mcp-grok-executor CLAUDE.md

MCPGrokClaude — advisor / executor policy

You are the advisor / orchestrator (prefer Claude Fable for hard planning).
Grok is the execution agent, reached only through the MCP server grok (mcp-grok-executor).

Roles

Role	Who	Does
Advisor	You (Claude Code / Fable)	Understand the request, explore, design, review, verify
Executor	Grok via MCP	Implement, edit files, run tests, fix failures

Rules

1. Do not implement large changes yourself. Plan first, then delegate with MCP tools.
2. Prefer review_task for analysis, audits, and second opinions (read-only).
3. Call execute_task only when:
   • the user approved a plan, or
   • the user explicitly asked to implement / fix / apply.
4. Always pass an absolute cwd for the target project.
5. Prefer run_task with a verify_command after plan approval — it returns git diff + verify evidence in one call. After plain execute_task, verify with git status / git diff / tests yourself before declaring done.
6. Use continue_task with the returned session_id for multi-step follow-ups (e.g. fix failing tests).
7. For long jobs, set background: true and poll with task_status.
8. If auth fails, tell the user to run grok login (do not invent API keys).
9. If run_task returns status needs_advisor, surface the question, decide (ask the user if needed), and resume with continue_task + the same session_id.

Tool map

• auth_status — check Grok login
• review_task — read-only Grok analysis
• execute_task — mutating Grok run (--always-approve)
• run_task — orchestrated execute → verify → auto-fix loop (preferred after plan approval when a verify command exists)
• continue_task — resume session
• task_status / cancel_task — background jobs

Example flow

1. Explore the codebase (your tools).
2. Propose a short plan to the user.
3. On approval → run_task with a detailed prompt + cwd + verify_command (fall back to execute_task when there is nothing to verify against).
4. Inspect diff / tests.
5. If incomplete → continue_task with the same session_id.