Council reviews are a multi-model quality gate where external AI models review work done by Claude Code agents. They are implemented via the second-opinion skill and are a standard part of the workflow -- not a last resort.
When debugging has failed after 2-3 attempts, get a fresh perspective from one model. Codex is the default because it can read actual repo files via CLI.
For architecture reviews, post-implementation audits, or design validation, run all models in parallel and compile a combined view.
| Backend | Method | Speed | Cost | Strengths |
|---|---|---|---|---|
| Codex | CLI (codex exec, read-only) |
~120-140s | Free (ChatGPT sub) | Strategic gaps, control-plane issues; reads actual repo files |
| Gemini | REST API (GEMINI_API_KEY) | ~1-5s | Free (AI Studio) | Operational risks, security, env/config, deployment concerns |
| Kimi | REST API (KIMI_CODE_API_KEY) | ~3-10s | Free (Kimi Coding sub) | Detailed code-level review, finds inconsistencies |
| Kimi-CLI | CLI subprocess | ~100-200s | Free | Same as Kimi but reads local files (use only when needed) |
| Minimax | REST API (MINIMAX_API_KEY) | ~3-5s | Free (MiniMax sub) | Independent perspective, different training data, catches groupthink |
| Opus | Subagent (model: opus) | ~55s | Claude usage | Nuanced assessment, race conditions, subtle bugs |
All backends are free (existing subscriptions or free-tier API keys). No paid API calls.
These triggers are mandatory -- agents invoke them without waiting to be asked:
| Trigger | Minimum Models | Timing |
|---|---|---|
| Dev loop completes, 3+ files changed | Gemini + Codex | After reviewer approves, before marking done |
| Infrastructure changes (systemd, cron, scheduler, SSH) | Gemini + Codex | Before marking done |
| Architecture decision (new node, service, integration) | All models | Before committing direction |
| Security-touching changes (auth, tokens, keys, permissions) | Gemini + Codex | Non-negotiable |
| Larry ticket review finds CRITICAL/IMPORTANT | Gemini | Before resetting ticket to todo |
| Monthly architecture review | All models | With Solutions Architect |
Rule: If about to mark a significant task "done" and council has not run -- run it first. Takes 30-120 seconds.
python ~/.claude/skills/second-opinion/scripts/ask_second_opinion.py -b codex "PROMPT".txt prompt file (architecture summary + implementation + questions)11-agent-relationships/lucienne|Lucienne can overrule council findings when:
Overrides must always include an explanation of why.
When 11-agent-relationships/larry|Larry completes a ticket:
review_larry_work() runs a code reviewer on the git diff| Platform | Codex | Gemini | Kimi |
|---|---|---|---|
| Windows | CLI via bash (use -f for long prompts) |
API | CLI via cmd.exe bat script |
| Linux (Luci) | CLI directly | API | CLI directly |
| macOS | CLI directly | API | CLI directly |
~/.claude/skills/second-opinion/SKILL.md -- full skill definition~/.claude/skills/second-opinion/scripts/ask_second_opinion.py -- the invocation scriptMission Control is the board for your delegated work: requests come in, Luci coordinates the next step, and evidence stays visible for review.
Luci is your always-on assistant for routing, status updates, and follow-through. Operators can still open deeper evidence when needed.