feat(solo): A/B/C plan gate + plan-critic specialist (#34)

* feat(solo): add council-style plan gate

Introduce a prompt-level A/B/C planning gate so Solo mode requires Planner, Reviewer, and Critic perspectives before implementation, while preserving the existing staged workflow and human approval gate.

Tested: bun test src/coordinator/soloPipelinePrompt.test.ts
Tested: bun test src/server/__tests__/conversations.test.ts -t "Solo Pipeline system prompt"
Tested: bun test src/server/__tests__/conversations.test.ts -t "routes coordinator and Solo"
Tested: cd desktop && bun run lint
Confidence: high
Scope-risk: narrow

Co-Authored-By: Claude GPT-5.5 <noreply@anthropic.com>

* feat(agents): add plan critic specialist

Add a read-only plan-critic specialist with parseable PLAN_REVIEW verdicts, register it in built-in and coordinator agent registries, and teach Solo's plan gate to use Plan plus plan-critic when available before synthesis.

Tested: bun test src/tools/AgentTool/builtInAgents.test.ts src/coordinator/workerAgent.test.ts src/coordinator/soloPipelinePrompt.test.ts
Tested: bun test src/server/__tests__/conversations.test.ts -t "Solo Pipeline system prompt"
Tested: bun test src/server/__tests__/conversations.test.ts -t "routes coordinator and Solo"
Not-tested: Full bun test remains blocked by unrelated existing failures in attribution header, shell PATH/env, release workflow, and desktop Vitest compatibility tests.
Confidence: high
Scope-risk: narrow

Co-Authored-By: Claude GPT-5.5 <noreply@anthropic.com>

* chore(ci): retrigger PR checks after applying policy labels

Co-Authored-By: Claude GPT-5.5 <noreply@anthropic.com>

---------

Co-authored-by: 你的姓名 <you@example.com>
Co-authored-by: Claude GPT-5.5 <noreply@anthropic.com>

desktop/src/i18n/locales/en.ts

@@ -1276,7 +1276,7 @@ export const en = {
   'session.coordinatorChip': 'Orchestration',
   'session.coordinatorChipTooltip': 'Orchestration mode is on for this session. Tasks fan out to specialist worker agents and the main thread coordinates them. Toggle it from the + menu in the composer.',
   'session.soloPipelineChip': 'Solo pipeline',
-  'session.soloPipelineChipTooltip': 'Solo Pipeline mode is on for this session. The AI runs a 5-stage workflow (planner → builder → tester → reviewer → integrator) on its own. Toggle it from the + menu in the composer.',
+  'session.soloPipelineChipTooltip': 'Solo Pipeline mode is on for this session. Before implementation, the AI runs an A/B/C plan gate (Planner → Reviewer → Critic), then continues through implement → test → review → land. Toggle it from the + menu in the composer.',
   'empty.recentActivity.handoff.branchLine': 'Last session on {branch}: "{title}".',
   'empty.recentActivity.handoff.titleLine': 'Last session: "{title}".',
   'empty.recentActivity.handoff.filesLine': 'Files touched: {files}{more}.',
@@ -1297,7 +1297,7 @@ export const en = {
   'solo.suggest.finishWip.title': 'Finish {count} pending change(s)',
   'solo.suggest.finishWip.detail': 'Uncommitted: {sample}',
   'solo.suggest.finishWip.detailForeign': 'Uncommitted (some may be from another agent): {sample}',
-  'solo.suggest.finishWip.taskPrompt': 'Wrap up the {count} uncommitted change(s) in this worktree ({files}). Walk them through plan → implement → test → review → land.',
+  'solo.suggest.finishWip.taskPrompt': 'Wrap up the {count} uncommitted change(s) in this worktree ({files}). Run the A/B/C plan gate first, then implement → test → review → land.',
 
   'solo.suggest.shipAhead.title': 'Ship {branch} ({count} commit(s) ahead)',
   'solo.suggest.shipAhead.detail': 'Local is {count} commit(s) ahead of upstream. The work is done — review and land.',

desktop/src/i18n/locales/jp.ts

@@ -1269,7 +1269,7 @@ export const jp: Record<TranslationKey, string> = {
   'session.coordinatorChip': 'オーケストレーション',
   'session.coordinatorChipTooltip': 'このセッションはオーケストレーションモードです。タスクは専門 worker エージェントへ分散し、メインスレッドが調整します。コンポーザーの + メニューから切り替えできます。',
   'session.soloPipelineChip': 'ソロパイプライン',
-  'session.soloPipelineChipTooltip': 'このセッションはソロパイプラインモードです。AI は 5 段階 (プランナー → 実装 → テスト → レビュー → 統合) を単独で実行します。コンポーザーの + メニューから切り替えできます。',
+  'session.soloPipelineChipTooltip': 'このセッションはソロパイプラインモードです。実装前に AI は A/B/C 計画ゲート（Planner → Reviewer → Critic）を実行し、その後 実装 → テスト → レビュー → リリース に進みます。コンポーザーの + メニューから切り替えできます。',
   'empty.recentActivity.handoff.branchLine': '前回のセッション ({branch}): 「{title}」。',
   'empty.recentActivity.handoff.titleLine': '前回のセッション:「{title}」。',
   'empty.recentActivity.handoff.filesLine': '触ったファイル: {files}{more}。',
@@ -1283,7 +1283,7 @@ export const jp: Record<TranslationKey, string> = {
   'solo.suggest.finishWip.title': '{count} 件の未コミット変更を仕上げる',
   'solo.suggest.finishWip.detail': '未コミット: {sample}',
   'solo.suggest.finishWip.detailForeign': '未コミット（他のエージェントの作業を含む可能性）: {sample}',
-  'solo.suggest.finishWip.taskPrompt': '現在の worktree にある {count} 件の未コミット変更（{files}）を仕上げてください。plan → implement → test → review → land を一通り通します。',
+  'solo.suggest.finishWip.taskPrompt': '現在の worktree にある {count} 件の未コミット変更（{files}）を仕上げてください。まず A/B/C 計画ゲートを通し、その後 実装 → テスト → レビュー → リリース に進みます。',
 
   'solo.suggest.shipAhead.title': '{branch} をリリース（{count} 件先行）',
   'solo.suggest.shipAhead.detail': 'ローカルが上流より {count} 件先行しており、作業はすでに完了しています — レビュー + リリース。',

desktop/src/i18n/locales/kr.ts

@@ -1269,7 +1269,7 @@ export const kr: Record<TranslationKey, string> = {
   'session.coordinatorChip': '오케스트레이션',
   'session.coordinatorChipTooltip': '이 세션은 오케스트레이션 모드입니다. 작업은 전문 worker 에이전트로 분산되고 메인 스레드가 조정합니다. 컴포저의 + 메뉴에서 전환할 수 있습니다.',
   'session.soloPipelineChip': '솔로 파이프라인',
-  'session.soloPipelineChipTooltip': '이 세션은 솔로 파이프라인 모드입니다. AI가 5 단계 (계획 → 구현 → 테스트 → 리뷰 → 통합)를 단독으로 실행합니다. 컴포저의 + 메뉴에서 전환할 수 있습니다.',
+  'session.soloPipelineChipTooltip': '이 세션은 솔로 파이프라인 모드입니다. 구현 전에 AI가 A/B/C 계획 게이트(Planner → Reviewer → Critic)를 먼저 실행한 뒤 구현 → 테스트 → 리뷰 → 통합으로 진행합니다. 컴포저의 + 메뉴에서 전환할 수 있습니다.',
   'empty.recentActivity.handoff.branchLine': '지난 세션({branch}): "{title}".',
   'empty.recentActivity.handoff.titleLine': '지난 세션: "{title}".',
   'empty.recentActivity.handoff.filesLine': '수정한 파일: {files}{more}.',
@@ -1283,7 +1283,7 @@ export const kr: Record<TranslationKey, string> = {
   'solo.suggest.finishWip.title': '{count}개의 미커밋 변경 마무리',
   'solo.suggest.finishWip.detail': '미커밋: {sample}',
   'solo.suggest.finishWip.detailForeign': '미커밋 (다른 에이전트의 작업 포함 가능): {sample}',
-  'solo.suggest.finishWip.taskPrompt': '현재 worktree의 {count}개 미커밋 변경({files})을 마무리해 주세요. plan → implement → test → review → land 한 사이클을 진행합니다.',
+  'solo.suggest.finishWip.taskPrompt': '현재 worktree의 {count}개 미커밋 변경({files})을 마무리해 주세요. 먼저 A/B/C 계획 게이트를 거친 뒤 구현 → 테스트 → 리뷰 → 통합을 진행합니다.',
 
   'solo.suggest.shipAhead.title': '{branch} 출시 ({count}개 커밋 앞섬)',
   'solo.suggest.shipAhead.detail': '로컬이 업스트림보다 {count}개 앞서 있고 작업은 이미 완료됨 — 리뷰 + 출시.',

desktop/src/i18n/locales/zh-TW.ts

@@ -1269,7 +1269,7 @@ export const zh: Record<TranslationKey, string> = {
   'session.coordinatorChip': '編排模式',
   'session.coordinatorChipTooltip': '該會話已開啟編排模式。任務會扇出到專家 worker 代理執行，主執行緒負責協調。在輸入框 + 選單裡切換。',
   'session.soloPipelineChip': '獨立流水線',
-  'session.soloPipelineChipTooltip': '該會話已開啟獨立流水線模式。AI 會按 5 個階段（規劃 → 實作 → 測試 → 審查 → 落地）獨立執行。在輸入框 + 選單裡切換。',
+  'session.soloPipelineChipTooltip': '該會話已開啟獨立流水線模式。實作前 AI 會先走 A/B/C 計畫閘門（Planner → Reviewer → Critic），再繼續實作 → 測試 → 審查 → 落地。在輸入框 + 選單裡切換。',
   'empty.recentActivity.handoff.branchLine': '上次在 {branch} 上做了：「{title}」。',
   'empty.recentActivity.handoff.titleLine': '上次會話：「{title}」。',
   'empty.recentActivity.handoff.filesLine': '動了這些檔案：{files}{more}。',
@@ -1283,7 +1283,7 @@ export const zh: Record<TranslationKey, string> = {
   'solo.suggest.finishWip.title': '收尾 {count} 個未提交改動',
   'solo.suggest.finishWip.detail': '未提交：{sample}',
   'solo.suggest.finishWip.detailForeign': '未提交（可能含其他 agent 的改動）：{sample}',
-  'solo.suggest.finishWip.taskPrompt': '把當前 worktree 裡 {count} 個未提交改動（{files}）收尾。走完一輪 plan → implement → test → review → land。',
+  'solo.suggest.finishWip.taskPrompt': '把當前 worktree 裡 {count} 個未提交改動（{files}）收尾。先走 A/B/C 計畫閘門，再實作 → 測試 → 審查 → 落地。',
 
   'solo.suggest.shipAhead.title': '把 {branch} 落地（{count} 個領先提交）',
   'solo.suggest.shipAhead.detail': '本地領先上游 {count} 個提交，工作已經做完——評審 + 落地。',

desktop/src/i18n/locales/zh.ts

@@ -1269,7 +1269,7 @@ export const zh: Record<TranslationKey, string> = {
   'session.coordinatorChip': '编排模式',
   'session.coordinatorChipTooltip': '该会话已开启编排模式。任务会扇出到专家 worker 代理执行，主线程负责协调。在输入框 + 菜单里切换。',
   'session.soloPipelineChip': '独立流水线',
-  'session.soloPipelineChipTooltip': '该会话已开启独立流水线模式。AI 会按 5 个阶段（规划 → 实现 → 测试 → 审查 → 落地）独立运行。在输入框 + 菜单里切换。',
+  'session.soloPipelineChipTooltip': '该会话已开启独立流水线模式。实现前 AI 会先走 A/B/C 计划闸门（Planner → Reviewer → Critic），再继续实现 → 测试 → 审查 → 落地。在输入框 + 菜单里切换。',
   'empty.recentActivity.handoff.branchLine': '上次在 {branch} 上做了:"{title}"。',
   'empty.recentActivity.handoff.titleLine': '上次会话:"{title}"。',
   'empty.recentActivity.handoff.filesLine': '动了这些文件:{files}{more}。',
@@ -1283,7 +1283,7 @@ export const zh: Record<TranslationKey, string> = {
   'solo.suggest.finishWip.title': '收尾 {count} 个未提交改动',
   'solo.suggest.finishWip.detail': '未提交:{sample}',
   'solo.suggest.finishWip.detailForeign': '未提交(可能含其他 agent 的改动):{sample}',
-  'solo.suggest.finishWip.taskPrompt': '把当前 worktree 里 {count} 个未提交改动({files})收尾。走完一轮 plan → implement → test → review → land。',
+  'solo.suggest.finishWip.taskPrompt': '把当前 worktree 里 {count} 个未提交改动({files})收尾。先走 A/B/C 计划闸门，再实现 → 测试 → 审查 → 落地。',
 
   'solo.suggest.shipAhead.title': '把 {branch} 落地({count} 个领先提交)',
   'solo.suggest.shipAhead.detail': '本地领先上游 {count} 个提交,工作已经做完——评审 + 落地。',

src/coordinator/coordinatorMode.ts

@@ -213,6 +213,7 @@ When calling ${AGENT_TOOL_NAME}, always pass \`subagent_type\`. Two flavours:
 | \`commit-pr\` | Draft commit messages or PR descriptions from a real diff. |
 | \`Explore\` | Fast read-only codebase search. Specify thoroughness ("quick" / "medium" / "very thorough"). |
 | \`Plan\` | Read-only architect that returns an implementation plan and the critical files for it. |
+| \`plan-critic\` | Read-only challenge pass for an implementation plan before code is written. Ends in \`PLAN_REVIEW: APPROVE\` / \`CHANGES_NEEDED\`. |
 | \`verification\` | Independent adversarial verification of completed work. Use after non-trivial implementation lands. |
 
 ${workerCapabilities}

src/coordinator/soloPipelinePrompt.test.ts

@@ -72,14 +72,27 @@ describe('getSoloPipelineSystemPrompt — invariants', () => {
     expect(prompt).toContain('Never run the pipeline for chat')
   })
 
-  it('declares all five stages with the planned specialists', () => {
+  it('declares all five Solo pipeline stages', () => {
-    expect(prompt).toMatch(/1\.\s+\*\*PLAN\*\*/)
+    expect(prompt).toMatch(/1\.\s+\*\*PLAN GATE/)
     expect(prompt).toMatch(/2\.\s+\*\*IMPLEMENT\*\*/)
     expect(prompt).toMatch(/3\.\s+\*\*TEST\*\*/)
     expect(prompt).toMatch(/4\.\s+\*\*REVIEW\*\*/)
     expect(prompt).toMatch(/5\.\s+\*\*LAND\*\*/)
   })
 
+  it('requires the prompt-level A/B/C Plan Gate before implementation', () => {
+    expect(prompt).toContain('A/B/C Plan Gate')
+    expect(prompt).toContain('Solo Council')
+    expect(prompt).toContain('A = Planner')
+    expect(prompt).toContain('B = Reviewer')
+    expect(prompt).toContain('C = Critic')
+    expect(prompt).toContain('final execution plan')
+    expect(prompt).toContain('use the `Plan` specialist when available')
+    expect(prompt).toContain('use the `plan-critic` specialist when available')
+    expect(prompt).toContain('otherwise run the same critique as a prompt-level role')
+    expect(prompt).toContain('Do not enter Stage 2')
+  })
+
   it('keeps the HUMAN GATE in Stage 4 (the safety contract)', () => {
     // Auto-merging review without explicit user approval would
     // defeat the whole point of Solo. Anyone who edits the prompt

src/coordinator/soloPipelinePrompt.ts

@@ -4,8 +4,9 @@
  * Sibling to `coordinatorMode.ts`. While coordinator mode does
  * dynamic fan-out (the lead AI decides what to delegate, when, and
  * to whom), Solo Pipeline mode runs a fixed five-stage script with
- * gates between stages: plan → implement → test → review (HUMAN
+ * gates between stages: plan gate (A/B/C council) → implement → test
- * APPROVAL) → land.
+ * → review (HUMAN APPROVAL) → land. Stage 1 is a prompt-level
+ * council pass, not a separate worker protocol.
  *
  * This module owns ONLY the prompt text + the boolean predicate
  * (`isSoloPipelineMode`). Wiring (CLI flag plumbing,
@@ -86,10 +87,13 @@ const SOLO_PIPELINE_PROMPT = `# Solo Pipeline Mode
 
 You are operating in **Solo Pipeline mode**: a staged, gated workflow for
 turning one feature/change request into shipped, verified work. You drive
-five stages in order, delegating each to a specialist subagent, and you
+five stages in order. Before any implementation, you must run an internal
-pause at human gates. You are the conductor — you do NOT write the code,
+**A/B/C Plan Gate — Solo Council**: A = Planner proposes, B = Reviewer
-run the tests, or approve the work yourself; you route to the right
+audits, C = Critic challenges, then you synthesize the final execution
-specialist and enforce the gate between stages.
+plan. A/B/C are prompt-level roles, not separate workers or a
+worker-to-worker protocol unless the runtime explicitly provides one. You
+are the conductor — you do NOT write the code, run the tests, or approve
+the work yourself; you enforce the gate between stages.
 
 ## STAGE 0 — Intent triage (ALWAYS run this first, silently)
 
@@ -117,26 +121,42 @@ only justified for real build work.
 
 Open with a one-line, scannable plan, e.g.:
 
-  "Solo pipeline: ① plan → ② implement → ③ test → ④ review (your
+  "Solo pipeline: ① plan gate (A/B/C council) → ② implement → ③ test →
-   approval) → ⑤ land. Starting Stage 1: planning."
+   ④ review (your approval) → ⑤ land. Starting Stage 1: A/B/C planning."
 
 Keep the user oriented: at each stage transition, print a single status
 line "▸ Stage N/5: <name> — <one-line goal>".
 
 ## The five stages
 
-1. **PLAN** (specialist: planner / context-gatherer)
+1. **PLAN GATE — Solo Council** (A/B/C; no implementation yet)
-   - Inspect the real code first. Produce a short written plan: the
+   - Inspect the real code first. Read enough to ground the plan in actual
-     changed surface, the approach, the files to touch, the success
+     files, commands, constraints, and existing patterns.
-     criteria, and the verification commands.
+   - **A = Planner**: use the \`Plan\` specialist when available. Produce the
-   - Hand-off artifact: \`plan\` (the written plan).
+     concrete implementation plan — changed surface, candidate files,
-   - GATE → automatic. Advance when the plan names concrete files +
+     approach, acceptance criteria, verification commands, and known risks.
-     a verification command. If it can't, loop back with the user.
+   - **B = Reviewer**: audit A's plan for missing files, integration points,
+     test gaps, safety issues, persistence/security boundaries, and repository
+     conventions.
+   - **C = Critic**: use the \`plan-critic\` specialist when available;
+     otherwise run the same critique as a prompt-level role. Challenge
+     assumptions, look for a smaller / safer / more verifiable alternative,
+     identify overreach, and name what would make the plan fail. The Critic
+     must not merely agree.
+   - **Synthesis**: merge A/B/C into one **final execution plan**. Treat that
+     final execution plan as the source of truth for Stage 2.
+   - Hand-off artifact: \`final execution plan\` with concrete files,
+     implementation steps, success criteria, verification commands, and
+     residual risks.
+   - GATE → automatic but strict. Do not enter Stage 2 or modify files until
+     the final execution plan is complete. If B reports a blocking issue or C
+     reports a blocking objection, revise the plan or ask the user one short
+     clarifying question before implementation.
 
 2. **IMPLEMENT** (specialist: implementer)
-   - Receives the plan verbatim. Makes the narrowest change that
+   - Receives the final execution plan verbatim. Makes the narrowest change
-     satisfies it. Touches only files the plan named (or explains any
+     that satisfies it. Touches only files the final execution plan named (or
-     deviation).
+     explains any deviation before proceeding).
    - Hand-off artifact: \`diff\` (files changed + summary).
 
 3. **TEST** (specialist: verification)
@@ -179,10 +199,15 @@ suggestion that prefilled the prompt), respect the entry stage:
 
 ## Rules
 
+- The A/B/C Plan Gate is mandatory for TASK messages that enter at
+  \`entryStage = 'plan'\` or have no entry-stage shortcut. Planning may read
+  code, but it must not implement or modify files.
 - One stage at a time. Never skip the human gate at Stage 4 unless the
   entry-stage shortcut placed you past it.
 - If a stage's specialist reports it can't proceed, surface it to the
   user rather than forcing the next stage.
+- The Critic must look for a smaller or safer alternative; the final
+  execution plan is the source of truth for implementation.
 - If the user interrupts mid-pipeline with a new message, re-run Stage 0
   intent triage on it: a chat aside shouldn't derail the pipeline, a new
   task should ask whether to queue or restart.

src/coordinator/workerAgent.test.ts

@@ -23,6 +23,7 @@ describe('coordinator agent registry', () => {
       'commit-pr',
       'Explore',
       'Plan',
+      'plan-critic',
       'verification',
     ]
     for (const expected of required) {

src/coordinator/workerAgent.ts

@@ -32,6 +32,7 @@ import { EXPLORE_AGENT } from '../tools/AgentTool/built-in/exploreAgent.js'
 import { MIGRATION_AGENT } from '../tools/AgentTool/built-in/migrationAgent.js'
 import { PERFORMANCE_AGENT } from '../tools/AgentTool/built-in/performanceAgent.js'
 import { PLAN_AGENT } from '../tools/AgentTool/built-in/planAgent.js'
+import { PLAN_CRITIC_AGENT } from '../tools/AgentTool/built-in/planCriticAgent.js'
 import { REFACTOR_AGENT } from '../tools/AgentTool/built-in/refactorAgent.js'
 import { SECURITY_REVIEWER_AGENT } from '../tools/AgentTool/built-in/securityReviewerAgent.js'
 import { TEST_AUTHOR_AGENT } from '../tools/AgentTool/built-in/testAuthorAgent.js'
@@ -95,6 +96,7 @@ export function getCoordinatorAgents(): AgentDefinition[] {
     // research / verify
     EXPLORE_AGENT,
     PLAN_AGENT,
+    PLAN_CRITIC_AGENT,
     VERIFICATION_AGENT,
   ]
 }

src/server/__tests__/conversations.test.ts

@@ -265,6 +265,9 @@ describe('ConversationService', () => {
     // The injected text must be the Solo prompt — sniff for the unique
     // Stage-0 intent-triage phrasing locked in soloPipelinePrompt.test.ts.
     expect(args[idx + 1]).toContain('Solo Pipeline mode')
+    expect(args[idx + 1]).toContain('A/B/C Plan Gate')
+    expect(args[idx + 1]).toContain('Critic')
+    expect(args[idx + 1]).toContain('final execution plan')
   })
 
   it('should not append the Solo prompt when soloPipelineMode is off', () => {
@@ -287,7 +290,7 @@ describe('ConversationService', () => {
     expect(flagIndices.length).toBe(2)
     const texts = flagIndices.map((i) => args[i + 1] ?? '')
     expect(texts.some((t) => t.includes(ORCHESTRATION_PROMPT_MARKER))).toBe(true)
-    expect(texts.some((t) => t.includes('Solo Pipeline mode'))).toBe(true)
+    expect(texts.some((t) => t.includes('A/B/C Plan Gate'))).toBe(true)
   })
 
   // Locks the orchestrator's "propagate project tool rules into every dispatched

src/tools/AgentTool/built-in/planCriticAgent.ts

@@ -0,0 +1,92 @@
+import { BASH_TOOL_NAME } from 'src/tools/BashTool/toolName.js'
+import { EXIT_PLAN_MODE_TOOL_NAME } from 'src/tools/ExitPlanModeTool/constants.js'
+import { FILE_EDIT_TOOL_NAME } from 'src/tools/FileEditTool/constants.js'
+import { FILE_WRITE_TOOL_NAME } from 'src/tools/FileWriteTool/prompt.js'
+import { NOTEBOOK_EDIT_TOOL_NAME } from 'src/tools/NotebookEditTool/constants.js'
+import { AGENT_TOOL_NAME } from '../constants.js'
+import type { BuiltInAgentDefinition } from '../loadAgentsDir.js'
+
+const PLAN_CRITIC_SYSTEM_PROMPT = `You are a plan critic specialist for Claude Code. Your job is to challenge an implementation plan before any code is written. You are not a code reviewer: code reviewers inspect diffs after implementation; you inspect the proposed plan for feasibility, scope, risk, and whether a smaller or safer path exists.
+
+=== CRITICAL: READ-ONLY MODE - NO FILE MODIFICATIONS ===
+This is a READ-ONLY plan critique task. You are STRICTLY PROHIBITED from:
+- Creating, modifying, or deleting any files (no Write, Edit, touch, rm, mv, cp)
+- Using redirect operators (>, >>) or heredocs to write files
+- Running git write operations (add, commit, push) or installing dependencies
+- Running ANY command that changes system state
+You do NOT have file-editing tools. Use ${BASH_TOOL_NAME} only for read-only inspection (git diff, git log, git status, ls, cat, head, tail, find, grep). Your output is a critique report, not edits — the parent agent revises or applies the plan.
+
+=== WHAT YOU RECEIVE ===
+You receive the user's requested outcome, a proposed implementation plan, and optionally relevant files or constraints. If important context is missing, inspect the codebase read-only to verify whether the plan is grounded in real files and existing patterns.
+
+=== CRITIQUE DIMENSIONS ===
+Challenge the plan in this order:
+
+**1. Feasibility**
+- Does the plan name the right files, APIs, data flow, and integration points?
+- Does it depend on behavior that does not exist or contradict current code?
+- Are sequencing and dependencies realistic?
+
+**2. Minimality and alternatives**
+- Is there a smaller change that satisfies the same user outcome?
+- Is the plan over-engineered, speculative, or adding abstractions before they are needed?
+- Can a safer staged rollout or narrower scope reduce risk?
+
+**3. Risk and contracts**
+- Does it touch auth, permissions, persistence, networking, shell execution, external providers, or public APIs?
+- Does it require migration, compatibility, or user-visible copy updates?
+- Are there hidden cross-process or cross-package effects?
+
+**4. Verification gaps**
+- Are success criteria concrete and testable?
+- Are the proposed checks the narrowest meaningful checks?
+- Is there an E2E/manual smoke path when unit tests cannot prove the behavior?
+
+=== OUTPUT FORMAT (REQUIRED) ===
+Return a concise report with these sections:
+
+1. Blocking objections
+- List only issues that should prevent implementation until the plan changes.
+- Use "None" if there are no blocking objections.
+
+2. Non-blocking concerns
+- List risks or improvements that can be handled during implementation.
+- Use "None" if there are no non-blocking concerns.
+
+3. Smaller / safer alternative
+- Name the best smaller or safer path if one exists.
+- If the proposed plan is already minimal, say so explicitly.
+
+4. Verification notes
+- State which checks or smoke paths should prove the final implementation.
+
+End with exactly one summary line the caller can parse:
+
+PLAN_REVIEW: APPROVE
+or
+PLAN_REVIEW: CHANGES_NEEDED
+
+Use CHANGES_NEEDED if any blocking objection remains. Use APPROVE only when the plan is feasible enough to implement.`
+
+const PLAN_CRITIC_WHEN_TO_USE =
+  'Use this agent to critique an implementation plan before code is written. It is read-only and challenges feasibility, scope, risks, verification gaps, and whether a smaller or safer plan exists. Ends with PLAN_REVIEW: APPROVE / CHANGES_NEEDED.'
+
+export const PLAN_CRITIC_AGENT: BuiltInAgentDefinition = {
+  agentType: 'plan-critic',
+  whenToUse: PLAN_CRITIC_WHEN_TO_USE,
+  color: 'yellow',
+  disallowedTools: [
+    AGENT_TOOL_NAME,
+    EXIT_PLAN_MODE_TOOL_NAME,
+    FILE_EDIT_TOOL_NAME,
+    FILE_WRITE_TOOL_NAME,
+    NOTEBOOK_EDIT_TOOL_NAME,
+  ],
+  source: 'built-in',
+  baseDir: 'built-in',
+  model: 'inherit',
+  omitClaudeMd: true,
+  getSystemPrompt: () => PLAN_CRITIC_SYSTEM_PROMPT,
+  criticalSystemReminder_EXPERIMENTAL:
+    'CRITICAL: This is a READ-ONLY plan critique task. You CANNOT edit, write, or create files. You MUST end with PLAN_REVIEW: APPROVE or PLAN_REVIEW: CHANGES_NEEDED.',
+}

src/tools/AgentTool/builtInAgents.test.ts

@@ -6,6 +6,7 @@ import {
   areExplorePlanAgentsEnabled,
   getBuiltInAgents,
 } from './builtInAgents.js'
+import { PLAN_CRITIC_AGENT } from './built-in/planCriticAgent.js'
 
 const originalDisableBuiltIns =
   process.env.CLAUDE_AGENT_SDK_DISABLE_BUILTIN_AGENTS
@@ -48,6 +49,7 @@ describe('built-in agents', () => {
     expect(agentTypes).toContain('docs-writer')
     expect(agentTypes).toContain('performance')
     expect(agentTypes).toContain('commit-pr')
+    expect(agentTypes).toContain('plan-critic')
     // game-developer is a project-level agent (.claude/agents), not a built-in.
     expect(agentTypes).not.toContain('game-developer')
   })
@@ -68,6 +70,25 @@ describe('built-in agents', () => {
     expect(agentTypes).toContain('docs-writer')
     expect(agentTypes).toContain('performance')
     expect(agentTypes).toContain('commit-pr')
+    expect(agentTypes).toContain('plan-critic')
+  })
+
+  test('plan-critic is read-only and returns a parseable verdict', () => {
+    expect(PLAN_CRITIC_AGENT.agentType).toBe('plan-critic')
+    expect(PLAN_CRITIC_AGENT.disallowedTools).toEqual(expect.arrayContaining([
+      'Agent',
+      'ExitPlanMode',
+      'Edit',
+      'Write',
+      'NotebookEdit',
+    ]))
+
+    const prompt = PLAN_CRITIC_AGENT.getSystemPrompt({} as never)
+    expect(prompt).toContain('READ-ONLY plan critique task')
+    expect(prompt).toContain('PLAN_REVIEW: APPROVE')
+    expect(prompt).toContain('PLAN_REVIEW: CHANGES_NEEDED')
+    expect(prompt).toContain('smaller or safer path')
+    expect(PLAN_CRITIC_AGENT.criticalSystemReminder_EXPERIMENTAL).toContain('PLAN_REVIEW: APPROVE')
   })
 
   test('preserves SDK opt-out in noninteractive sessions', () => {

src/tools/AgentTool/builtInAgents.ts

@@ -12,6 +12,7 @@ import { GENERAL_PURPOSE_AGENT } from './built-in/generalPurposeAgent.js'
 import { MIGRATION_AGENT } from './built-in/migrationAgent.js'
 import { PERFORMANCE_AGENT } from './built-in/performanceAgent.js'
 import { PLAN_AGENT } from './built-in/planAgent.js'
+import { PLAN_CRITIC_AGENT } from './built-in/planCriticAgent.js'
 import { REFACTOR_AGENT } from './built-in/refactorAgent.js'
 import { SECURITY_REVIEWER_AGENT } from './built-in/securityReviewerAgent.js'
 import { STATUSLINE_SETUP_AGENT } from './built-in/statuslineSetup.js'
@@ -76,6 +77,7 @@ export function getBuiltInAgents(): AgentDefinition[] {
     DOCS_WRITER_AGENT,
     PERFORMANCE_AGENT,
     COMMIT_PR_AGENT,
+    PLAN_CRITIC_AGENT,
   ]
 
   if (areExplorePlanAgentsEnabled()) {