codex-planexec

Inspired by Claude Code's opusplan mode: use a stronger model for planning/review and a cheaper or faster CLI agent for bounded execution.

codex-planexec is a harness, not a new agent. Codex owns the task spec, review boundary, and final accept/reject decision. A worker CLI such as Kimi performs the implementation turn, while the harness captures artifacts, enforces write policy, runs validation, and returns compressed evidence.

Current MVP

• Codex plugin scaffold under plugins/codex-planexec/
• codex-planexec run --task <task.json> --repo <repo> --out <run-dir>
• Kimi-compatible worker invocation through non-interactive print mode
• Git checkpoint capture before and after worker execution
• Allowed/blocked path policy gate
• Validation command runner with full logs saved as artifacts
• Compressed review.json with changed files, policy/validation summaries, selected diff hunks, worker report, and artifact paths

Usage

Create a task file:

{
  "id": "demo",
  "goal": "Implement the requested change.",
  "instructions": "Only touch files needed for this task.",
  "allowed_write_paths": ["src/**", "tests/**"],
  "validation_commands": ["npm test"],
  "worker": {
    "kind": "kimi",
    "command": "kimi",
    "max_steps_per_turn": 20,
    "extra_args": []
  }
}

Run the harness:

codex-planexec run \
  --task ./task.json \
  --repo /path/to/repo \
  --out /path/to/repo/.codex-planexec/runs/demo

The command prints the compressed review packet and writes full artifacts to the run directory.

Review Model

Codex should read review.json first. Full logs and diffs are kept on disk for drill-down review, but they are not included in the default review packet. This keeps Codex focused on evidence instead of replaying the worker's entire context.