Make Claude Code argue with itself (and 3 local LLMs) before it ships a plan.
A hook system that forces a 5-LLM peer review on every implementation plan Claude Code writes — before it's allowed to leave plan mode. The main session cannot approve an unreviewed plan. There is no bypass.
When plan-agent writes a plan to ~/.claude/plans/<slug>.md, a 4-phase pipeline fires automatically:
| Phase | Job | Model(s) |
|---|---|---|
| 0 | Grounding — verify every file path, symbol, and claim against the actual codebase | Haiku 4.5 (with Read/Grep/Glob) |
| 1 | 5 independent reviewers critique the plan | Haiku 4.5 + Sonnet 4.6 + 3 local Ollama |
| 2 | Same 5 models vote on a consensus document (APPROVE / REVISE / REJECT) | Haiku 4.5 + Sonnet 4.6 + 3 local Ollama |
| 3 | Final synthesis — actionable brief for the architect | Opus |
The main Opus session reads the consensus, incorporates feedback, and writes <!-- REVIEWED --> to line 1 of the plan. A PreToolUse hook on ExitPlanMode blocks the call until that marker is present.
You cannot escape the review loop. That's the point.
Plans drift. LLMs hallucinate file paths, propose changes that are already implemented, miss edge cases, over-engineer, or skip prerequisite steps. A single reviewer (even Opus) misses things a panel doesn't.
This repo makes that panel automatic, local-first, and enforced — not an opt-in suggestion.
User asks for a plan
│
▼
Agent(subagent_type="plan-agent")
│
▼
plan-agent writes ~/.claude/plans/<slug>.md
│
▼ (PostToolUse hook)
review_plan_consensus.sh
├─ Phase 0: Haiku grounding (Read, Grep, Glob — verifies plan claims)
├─ Phase 1: Haiku (bg) ║ Sonnet (bg) ║ model-a → model-b → model-c
├─ Phase 2: same 5 models vote on consensus
└─ Phase 3: Opus reads Phase 2 → synthesis brief
│
▼
All phases archived to ~/.claude/reviews/
│
▼
plan-agent reads the synthesis, revises, adds <!-- REVIEWED -->
│
▼ (PreToolUse hook)
ExitPlanMode ─► enforce_plan_agent.py
└─ BLOCKS unless <!-- REVIEWED --> is on line 1
git clone https://github.com/robomello/claude-plan-consensus
cd claude-plan-consensus
bash install.sh ~/.claude/hooks/plan-consensus # user-local, no sudoThen merge the generated settings-fragment-generated.json into ~/.claude/settings.json — the installer prints a Python one-liner that does it for you.
That's it. Next time plan-agent writes a plan, the pipeline fires automatically.
| Tool | Why |
|---|---|
claude CLI |
Phase 0/1/2 (Haiku, Sonnet) + Phase 3 (Opus) |
ollama |
3 local reviewer models (Phase 1, 2) |
jq |
JSON parsing in shell hooks |
curl |
Ollama API |
python3 |
Python hooks |
Defaults are the models used on the original server. Override via env:
export OLLAMA_MODEL_A="llama3.3:70b"
export OLLAMA_MODEL_B="mistral:7b"
export OLLAMA_MODEL_C="deepseek-r1:32b"Or edit OLLAMA_MODEL_A/B/C at the top of hooks/review_plan_consensus.sh.
export OLLAMA_BASE_URL="http://localhost:11434" # defaultAt the top of review_plan_consensus.sh:
HAIKU_TIMEOUT=180 # seconds — Haiku grounding + review
SONNET_TIMEOUT=300 # seconds — Sonnet review/consensus
OPUS_TIMEOUT=300 # seconds — Opus synthesis
OLLAMA_TIMEOUT=600 # seconds — per Ollama model call (bump for 70B+)enforce_plan_agent.py is a PreToolUse hook on ExitPlanMode. It reads the most recently modified *.md in ~/.claude/plans/ and denies the tool call if <!-- REVIEWED --> is not on line 1.
This means: the consensus pipeline MUST complete and plan-agent MUST incorporate feedback before the plan can be presented to the user.
enforce_plan_agent.sh fires on EnterPlanMode and plan-file writes — it warns (does not block) if the main session tries to write a plan directly instead of delegating to plan-agent. The hard block is at ExitPlanMode.
claude-plan-consensus/
├── hooks/
│ ├── review_plan_consensus.sh ← 4-phase consensus pipeline
│ ├── enforce_plan_agent.py ← BLOCKS ExitPlanMode without REVIEWED marker
│ ├── enforce_plan_agent.sh ← warns against bypassing plan-agent
│ ├── archive_plan.py ← archives approved plans by date
│ ├── validate_team_section.py ← warns on missing/stale Team Structure
│ └── plan_update_hook.py ← delegates plan saves to background Haiku
├── agents/
│ └── plan-agent.md ← installs to ~/.claude/agents/
├── rules/
│ └── planning-and-agents.md ← workflow rules for your CLAUDE.md
├── settings-fragment.json ← hook wiring template
└── install.sh ← setup script
Every plan produced by plan-agent includes these sections, in order:
- Context — why this change is needed
- Proposed changes — files and line numbers
- Task breakdown — tasks with explicit dependencies
- Prerequisites — packages/images/downloads needed before execution
- Verification steps — how to confirm success
- Review Notes — consensus findings incorporated or dismissed (added post-review)
- Team Structure — table of teammates, models, roles (REQUIRED)
The <!-- REVIEWED --> marker is line 1, before everything else.
Phase artifacts land in ~/.claude/reviews/:
~/.claude/reviews/
20260429_143022-plan-phase0-grounding.md
20260429_143022-plan-phase1.md
20260429_143022-plan-phase2-consensus.md
20260429_143022-plan-phase3-opus.md
The plan-agent receives the Opus synthesis prepended to the full Phase 2 details.
Q: Doesn't this slow Claude down? Yes — by 2 to 5 minutes per plan, depending on Ollama model size. That's the price of catching a 30-minute mistake before it ships.
Q: Can I disable a phase?
Set the matching timeout to 1 and the phase will fail-soft. The pipeline degrades gracefully if reviewers fail or time out (MIN_VALID_REVIEWS=2 floor).
Q: What if Ollama is down? Phase 1/2 will run with just Haiku + Sonnet. The pipeline still completes; you just lose the local reviewers.
Q: Can I use different cloud models?
Edit the --model flags in review_plan_consensus.sh. Anything the claude CLI accepts works.
Q: Why local Ollama instead of more cloud calls? Cost, latency, and diversity of opinion. Local 70B models often catch things cloud models miss because they were trained on different data slices.
MIT — do whatever you want, just don't blame me when your plan-agent calls you out.