🏛️ Code Review Council¶

AI reviews AI-generated code. Evidence required. Every blocker must prove itself.

AI is writing your code — who is reviewing it?

⚡ How It Works¶

Every PR passes through a five-stage pipeline — deterministic checks first, LLM analysis only where it adds value:

Stage	Name	What Happens	Cost
0	🔒 Gate Zero	Secrets, lint, types, missing docs — under 2 seconds	Zero
1	✂️ Diff Preprocessor	Filters lockfiles and generated code, enforces token budgets	Zero
2	📦 ReviewPack	Builds structured context: symbols, test map, policy violations	Zero
3	🤖 Reviewer Panel	SecOps, QA, Architect, Docs — parallel, evidence required	4× LLM
4	🪑 Council Chair	Synthesises findings, requires exploit chain for blockers, renders verdict	1× LLM

5-stage pipeline and two output modes

🎯 Two Outputs, One Review Engine¶

	🧑‍💻 Developer	🧑‍💼 Owner
Audience	Engineers	Product / Leadership
Focus	File/line findings, evidence, fix suggestions	Plain-English risk, ship/no-ship recommendation
Extra	Policy references, Chair rationale	Copy-paste fix prompt for AI coding agent
Review strength	Full	Full — same engine, different presentation

📊 Real Verdict¶

This is actual output from Council reviewing its own PR:

Overall verdict: PASS  (confidence: 0.90)

  secops     PASS   0 findings
  qa         PASS   2 findings  (2 warnings accepted)
  architect  PASS   0 findings
  docs       PASS   0 findings

Runtime: 48 seconds

→ See the full breakdown on the Self Review page.

⚡ Try It Locally¶

git clone https://github.com/vishal8shah/code-review-council
cd code-review-council
pip install .
council init
export GOOGLE_API_KEY=...
council doctor --branch main
council review --branch main

Quality gate, not a guarantee

Council is not a substitute for human engineering judgment. LLM findings are evidence-backed but not infallible — hallucination risk is reduced by design (structured context, mandatory file/line references, exploit chain requirements) but not eliminated. Cost and latency vary by model and diff size. Use restricted BYOK keys on repos you control.

🗺️ Explore the Docs¶

Page	What You'll Find
Overview	Why this exists, the multi-LLM approach, the autonomous loop vision
Getting Started	Install, init, first review, CI setup
Design	Architecture decisions, ReviewPack, evidence-based Chair
Security	BYOK model, key scoping, threat mitigations, merge gates
Workflows	PR workflow vs BYOK workflow, artifact locations
Self Review	Council reviewing its own PR — real output, 26 fixes, 286 collected tests
FAQ	Fork PRs, model config, cost tuning, PR comments
Contributing	Setup, tests, adding a new reviewer persona