Digital-Threads · Shahinyanm · Jul 26, 2026 · Jul 26, 2026
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
@@ -6,14 +6,14 @@
   },
   "metadata": {
     "description": "Token Pilot — save 60-90% tokens when AI reads code",
-    "version": "0.49.0"
+    "version": "0.50.0"
   },
   "plugins": [
     {
       "name": "token-pilot",
       "source": "./",
       "description": "Reduces token consumption by 60-90% via AST-aware lazy file reading, structural symbol navigation, and cross-session tool-usage analytics. 25 MCP tools + 25 subagents + budget watchdog hooks.",
-      "version": "0.49.0",
+      "version": "0.50.0",
       "author": {
         "name": "Digital-Threads"
       },

diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "token-pilot",
-  "version": "0.49.0",
+  "version": "0.50.0",
   "description": "Saves 60-90% tokens on AI code reading. AST-aware lazy reads, symbol navigation, find_usages, structural git diff/log, edit-safety guard, Task-routing matcher, cross-session telemetry (errors + diagnostics), 25 tp-* subagents tiered to haiku/sonnet/opus with budget watchdog.",
   "author": {
     "name": "Digital-Threads",

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -5,6 +5,38 @@ All notable changes to Token Pilot will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.50.0] - 2026-07-26
+
+### Fixed — dispatch routing now enforces instead of suggesting
+
+Sending a task to `general-purpose` when a tp-* specialist covers it is
+the single most expensive mistake in a session. Measured on the same
+question: **57,921 tokens through `general-purpose` against 18,677
+through `tp-run`** — 3.1x, about 39k wasted per launch. Most of that is
+startup: a general agent pays ~35k just to exist, before doing any work.
+
+token-pilot has matched dispatches to specialists since v0.31.0, and the
+matcher works — it named the right agent on five of seven real
+descriptions. The suggestion simply never landed. It was delivered as
+`permissionDecision=allow` with the advice in the reason, and an allow is
+something the model may ignore. It did: over one measured session, ten of
+eleven dispatches went to `general-purpose` anyway.
+
+Reads are disciplined because a wasteful read comes back **denied**.
+Dispatches were not, because they came back allowed. So deny mode — the
+default — now blocks a high-confidence match and names the agent to use.
+
+Matching also reads the `prompt`, not just the `description`. Real
+descriptions are two or three words ("Reuse check") and score as low
+confidence; the prompt carries the actual task and takes the same
+dispatch to high. Without that, the blocking tier would almost never fire.
+
+Unchanged: a low-confidence match still only advises, because a weak
+keyword must never cost you a dispatch. Escape phrases ("ad-hoc",
+"open-ended") still pass — now recognised in the prompt as well —
+`TOKEN_PILOT_MODE=advisory` still turns blocking off entirely, and tp-*
+dispatches are never touched.
+
 ## [0.49.0] - 2026-07-26
 
 ### Fixed — the tp-* response-budget watchdog was dead

diff --git a/agents/tp-api-surface-tracker.md b/agents/tp-api-surface-tracker.md
@@ -9,7 +9,7 @@ tools:
   - mcp__token-pilot__read_symbol
   - Bash
 model: haiku
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: dd184501203fa7f3c73f419c4ffbe33c4be75400cb64a7a51733a3fe23f6e085
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-audit-scanner.md b/agents/tp-audit-scanner.md
@@ -11,7 +11,7 @@ tools:
   - Grep
   - Read
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: d172f600bf32277ea6eb4cbbee4542ddd698a986dcd96997d33930561964569b
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-commit-writer.md b/agents/tp-commit-writer.md
@@ -8,7 +8,7 @@ tools:
   - mcp__token-pilot__test_summary
   - mcp__token-pilot__outline
   - Bash
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: de64a406b5176de19f7422619c7de7949b1f28865f225402c9cea9255f377428
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-context-engineer.md b/agents/tp-context-engineer.md
@@ -13,7 +13,7 @@ tools:
   - Edit
   - Glob
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 68b32af2dacd82ebe52c4eec93edb903d452688274c3065218270627c564d8b0
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-dead-code-finder.md b/agents/tp-dead-code-finder.md
@@ -11,7 +11,7 @@ tools:
   - Grep
   - Read
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: d9b7f5b7ae6f4ae21305c775361bcab097cc774370a6d976c093571d46d55021
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-debugger.md b/agents/tp-debugger.md
@@ -12,7 +12,7 @@ tools:
   - Read
   - Bash
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 052413de8d92377edcde6ae5c823f5378db304baccfa29e8866467f42553a500
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-dep-health.md b/agents/tp-dep-health.md
@@ -9,7 +9,7 @@ tools:
   - Bash
   - Read
 model: haiku
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: e14dc57493d816f8c2e017963e2ef5f66bea50fd0b805a80e8a0d97c968427e7
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-doc-writer.md b/agents/tp-doc-writer.md
@@ -13,7 +13,7 @@ tools:
   - Edit
   - Glob
 model: haiku
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 57d741794ab40e31a7ac49c68ea39a9088f5827cdef866ce81bfca1b7c9180cf
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-history-explorer.md b/agents/tp-history-explorer.md
@@ -10,7 +10,7 @@ tools:
   - Bash
   - Read
 model: haiku
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 7b70fa76a60e3c58a1de4f56c32c0f166424137e203a0cf1c8654e7c9235d904
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-impact-analyzer.md b/agents/tp-impact-analyzer.md
@@ -12,7 +12,7 @@ tools:
   - mcp__token-pilot__read_symbols
   - Read
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 351a987e11eba63852f5431a16d8eb53104f4f689f82fdcc5a2bf4db948ba92f
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-incident-timeline.md b/agents/tp-incident-timeline.md
@@ -8,7 +8,7 @@ tools:
   - mcp__token-pilot__read_symbol
   - Bash
 model: inherit
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: de5722bfea374eaab096c1ae635c37879e7a91370ee3cd0532f4240be03c91eb
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-incremental-builder.md b/agents/tp-incremental-builder.md
@@ -13,7 +13,7 @@ tools:
   - Edit
   - Bash
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 375a824d0d847bb5453ec594c7a62ad566ee7e4d92717b0473f771f1a0477c60
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-migration-scout.md b/agents/tp-migration-scout.md
@@ -11,7 +11,7 @@ tools:
   - Grep
   - Glob
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 0334de1bf99b431b65359637d125cda7c44c6f780eb92c57cc538715b1939536
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-onboard.md b/agents/tp-onboard.md
@@ -10,7 +10,7 @@ tools:
   - mcp__token-pilot__smart_read
   - mcp__token-pilot__smart_read_many
   - mcp__token-pilot__read_section
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 832e95633fbc8e9b0c10f3e540a327d4be062fb4b3f17a6cce6be13f414e2927
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-performance-profiler.md b/agents/tp-performance-profiler.md
@@ -11,7 +11,7 @@ tools:
   - Bash
   - Read
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: b61f06380d80798fa2e49d37bcba0653495bee04dd6bdbc1feff9a75607b0508
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-pr-reviewer.md b/agents/tp-pr-reviewer.md
@@ -11,7 +11,7 @@ tools:
   - mcp__token-pilot__read_for_edit
   - Read
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: f83f50d05b4f70285ae7afed2b1a406fc436df56e61a0aedbfb31edc7f2b6e66
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-refactor-planner.md b/agents/tp-refactor-planner.md
@@ -8,7 +8,7 @@ tools:
   - mcp__token-pilot__outline
   - mcp__token-pilot__read_symbol
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: c5f6fc122c89e16e5cf774045f92169ee3468555320b898171ba13eca5323550
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-review-impact.md b/agents/tp-review-impact.md
@@ -9,7 +9,7 @@ tools:
   - mcp__token-pilot__module_info
   - Bash
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 8ef3c3341cbfed4eb8dd130126a9683edc57e378c92ff0ca764d584fd941c55c
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-run.md b/agents/tp-run.md
@@ -16,7 +16,7 @@ tools:
   - Glob
   - Bash
 model: haiku
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 2b08618d34a61f00aafccbda9fed6d83243296dedb83440edbd2d5c28bb6dbc4
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-session-restorer.md b/agents/tp-session-restorer.md
@@ -9,7 +9,7 @@ tools:
   - mcp__token-pilot__session_budget
   - Bash
   - Read
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 529374ed728f5eed5b758b3be3da65624783c0bf0c1a253d7d661a843eb5f767
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-ship-coordinator.md b/agents/tp-ship-coordinator.md
@@ -11,7 +11,7 @@ tools:
   - Read
   - Grep
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: a60f6ae110eb3138064bce074e8ba26fa0ce5f4659df1624a9d9d3646803391b
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-spec-writer.md b/agents/tp-spec-writer.md
@@ -9,7 +9,7 @@ tools:
   - Read
   - Write
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: c7a4e8b39228fd5158528f389c924c5ff2d98c4b9b05ee0106d54a26c5dc1350
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-test-coverage-gapper.md b/agents/tp-test-coverage-gapper.md
@@ -10,7 +10,7 @@ tools:
   - mcp__token-pilot__test_summary
   - Glob
   - Grep
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: be81eed53a3720d146cf89e4a14a7a56577633f7c84c234c412ab70d64c05b11
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-test-triage.md b/agents/tp-test-triage.md
@@ -8,7 +8,7 @@ tools:
   - mcp__token-pilot__find_usages
   - mcp__token-pilot__read_symbol
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 362ecf4cb03b059421ea26933473700900073dc38b3a7fe271208dfb1ae14f90
 requiredMcpServers:
   - "token-pilot"

diff --git a/agents/tp-test-writer.md b/agents/tp-test-writer.md
@@ -13,7 +13,7 @@ tools:
   - Edit
   - Bash
 model: sonnet
-token_pilot_version: "0.49.0"
+token_pilot_version: "0.50.0"
 token_pilot_body_hash: 269f2fe22ff4517c277d3f56ca67d8a5527b93290ab21079a83ba7af22c1b5a9
 requiredMcpServers:
   - "token-pilot"

diff --git a/package-lock.json b/package-lock.json
diff --git a/package.json b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "token-pilot",
-  "version": "0.49.0",
+  "version": "0.50.0",
   "description": "Save up to 80% tokens when AI reads code \u2014 MCP server for token-efficient code navigation, AST-aware structural reading instead of dumping full files into context window",
   "type": "module",
   "main": "dist/index.js",

diff --git a/server.json b/server.json
@@ -6,12 +6,12 @@
     "url": "https://github.com/Digital-Threads/token-pilot",
     "source": "github"
   },
-  "version": "0.49.0",
+  "version": "0.50.0",
   "packages": [
     {
       "registryType": "npm",
       "identifier": "token-pilot",
-      "version": "0.49.0",
+      "version": "0.50.0",
       "transport": {
         "type": "stdio"
       },

diff --git a/src/hooks/pre-task.ts b/src/hooks/pre-task.ts
@@ -133,6 +133,15 @@ export function decidePreTask(
     };
   }
 
+  // v0.50.0 — match against the prompt as well as the description.
+  // Descriptions are a few words ("Reuse check") and score 1, which is
+  // always low confidence; the prompt carries the actual task and takes
+  // the same dispatch to 3. Without this the high-confidence tier below
+  // almost never fires on real traffic.
+  const prompt =
+    typeof input.tool_input?.prompt === "string" ? input.tool_input.prompt : "";
+  const haystack = prompt ? `${description} ${prompt}` : description;
+
   // No description → nothing to match against. Inject the generic
   // tool-guide so the subagent still picks tp-tools (B14).
   if (!description || description.length === 0) {
@@ -141,11 +150,13 @@ export function decidePreTask(
 
   // Author-blessed escape clauses — user is explicitly saying
   // "this is broad". Inject the tool-guide but no agent suggestion.
-  if (containsEscape(description)) {
+  // Checked across the prompt too, so an escape written there is honoured
+  // now that the prompt can trigger a block.
+  if (containsEscape(haystack)) {
     return { kind: "advise", message: SUBAGENT_TOOL_GUIDE };
   }
 
-  const hit = matchTpAgent(description, ctx.agentIndex);
+  const hit = matchTpAgent(haystack, ctx.agentIndex);
   if (!hit) {
     // No specific tp-* match. Still send the generic tool-guide so
     // the subagent learns about smart_read / read_symbol — covers the
@@ -162,10 +173,21 @@ export function decidePreTask(
     `TOKEN_PILOT_MODE=advisory for warn-only behaviour.\n\n` +
     SUBAGENT_TOOL_GUIDE;
 
+  // v0.50.0 — deny mode now blocks a high-confidence match instead of
+  // only advising. An advisory rides along as permissionDecision=allow,
+  // which the model is free to ignore, and it did: measured over one
+  // session, ten of eleven dispatches went to general-purpose while the
+  // matcher named a specialist every time. The same question costs 57,921
+  // tokens through general-purpose and 18,677 through tp-run. Reads are
+  // disciplined because they come back denied; dispatches were not,
+  // because they came back allowed.
+  //
+  // Low-confidence matches still only advise — a weak keyword must never
+  // cost someone their dispatch — and both escape routes are unchanged.
   const hardBlock =
     ctx.force ||
     ctx.mode === "strict" ||
-    (ctx.mode === "deny" && hit.confidence === "high" && ctx.force);
+    (ctx.mode === "deny" && hit.confidence === "high");
 
   if (hardBlock) {
     return {