councilreview.md

councilreview.md

name	allowed-tools	description
councilreview	Bash(codex *), Bash(git *), Bash(cat *), Bash(mktemp *), Bash(rm *), Read, Glob, Grep	Multi-model council review for high-stakes decisions

Council Review — Multi-Model Expert Feedback

You are orchestrating a council review: collecting blunt, independent expert feedback from multiple AI models (yourself — Claude — and OpenAI via Codex CLI), then synthesizing a comparison.

Step 1: Determine Input

Check $ARGUMENTS:

• If it contains a file path → read that file as the review target.
• If empty → review the current branch diff against the main branch (master or main).

For branch diff mode, run:

git log --oneline master..HEAD
git diff master...HEAD

For file mode, read the file with the Read tool.

Capture the full content to review as REVIEW_CONTENT.

Step 2: Ask Review Focus (optional)

Use AskUserQuestion to ask:

"What kind of review does this need?"

Options:

• Architecture & design — API boundaries, coupling, extensibility, naming
• Correctness & edge cases — bugs, race conditions, missing error handling
• Production readiness — observability, rollback, perf, security
• All of the above (default)

Step 3: Claude Review (inline)

You ARE Claude. Provide your own expert review directly. Be blunt, opinionated, and specific.

Structure your review as:

Claude Review

Verdict: APPROVE / REQUEST_CHANGES / NEEDS_DISCUSSION

Summary: 1-2 sentences.

Critical Issues (blocking):

• Issue → why it matters → suggested fix

Concerns (non-blocking):

• Concern → tradeoff → suggestion

Strengths:

• What's done well

Architecture Opinion:

• Does the approach make sense? What would you do differently?

Risk Assessment: LOW / MEDIUM / HIGH — explain why.

Save your full review text mentally — you'll need it for comparison.

Step 4: Codex Review (via CLI)

Create a temporary file for the review content and run Codex non-interactively.

# Write content to temp file
TMPFILE=$(mktemp /tmp/council-review-XXXXXX.txt)
# Write REVIEW_CONTENT to TMPFILE using the Write tool or bash

# Run codex
codex exec \
  -m gpt-5.3-codex \
  --approval-mode full-auto \
  -o /tmp/codex-council-output.txt \
  "You are a senior staff engineer doing a blunt code/design review. Be opinionated and specific. No pleasantries.

Review the following content and provide:

**Verdict:** APPROVE / REQUEST_CHANGES / NEEDS_DISCUSSION

**Summary:** 1-2 sentences.

**Critical Issues** (blocking):
- Issue → why it matters → suggested fix

**Concerns** (non-blocking):
- Concern → tradeoff → suggestion

**Strengths:**
- What's done well

**Architecture Opinion:**
- Does the approach make sense? What would you do differently?

**Risk Assessment:** LOW / MEDIUM / HIGH — explain why.

Here is the content to review:

$(cat $TMPFILE)"

rm $TMPFILE

Read the output from /tmp/codex-council-output.txt.

IMPORTANT: If codex exec fails or times out, note the failure and proceed with Claude-only review + a note that Codex was unavailable.

Step 5: Synthesize Comparison Table

Present the final synthesis as:

Council Review Summary

Dimension	Claude (Opus)	Codex (GPT-5.3)
Verdict	...	...
Risk	...	...
Top Critical Issue	...	...
Top Concern	...	...
Architecture Take	...	...

Agreement

• Where both models agree (high confidence signals)

Disagreement

• Where they diverge (needs human judgment)

Recommended Actions

Numbered list of concrete next steps, ordered by priority, combining insights from both reviewers.

---

Rules

• Be blunt. No sugar-coating from either reviewer.
• Each reviewer operates independently — don't let one influence the other.
• The comparison table is the most valuable artifact. Make it scannable.
• If reviewing a PR/diff, reference specific files and line ranges.
• If reviewing a document, reference specific sections.
• Keep total output focused — the table + disagreements + actions are what matter most.
• Clean up temp files when done.