How to use OpenAI Codex via MCP tools for code generation, plan review, and collaborative implementation loops. Use when calling Codex for implementation, review, or any cross-model collaboration.
Master task dispatcher. Simple tasks (≤2 files, ≤30 lines, clear spec) are done inline by the main agent. Complex tasks are routed to bg-codex-implementer as a background agent. Triggers on: /dispatch, implement this, fix this, debug this, dispatch task.
Benchmark one or more model checkpoints (2000 games/side) with auto-logging. Triggers on: /bench, benchmark this model, run benchmark, bench checkpoint.
Analyze a slow process, identify bottleneck, write spec, delegate optimization to Codex. Triggers on: /speedup-analysis, speed this up, optimize performance, this is too slow, why is this slow.
Chain training → export → benchmark as one background pipeline. Triggers on: /train-and-bench, train and benchmark, run training pipeline, train this model.
Full system status: background tasks, running processes, GPU/CPU/memory utilization, lock files. Triggers on: /check-tasks, /status, task status, what's running, system status, check background.
Batch 5-20 rules/card questions into one Haiku agent call that reads PDF + docs once and answers all. Replaces per-question rules-lawyer calls. Triggers on: /rules-batcher, batch rules questions, look up these cards, verify these rules.
Explore codebase with Haiku, synthesize a tight machine-readable implementation spec. Output feeds directly into /feature-coder or /tdd-fixer — no plan-audit loop. Triggers on: /spec-writer, write a spec, spec out this feature, spec before coding.