External review

Use a fresh process to challenge the primary agent's implementation or design. The result is advisory: verify every finding locally before changing code or reporting it as fact.

Core workflow

1. Bound the target to the smallest useful diff, branch, PR, issue, file, or artifact. State any scope assumption.
2. Build a review packet with the request and intent, constraints, changed-file list and diffstat, focused diff or excerpts, validation already run, known failures, and the user's specific concerns. Exclude unrelated context and secrets.
3. Use the requested runner and model when provided. Otherwise require EXTERNAL_REVIEW_MODEL and use EXTERNAL_REVIEW_AGENT (default: pi). Do not silently downgrade the model.
4. Request a concise opinion; material, evidence-backed findings ordered by severity; trade-offs and fix direction; validation gaps; open questions; and a pass, revise, or blocked verdict.
5. Verify each finding against local code and classify it as must-fix, optional/deferred, false positive, or human decision. Explain rejected findings briefly.
6. If fixes are requested, implement only confirmed findings and rerun focused validation.

Launch command

: "${EXTERNAL_REVIEW_MODEL:?Set EXTERNAL_REVIEW_MODEL to a frontier model}"

case "${EXTERNAL_REVIEW_AGENT:-pi}" in
  pi)
    pi --no-session --no-tools --model "$EXTERNAL_REVIEW_MODEL" --thinking high -p "$REVIEW_PROMPT"
    ;;
  codex)
    codex exec --sandbox read-only --ask-for-approval never --model "$EXTERNAL_REVIEW_MODEL" "$REVIEW_PROMPT"
    ;;
  claude)
    claude -p "$REVIEW_PROMPT" --model "$EXTERNAL_REVIEW_MODEL" --permission-mode plan
    ;;
  *)
    printf 'Unsupported external reviewer: %s\n' "$EXTERNAL_REVIEW_AGENT" >&2
    exit 2
    ;;
esac

Keep tools disabled or read-only unless the user explicitly requests broader access.

Review packet

Use this compact shape for $REVIEW_PROMPT:

You are an independent reviewer in a fresh, read-only context.

Target and intent:
<bounded scope, requested behavior, constraints, trade-offs>

Evidence:
<changed files, diffstat, focused diff/excerpts, validation and known failures>

Focus:
<user concerns or "material correctness, security, API, test, and scope risks">

Return:
- Opinion: whether the change achieves its intent
- Findings: blocker|major|minor, confidence, exact evidence, impact, fix direction
- Validation gaps and open questions
- Verdict: pass|revise|blocked
Report only material findings; no generic checklist or empty praise.

Split oversized packets by subsystem or risk rather than sending an entire repository.

Guardrails

Do not expose credentials or unrelated proprietary context, accept findings without local verification, paste raw reviewer output as the final answer unless asked, or treat the reviewer as the final merge authority.