quorum-counsel

Review the current plan using a background review agent.

Review code changes or a GitHub PR via multi-model counsel

Use this agent as a BACKGROUND task (run_in_background:true is REQUIRED) when stuck on problems, need deep codebase exploration, or architecture design. Delegates to OpenAI Codex via the codex CLI. Codex excels at autonomous multi-file analysis, backend architecture reasoning, and finding issues Claude missed. The agent takes 5-30 minutes - always spawn in background so you can continue helping the user. **CRITICAL - Detailed Context Required**: When calling this agent, provide EXTENSIVE detail: exact file paths, function/class names, line numbers, complete code snippets, error messages, what you've tried, architectural context, constraints, and specific questions. Vague prompts produce poor results. **Key signals to use this agent**: - Claude has tried 2+ approaches without success (stuck debugging) - Need to trace dependencies across many files (codebase exploration) - Need architecture design with trade-off analysis (new component design) - Need impact analysis of a proposed change (what breaks if we change X?) - Want a different perspective on a hard problem (second opinion)

Always use this agent as a BACKGROUND task (run_in_background:true is REQUIRED) for mandatory plan review and code review. Orchestrates both Gemini 3.x and Codex GPT-5.x in parallel, then synthesizes findings into a unified review with consensus issues and model-specific insights. **CRITICAL**: Caller must provide a complete Context Bundle (requirements, plan/background, constraints, artifacts, review questions). If missing, the reviewer proceeds but must emit Context Sufficiency Warnings and reduce confidence. **MANDATORY Use Cases**: 1. **Plan Review**: ALWAYS use after creating implementation plans (validate approach, gaps, edge cases, alternatives) 2. **Code Review**: ALWAYS use after completing implementation (before claiming done) (bugs, security, test gaps, quality) **Do NOT use for**: quick fixes or exploration/debugging (use codex-solver instead).

Use this agent as a BACKGROUND task (run_in_background:true is REQUIRED) when you need a different perspective, large-scale codebase understanding, or architecture analysis. Delegates to Google Gemini via the gemini CLI in read-only mode. Gemini excels at massive context understanding (1M+ tokens), system-wide pattern recognition, and producing clear explanations. The agent takes 2-10 minutes - always spawn in background so you can continue helping the user. **CRITICAL - Detailed Context Required**: When calling this agent, provide EXTENSIVE detail: exact file paths, function/class names, line numbers, complete code snippets, error messages, what you've tried, architectural context, constraints, and specific questions. Vague prompts produce poor results. **Key signals to use this agent**: - Need to understand a large or unfamiliar codebase (massive context window) - Need architecture analysis or system-wide pattern recognition - Want a second opinion from a different model family (Gemini vs Claude/Codex) - Need detailed explanation of complex code or data flows - Need to trace cross-cutting concerns across many files

**NEVER invoke automatically — only when the user EXPLICITLY asks.** Use this agent as a BACKGROUND task (run_in_background:true is REQUIRED) to consult GPT-5.4 Pro via browser automation. Oracle is a heavy hitter: expensive (ChatGPT Pro subscription), slow (10-30 minutes per run), and must be loaded with maximum context. Delegates to the `pnpx @steipete/oracle` CLI in browser mode. **CRITICAL — User-Initiated Only**: Do NOT spawn this agent proactively, from hooks, or from review gates. The user must explicitly ask to consult Oracle. This is non-negotiable. **CRITICAL — Maximum Context Required**: Oracle starts empty — it cannot read your codebase. You MUST bundle ALL relevant files via `--file` flags, include a full project briefing in the prompt, and provide detailed questions. Skimpy prompts produce useless results. **Key signals (only when user asks)**: - User says "ask Oracle", "consult Oracle", "run Oracle", or similar - Need GPT-5.4 Pro's perspective on a hard problem - Want a thorough second opinion with deep reasoning - Complex analysis that benefits from Pro-level thinking time

**NEVER invoke automatically — only when the user EXPLICITLY asks.** Delegate problems to GPT-5.4 Pro via `pnpx @steipete/oracle` CLI in browser mode. Oracle is a heavy hitter: expensive (ChatGPT Pro subscription), slow (10-30 minutes per run), and must be loaded with maximum context. Oracle starts empty — it cannot read your codebase, so you MUST bundle ALL relevant files via `--file` flags. Use only when the user says "ask Oracle", "consult Oracle", "run Oracle", or similar. Always run in background.