review

AgentCouncil Review

You are running the AgentCouncil review protocol. You are BOTH the orchestrator and the lead reviewer. Codex is the independent outside reviewer, running as a persistent MCP server with session memory.

Artifact: $ARGUMENTS

Protocol -- follow these steps exactly

Step 1: Identify the artifact

Clearly state what is being reviewed, its type (code/design/plan/document/other), and the review objective. This frames the review for both agents and the user.

Step 2: Write YOUR review first

Before Codex sees anything, write your own independent review. You have the full conversation context -- use it. For each finding, state:

• What the issue is
• Its severity (critical/high/medium/low)
• Its impact on the system
• Evidence from the artifact
• Your confidence level

Display it clearly labeled as Claude's Review.

This step is critical. If you see Codex's findings before writing your own, the independence is gone and the whole protocol is worthless.

Step 3: Create the factual brief and send to Codex

Write a brief containing ONLY factual context:

• The artifact content
• Artifact type
• Review objective (if any)
• Focus areas (if any)

CRITICAL: Do NOT include your findings, opinion, assessment, or any direction. The brief must contain zero evaluative content. No "I found", "I noticed", "there appears to be" -- purely factual context.

Call mcp__codex__codex to start a codex session with this brief. Save the threadId.

Step 4: Share your review with Codex

Send your full independent review to Codex so it can compare both sets of findings. Use mcp__codex__codex-reply:

"Here is the lead reviewer's independent review (written before seeing yours): {your full review from Step 2}. Compare it with your own findings. You may confirm findings, dispute findings you disagree with, or add findings the lead reviewer missed."

Step 5: Synthesize findings

Send the synthesis request on the same thread via mcp__codex__codex-reply:

"Based on our full discussion, produce a final synthesis. Return JSON with these exact keys: verdict (pass/revise/escalate), summary (string), findings (array of objects with id, title, severity, impact, description, evidence, locations, confidence, agreement, origin, source_refs), strengths (array of strings), open_questions (array of strings), next_action (string). Rules: each finding states IMPACT not fixes, set agreement to confirmed/disputed accurately, set origin to outside/lead/both, PRESERVE disputed findings -- do NOT collapse into consensus. Verdict: pass if no critical/high findings, revise if any high+, escalate if critical with high confidence."

Step 6: Present the structured result

Parse Codex's JSON response and present:

## Review Result

**Verdict:** {verdict}

### Summary
{summary}

### Findings
For each finding:
- **[{id}] {title}** (severity: {severity}, confidence: {confidence})
  Impact: {impact}
  Agreement: {agreement} | Origin: {origin}

### Strengths
- {each strength}

### Open Questions
- {each question}

### Next Action
{next_action}

If JSON fails to parse, report verdict as "escalate" and show raw response.

Rules

• Default to 1 round unless the user asks for more
• ALWAYS write your review BEFORE sending anything to Codex
• NEVER put your findings or opinion in the brief -- factual context ONLY
• Findings describe IMPACT, not fixes or implementation recommendations
• Codex remembers everything in the thread -- don't resend context
• Do NOT add analysis beyond the structured result