event(MouseEvent {
+        kind: MouseEventKind::Drag(MouseButton::Left),
+        column: col,
+        row: target_row,
+        modifiers: KeyModifiers::empty(),
+    });
+
+    // No autoscroll should be armed: there is nothing below to pull in.
+    assert!(
+        !crate::tui::TuiState::copy_selection_edge_autoscroll_active(&app),
+        "edge autoscroll must not arm when pinned to the bottom with no content below"
+    );
+    assert_eq!(
+        app.scroll_offset(),
+        before_scroll,
+        "dragging into the bottom band while pinned must not scroll"
+    );
+
+    // The selection end should land on the exact line under the cursor, not snap
+    // to the very last line of the transcript.
+    let range = app.normalized_copy_selection().expect("normalized range");
+    assert_eq!(
+        range.end.abs_line, target_line,
+        "selection should extend to the line under the cursor, not snap to the last line"
+    );
+
+    app.handle_mouse_event(MouseEvent {
+        kind: MouseEventKind::Up(MouseButton::Left),
+        column: col,
+        row: target_row,
+        modifiers: KeyModifiers::empty(),
+    });
+}
+
+#[test]
+fn test_copy_selection_drag_below_last_line_fully_selects_last_line() {
+    // Dragging *past* the last content line (into the empty area below the
+    // chat pane) should fully select that last line through its end, just like
+    // native terminal and browser selection. The chat pane is sized to its
+    // content, so a downward drag that overshoots reports a row at/below the
+    // bottom boundary that maps to no line at all; that used to silently drop
+    // the extension so the bottom line could never be fully highlighted.
+    let _render_lock = scroll_render_test_lock();
+    let mut app = create_test_app();
+
+    let lines = (1..=6)
+        .map(|idx| format!("line {idx:03}"))
+        .collect::<Vec<_>>()
+        .join("\n");
+    app.display_messages = vec![DisplayMessage {
+        role: "assistant".to_string(),
+        content: lines,
+        tool_calls: vec![],
+        duration_secs: None,
+        title: None,
+        tool_data: None,
+    }];
+    app.bump_display_messages_version();
+    app.scroll_offset = 0;
+    app.auto_scroll_paused = false;
+    app.is_processing = false;
+    app.streaming_text.clear();
+    app.status = ProcessingStatus::Idle;
+
+    // Tall terminal so there is empty space below the content-sized chat pane.
+    let backend = ratatui::backend::TestBackend::new(60, 20);
+    let mut terminal = ratatui::Terminal::new(backend).expect("failed to create test terminal");
+    render_and_snap(&app, &mut terminal);
+
+    app.handle_key(KeyCode::Char('y'), KeyModifiers::ALT)
+        .unwrap();
+
+    let (visible_start, visible_end) =
+        crate::tui::ui::copy_viewport_visible_range().expect("visible copy range");
+    let line_count = crate::tui::ui::copy_viewport_line_count().expect("line count");
+    assert_eq!(visible_end, line_count, "view must be pinned to the bottom");
+
+    // The last line that maps to a real screen point.
+    let last_line = (visible_start..visible_end)
+        .rev()
+        .find(|&ln| {
+            crate::tui::ui::copy_viewport_line_text(ln)
+                .map(|t| unicode_width::UnicodeWidthStr::width(t.as_str()) > 0)
+                .unwrap_or(false)
+        })
+        .expect("a non-empty visible content line");
+    let last_text = crate::tui::ui::copy_viewport_line_text(last_line).unwrap_or_default();
+    let last_width = unicode_width::UnicodeWidthStr::width(last_text.as_str());
+
+    let layout = crate::tui::ui::last_layout_snapshot().expect("layout snapshot");
+    let area = layout.messages_area;
+
+    // Anchor on a valid cell at the START of the last content line.
+    let last_content_row = area.y + (last_line - visible_start) as u16;
+    let anchor_x = (area.x..area.x + area.width)
+        .find(|&x| {
+            crate::tui::ui::copy_viewport_point_from_screen(x, last_content_row)
+                .map(|p| p.abs_line == last_line)
+                .unwrap_or(false)
+        })
+        .expect("a screen column mapping to the last content line");
+    app.handle_mouse_event(MouseEvent {
+        kind: MouseEventKind::Down(MouseButton::Left),
+        column: anchor_x,
+        row: last_content_row,
+        modifiers: KeyModifiers::empty(),
+    });
+
+    // Drag straight down, past the bottom of the pane, with the cursor x landing
+    // partway through (not at the end of) the last line. Even so the whole last
+    // line should be selected, because we have overshot it vertically.
+    let mid_x = anchor_x + 1;
+    let below_row = (area.y + area.height + 2).min(terminal.backend().size().unwrap().height - 1);
+    assert!(
+        below_row > last_content_row,
+        "test must drag strictly below the last content row"
+    );
+    let before_scroll = app.scroll_offset();
+    app.handle_mouse_event(MouseEvent {
+        kind: MouseEventKind::Drag(MouseButton::Left),
+        column: mid_x,
+        row: below_row,
+        modifiers: KeyModifiers::empty(),
+    });
+
+    // No autoscroll (nothing below), and no scroll movement.
+    assert!(
+        !crate::tui::TuiState::copy_selection_edge_autoscroll_active(&app),
+        "edge autoscroll must not arm dragging past the last line"
+    );
+    assert_eq!(app.scroll_offset(), before_scroll, "must not scroll");
+
+    // The selection should now extend through the END of the last line.
+    let range = app.normalized_copy_selection().expect("normalized range");
+    assert_eq!(
+        range.end.abs_line, last_line,
+        "selection should extend to the last content line"
+    );
+    assert_eq!(
+        range.end.column, last_width,
+        "selection should cover the full last line (through its end)"
+    );
+    let selected = app
+        .current_copy_selection_text()
+        .expect("expected selection text");
+    assert!(
+        selected.contains(last_text.trim_end()),
+        "selection should include the full last line text: got {selected:?}"
+    );
+
+    app.handle_mouse_event(MouseEvent {
+        kind: MouseEventKind::Up(MouseButton::Left),
+        column: mid_x,
+        row: below_row,
+        modifiers: KeyModifiers::empty(),
+    });
+}
+
 #[test]
 fn test_alt_a_copies_chat_viewport_with_context_when_input_empty() {
     let _render_lock = scroll_render_test_lock();
diff --git a/crates/jcode-tui/src/tui/app/tests/state_model_poke_02/part_01.rs b/crates/jcode-tui/src/tui/app/tests/state_model_poke_02/part_01.rs
index 693b17d88..086662f47 100644
--- a/crates/jcode-tui/src/tui/app/tests/state_model_poke_02/part_01.rs
+++ b/crates/jcode-tui/src/tui/app/tests/state_model_poke_02/part_01.rs
@@ -910,6 +910,7 @@ fn test_context_command_reports_session_context_snapshot() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "one".to_string(),
                 content: "Inspect context summary".to_string(),
                 status: "pending".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/tests/state_model_poke_03.rs b/crates/jcode-tui/src/tui/app/tests/state_model_poke_03.rs
index 014d543e9..f12848a2a 100644
--- a/crates/jcode-tui/src/tui/app/tests/state_model_poke_03.rs
+++ b/crates/jcode-tui/src/tui/app/tests/state_model_poke_03.rs
@@ -1858,6 +1858,7 @@ fn test_poke_arms_auto_poke_until_todos_are_done() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Finish the remaining task".to_string(),
                 status: "pending".to_string(),
@@ -1888,6 +1889,7 @@ fn test_poke_status_reports_current_state() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Finish the remaining task".to_string(),
                 status: "pending".to_string(),
@@ -1940,6 +1942,7 @@ fn test_poke_off_disarms_and_clears_queued_followup() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Keep going".to_string(),
                 status: "pending".to_string(),
@@ -1987,6 +1990,7 @@ fn test_poke_queues_when_turn_is_in_progress() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Finish the remaining task".to_string(),
                 status: "pending".to_string(),
@@ -2021,6 +2025,7 @@ fn test_poke_queues_when_turn_is_in_progress() {
             &app.session.id,
             &[
                 crate::todo::TodoItem {
+                    group: None,
                     id: "todo-1".to_string(),
                     content: "Finish the remaining task".to_string(),
                     status: "pending".to_string(),
@@ -2031,6 +2036,7 @@ fn test_poke_queues_when_turn_is_in_progress() {
                     completion_confidence: None,
                 },
                 crate::todo::TodoItem {
+                    group: None,
                     id: "todo-2".to_string(),
                     content: "Pick up the newly discovered task".to_string(),
                     status: "pending".to_string(),
@@ -2088,6 +2094,7 @@ fn test_finish_turn_auto_pokes_again_when_todos_remain() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Keep going".to_string(),
                 status: "in_progress".to_string(),
@@ -2118,6 +2125,7 @@ fn test_finish_turn_auto_poke_queues_confidence_summary_when_todos_done() {
             &app.session.id,
             &[
                 crate::todo::TodoItem {
+                    group: None,
                     id: "todo-1".to_string(),
                     content: "Finish risky provider path".to_string(),
                     status: "completed".to_string(),
@@ -2128,6 +2136,7 @@ fn test_finish_turn_auto_poke_queues_confidence_summary_when_todos_done() {
                     completion_confidence: Some(80),
                 },
                 crate::todo::TodoItem {
+                    group: None,
                     id: "todo-2".to_string(),
                     content: "Document straightforward behavior".to_string(),
                     status: "completed".to_string(),
@@ -2191,6 +2200,7 @@ fn test_finish_turn_without_auto_poke_does_not_queue_confidence_summary() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Done without poke".to_string(),
                 status: "completed".to_string(),
@@ -2224,6 +2234,7 @@ fn test_finish_turn_auto_poke_preserves_visible_turn_started() {
         crate::todo::save_todos(
             &app.session.id,
             &[crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Keep going".to_string(),
                 status: "in_progress".to_string(),
diff --git a/crates/jcode-tui/src/tui/app/todos_view.rs b/crates/jcode-tui/src/tui/app/todos_view.rs
index 8644955b8..5f04f0d93 100644
--- a/crates/jcode-tui/src/tui/app/todos_view.rs
+++ b/crates/jcode-tui/src/tui/app/todos_view.rs
@@ -275,6 +275,29 @@ fn build_todos_view_markdown(session_id: Option<&str>, todos: &[TodoItem]) -> St
         ("cancelled", "Cancelled"),
     ];
 
+    if let Some(groups) = grouped_todos_view(todos) {
+        for (group, items) in groups {
+            let group_name = group.as_deref().unwrap_or("Other");
+            let group_total = items.len();
+            let group_done = items.iter().filter(|t| t.status == "completed").count();
+            markdown.push_str(&format!(
+                "\n## {} ({}/{})\n",
+                group_name, group_done, group_total
+            ));
+            for (status, heading) in sections {
+                let status_items = sorted_group_items_for_status(&items, status);
+                if status_items.is_empty() {
+                    continue;
+                }
+                markdown.push_str(&format!("\n### {}\n\n", heading));
+                for todo in status_items {
+                    markdown.push_str(&format_todo_markdown(todo));
+                }
+            }
+        }
+        return markdown;
+    }
+
     for (status, heading) in sections {
         let items = sorted_todos_for_status(todos, status);
         if items.is_empty() {
@@ -289,6 +312,49 @@ fn build_todos_view_markdown(session_id: Option<&str>, todos: &[TodoItem]) -> St
     markdown
 }
 
+/// Group key for the side-panel view, treating empty/whitespace as ungrouped.
+fn todo_group_key(todo: &TodoItem) -> Option<String> {
+    todo.group
+        .as_deref()
+        .map(str::trim)
+        .filter(|group| !group.is_empty())
+        .map(|group| group.to_string())
+}
+
+/// Partition todos into ordered groups (first-seen order, ungrouped last).
+/// Returns `None` when no todo declares a group so callers keep the flat layout.
+fn grouped_todos_view(todos: &[TodoItem]) -> Option<Vec<(Option<String>, Vec<&TodoItem>)>> {
+    if !todos.iter().any(|todo| todo_group_key(todo).is_some()) {
+        return None;
+    }
+    let mut groups: Vec<(Option<String>, Vec<&TodoItem>)> = Vec::new();
+    for todo in todos {
+        let key = todo_group_key(todo);
+        if let Some(entry) = groups.iter_mut().find(|(existing, _)| *existing == key) {
+            entry.1.push(todo);
+        } else {
+            groups.push((key, vec![todo]));
+        }
+    }
+    groups.sort_by_key(|(key, _)| key.is_none());
+    Some(groups)
+}
+
+fn sorted_group_items_for_status<'a>(items: &[&'a TodoItem], status: &str) -> Vec<&'a TodoItem> {
+    let mut filtered: Vec<&TodoItem> = items
+        .iter()
+        .copied()
+        .filter(|todo| todo.status == status)
+        .collect();
+    filtered.sort_by(|a, b| {
+        priority_rank(&a.priority)
+            .cmp(&priority_rank(&b.priority))
+            .then_with(|| a.content.cmp(&b.content))
+            .then_with(|| a.id.cmp(&b.id))
+    });
+    filtered
+}
+
 fn sorted_todos_for_status<'a>(todos: &'a [TodoItem], status: &str) -> Vec<&'a TodoItem> {
     let mut items: Vec<&TodoItem> = todos.iter().filter(|todo| todo.status == status).collect();
     items.sort_by(|a, b| {
@@ -405,6 +471,7 @@ fn hash_todos_payload(session_id: Option<&str>, todos: &[TodoItem]) -> u64 {
         todo.content.hash(&mut hasher);
         todo.status.hash(&mut hasher);
         todo.priority.hash(&mut hasher);
+        todo.group.hash(&mut hasher);
         todo.confidence.hash(&mut hasher);
         todo.completion_confidence.hash(&mut hasher);
         todo.blocked_by.hash(&mut hasher);
@@ -441,6 +508,7 @@ mod tests {
             content: content.to_string(),
             status: status.to_string(),
             priority: priority.to_string(),
+            group: None,
             confidence,
             completion_confidence,
             blocked_by: Vec::new(),
@@ -495,4 +563,40 @@ mod tests {
 
         assert_ne!(before, after);
     }
+
+    #[test]
+    fn todos_view_markdown_groups_items_under_group_headers() {
+        let mut grouped_a = todo("g1", "Cut frame allocs", "in_progress", "high", Some(80), None);
+        grouped_a.group = Some("optimize rendering".to_string());
+        let mut grouped_b = todo("g2", "Batch draw calls", "completed", "medium", Some(70), Some(90));
+        grouped_b.group = Some("optimize rendering".to_string());
+        let mut other = todo("o1", "Fix scrollback", "pending", "low", Some(60), None);
+        other.group = Some("scrollback".to_string());
+        let ungrouped = todo("u1", "Misc cleanup", "pending", "low", Some(60), None);
+
+        let markdown = build_todos_view_markdown(
+            Some("session_test"),
+            &[grouped_a, grouped_b, other, ungrouped],
+        );
+
+        assert!(markdown.contains("## optimize rendering (1/2)"), "{markdown}");
+        assert!(markdown.contains("## scrollback (0/1)"), "{markdown}");
+        assert!(markdown.contains("## Other (0/1)"), "{markdown}");
+        // Status sub-headings nest under groups.
+        assert!(markdown.contains("### In progress"), "{markdown}");
+        // First-seen group order, ungrouped bucket last.
+        let opt = markdown.find("## optimize rendering").unwrap();
+        let scroll = markdown.find("## scrollback").unwrap();
+        let other_idx = markdown.find("## Other").unwrap();
+        assert!(opt < scroll && scroll < other_idx, "{markdown}");
+    }
+
+    #[test]
+    fn todos_view_hash_changes_when_group_changes() {
+        let mut todos = vec![todo("g", "Group hash", "pending", "high", Some(80), None)];
+        let before = hash_todos_payload(Some("session_test"), &todos);
+        todos[0].group = Some("rendering".to_string());
+        let after = hash_todos_payload(Some("session_test"), &todos);
+        assert_ne!(before, after);
+    }
 }
diff --git a/crates/jcode-tui/src/tui/app/tui_lifecycle.rs b/crates/jcode-tui/src/tui/app/tui_lifecycle.rs
index be1ff75cb..f561dfeab 100644
--- a/crates/jcode-tui/src/tui/app/tui_lifecycle.rs
+++ b/crates/jcode-tui/src/tui/app/tui_lifecycle.rs
@@ -374,6 +374,9 @@ impl App {
             reasoning_streaming: false,
             reasoning_pending_line: String::new(),
             reasoning_partial_len: 0,
+            reasoning_block_start: None,
+            reasoning_block_started_at: None,
+            reasoning_collapse: None,
             reload_requested: None,
             rebuild_requested: None,
             update_requested: None,
@@ -780,6 +783,9 @@ impl App {
             reasoning_streaming: false,
             reasoning_pending_line: String::new(),
             reasoning_partial_len: 0,
+            reasoning_block_start: None,
+            reasoning_block_started_at: None,
+            reasoning_collapse: None,
             reload_requested: None,
             rebuild_requested: None,
             update_requested: None,
diff --git a/crates/jcode-tui/src/tui/app/tui_state.rs b/crates/jcode-tui/src/tui/app/tui_state.rs
index 1d6e20fb8..ceffa3969 100644
--- a/crates/jcode-tui/src/tui/app/tui_state.rs
+++ b/crates/jcode-tui/src/tui/app/tui_state.rs
@@ -459,9 +459,9 @@ impl crate::tui::TuiState for App {
         if self.is_remote {
             self.remote_header_provider_name().unwrap_or_default()
         } else {
-            self.remote_provider_name.clone().unwrap_or_else(|| {
-                crate::provider_catalog::runtime_provider_display_name(self.provider.name())
-            })
+            self.remote_provider_name
+                .clone()
+                .unwrap_or_else(|| self.provider.display_name())
         }
     }
 
@@ -596,6 +596,10 @@ impl crate::tui::TuiState for App {
         self.mouse_scroll_queue != 0
     }
 
+    fn reasoning_collapse_animating(&self) -> bool {
+        self.reasoning_collapse_active()
+    }
+
     fn total_session_tokens(&self) -> Option<(u64, u64)> {
         // In remote mode, use tokens from server
         // Independent mode doesn't currently track total tokens
@@ -1028,6 +1032,7 @@ impl crate::tui::TuiState for App {
                     status: item.status.clone(),
                     priority: item.priority.clone(),
                     id: item.id.clone(),
+                    group: None,
                     blocked_by: item.blocked_by.clone(),
                     assigned_to: item.assigned_to.clone(),
                     confidence: None,
@@ -1294,9 +1299,9 @@ impl crate::tui::TuiState for App {
             provider_name: if uses_remote_widget_metadata {
                 self.remote_provider_name
                     .clone()
-                    .or_else(|| Some(self.provider.name().to_string()))
+                    .or_else(|| Some(self.provider.display_name()))
             } else {
-                Some(self.provider.name().to_string())
+                Some(self.provider.display_name())
             },
             auth_method,
             upstream_provider: self.upstream_provider.clone(),
@@ -1506,19 +1511,22 @@ impl crate::tui::TuiState for App {
             return None;
         }
 
-        let text = self.current_copy_selection_text().unwrap_or_default();
-        let has_selection = !text.is_empty();
+        // Compute selection metrics without building the full selected string,
+        // which previously re-allocated the entire selection on every render
+        // frame and drag move (O(selection) per frame; a "select all" rebuilt
+        // the whole transcript text repeatedly).
+        let (selected_chars, selected_lines) = self
+            .normalized_copy_selection()
+            .and_then(crate::tui::ui::copy_selection_metrics)
+            .unwrap_or((0, 0));
+        let has_selection = selected_chars > 0;
         Some(crate::tui::CopySelectionStatus {
             pane: self
                 .current_copy_selection_pane()
                 .unwrap_or(crate::tui::CopySelectionPane::Chat),
             has_action: has_selection,
-            selected_chars: text.chars().count(),
-            selected_lines: if has_selection {
-                text.lines().count().max(1)
-            } else {
-                0
-            },
+            selected_chars,
+            selected_lines: if has_selection { selected_lines.max(1) } else { 0 },
             dragging: self.copy_selection_dragging,
         })
     }
diff --git a/crates/jcode-tui/src/tui/app/turn.rs b/crates/jcode-tui/src/tui/app/turn.rs
index 6e78f7771..9f278015a 100644
--- a/crates/jcode-tui/src/tui/app/turn.rs
+++ b/crates/jcode-tui/src/tui/app/turn.rs
@@ -268,6 +268,8 @@ impl App {
                         if let Some(chunk) = self.stream_buffer.flush_smooth_frame() {
                             self.append_streaming_text(&chunk);
                         }
+                        // Advance the "current reasoning collapses away" animation.
+                        self.advance_reasoning_collapse();
                         // Poll for background compaction completion during streaming
                         self.poll_compaction_completion();
                         status_spinner_renderer.draw_full(self, terminal)?;
diff --git a/crates/jcode-tui/src/tui/info_widget_overview.rs b/crates/jcode-tui/src/tui/info_widget_overview.rs
index be4e1b08a..a1df179b8 100644
--- a/crates/jcode-tui/src/tui/info_widget_overview.rs
+++ b/crates/jcode-tui/src/tui/info_widget_overview.rs
@@ -262,6 +262,7 @@ mod tests {
     fn compute_page_layout_keeps_multiple_expanded_pages_when_height_allows() {
         let data = InfoWidgetData {
             todos: vec![TodoItem {
+                group: None,
                 content: "ship refactor".to_string(),
                 status: "pending".to_string(),
                 priority: "high".to_string(),
diff --git a/crates/jcode-tui/src/tui/info_widget_tests.rs b/crates/jcode-tui/src/tui/info_widget_tests.rs
index 31f08b08d..0849b49fb 100644
--- a/crates/jcode-tui/src/tui/info_widget_tests.rs
+++ b/crates/jcode-tui/src/tui/info_widget_tests.rs
@@ -99,6 +99,7 @@ fn todos_widgets_show_item_and_aggregate_confidence() {
     let data = InfoWidgetData {
         todos: vec![
             crate::todo::TodoItem {
+                group: None,
                 id: "todo-1".to_string(),
                 content: "Validate confidence UI".to_string(),
                 status: "in_progress".to_string(),
@@ -109,6 +110,7 @@ fn todos_widgets_show_item_and_aggregate_confidence() {
                 assigned_to: None,
             },
             crate::todo::TodoItem {
+                group: None,
                 id: "todo-2".to_string(),
                 content: "Ship completed item".to_string(),
                 status: "completed".to_string(),
@@ -136,9 +138,68 @@ fn todos_widgets_show_item_and_aggregate_confidence() {
     assert!(compact_text.contains("86%"));
 }
 
+#[test]
+fn todos_widgets_render_group_headers_when_groups_present() {
+    let mk = |group: Option<&str>, id: &str, status: &str| crate::todo::TodoItem {
+        group: group.map(|g| g.to_string()),
+        id: id.to_string(),
+        content: format!("task {id}"),
+        status: status.to_string(),
+        priority: "medium".to_string(),
+        confidence: Some(80),
+        completion_confidence: None,
+        blocked_by: Vec::new(),
+        assigned_to: None,
+    };
+    let data = InfoWidgetData {
+        todos: vec![
+            mk(Some("optimize rendering"), "a", "completed"),
+            mk(Some("optimize rendering"), "b", "in_progress"),
+            mk(Some("fix scrollback"), "c", "pending"),
+            mk(None, "d", "pending"),
+        ],
+        ..Default::default()
+    };
+
+    let expanded = lines_text(&render_todos_expanded(&data, Rect::new(0, 0, 80, 14)));
+    // Group headers appear with per-group progress counters, first-seen order,
+    // and the ungrouped bucket renders under "Other".
+    assert!(expanded.contains("optimize rendering"), "{expanded}");
+    assert!(expanded.contains("1/2"), "{expanded}");
+    assert!(expanded.contains("fix scrollback"), "{expanded}");
+    assert!(expanded.contains("Other"), "{expanded}");
+    let opt_idx = expanded.find("optimize rendering").unwrap();
+    let fix_idx = expanded.find("fix scrollback").unwrap();
+    let other_idx = expanded.find("Other").unwrap();
+    assert!(opt_idx < fix_idx, "first-seen group order: {expanded}");
+    assert!(fix_idx < other_idx, "ungrouped bucket last: {expanded}");
+}
+
+#[test]
+fn todos_widgets_stay_flat_without_groups() {
+    let mk = |id: &str, status: &str| crate::todo::TodoItem {
+        group: None,
+        id: id.to_string(),
+        content: format!("task {id}"),
+        status: status.to_string(),
+        priority: "medium".to_string(),
+        confidence: Some(80),
+        completion_confidence: None,
+        blocked_by: Vec::new(),
+        assigned_to: None,
+    };
+    let data = InfoWidgetData {
+        todos: vec![mk("a", "completed"), mk("b", "pending")],
+        ..Default::default()
+    };
+    let expanded = lines_text(&render_todos_expanded(&data, Rect::new(0, 0, 80, 14)));
+    assert!(!expanded.contains("Other"), "no group bucket: {expanded}");
+}
+
 #[test]
 fn todos_widget_renders_exact_pips_for_small_lists() {
     let mk = |status: &str| crate::todo::TodoItem {
+        group: None,
         id: status.to_string(),
         content: format!("item {status}"),
         status: status.to_string(),
@@ -991,6 +1052,7 @@ fn placements_never_include_border_only_widgets() {
             ..Default::default()
         }),
         todos: vec![crate::todo::TodoItem {
+            group: None,
             content: "ship patch".to_string(),
             status: "in_progress".to_string(),
             priority: "high".to_string(),
diff --git a/crates/jcode-tui/src/tui/info_widget_todos.rs b/crates/jcode-tui/src/tui/info_widget_todos.rs
index afae1453a..c25e6c496 100644
--- a/crates/jcode-tui/src/tui/info_widget_todos.rs
+++ b/crates/jcode-tui/src/tui/info_widget_todos.rs
@@ -161,6 +161,189 @@ fn push_aggregate_confidence_suffix(spans: &mut Vec<Span<'static>>, data: &InfoW
     ));
 }
 
+/// Normalize a todo's group label, treating empty/whitespace as ungrouped.
+fn todo_group_key(todo: &crate::todo::TodoItem) -> Option<String> {
+    todo.group
+        .as_deref()
+        .map(str::trim)
+        .filter(|group| !group.is_empty())
+        .map(|group| group.to_string())
+}
+
+/// Partition todos into ordered groups, preserving the order groups first
+/// appear. Ungrouped items collapse into a trailing `None` bucket. Returns
+/// `None` when no todo declares a group, so callers fall back to the flat list.
+fn grouped_todos(
+    todos: &[crate::todo::TodoItem],
+) -> Option<Vec<(Option<String>, Vec<&crate::todo::TodoItem>)>> {
+    if !todos.iter().any(|todo| todo_group_key(todo).is_some()) {
+        return None;
+    }
+    let mut groups: Vec<(Option<String>, Vec<&crate::todo::TodoItem>)> = Vec::new();
+    for todo in todos {
+        let key = todo_group_key(todo);
+        if let Some(entry) = groups.iter_mut().find(|(existing, _)| *existing == key) {
+            entry.1.push(todo);
+        } else {
+            groups.push((key, vec![todo]));
+        }
+    }
+    // Keep the ungrouped bucket last; sort_by_key is stable so named groups
+    // retain their first-seen order.
+    groups.sort_by_key(|(key, _)| key.is_none());
+    Some(groups)
+}
+
+fn status_sort_rank(status: &str) -> u8 {
+    match status {
+        "in_progress" => 0,
+        "pending" => 1,
+        "completed" => 2,
+        "cancelled" => 3,
+        _ => 4,
+    }
+}
+
+fn sort_todos_by_status<'a>(todos: &[&'a crate::todo::TodoItem]) -> Vec<&'a crate::todo::TodoItem> {
+    let mut sorted: Vec<&crate::todo::TodoItem> = todos.to_vec();
+    sorted.sort_by(|a, b| status_sort_rank(&a.status).cmp(&status_sort_rank(&b.status)));
+    sorted
+}
+
+fn push_group_header(
+    lines: &mut Vec<Line<'static>>,
+    name: &str,
+    items: &[&crate::todo::TodoItem],
+    inner: Rect,
+) {
+    let total = items.len();
+    let completed = items.iter().filter(|t| t.status == "completed").count();
+    let counter = format!(" {}/{}", completed, total);
+    let max_name = inner
+        .width
+        .saturating_sub(counter.len() as u16)
+        .max(4) as usize;
+    let highlight = items.iter().any(|t| t.status == "in_progress");
+    let name_style = if highlight {
+        Style::default().fg(rgb(255, 210, 130)).bold()
+    } else {
+        Style::default().fg(rgb(170, 175, 205)).bold()
+    };
+    lines.push(Line::from(vec![
+        Span::styled(truncate_smart(name, max_name), name_style),
+        Span::styled(counter, Style::default().fg(rgb(120, 120, 140))),
+    ]));
+}
+
+/// Render one todo as a line. `show_priority_marker` adds the `!` high-priority
+/// marker (used by the expanded widget); `indent` is the leading-space depth
+/// used when items sit under a group header.
+fn push_todo_item_line(
+    lines: &mut Vec<Line<'static>>,
+    todo: &crate::todo::TodoItem,
+    inner: Rect,
+    show_priority_marker: bool,
+    indent: usize,
+) {
+    let is_blocked = !todo.blocked_by.is_empty();
+    let (icon, status_color) = if is_blocked && todo.status != "completed" {
+        ("⊳", rgb(180, 140, 100))
+    } else {
+        match todo.status.as_str() {
+            "completed" => ("✓", rgb(100, 180, 100)),
+            "in_progress" => ("▶", rgb(255, 200, 100)),
+            "cancelled" => ("✗", rgb(120, 80, 80)),
+            _ => ("○", rgb(120, 120, 130)),
+        }
+    };
+
+    let priority_marker = if show_priority_marker {
+        match todo.priority.as_str() {
+            "high" => ("!", rgb(255, 120, 100)),
+            _ => ("", rgb(120, 120, 130)),
+        }
+    } else {
+        ("", rgb(120, 120, 130))
+    };
+
+    let suffix = if is_blocked && todo.status != "completed" {
+        " (blocked)"
+    } else {
+        ""
+    };
+
+    let reserved = indent as u16
+        + 3
+        + priority_marker.0.len() as u16
+        + suffix.len() as u16
+        + todo_confidence_suffix_width(todo);
+    let max_len = inner.width.saturating_sub(reserved) as usize;
+    let content = truncate_smart(&todo.content, max_len);
+
+    let text_color = if todo.status == "completed" {
+        rgb(100, 100, 110)
+    } else if is_blocked {
+        rgb(120, 120, 130)
+    } else if todo.status == "in_progress" {
+        rgb(200, 200, 210)
+    } else {
+        rgb(160, 160, 170)
+    };
+
+    let mut spans = Vec::new();
+    if indent > 0 {
+        spans.push(Span::raw(" ".repeat(indent)));
+    }
+    spans.push(Span::styled(
+        format!("{} ", icon),
+        Style::default().fg(status_color),
+    ));
+    if !priority_marker.0.is_empty() {
+        spans.push(Span::styled(
+            priority_marker.0,
+            Style::default().fg(priority_marker.1),
+        ));
+    }
+    spans.push(Span::styled(content, Style::default().fg(text_color)));
+    push_todo_confidence_suffix(&mut spans, todo);
+    if !suffix.is_empty() {
+        spans.push(Span::styled(
+            suffix.to_string(),
+            Style::default().fg(rgb(100, 100, 110)),
+        ));
+    }
+    lines.push(Line::from(spans));
+}
+
+/// Render todos partitioned by group, honoring a `max_lines` budget that counts
+/// both group headers and item rows. Returns the rendered lines plus the number
+/// of todo items actually shown (so callers can render a "+N more" footer).
+fn render_grouped_todo_lines(
+    groups: &[(Option<String>, Vec<&crate::todo::TodoItem>)],
+    inner: Rect,
+    show_priority_marker: bool,
+    max_lines: usize,
+) -> (Vec<Line<'static>>, usize) {
+    let mut lines: Vec<Line<'static>> = Vec::new();
+    let mut shown = 0usize;
+    for (group, items) in groups {
+        if lines.len() >= max_lines {
+            break;
+        }
+        let header_name = group.as_deref().unwrap_or("Other");
+        push_group_header(&mut lines, header_name, items, inner);
+        for todo in sort_todos_by_status(items) {
+            if lines.len() >= max_lines {
+                break;
+            }
+            push_todo_item_line(&mut lines, todo, inner, show_priority_marker, 2);
+            shown += 1;
+        }
+    }
+    (lines, shown)
+}
+
+
 /// Render todos widget content
 pub(super) fn render_todos_widget(data: &InfoWidgetData, inner: Rect) -> Vec<Line<'static>> {
     if data.todos.is_empty() {
@@ -193,71 +376,33 @@ pub(super) fn render_todos_widget(data: &InfoWidgetData, inner: Rect) -> Vec<Lin
     push_aggregate_confidence_suffix(&mut header, data);
     lines.push(Line::from(header));
 
+    let available_lines = inner.height.saturating_sub(1) as usize; // Account for header
+    let budget = available_lines.min(5).max(1);
+
+    // Grouped layout when any todo declares a group; otherwise the flat list.
+    if let Some(groups) = grouped_todos(&data.todos) {
+        let (group_lines, shown) = render_grouped_todo_lines(&groups, inner, false, budget);
+        lines.extend(group_lines);
+        if total > shown {
+            lines.push(Line::from(vec![Span::styled(
+                format!("  +{} more", total - shown),
+                Style::default().fg(rgb(100, 100, 110)),
+            )]));
+        }
+        return lines;
+    }
+
     // Sort todos: in_progress first, then pending, then completed
     let mut sorted_todos: Vec<&crate::todo::TodoItem> = data.todos.iter().collect();
-    sorted_todos.sort_by(|a, b| {
-        let order = |s: &str| match s {
-            "in_progress" => 0,
-            "pending" => 1,
-            "completed" => 2,
-            "cancelled" => 3,
-            _ => 4,
-        };
-        order(&a.status).cmp(&order(&b.status))
-    });
+    sorted_todos.sort_by(|a, b| status_sort_rank(&a.status).cmp(&status_sort_rank(&b.status)));
 
     // Render todos (limit based on available height)
-    let available_lines = inner.height.saturating_sub(1) as usize; // Account for header
-    for todo in sorted_todos.iter().take(available_lines.min(5)) {
-        let is_blocked = !todo.blocked_by.is_empty();
-        let (icon, status_color) = if is_blocked && todo.status != "completed" {
-            ("⊳", rgb(180, 140, 100))
-        } else {
-            match todo.status.as_str() {
-                "completed" => ("✓", rgb(100, 180, 100)),
-                "in_progress" => ("▶", rgb(255, 200, 100)),
-                "cancelled" => ("✗", rgb(120, 80, 80)),
-                _ => ("○", rgb(120, 120, 130)),
-            }
-        };
-
-        let suffix = if is_blocked && todo.status != "completed" {
-            " (blocked)"
-        } else {
-            ""
-        };
-        let max_len = inner
-            .width
-            .saturating_sub(3 + suffix.len() as u16 + todo_confidence_suffix_width(todo))
-            as usize;
-        let content = truncate_smart(&todo.content, max_len);
-
-        let text_color = if todo.status == "completed" {
-            rgb(100, 100, 110)
-        } else if is_blocked {
-            rgb(120, 120, 130)
-        } else if todo.status == "in_progress" {
-            rgb(200, 200, 210)
-        } else {
-            rgb(160, 160, 170)
-        };
-
-        let mut spans = vec![
-            Span::styled(format!("{} ", icon), Style::default().fg(status_color)),
-            Span::styled(content, Style::default().fg(text_color)),
-        ];
-        push_todo_confidence_suffix(&mut spans, todo);
-        if !suffix.is_empty() {
-            spans.push(Span::styled(
-                suffix.to_string(),
-                Style::default().fg(rgb(100, 100, 110)),
-            ));
-        }
-        lines.push(Line::from(spans));
+    for todo in sorted_todos.iter().take(budget) {
+        push_todo_item_line(&mut lines, todo, inner, false, 0);
     }
 
     // Show count of remaining items
-    let shown = available_lines.min(5).min(sorted_todos.len());
+    let shown = budget.min(sorted_todos.len());
     if data.todos.len() > shown {
         let remaining = data.todos.len() - shown;
         lines.push(Line::from(vec![Span::styled(
@@ -301,86 +446,28 @@ pub(super) fn render_todos_expanded(data: &InfoWidgetData, inner: Rect) -> Vec<L
     push_aggregate_confidence_suffix(&mut header, data);
     lines.push(Line::from(header));
 
-    // Sort todos: in_progress first, then pending, then completed
-    let mut sorted_todos: Vec<&crate::todo::TodoItem> = data.todos.iter().collect();
-    sorted_todos.sort_by(|a, b| {
-        let order = |s: &str| match s {
-            "in_progress" => 0,
-            "pending" => 1,
-            "completed" => 2,
-            "cancelled" => 3,
-            _ => 4,
-        };
-        order(&a.status).cmp(&order(&b.status))
-    });
-
-    // Render todos with priority colors
     let available_lines = MAX_TODO_LINES.saturating_sub(1); // Account for header
-    for todo in sorted_todos.iter().take(available_lines) {
-        let is_blocked = !todo.blocked_by.is_empty();
-        let (icon, status_color) = if is_blocked && todo.status != "completed" {
-            ("⊳", rgb(180, 140, 100))
-        } else {
-            match todo.status.as_str() {
-                "completed" => ("✓", rgb(100, 180, 100)),
-                "in_progress" => ("▶", rgb(255, 200, 100)),
-                "cancelled" => ("✗", rgb(120, 80, 80)),
-                _ => ("○", rgb(120, 120, 130)),
-            }
-        };
-
-        // Priority indicator
-        let priority_marker = match todo.priority.as_str() {
-            "high" => ("!", rgb(255, 120, 100)),
-            "medium" => ("", rgb(200, 180, 100)),
-            _ => ("", rgb(120, 120, 130)),
-        };
-
-        let suffix = if is_blocked && todo.status != "completed" {
-            " (blocked)"
-        } else {
-            ""
-        };
-        let max_len = inner
-            .width
-            .saturating_sub(4 + suffix.len() as u16 + todo_confidence_suffix_width(todo))
-            as usize;
-        let content = truncate_smart(&todo.content, max_len);
-
-        // Dim completed and blocked items
-        let text_color = if todo.status == "completed" {
-            rgb(100, 100, 110)
-        } else if is_blocked {
-            rgb(120, 120, 130)
-        } else if todo.status == "in_progress" {
-            rgb(200, 200, 210)
-        } else {
-            rgb(160, 160, 170)
-        };
-
-        let mut spans = vec![Span::styled(
-            format!("{} ", icon),
-            Style::default().fg(status_color),
-        )];
-
-        if !priority_marker.0.is_empty() {
-            spans.push(Span::styled(
-                priority_marker.0,
-                Style::default().fg(priority_marker.1),
-            ));
-        }
-
-        spans.push(Span::styled(content, Style::default().fg(text_color)));
-        push_todo_confidence_suffix(&mut spans, todo);
 
-        if !suffix.is_empty() {
-            spans.push(Span::styled(
-                suffix.to_string(),
+    // Grouped layout when any todo declares a group; otherwise the flat list.
+    if let Some(groups) = grouped_todos(&data.todos) {
+        let (group_lines, shown) = render_grouped_todo_lines(&groups, inner, true, available_lines);
+        lines.extend(group_lines);
+        if total > shown {
+            lines.push(Line::from(vec![Span::styled(
+                format!("  +{} more", total - shown),
                 Style::default().fg(rgb(100, 100, 110)),
-            ));
+            )]));
         }
+        return lines;
+    }
+
+    // Sort todos: in_progress first, then pending, then completed
+    let mut sorted_todos: Vec<&crate::todo::TodoItem> = data.todos.iter().collect();
+    sorted_todos.sort_by(|a, b| status_sort_rank(&a.status).cmp(&status_sort_rank(&b.status)));
 
-        lines.push(Line::from(spans));
+    // Render todos with priority colors
+    for todo in sorted_todos.iter().take(available_lines) {
+        push_todo_item_line(&mut lines, todo, inner, true, 0);
     }
 
     // Show count of remaining items
diff --git a/crates/jcode-tui/src/tui/mod.rs b/crates/jcode-tui/src/tui/mod.rs
index 062ecbd3c..d517b1245 100644
--- a/crates/jcode-tui/src/tui/mod.rs
+++ b/crates/jcode-tui/src/tui/mod.rs
@@ -219,6 +219,11 @@ pub trait TuiState {
     fn has_pending_mouse_scroll_animation(&self) -> bool {
         false
     }
+    /// Whether a "current reasoning collapses away" animation is in progress and
+    /// the redraw loop must keep ticking to advance it.
+    fn reasoning_collapse_animating(&self) -> bool {
+        false
+    }
     /// Optional configured keybinding label for external dictation.
     fn dictation_key_label(&self) -> Option<String>;
     /// Time since app started (for startup animations)
@@ -1294,6 +1299,7 @@ pub(crate) fn redraw_interval_with_policy(
         || !state.streaming_text().is_empty()
         || state.status_notice().is_some()
         || state.has_pending_mouse_scroll_animation()
+        || state.reasoning_collapse_animating()
         || state.copy_selection_edge_autoscroll_active()
         || state.has_notification()
         || rate_limit_countdown_redraw_active(state)
@@ -1353,6 +1359,7 @@ pub(crate) fn periodic_redraw_required(state: &dyn TuiState) -> bool {
         || !state.streaming_text().is_empty()
         || state.status_notice().is_some()
         || state.has_pending_mouse_scroll_animation()
+        || state.reasoning_collapse_animating()
         || state.copy_selection_edge_autoscroll_active()
         || state.chat_overscroll_active()
         || state.has_notification()
diff --git a/crates/jcode-tui/src/tui/session_picker.rs b/crates/jcode-tui/src/tui/session_picker.rs
index 1d95efd4b..eae58d398 100644
--- a/crates/jcode-tui/src/tui/session_picker.rs
+++ b/crates/jcode-tui/src/tui/session_picker.rs
@@ -34,7 +34,7 @@ mod render;
 #[cfg(test)]
 use loading::collect_recent_session_stems;
 pub(crate) use loading::latest_external_cli_session_secs;
-pub(crate) use loading::load_external_cli_sessions_grouped;
+pub(crate) use loading::load_external_cli_sessions_grouped_multi;
 use loading::{build_messages_preview, build_search_index, crashed_sessions_from_all_sessions};
 pub use loading::{
     invalidate_session_list_cache, load_cached_sessions_grouped, load_servers, load_sessions,
@@ -525,6 +525,16 @@ impl SessionPicker {
             .filter_map(|session_ref| self.session_by_ref(*session_ref))
     }
 
+    /// Test-only accessor: the source classification of every currently visible
+    /// session. Used by onboarding tests to assert the combined external-CLI
+    /// picker surfaces both Codex and Claude Code transcripts.
+    #[cfg(test)]
+    pub(crate) fn visible_session_iter_for_test(
+        &self,
+    ) -> impl Iterator<Item = &SessionInfo> + '_ {
+        self.visible_session_iter()
+    }
+
     fn load_preview_for_target(
         resume_target: ResumeTarget,
         external_path: Option<String>,
@@ -552,10 +562,6 @@ impl SessionPicker {
             ResumeTarget::OpenCodeSession { .. } => external_path.as_deref().and_then(|path| {
                 loading::load_opencode_preview_from_path(std::path::Path::new(path))
             }),
-            // Foreign providers: we don't have a generic preview loader
-            // (each provider has its own transcript format). The TUI
-            // falls back to the metadata-only preview that the
-            // SessionInfo already carries.
             ResumeTarget::ForeignSession { .. } => None,
         }
     }
diff --git a/crates/jcode-tui/src/tui/session_picker/filter.rs b/crates/jcode-tui/src/tui/session_picker/filter.rs
index f75349757..736ec54d0 100644
--- a/crates/jcode-tui/src/tui/session_picker/filter.rs
+++ b/crates/jcode-tui/src/tui/session_picker/filter.rs
@@ -39,8 +39,11 @@ impl SessionPicker {
 
         let can_narrow_cached = !self.cached_search_query.is_empty()
             && normalized.starts_with(&self.cached_search_query);
+        // When narrowing, reuse the previous match set in place via mem::take
+        // instead of cloning it into `candidates` and then cloning the new
+        // matches back into the cache (two full-list clones per keystroke).
         let candidates = if can_narrow_cached {
-            self.cached_search_refs.clone()
+            std::mem::take(&mut self.cached_search_refs)
         } else {
             self.all_session_refs()
         };
@@ -146,6 +149,9 @@ impl SessionPicker {
             SessionFilterMode::Codex => Self::session_is_codex(session),
             SessionFilterMode::Pi => Self::session_is_pi(session),
             SessionFilterMode::OpenCode => Self::session_is_open_code(session),
+            SessionFilterMode::ExternalClis => {
+                Self::session_is_codex(session) || Self::session_is_claude_code(session)
+            }
         }
     }
 
@@ -220,43 +226,41 @@ impl SessionPicker {
         }
 
         if !self.all_server_groups.is_empty() {
-            let grouped_sections: Vec<(String, String, String, Vec<SessionRef>)> = self
-                .all_server_groups
-                .iter()
-                .enumerate()
-                .filter_map(|(group_idx, group)| {
-                    let visible: Vec<SessionRef> = filtered_refs
-                        .iter()
-                        .copied()
-                        .filter(|session_ref| match session_ref {
-                            SessionRef::Group {
-                                group_idx: ref_group_idx,
-                                session_idx,
-                            } => {
-                                if *ref_group_idx != group_idx {
-                                    return false;
-                                }
-                                group
-                                    .sessions
-                                    .get(*session_idx)
-                                    .is_some_and(|session| !saved_ids.contains(&session.id))
-                            }
-                            _ => false,
-                        })
-                        .collect();
-
-                    if visible.is_empty() {
-                        None
-                    } else {
-                        Some((
-                            group.name.clone(),
-                            group.icon.clone(),
-                            group.version.clone(),
-                            visible,
-                        ))
-                    }
-                })
-                .collect();
+            // Partition the filtered refs by group in a single pass instead of
+            // rescanning every filtered ref once per group. The previous code
+            // was O(groups * filtered_refs); with many remote/server groups and
+            // many sessions this scaled poorly on every search keystroke. One
+            // bucketing pass is O(filtered_refs), then emitting is O(groups).
+            let mut group_buckets: Vec<Vec<SessionRef>> =
+                vec![Vec::new(); self.all_server_groups.len()];
+            for session_ref in filtered_refs.iter().copied() {
+                if let SessionRef::Group {
+                    group_idx,
+                    session_idx,
+                } = session_ref
+                    && let Some(group) = self.all_server_groups.get(group_idx)
+                    && group
+                        .sessions
+                        .get(session_idx)
+                        .is_some_and(|session| !saved_ids.contains(&session.id))
+                {
+                    group_buckets[group_idx].push(session_ref);
+                }
+            }
+
+            let mut grouped_sections: Vec<(String, String, String, Vec<SessionRef>)> = Vec::new();
+            for (group_idx, group) in self.all_server_groups.iter().enumerate() {
+                let visible = std::mem::take(&mut group_buckets[group_idx]);
+                if visible.is_empty() {
+                    continue;
+                }
+                grouped_sections.push((
+                    group.name.clone(),
+                    group.icon.clone(),
+                    group.version.clone(),
+                    visible,
+                ));
+            }
 
             for (name, icon, version, visible) in grouped_sections {
                 self.items.push(PickerItem::ServerHeader {
diff --git a/crates/jcode-tui/src/tui/session_picker/loading.rs b/crates/jcode-tui/src/tui/session_picker/loading.rs
index 6e99abae0..670bb1b94 100644
--- a/crates/jcode-tui/src/tui/session_picker/loading.rs
+++ b/crates/jcode-tui/src/tui/session_picker/loading.rs
@@ -2413,6 +2413,14 @@ pub(crate) fn load_external_cli_sessions_grouped(
     (Vec::new(), sessions)
 }
 
+pub(crate) fn load_external_cli_sessions_grouped_multi(
+    _clis: &[crate::tui::app::onboarding_flow::ExternalCli],
+) -> (Vec<ServerGroup>, Vec<SessionInfo>) {
+    let scan_limit = session_scan_limit();
+    let sessions = load_external_casr_sessions(scan_limit);
+    (Vec::new(), sessions)
+}
+
 #[cfg(test)]
 #[path = "loading_tests.rs"]
 mod tests;
diff --git a/crates/jcode-tui/src/tui/session_picker/loading_tests.rs b/crates/jcode-tui/src/tui/session_picker/loading_tests.rs
index 7ad9e69f6..5ffbe0e9e 100644
--- a/crates/jcode-tui/src/tui/session_picker/loading_tests.rs
+++ b/crates/jcode-tui/src/tui/session_picker/loading_tests.rs
@@ -318,6 +318,88 @@ fn load_codex_preview_preserves_blank_line_between_tool_transcript_and_followup_
     );
 }
 
+#[test]
+fn load_codex_preview_reads_only_tail_of_large_transcript() {
+    // A transcript far larger than the tail cap should still produce a preview
+    // of the most-recent messages, parsed from only the tail slice. This is the
+    // regression guard for the picker-navigation lag: previews must not depend
+    // on parsing the whole (multi-MB) file.
+    let temp = tempfile::tempdir().expect("temp dir");
+    let transcript_path = temp.path().join("rollout-big.jsonl");
+
+    let mut contents = String::new();
+    // session_meta header line (always skipped).
+    contents.push_str(
+        "{\"timestamp\":\"2026-04-10T19:05:54.536Z\",\"type\":\"session_meta\",\"payload\":{\"id\":\"019d-big\"}}\n",
+    );
+    // Padding messages near the head that must NOT appear in the preview once
+    // the file exceeds the tail cap.
+    for i in 0..50_000 {
+        contents.push_str(&format!(
+            "{{\"type\":\"response_item\",\"payload\":{{\"type\":\"message\",\"role\":\"assistant\",\"content\":[{{\"type\":\"output_text\",\"text\":\"old padding message {i} aaaaaaaaaaaaaaaaaaaaaaaaaaaaaa\"}}]}}}}\n",
+        ));
+    }
+    assert!(
+        contents.len() as u64 > EXTERNAL_PREVIEW_TAIL_BYTES,
+        "test transcript must exceed the tail cap"
+    );
+    // Distinctive recent messages at the very end.
+    contents.push_str(
+        "{\"type\":\"response_item\",\"payload\":{\"type\":\"message\",\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"RECENT_USER_MARKER\"}]}}\n",
+    );
+    contents.push_str(
+        "{\"type\":\"response_item\",\"payload\":{\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"output_text\",\"text\":\"RECENT_ASSISTANT_MARKER\"}]}}\n",
+    );
+    std::fs::write(&transcript_path, &contents).expect("write big transcript");
+
+    let preview = load_codex_preview_from_path(&transcript_path).expect("preview");
+    // Preview is capped at 20 messages.
+    assert!(preview.len() <= 20, "preview should be capped, got {}", preview.len());
+    // The most-recent markers must be present.
+    let last_two = &preview[preview.len().saturating_sub(2)..];
+    assert!(last_two.iter().any(|m| m.content.contains("RECENT_USER_MARKER")));
+    assert!(last_two.iter().any(|m| m.content.contains("RECENT_ASSISTANT_MARKER")));
+    // The head padding must have been skipped (not parsed from the tail slice).
+    assert!(
+        !preview.iter().any(|m| m.content.contains("old padding message 0 ")),
+        "head messages should not appear when only the tail is read"
+    );
+}
+
+#[test]
+fn load_claude_code_preview_reads_only_tail_of_large_transcript() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let transcript_path = temp.path().join("claude-big.jsonl");
+
+    let mut contents = String::new();
+    for i in 0..50_000 {
+        contents.push_str(&format!(
+            "{{\"type\":\"assistant\",\"uuid\":\"a{i}\",\"message\":{{\"role\":\"assistant\",\"content\":[{{\"type\":\"text\",\"text\":\"old padding message {i} bbbbbbbbbbbbbbbbbbbbbbbbbbbbbb\"}}]}}}}\n",
+        ));
+    }
+    assert!(
+        contents.len() as u64 > EXTERNAL_PREVIEW_TAIL_BYTES,
+        "test transcript must exceed the tail cap"
+    );
+    contents.push_str(
+        "{\"type\":\"user\",\"uuid\":\"u_last\",\"message\":{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"RECENT_USER_MARKER\"}]}}\n",
+    );
+    contents.push_str(
+        "{\"type\":\"assistant\",\"uuid\":\"a_last\",\"message\":{\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"RECENT_ASSISTANT_MARKER\"}]}}\n",
+    );
+    std::fs::write(&transcript_path, &contents).expect("write big transcript");
+
+    let preview = load_claude_code_preview_from_path(&transcript_path).expect("preview");
+    assert!(preview.len() <= 20, "preview should be capped, got {}", preview.len());
+    let last_two = &preview[preview.len().saturating_sub(2)..];
+    assert!(last_two.iter().any(|m| m.content.contains("RECENT_USER_MARKER")));
+    assert!(last_two.iter().any(|m| m.content.contains("RECENT_ASSISTANT_MARKER")));
+    assert!(
+        !preview.iter().any(|m| m.content.contains("old padding message 0 ")),
+        "head messages should not appear when only the tail is read"
+    );
+}
+
 #[test]
 fn load_sessions_prefers_custom_title_over_generated_title() {
     let _env_lock = crate::storage::lock_test_env();
diff --git a/crates/jcode-tui/src/tui/ui.rs b/crates/jcode-tui/src/tui/ui.rs
index 30b2fe5f9..96742bac0 100644
--- a/crates/jcode-tui/src/tui/ui.rs
+++ b/crates/jcode-tui/src/tui/ui.rs
@@ -148,8 +148,8 @@ use memory_ui::{group_into_tiles, render_memory_tiles, split_by_display_width};
 use messages::get_cached_message_lines;
 #[cfg_attr(test, allow(unused_imports))]
 pub(crate) use messages::{
-    render_assistant_message, render_background_task_message, render_swarm_message,
-    render_system_message, render_tool_message, render_usage_message,
+    render_assistant_message, render_background_task_message, render_reasoning_message,
+    render_swarm_message, render_system_message, render_tool_message, render_usage_message,
 };
 pub use pinned_ui::{
     SidePanelDebugStats, SidePanelMermaidProbe, SidePanelMermaidProbeRect,
@@ -1632,9 +1632,17 @@ pub(crate) fn copy_pane_vertical_edge_point(
     let zone = edge_autoscroll_zone_rows(area.height);
     let top_trigger = area.y.saturating_add(zone);
     let bottom_trigger = last_row.saturating_sub(zone);
-    let (edge_row, upward) = if row <= top_trigger {
+    // Only engage the hot zone when there is actually more transcript to pull in
+    // that direction. Otherwise dragging into the bottom band while the view is
+    // already pinned to the end (the common case) would snap the selection to the
+    // last visible line and fight precise highlighting of the bottom rows. When
+    // there is nothing to scroll, fall through (`None`) so the caller extends the
+    // selection to the exact cell under the cursor instead.
+    let can_scroll_up = snapshot.scroll > 0;
+    let can_scroll_down = snapshot.visible_end < snapshot.wrapped_plain_line_count();
+    let (edge_row, upward) = if row <= top_trigger && can_scroll_up {
         (area.y, true)
-    } else if row >= bottom_trigger {
+    } else if row >= bottom_trigger && can_scroll_down {
         (last_row, false)
     } else {
         return None;
@@ -1645,6 +1653,77 @@ pub(crate) fn copy_pane_vertical_edge_point(
     copy_point_from_snapshot(&snapshot, clamped_col, edge_row).map(|point| (point, upward))
 }
 
+/// Resolve the selection point for a drag at `(column, row)`, clamping vertical
+/// overshoot to the nearest in-bounds line edge.
+///
+/// Terminals report a drag that "leaves" the pane on the boundary row, but a
+/// drag *into the empty space below the last content line* (common with short
+/// transcripts that leave blank rows underneath) lands on a row that maps to no
+/// line at all, so `copy_point_from_screen` returns `None`. Native terminal and
+/// browser selection treat that as "select through the end of the last line".
+/// This mirrors that: dragging below the last visible line snaps to the end of
+/// that line, and dragging above the first visible line snaps to its start, so
+/// the boundary line is fully covered even when there is nothing more to scroll.
+pub(crate) fn copy_pane_drag_point(
+    pane: crate::tui::CopySelectionPane,
+    column: u16,
+    row: u16,
+) -> Option<crate::tui::CopySelectionPoint> {
+    let snapshot = copy_snapshot_for_pane(pane)?;
+    let area = snapshot.content_area;
+    if area.width == 0 || area.height == 0 {
+        return None;
+    }
+
+    // A direct hit on a real line wins: precise per-cell selection.
+    if let Some(point) = copy_point_from_snapshot(&snapshot, column, row) {
+        return Some(point);
+    }
+
+    let line_count = snapshot.wrapped_plain_line_count();
+    if line_count == 0 {
+        return None;
+    }
+    let last_line = line_count.saturating_sub(1);
+    let last_visible_line = snapshot.visible_end.saturating_sub(1).min(last_line);
+    let first_visible_line = snapshot.scroll.min(last_line);
+
+    let last_row = area.y.saturating_add(area.height).saturating_sub(1);
+    let clamped_col = column.clamp(area.x, area.x.saturating_add(area.width).saturating_sub(1));
+
+    // Below the visible content: snap to the end of the last visible line.
+    if row >= last_row {
+        let text = snapshot.wrapped_plain_line(last_visible_line).unwrap_or("");
+        return Some(crate::tui::CopySelectionPoint {
+            pane,
+            abs_line: last_visible_line,
+            column: line_display_width(text),
+        });
+    }
+
+    // Above the visible content: snap to the start of the first visible line.
+    if row <= area.y {
+        return Some(crate::tui::CopySelectionPoint {
+            pane,
+            abs_line: first_visible_line,
+            column: snapshot.wrapped_copy_offset(first_visible_line).unwrap_or(0),
+        });
+    }
+
+    // Interior row that maps to no line (e.g. a blank gap row between/after
+    // content within the visible band): fall back to the boundary-clamped point.
+    copy_point_from_snapshot(
+        &snapshot,
+        clamped_col,
+        row.clamp(area.y, last_row),
+    )
+    .or(Some(crate::tui::CopySelectionPoint {
+        pane,
+        abs_line: last_visible_line,
+        column: line_display_width(snapshot.wrapped_plain_line(last_visible_line).unwrap_or("")),
+    }))
+}
+
 /// Edge point for tick-driven continuous auto-scroll, where there is no live
 /// mouse position. Uses the top/bottom boundary row of the pane and its left
 /// content column so the selection keeps extending to the freshly revealed line.
@@ -1805,6 +1884,72 @@ pub(crate) fn copy_selection_text(range: crate::tui::CopySelectionRange) -> Opti
     Some(out)
 }
 
+/// Compute `(char_count, line_count)` for the current copy selection without
+/// allocating the full joined selection string. Mirrors `copy_selection_text`
+/// so the status line "N chars · M lines" matches what would be copied, but is
+/// allocation-free so it can run cheaply on every render frame / drag move.
+pub(crate) fn copy_selection_metrics(
+    range: crate::tui::CopySelectionRange,
+) -> Option<(usize, usize)> {
+    if range.start.pane != range.end.pane {
+        return None;
+    }
+    let snapshot = copy_snapshot_for_pane(range.start.pane)?;
+    let (start, end) =
+        if (range.start.abs_line, range.start.column) <= (range.end.abs_line, range.end.column) {
+            (range.start, range.end)
+        } else {
+            (range.end, range.start)
+        };
+
+    if start.abs_line >= snapshot.wrapped_plain_line_count()
+        || end.abs_line >= snapshot.wrapped_plain_line_count()
+    {
+        return None;
+    }
+
+    if let Some(metrics) =
+        copy_selection::copy_selection_metrics_from_raw_lines(&snapshot, start, end)
+    {
+        return Some(metrics);
+    }
+
+    let mut chars = 0usize;
+    let mut lines = 0usize;
+    for abs_line in start.abs_line..=end.abs_line {
+        if abs_line > start.abs_line {
+            chars += 1; // joining '\n'
+        }
+        lines += 1;
+        let text = snapshot.wrapped_plain_line(abs_line)?;
+        if abs_line != start.abs_line && abs_line != end.abs_line {
+            let copy_start = snapshot.wrapped_copy_offset(abs_line).unwrap_or(0);
+            if copy_start == 0 {
+                chars += text.chars().count();
+                continue;
+            }
+        }
+        let line_width = line_display_width(&text);
+        let copy_start = snapshot.wrapped_copy_offset(abs_line).unwrap_or(0);
+        let start_col = if abs_line == start.abs_line {
+            clamp_display_col(&text, start.column).max(copy_start)
+        } else {
+            copy_start
+        };
+        let end_col = if abs_line == end.abs_line {
+            clamp_display_col(&text, end.column).max(copy_start)
+        } else {
+            line_width
+        };
+        if end_col < start_col {
+            continue;
+        }
+        chars += display_col_slice(&text, start_col, end_col).chars().count();
+    }
+
+    Some((chars, lines.max(1)))
+}
+
 pub(crate) fn link_target_from_screen(column: u16, row: u16) -> Option<String> {
     let point = copy_point_from_screen(column, row)?;
     let snapshot = copy_snapshot_for_pane(point.pane)?;
diff --git a/crates/jcode-tui/src/tui/ui/copy_selection.rs b/crates/jcode-tui/src/tui/ui/copy_selection.rs
index 31b5e41d6..2ec4020bd 100644
--- a/crates/jcode-tui/src/tui/ui/copy_selection.rs
+++ b/crates/jcode-tui/src/tui/ui/copy_selection.rs
@@ -102,6 +102,59 @@ pub(super) fn copy_selection_text_from_raw_lines(
     Some(out)
 }
 
+/// Selection metrics (character count and line count) for the raw-lines path,
+/// computed without allocating the full joined selection string. Mirrors the
+/// slicing in [`copy_selection_text_from_raw_lines`] exactly so the displayed
+/// "N chars · M lines" matches what would actually be copied.
+pub(super) fn copy_selection_metrics_from_raw_lines(
+    snapshot: &CopyViewportSnapshot,
+    start: crate::tui::CopySelectionPoint,
+    end: crate::tui::CopySelectionPoint,
+) -> Option<(usize, usize)> {
+    if snapshot.raw_plain_line_count() == 0 || snapshot.wrapped_line_map(start.abs_line).is_none() {
+        return None;
+    }
+
+    let start = raw_selection_point(snapshot, start)?;
+    let end = raw_selection_point(snapshot, end)?;
+    if start.raw_line >= snapshot.raw_plain_line_count()
+        || end.raw_line >= snapshot.raw_plain_line_count()
+    {
+        return None;
+    }
+
+    let mut chars = 0usize;
+    let mut lines = 0usize;
+    for raw_line in start.raw_line..=end.raw_line {
+        if raw_line > start.raw_line {
+            chars += 1; // the joining '\n'
+        }
+        lines += 1;
+        let text = snapshot.raw_plain_line(raw_line)?;
+        if raw_line != start.raw_line && raw_line != end.raw_line {
+            chars += text.chars().count();
+            continue;
+        }
+        let line_width = line_display_width(&text);
+        let start_col = if raw_line == start.raw_line {
+            clamp_display_col(&text, start.column)
+        } else {
+            0
+        };
+        let end_col = if raw_line == end.raw_line {
+            clamp_display_col(&text, end.column)
+        } else {
+            line_width
+        };
+        if end_col < start_col {
+            continue;
+        }
+        chars += display_col_slice(&text, start_col, end_col).chars().count();
+    }
+
+    Some((chars, lines.max(1)))
+}
+
 pub(super) fn link_target_from_snapshot(
     snapshot: &CopyViewportSnapshot,
     point: crate::tui::CopySelectionPoint,
diff --git a/crates/jcode-tui/src/tui/ui_header.rs b/crates/jcode-tui/src/tui/ui_header.rs
index f780dceb7..888d593e6 100644
--- a/crates/jcode-tui/src/tui/ui_header.rs
+++ b/crates/jcode-tui/src/tui/ui_header.rs
@@ -45,9 +45,22 @@ pub(crate) fn capitalize(s: &str) -> String {
     }
 }
 
-fn format_model_name(short: &str) -> String {
+fn format_model_name(short: &str, provider_name: &str) -> String {
     if short.contains('/') {
-        return format!("OpenRouter: {}", short);
+        // Slashed model ids (e.g. `nvidia/nemotron-...`) are served by the
+        // OpenRouter slot, which also fronts direct OpenAI-compatible profiles
+        // such as NVIDIA NIM or DeepSeek. Label the line with the active
+        // provider's display name instead of hard-coding "OpenRouter" so the
+        // header matches the profile the user actually selected.
+        let label = {
+            let trimmed = provider_name.trim();
+            if trimmed.is_empty() {
+                "OpenRouter".to_string()
+            } else {
+                trimmed.to_string()
+            }
+        };
+        return format!("{}: {}", label, short);
     }
     if short.contains("opus") {
         if short.contains("4.5") {
@@ -389,7 +402,7 @@ pub(super) fn build_persistent_header(app: &dyn TuiState, width: u16) -> Vec<Lin
     let short_model = shorten_model_name(&model);
     let icon = connection_type_icon(app.connection_type().as_deref())
         .unwrap_or_else(|| crate::id::session_icon(&session_name));
-    let nice_model = format_model_name(&short_model);
+    let nice_model = format_model_name(&short_model, &app.provider_name());
     let build_info = binary_age().unwrap_or_else(|| "unknown".to_string());
     let align = Alignment::Center;
     let mut lines: Vec<Line> = Vec::new();
@@ -1028,4 +1041,26 @@ mod tests {
         let line = build_auth_status_line(&AuthStatus::default(), 120);
         assert!(line.spans.is_empty(), "line should be empty: {line:?}");
     }
+
+    #[test]
+    fn format_model_name_labels_slashed_models_with_active_provider() {
+        // Regression for issue #329: a NVIDIA NIM model must be labeled with the
+        // active provider's display name, not the fixed "OpenRouter" aggregator.
+        assert_eq!(
+            format_model_name("nvidia/nemotron-3-super-120b-a12b", "NVIDIA NIM"),
+            "NVIDIA NIM: nvidia/nemotron-3-super-120b-a12b"
+        );
+        // The public aggregator still reads "OpenRouter".
+        assert_eq!(
+            format_model_name("anthropic/claude-sonnet-4", "OpenRouter"),
+            "OpenRouter: anthropic/claude-sonnet-4"
+        );
+        // Missing provider name falls back to "OpenRouter" rather than an empty label.
+        assert_eq!(
+            format_model_name("deepseek/deepseek-chat", ""),
+            "OpenRouter: deepseek/deepseek-chat"
+        );
+        // Non-slashed models are unaffected by the provider label.
+        assert_eq!(format_model_name("claude-opus-4-6", "OpenRouter"), "Claude Opus");
+    }
 }
diff --git a/crates/jcode-tui/src/tui/ui_messages.rs b/crates/jcode-tui/src/tui/ui_messages.rs
index 2b94ae673..80b8f1b11 100644
--- a/crates/jcode-tui/src/tui/ui_messages.rs
+++ b/crates/jcode-tui/src/tui/ui_messages.rs
@@ -72,6 +72,23 @@ pub(crate) fn render_assistant_message(
     lines
 }
 
+/// Render a collapsed/collapsing reasoning trace ("current" mode). The content is
+/// sentinel-wrapped dim+italic markup (reasoning lines and/or a `▸ thought for Xs`
+/// summary), so it reuses the standard markdown path that styles those runs dim.
+pub(crate) fn render_reasoning_message(
+    msg: &DisplayMessage,
+    width: u16,
+    _diff_mode: crate::config::DiffDisplayMode,
+) -> Vec<Line<'static>> {
+    let centered = markdown::center_code_blocks();
+    let wrap_width = centered_wrap_width(width, centered, 96);
+    let mut lines = markdown::render_markdown_with_width(&msg.content, Some(wrap_width));
+    if centered {
+        left_pad_lines_for_centered_mode(&mut lines, width);
+    }
+    lines
+}
+
 fn render_assistant_tool_call_lines(
     tool_calls: &[String],
     width: usize,
diff --git a/crates/jcode-tui/src/tui/ui_prepare.rs b/crates/jcode-tui/src/tui/ui_prepare.rs
index 07160f396..4c73baa07 100644
--- a/crates/jcode-tui/src/tui/ui_prepare.rs
+++ b/crates/jcode-tui/src/tui/ui_prepare.rs
@@ -202,6 +202,46 @@ fn is_error_copy_content(content: &str) -> bool {
     trimmed.starts_with("Error:") || trimmed.starts_with("error:") || trimmed.starts_with("Failed:")
 }
 
+/// Build the image regions for an image/mermaid placeholder in `wrapped_lines`,
+/// where each placeholder "owns" the run of blank lines that follow it.
+///
+/// Done in a single reverse pass that precomputes, for every line, the length
+/// of the blank run starting at that line. The previous implementation scanned
+/// forward through the trailing blanks for every placeholder, which is O(L^2)
+/// when a message has many placeholders each followed by long blank runs.
+fn compute_image_regions(wrapped_lines: &[ratatui::text::Line<'static>]) -> Vec<ImageRegion> {
+    fn is_blank_line(line: &ratatui::text::Line<'static>) -> bool {
+        line.spans.is_empty()
+            || (line.spans.len() == 1 && line.spans[0].content.is_empty())
+    }
+
+    let len = wrapped_lines.len();
+    // blank_run[i] = number of consecutive blank lines starting at index i.
+    let mut blank_run = vec![0usize; len + 1];
+    for idx in (0..len).rev() {
+        blank_run[idx] = if is_blank_line(&wrapped_lines[idx]) {
+            blank_run[idx + 1] + 1
+        } else {
+            0
+        };
+    }
+
+    let mut image_regions = Vec::new();
+    for (idx, line) in wrapped_lines.iter().enumerate() {
+        if let Some(hash) = super::super::mermaid::parse_image_placeholder(line) {
+            // The placeholder line plus the blank run immediately after it.
+            let height = (1 + blank_run[idx + 1]).min(u16::MAX as usize) as u16;
+            image_regions.push(ImageRegion {
+                abs_line_idx: idx,
+                end_line: idx + height as usize,
+                hash,
+                height,
+            });
+        }
+    }
+    image_regions
+}
+
 fn error_copy_target(content: &str, rendered_line_count: usize) -> Option<RawCopyTarget> {
     copy_target_for_kind(CopyTargetKind::Error, content, rendered_line_count)
 }
@@ -752,10 +792,13 @@ pub(super) fn prepare_body_incremental(
     let pending_count = input_ui::pending_prompt_count(app);
     let prompt_number_offset = app.compacted_hidden_user_prompts();
 
-    let mut prompt_num = messages[..prev_msg_count]
-        .iter()
-        .filter(|m| m.effective_role() == "user")
-        .count();
+    // The number of user prompts already rendered equals the number of cached
+    // user prompt texts. Re-counting `messages[..prev_msg_count]` here on every
+    // incremental append rescans the whole prior transcript, making a session
+    // that grows one message at a time O(n^2). `prev.user_prompt_texts` is
+    // extended in lockstep with each rendered user message, so its length is the
+    // exact prior prompt count.
+    let mut prompt_num = prev.user_prompt_texts.len();
 
     let mut new_lines: Vec<Line> = Vec::new();
     let mut new_user_line_indices: Vec<usize> = Vec::new();
@@ -916,6 +959,20 @@ pub(super) fn prepare_body_incremental(
                     new_line_copy_offsets.push(0);
                 }
             }
+            "reasoning" => {
+                let content_width = width.saturating_sub(4);
+                let cached = get_cached_message_lines(
+                    msg,
+                    content_width,
+                    app.diff_mode(),
+                    render_reasoning_message,
+                );
+                for line in cached {
+                    new_lines.push(align_if_unset(line, align));
+                    new_line_raw_overrides.push(None);
+                    new_line_copy_offsets.push(0);
+                }
+            }
             "background_task" => {
                 let content_width = width.saturating_sub(4);
                 let cached = get_cached_message_lines(
@@ -1386,6 +1443,20 @@ pub(super) fn prepare_body(
                     line_copy_offsets.push(0);
                 }
             }
+            "reasoning" => {
+                let content_width = width.saturating_sub(4);
+                let cached = get_cached_message_lines(
+                    msg,
+                    content_width,
+                    app.diff_mode(),
+                    render_reasoning_message,
+                );
+                for line in cached {
+                    lines.push(align_if_unset(line, align));
+                    line_raw_overrides.push(None);
+                    line_copy_offsets.push(0);
+                }
+            }
             "background_task" => {
                 let content_width = width.saturating_sub(4);
                 let cached = get_cached_message_lines(
@@ -1614,27 +1685,7 @@ fn wrap_lines(
         wrapped_idx += count;
     }
 
-    let mut image_regions = Vec::new();
-    for (idx, line) in wrapped_lines.iter().enumerate() {
-        if let Some(hash) = super::super::mermaid::parse_image_placeholder(line) {
-            let mut height = 1u16;
-            for subsequent in wrapped_lines.iter().skip(idx + 1) {
-                if subsequent.spans.is_empty()
-                    || (subsequent.spans.len() == 1 && subsequent.spans[0].content.is_empty())
-                {
-                    height += 1;
-                } else {
-                    break;
-                }
-            }
-            image_regions.push(ImageRegion {
-                abs_line_idx: idx,
-                end_line: idx + height as usize,
-                hash,
-                height,
-            });
-        }
-    }
+    let image_regions = compute_image_regions(&wrapped_lines);
 
     let wrapped_plain_lines = Arc::new(wrapped_lines.iter().map(ui::line_plain_text).collect());
 
@@ -1733,27 +1784,7 @@ fn wrap_lines_with_map(
     }
     raw_to_wrapped.push(wrapped_idx);
 
-    let mut image_regions = Vec::new();
-    for (idx, line) in wrapped_lines.iter().enumerate() {
-        if let Some(hash) = super::super::mermaid::parse_image_placeholder(line) {
-            let mut height = 1u16;
-            for subsequent in wrapped_lines.iter().skip(idx + 1) {
-                if subsequent.spans.is_empty()
-                    || (subsequent.spans.len() == 1 && subsequent.spans[0].content.is_empty())
-                {
-                    height += 1;
-                } else {
-                    break;
-                }
-            }
-            image_regions.push(ImageRegion {
-                abs_line_idx: idx,
-                end_line: idx + height as usize,
-                hash,
-                height,
-            });
-        }
-    }
+    let image_regions = compute_image_regions(&wrapped_lines);
 
     let mut edit_tool_ranges = Vec::new();
     for (msg_idx, file_path, raw_start, raw_end, expandable) in edit_ranges {
diff --git a/crates/jcode-tui/src/tui/ui_tests/prepare.rs b/crates/jcode-tui/src/tui/ui_tests/prepare.rs
index bcc2f7e08..8fc1e8838 100644
--- a/crates/jcode-tui/src/tui/ui_tests/prepare.rs
+++ b/crates/jcode-tui/src/tui/ui_tests/prepare.rs
@@ -772,3 +772,60 @@ fn test_render_tool_message_batch_subcall_lines_alignment_unset() {
     }
     crate::tui::markdown::set_center_code_blocks(false);
 }
+
+#[test]
+fn test_prepare_messages_renders_reasoning_role_dim_italic_without_sentinel() {
+    let _guard = crate::storage::lock_test_env();
+    clear_test_render_state_for_tests();
+
+    // A collapsing reasoning message carries sentinel-wrapped dim/italic markup.
+    let mut content = String::new();
+    content.push_str(&jcode_tui_markdown::reasoning_line_markup("weighing the options"));
+    content.push_str(&jcode_tui_markdown::reasoning_line_markup("▸ thought for 3s"));
+
+    let state = TestState {
+        display_messages: vec![
+            DisplayMessage::user("hi"),
+            DisplayMessage::reasoning(content),
+        ],
+        ..Default::default()
+    };
+
+    let prepared = prepare::prepare_messages(&state, 100, 30);
+    let lines = prepared.materialize_all_lines();
+
+    // The visible reasoning body is present, dim+italic, and sentinel-free.
+    let body = lines
+        .iter()
+        .find(|l| {
+            let joined: String = l.spans.iter().map(|s| s.content.as_ref()).collect();
+            joined.contains("weighing the options")
+        })
+        .expect("reasoning body line present");
+    let rendered: String = body.spans.iter().map(|s| s.content.as_ref()).collect();
+    assert!(
+        !rendered.contains(jcode_tui_markdown::REASONING_SENTINEL),
+        "sentinel must be stripped from visible reasoning: {rendered:?}"
+    );
+    let span = body
+        .spans
+        .iter()
+        .find(|s| s.content.as_ref().contains("weighing"))
+        .expect("body span");
+    assert!(
+        span.style
+            .add_modifier
+            .contains(ratatui::style::Modifier::ITALIC),
+        "reasoning body should be italic: {:?}",
+        span.style
+    );
+
+    // The summary line is present too.
+    assert!(
+        lines.iter().any(|l| {
+            let joined: String = l.spans.iter().map(|s| s.content.as_ref()).collect();
+            joined.contains("thought for 3s")
+        }),
+        "summary line should render"
+    );
+}
diff --git a/docs/GMAIL_COMPOSIO_BACKEND.md b/docs/GMAIL_COMPOSIO_BACKEND.md
new file mode 100644
index 000000000..8a3164221
--- /dev/null
+++ b/docs/GMAIL_COMPOSIO_BACKEND.md
@@ -0,0 +1,96 @@
+# Gmail Tool: Composio Managed Backend
+
+The native `gmail` tool can source credentials and transport from one of two
+backends. The tool interface, confirmation gating, access-tier logic, and
+token-lean output formatting are identical across backends; only the
+auth/transport layer changes.
+
+## Backends
+
+| Backend | Auth | Pros | Cons |
+|---|---|---|---|
+| `direct` (default) | Local Google OAuth tokens (`jcode login google`) | No third party in the loop | Unverified-app warning; 7-day refresh-token expiry in Google "Testing" mode |
+| `composio` | Composio-managed OAuth (Google-verified app) | No unverified-app warning, no 7-day expiry, no per-user Google Cloud project | Composio brokers Gmail token custody; external dependency/cost |
+
+Both backends call the *same* Gmail REST endpoints
+(`https://gmail.googleapis.com/gmail/v1/users/me/...`). The Composio backend
+routes those calls through Composio's
+[`proxy-execute`](https://docs.composio.dev/reference/api-reference/tools/postToolsExecuteProxy)
+endpoint, which attaches the managed Gmail credentials. Because the upstream
+response shape is unchanged, all existing typed parsing and output formatting
+is reused.
+
+## Selecting the backend
+
+The backend is resolved from environment at `GmailClient::new()`:
+
+- `JCODE_GMAIL_BACKEND=direct` (or unset) -> direct Google backend.
+- `JCODE_GMAIL_BACKEND=composio` -> Composio backend (requires `COMPOSIO_API_KEY`).
+
+If `composio` is requested but `COMPOSIO_API_KEY` is missing, jcode warns and
+falls back to `direct`.
+
+### Composio environment variables
+
+| Variable | Required | Description |
+|---|---|---|
+| `COMPOSIO_API_KEY` | Yes | Project API key from <https://platform.composio.dev> |
+| `COMPOSIO_BASE_URL` | No | Override API base (default `https://backend.composio.dev/api/v3.1`) |
+| `COMPOSIO_GMAIL_AUTH_CONFIG_ID` | For `connect` | Gmail auth config id (`ac_...`) from the Composio dashboard. Defines the OAuth blueprint/scopes used by the connect flow. |
+| `COMPOSIO_GMAIL_CONNECTED_ACCOUNT_ID` | No | Pin a specific connected account (`ca_...`). Normally set automatically after `connect`. |
+| `COMPOSIO_GMAIL_USER_ID` / `COMPOSIO_USER_ID` | No | End-user id for multi-user connected accounts (defaults to `default`) |
+
+## Connecting a Gmail account (in-agent OAuth)
+
+Once `COMPOSIO_API_KEY` and `COMPOSIO_GMAIL_AUTH_CONFIG_ID` are set, the user
+(or the agent) runs the gmail tool with `action: "connect"`:
+
+1. jcode calls Composio's `POST /connected_accounts/link` (hosted "Connect
+   Link" flow) to start an OAuth session.
+2. The returned `redirect_url` is opened in the system browser (printed to
+   stderr as a fallback, e.g. over SSH).
+3. The user approves Gmail access on Google's consent screen. Because Composio
+   owns a Google-verified app, there is no "unverified app" warning.
+4. jcode polls `GET /connected_accounts/{id}` until the connection is `ACTIVE`,
+   then persists it to `~/.jcode/composio_gmail.json`.
+
+Future sessions load the persisted `connected_account_id`, so the connect step
+is a one-time action per account. Tool calls before a connection exists return
+a hint telling the agent to run `action: "connect"` first.
+
+> Note: Composio is retiring `initiate()` for managed OAuth in favor of the
+> Connect Link `link()` flow used here, so this path is the supported one going
+> forward.
+
+## One-time Composio setup
+
+1. Sign in at <https://platform.composio.dev> and copy your project API key.
+2. Connect a Gmail account (Composio's hosted OAuth, no unverified-app warning).
+   Note the resulting `connected_account_id` if you want to pin it.
+3. Export the variables:
+   ```bash
+   export JCODE_GMAIL_BACKEND=composio
+   export COMPOSIO_API_KEY="ck_..."
+   # optional:
+   export COMPOSIO_GMAIL_CONNECTED_ACCOUNT_ID="ca_..."
+   export COMPOSIO_GMAIL_USER_ID="me"
+   ```
+4. Ensure the `gmail` tool is enabled in `config.toml`:
+   ```toml
+   [tools]
+   enabled = ["*"]
+   ```
+
+## Access tiers
+
+- `direct`: honors the access tier chosen at `jcode login google`
+  (Read & Draft Only logins cannot send/trash, enforced at the OAuth scope level).
+- `composio`: connections request full Gmail scopes, so send/trash are
+  available. The tool still requires explicit `confirmed: true` for send,
+  send_draft, and trash.
+
+## Trust note
+
+With the Composio backend, Composio holds your Gmail OAuth grant and sees API
+traffic. This is the core tradeoff versus the direct backend. Disclose this to
+users before enabling it as a default.
diff --git a/scripts/stale_server_upgrade_sandbox.sh b/scripts/stale_server_upgrade_sandbox.sh
new file mode 100755
index 000000000..bf761180e
--- /dev/null
+++ b/scripts/stale_server_upgrade_sandbox.sh
@@ -0,0 +1,157 @@
+#!/usr/bin/env bash
+# Live end-to-end sandbox for the "current client, stale older server" fix.
+#
+#   Server: the REAL released v0.14.6 binary (downloaded from GitHub).
+#   Client: the freshly built current binary (target/debug/jcode, has the fix).
+#   Field state: shared-server channel pinned to OLD (v0.14.6); stable -> NEW.
+#
+# It starts the real old daemon, then runs the NEW client's `jcode server reload`
+# (which repairs the stale shared-server channel, then forces a reload). PASS iff
+# the resulting daemon is running v0.22.x.
+#
+# Usage:
+#   cargo build -p jcode --bin jcode
+#   scripts/stale_server_upgrade_sandbox.sh
+#
+# Linux x86_64 only (uses the published jcode-linux-x86_64 release asset).
+set -uo pipefail
+
+REPO_ROOT="$(cd -- "$(dirname -- "$0")/.." && pwd)"
+NEW_BIN="${NEW_BIN:-$REPO_ROOT/target/debug/jcode}"
+OLD_VERSION="${OLD_VERSION:-v0.14.6}"
+OLD_DIR="${OLD_DIR:-/tmp/jcode-sandbox}"
+OLD_WRAP="$OLD_DIR/jcode-linux-x86_64"
+
+[ -x "$NEW_BIN" ] || { echo "missing new client binary: $NEW_BIN (run: cargo build -p jcode --bin jcode)"; exit 2; }
+
+# Fetch + extract the real old release binary if it is not already present.
+if [ ! -x "$OLD_WRAP" ]; then
+  mkdir -p "$OLD_DIR"
+  url="$(curl -fsSL "https://api.github.com/repos/1jehuang/jcode/releases/tags/$OLD_VERSION" \
+        | grep -o 'https://[^"]*jcode-linux-x86_64.tar.gz' | head -1)"
+  [ -n "$url" ] || { echo "could not resolve $OLD_VERSION linux asset URL"; exit 2; }
+  echo "Downloading old server $OLD_VERSION ..."
+  curl -fsSL "$url" -o "$OLD_DIR/old.tar.gz"
+  tar -C "$OLD_DIR" -xzf "$OLD_DIR/old.tar.gz"
+fi
+[ -x "$OLD_WRAP" ] || { echo "missing old binary $OLD_WRAP after download"; exit 2; }
+
+SANDBOX="$(mktemp -d /tmp/jcode-stale-sandbox.XXXXXX)"
+export JCODE_HOME="$SANDBOX/home"
+export JCODE_RUNTIME_DIR="$SANDBOX/runtime"
+# Hard isolation: pin the socket explicitly so we can NEVER touch the real
+# global daemon at /run/user/<uid>/jcode.sock.
+export JCODE_SOCKET="$SANDBOX/runtime/jcode.sock"
+# Make the new client's clean release version comparable (debug build is dirty).
+export JCODE_TEST_CLIENT_VERSION_OVERRIDE="v0.22.0 (sandbox)"
+mkdir -p "$JCODE_HOME" "$JCODE_RUNTIME_DIR"
+
+BUILDS="$JCODE_HOME/builds"
+mkdir -p "$BUILDS/versions/0.14.6" "$BUILDS/versions/0.22.0" \
+         "$BUILDS/shared-server" "$BUILDS/stable" "$BUILDS/current"
+
+log() { printf '\n=== %s ===\n' "$*"; }
+
+# --- Install the OLD binary (with bundled libs) as version 0.14.6 ----------
+cp "$OLD_DIR/jcode-linux-x86_64.bin" "$OLD_DIR/libssl.so.10" \
+   "$OLD_DIR/libcrypto.so.10" "$BUILDS/versions/0.14.6/"
+cat > "$BUILDS/versions/0.14.6/jcode" <<'WRAP'
+#!/usr/bin/env sh
+set -eu
+real=$0
+if command -v readlink >/dev/null 2>&1; then
+  resolved=$(readlink -f -- "$0" 2>/dev/null || true)
+  [ -n "$resolved" ] && real=$resolved
+fi
+self_dir=$(CDPATH= cd -- "$(dirname -- "$real")" && pwd)
+export LD_LIBRARY_PATH="$self_dir:${LD_LIBRARY_PATH:-}"
+exec "$self_dir/jcode-linux-x86_64.bin" "$@"
+WRAP
+chmod +x "$BUILDS/versions/0.14.6/jcode"
+
+# --- Install the NEW binary as version 0.22.0 (newer mtime) ----------------
+cp "$NEW_BIN" "$BUILDS/versions/0.22.0/jcode"
+touch -d "+1 minute" "$BUILDS/versions/0.22.0/jcode"
+
+# --- Field state: shared-server -> OLD, stable/current -> NEW --------------
+ln -sf "../versions/0.14.6/jcode" "$BUILDS/shared-server/jcode"
+echo "0.14.6" > "$BUILDS/shared-server-version"
+ln -sf "../versions/0.22.0/jcode" "$BUILDS/stable/jcode"
+echo "0.22.0" > "$BUILDS/stable-version"
+ln -sf "../versions/0.22.0/jcode" "$BUILDS/current/jcode"
+echo "0.22.0" > "$BUILDS/current-version"
+
+log "Initial channel state (the field bug: shared-server pinned to OLD)"
+echo "shared-server-version: $(cat "$BUILDS/shared-server-version")"
+echo "stable-version:        $(cat "$BUILDS/stable-version")"
+
+SERVER_PID=""
+cleanup() {
+  [ -n "$SERVER_PID" ] && kill "$SERVER_PID" 2>/dev/null || true
+  "$NEW_BIN" --no-update server stop >/dev/null 2>&1 || true
+  pkill -f "$BUILDS/versions/0.14.6/jcode-linux-x86_64.bin" 2>/dev/null || true
+  pkill -f "$BUILDS/versions/0.22.0/jcode" 2>/dev/null || true
+  rm -rf "$SANDBOX"
+}
+trap cleanup EXIT
+
+server_version_via_socket() {
+  # Ask the running daemon (via the new client's debug surface) its version.
+  "$NEW_BIN" --no-update debug server:info 2>/dev/null \
+    | grep -oE '"version":[[:space:]]*"[^"]*"' | head -1
+}
+
+# --- 1) Start the REAL old v0.14.6 daemon ----------------------------------
+log "Starting OLD v0.14.6 daemon"
+"$BUILDS/shared-server/jcode" --no-update --provider antigravity serve \
+  >"$SANDBOX/server.log" 2>&1 &
+SERVER_PID=$!
+# Wait for the socket to appear.
+for _ in $(seq 1 40); do
+  [ -S "$JCODE_SOCKET" ] && break
+  sleep 0.25
+done
+sleep 1
+echo "old daemon pid=$SERVER_PID"
+echo "server.log tail:"; tail -8 "$SANDBOX/server.log" 2>/dev/null || true
+BEFORE="$(server_version_via_socket)"
+echo "server version BEFORE (via socket): ${BEFORE:-<none>}"
+
+# --- 2) New client: jcode server reload (repairs channel, then reloads) ----
+log "Running NEW client: jcode server reload"
+"$NEW_BIN" --no-update server reload 2>&1 | sed 's/^/[server reload] /' || true
+echo "shared-server-version after repair: $(cat "$BUILDS/shared-server-version")"
+
+# Give the handoff a moment.
+for _ in $(seq 1 40); do
+  [ -S "$JCODE_SOCKET" ] && break
+  sleep 0.25
+done
+sleep 2
+
+# --- 3) Verify the running daemon is now v0.22.x ---------------------------
+AFTER="$(server_version_via_socket)"
+echo "server version AFTER (via socket): ${AFTER:-<none>}"
+echo "server.log tail (post-reload):"; tail -8 "$SANDBOX/server.log" 2>/dev/null || true
+
+log "RESULT"
+echo "shared-server-version: before=0.14.6  after=$(cat "$BUILDS/shared-server-version")"
+echo "server version:        before=${BEFORE:-?}  after=${AFTER:-?}"
+
+ok_channel=0
+[ "$(cat "$BUILDS/shared-server-version")" = "0.22.0" ] && ok_channel=1
+
+ok_server=0
+echo "${AFTER:-}" | grep -q "0.22" && ok_server=1
+
+if [ "$ok_channel" = 1 ] && [ "$ok_server" = 1 ]; then
+  echo "PASS: new client repaired the channel AND the stale server upgraded to v0.22"
+  exit 0
+elif [ "$ok_channel" = 1 ]; then
+  echo "PARTIAL: channel repaired to 0.22.0, but server version probe inconclusive (AFTER=${AFTER:-none})"
+  echo "         (channel repair is the load-bearing fix; server exec depends on old daemon handoff)"
+  exit 0
+else
+  echo "FAIL: channel was not repaired"
+  exit 1
+fi
diff --git a/src/cli/commands.rs b/src/cli/commands.rs
index a257ce34e..df265f475 100644
--- a/src/cli/commands.rs
+++ b/src/cli/commands.rs
@@ -2350,6 +2350,34 @@ pub async fn run_server_reload_command(force: bool, emit_json: bool) -> Result<(
     }
 
     let mut client = crate::server::Client::connect().await?;
+
+    // Before asking the (possibly older) daemon to reload, repair a stale
+    // `shared-server` channel from the client side. The running server resolves
+    // its reload target from that channel; if it still points at the server's
+    // own old binary (the "current client, stale server" state, e.g. after a
+    // no-op `/update`), a forced reload would just re-exec the same old binary.
+    // Repointing shared-server -> stable when stable is strictly newer gives the
+    // reload a newer binary to exec into. Never downgrades; preserves a fresher
+    // self-dev pin. Best-effort: a failure here must not block the reload.
+    match crate::build::repair_stale_shared_server_channel() {
+        Ok(crate::build::SharedServerRepair::Repaired {
+            repaired_to,
+            previous,
+        }) => {
+            crate::logging::info(&format!(
+                "server reload: repaired stale shared-server channel {:?} -> {} before reload",
+                previous, repaired_to
+            ));
+        }
+        Ok(crate::build::SharedServerRepair::AlreadyCurrent) => {}
+        Err(err) => {
+            crate::logging::warn(&format!(
+                "server reload: shared-server channel repair failed (continuing): {}",
+                err
+            ));
+        }
+    }
+
     let request_id = client.reload_with_force(force).await?;
 
     let mut reloading = false;