pai-opencode: v3.0.1

3.0.1 (2026-04-13)

Bug Fixes

• migration: add missing closing brace in migrateSkills function (#147) (1625b41)
• update hero banner path in README (e7e1c92)

pai-opencode: v3.0.0

3.0.0 (2026-04-13)

⚠️ Breaking Changes

• wp-m1: model_tiers removed — PAI-OpenCode now runs on vanilla OpenCode from opencode.ai. The custom Steffen025/opencode fork is archived. Each agent in opencode.json has exactly one model field. See ADR-019 and UPGRADE.md.
• core: Bootstrap loading removed — the PAI Core Skill now loads via OpenCode's native skill system (tier: always). MINIMAL_BOOTSTRAP.md is deleted. See ADR-020.
• skills: skills/CORE renamed to skills/PAI.
• installer: Provider selection removed from install — configure post-install via /connect + switch-provider.ts.

Features

• OpenCode-native architecture — zero bootstrap, PAI/SKILL.md tier:always via native skill scanner
• Zen free out-of-box — all agents default to opencode/big-pickle, no API key required
• Interactive terminal install wizard for v3
• Two-step provider setup: /connect (credentials) → switch-provider.ts (opencode.json)
• PAI wrapper script — separate pai command from opencode
• Legacy config migration script for v2.x model_tiers → flat model
• Professional CI/CD workflows, CodeRabbit, upstream sync automation
• ADR-020 — Native OpenCode context loading architecture decision

Migration

Run bun run PAI-Install/engine/migrate-legacy-config.ts ~/.opencode/opencode.json to auto-convert existing v2.x configs. See UPGRADE.md.

PAI-OpenCode v2.0.0 — PAI 2.5 → PAI v3.0 / Algorithm v1.8.0

The Big Leap: PAI 2.5 → PAI v3.0

PAI-OpenCode v2.0.0 is the biggest release since launch. We jumped from PAI 2.5 (Algorithm v0.2.25) all the way to PAI v3.0 (Algorithm v1.8.0) — that's a complete generation change in how the Algorithm thinks, verifies, and learns.

This isn't an incremental update. The Algorithm was fundamentally rewritten. The old 3-depth system (FULL/ITERATION/MINIMAL) is gone. The new system has 8 effort levels, mechanical verification gates, constraint extraction, self-interrogation, and a wisdom accumulation system that gets smarter across sessions.

---

What Changed (PAI 2.5 → PAI v3.0)

The Algorithm Got Serious About Verification

Before (v0.2.25): The Algorithm could claim "8/8 PASSED" without actually calling TaskUpdate. Rubber-stamp verification was possible.

After (v1.8.0): The Verify Completion Gate makes this impossible. Every PASS claim requires an actual TaskUpdate call with evidence. This is NON-NEGOTIABLE at ALL effort levels. No more hallucinated verification.

8 Effort Levels Replace 3 Depth Modes

Old (v0.2.25)	New (v1.8.0)
FULL	Standard, Extended, Advanced, Deep, Comprehensive, Loop
ITERATION	Fast
MINIMAL	Instant

The Algorithm now scales its rigor to the task. A quick rename gets Instant (no ISC). A complex architecture decision gets Deep or Comprehensive with full self-interrogation, constraint extraction, and verification rehearsal.

Wisdom Frames — The Algorithm Learns Across Sessions

New in v1.8.0: Wisdom Frames are domain-specific knowledge stores in MEMORY/WISDOM/ that compound learning across sessions:

• OBSERVE reads wisdom frames → better ISC from past experience
• LEARN writes wisdom frames → knowledge accumulates over time
• Ships with 5 seed domains: development, deployment, security, architecture, communication
• WisdomFrameUpdater.ts CLI tool for management

This is the Algorithm's long-term memory — it doesn't just solve problems, it gets better at solving them.

Phase Separation Enforcement

Every phase now has a STOP marker — the Algorithm must explicitly pause between THINK, PLAN, BUILD, EXECUTE, and VERIFY. No more phase-blending where BUILD bleeds into VERIFY.

Effort-Scaled Gates

Not everything runs at full intensity:

• Self-Interrogation: Instant/Fast skip entirely, Standard answers 2 of 5 questions, Extended+ answers all 5
• Quality Gates QG6/QG7: Only enforced at Extended+ effort levels
• Constraint Fidelity Steps 6-8: Gated to Extended+ only

This prevents over-engineering on simple tasks while maintaining rigor on complex ones.

---

New Capabilities

10 New Skills (29 → 39 total)

Skill	What It Does
Cloudflare	Pages, Workers, R2, KV automation
ExtractWisdom	Fabric-style wisdom extraction from content
IterativeDepth	Progressive deepening with multiple lenses
Science	Hypothesis-driven experimentation methodology
Parser	Content parsing with entity extraction
Remotion	Programmatic video creation
Sales	Sales narrative and package creation
USMetrics	FRED economic data analysis
WorldThreatModelHarness	Geopolitical threat modeling
WriteStory	Long-form creative writing

6 New/Updated Plugin Handlers (14 → 20 total)

Handler	Purpose
algorithm-tracker	Monitors phase transitions and ISC progress
agent-execution-guard	Validates agent invocations
skill-guard	Ensures skill prerequisites
check-version	Algorithm version compatibility check
integrity-check	Session-end validation
format-reminder	Updated: 8-tier effort level detection

Infrastructure Fixes

• Security Validator: env var prefix stripping prevents false positives on export VAR=value (upstream #620)
• Rating Capture: 5/10 ratings skip learning file generation (ambiguous signal, still tracked in JSONL)
• GenerateSkillIndex: Now follows symlinks to skill directories
• SessionHarvester: CLAUDE_DIR → PAI_DIR variable rename

---

Breaking Changes

• Algorithm format: 8 effort levels replace FULL/ITERATION/MINIMAL depth
• Start symbol: ♻︎ replaces 🤖
• ISC naming: ISC-{Domain}-{N} with priority/confidence tags
• SKILL.md: Completely rewritten from Algorithm v0.2.25 to v1.8.0

---

Stats

• 149 files changed, 29,127 insertions, 347 deletions
• 39 skills (was 29)
• 20 plugin handlers (was 14)
• 16 agents with dynamic tier routing
• 5 Wisdom Frame domains seeded
• Algorithm jumped 7 major versions (v0.2.25 → v1.8.0)

---

Installation

# New users
git clone https://github.com/Steffen025/pai-opencode.git
cd pai-opencode
bun run .opencode/PAIOpenCodeWizard.ts
opencode

# Existing users
cd pai-opencode
git pull origin main
bun run .opencode/PAIOpenCodeWizard.ts  # Re-run wizard

---

Full Changelog

See CHANGELOG.md for the complete release notes including all v1.x releases.

---

Credits: Built on Daniel Miessler's PAI and Anomaly's OpenCode.

v1.3.0 — Multi-Provider Agent System

PAI-OpenCode v1.3.0 — Multi-Provider Agent System

v1.3.0 is a turning point for PAI-OpenCode. Up until now, we were essentially running a 1:1 port of vanilla PAI — proving the architecture works on OpenCode. That's done.

With v1.3.0, we're now leveraging what makes OpenCode unique: the ability to mix providers, mix models, and route different agents to different backends — all within the same session.

What's Different Here

Most AI coding tools either lock you to one provider (Claude Code → Anthropic, Copilot → GitHub Models) or let you manually switch models (Cursor, Aider). PAI-OpenCode does something none of them do: the orchestrator automatically routes each task to the right model at the right cost — and the same agent scales up or down dynamically.

• Your Architect runs on Kimi K2.5 for normal design work — but scales up to Claude Opus 4.6 when the architecture decision requires maximum reasoning
• Your Engineer uses Kimi K2.5 for feature implementation — but drops to GLM 4.7 for simple batch edits, and scales up to Claude Sonnet 4.5 for complex debugging
• Your DeepResearcher defaults to GLM 4.7 — but drops to MiniMax for quick lookups and scales up to Kimi K2.5 for deep analysis
• Your PerplexityResearcher queries Sonar for web search — but scales up to Sonar Deep Research when the orchestrator needs comprehensive investigation
• Your Intern stays on MiniMax M2.1 ($0.30/1M tokens) because grunt work is grunt work

The orchestrator decides per task. You always pay exactly what the task requires — no more, no less.

This is built by the PAI-OpenCode agent system on top of OpenCode's multi-provider support. As far as we can tell, no other AI coding assistant or agent framework currently offers this pattern of dynamic per-task model tier routing across provider boundaries.

15 Specialized Agents with Dynamic Tier Routing

Agent	Default	Scales Down To	Scales Up To
Architect	Kimi K2.5	GLM 4.7 (quick review)	Claude Opus 4.6 (complex architecture)
Engineer	Kimi K2.5	GLM 4.7 (batch edits)	Claude Sonnet 4.5 (complex debugging)
DeepResearcher	GLM 4.7	MiniMax (quick lookup)	Kimi K2.5 (deep analysis)
GeminiResearcher	Gemini 3 Flash	—	Gemini 3 Pro (deep research)
PerplexityResearcher	Sonar	—	Sonar Deep Research
GrokResearcher	Grok 4.1 Fast	—	Grok 4.1 (full analysis)
CodexResearcher	GPT-5.1 Codex Mini	—	GPT-5.2 Codex
Writer	Gemini 3 Flash	MiniMax (quick drafts)	Claude Sonnet 4.5 (premium copy)
Pentester	Kimi K2.5	GLM 4.7 (quick scan)	Claude Sonnet 4.5 (deep audit)
Intern	MiniMax M2.1	—	—
explore	MiniMax M2.1	—	—
QATester	GLM 4.7	—	—

3 Configuration Presets

Pick a strategy and go — no manual config editing:

• zen-paid (Recommended) — 75+ providers via Zen AI Gateway. Combine providers freely.
• openrouter — OpenRouter routing with familiar model names.
• local-ollama — Fully local with Ollama. Zero cloud, complete privacy.

Model Tiers

Tier	When	Cost
quick	Simple tasks, batch work	~$1.25/1M tokens
standard	Normal operations (default)	~$15/1M tokens
advanced	Complex reasoning, deep analysis	~$75/1M tokens

Research Skill with Cost Control

• Quick — Single agent, FREE ($0)
• Standard — 3 agents (Claude + Gemini + Perplexity), ~$0.01
• Extensive — 4-5 providers, ~$0.10-0.50, requires confirmation

Other Improvements

• Documentation overhaul — 47 gaps fixed across 11 files, PAIAGENTSYSTEM.md fully rewritten
• Image pipeline fixed — All images optimized, 79% size reduction (12.4 MB → 2.6 MB)
• Agent cleanup — ClaudeResearcher → DeepResearcher (provider-agnostic naming)

Stats

• 113 files changed, 2,824 insertions(+), 1,792 deletions(-)
• 15 agents with dynamic tier routing
• 3 presets ready to use

Breaking Changes

• Image file extensions changed from .png to .jpg — update any external references
• ClaudeResearcher agent renamed to DeepResearcher
• PerplexityProResearcher removed — use PerplexityResearcher with model_tier: advanced

Getting Started

git clone https://github.com/Steffen025/pai-opencode.git
cd pai-opencode
bun run .opencode/PAIOpenCodeWizard.ts
opencode

Full Changelog

See CHANGELOG.md for the complete history.

---

PAI-OpenCode is a template for building Personal AI Infrastructure on top of OpenCode. Based on Daniel Miessler's PAI framework.

v1.2.0 - Real-Time Observability Dashboard

🔭 Real-Time Observability Dashboard

Complete monitoring infrastructure for PAI-OpenCode with real-time event streaming, SQLite persistence, and a Vue 3 dashboard.

✨ Highlights

• Observability Server on port 8889 with REST API and SSE streaming
• Vue 3 Dashboard with GitHub Dark theme
• 14 Event Types captured across all plugin hooks
• SQLite Persistence with 30-day retention

🖥️ Dashboard Features

Page	Description
Dashboard	Real-time stats cards + live event stream
Events	Searchable/filterable browser with pagination
Sessions	Expandable session viewer with event details

📡 API Endpoints

Endpoint	Method	Purpose
/health	GET	Health check
/events	POST	Event ingestion
/api/events	GET	Query events
/api/events/stream	GET	SSE real-time
/api/sessions	GET	Query sessions
/api/stats	GET	Statistics

🚀 Quick Start

# Start the observability server
cd ~/.opencode/observability-server
bun run server.ts

# (Optional) Start the dashboard
cd dashboard
bun install && bun run dev
# Open http://localhost:5173

📊 Event Types Captured

• Session: session.start, session.end
• Tools: tool.execute, tool.blocked
• Security: security.block, security.warn
• Messages: message.user, message.assistant
• Ratings: rating.explicit, rating.implicit
• Agents: agent.spawn, agent.complete
• Voice: voice.sent
• Learning: learning.captured, isc.validated, context.loaded

📁 New Files

.opencode/observability-server/     # Server + Dashboard
.opencode/plugins/handlers/observability-emitter.ts
docs/specs/OBSERVABILITY-SERVER-SPEC.md

Full Changelog: v1.1.0...v1.2.0

PAI-OpenCode v1.1.0 - Full PAI 2.5 with Voice Server

🎉 PAI-OpenCode v1.1.0 - Full PAI 2.5 with Voice Server

This release brings full PAI 2.5 compatibility to OpenCode, including the complete voice notification system.

✨ Highlights

• Voice Server Included - TTS notifications via ElevenLabs, Google Cloud TTS, or macOS say
• Full PAI 2.5 Algorithm (v0.2.25) with 7-phase workflow
• 13 Plugin Handlers - voice, sentiment, tabstate, ISC tracking, and more
• 29 Skills, 16 Agents - Complete PAI infrastructure

🔊 Voice Server

The voice server is now included in .opencode/voice-server/:

# Start the voice server
cd .opencode/voice-server && bun run server.ts

Configure your TTS provider in .opencode/.env:

• ElevenLabs: ELEVENLABS_API_KEY=your_key
• Google Cloud TTS: GOOGLE_API_KEY=your_key
• Fallback: macOS say command works automatically

📦 What's Included

Component	Count	Description
Skills	29	PAI 2.5 skills adapted for OpenCode
Agents	16	Named AI personalities
Plugin Handlers	13	Security, context, voice, sentiment, etc.
Voice Server	1	TTS notifications (ElevenLabs/Google/macOS)

🆕 New in v1.1.0

• Voice notification handler with 3 TTS backends
• Implicit sentiment detection (AI-powered)
• Tab state management (Kitty terminal)
• ISC tracking & response capture
• Full 7-phase Algorithm (OBSERVE → LEARN)

⬆️ Upgrade

git pull origin main

Full Changelog: v1.0.1...v1.1.0

PAI-OpenCode v1.0.0 - Initial Release

PAI-OpenCode v1.0.0 🎉

Personal AI Infrastructure for OpenCode — Bringing Daniel Miessler's PAI scaffolding to any AI provider.

---

✨ Features

• 33 Skills — CORE, Agents, Art, Browser, Security, Research, and more
• Plugin System — Auto-discovery, no configuration needed
• TELOS Framework — Deep personalization for your AI
• Security Validation — Blocks dangerous commands by default
• 8 AI Providers — Anthropic, OpenAI, Google, Groq, AWS Bedrock, Azure, ZEN (free), Ollama (local)
• Interactive Wizard — 2-minute setup

---

🚀 Quick Start

git clone https://github.com/Steffen025/pai-opencode.git
cd pai-opencode
bun run .opencode/PAIOpenCodeWizard.ts
opencode

---

📋 What's in v1.0

Working	Planned for v1.1
✅ All 33 skills	🔜 Voice Server
✅ Plugin system	🔜 Observability Dashboard
✅ TELOS/USER context	🔜 Multi-channel notifications
✅ Security validation
✅ Memory structure

See KNOWN_LIMITATIONS.md for details.

---

🙏 Credits

• Daniel Miessler — Original PAI vision and architecture
• Anomaly — OpenCode runtime

---

Welcome to Personal AI Infrastructure, your way.