| name | fable-gpt |
| description | Use when a task involves heavy implementation, debugging, test fixing, refactoring, or multi-file edits that Codex could execute, or when orchestrating Claude + Codex together in a session. |
Fable-GPT: Claude orchestrates, Codex executes
Role split
- Claude (main thread): planning, repo understanding, architecture decisions, task decomposition, final review, small surgical edits.
- Codex (codex plugin): heavy implementation, debugging, test fixing, refactors, multi-file edits. Runs on the ChatGPT subscription — zero Claude tokens for execution. Global default is
gpt-5.5 at xhigh reasoning (~/.codex/config.toml); do not pass --model/--effort unless deliberately overriding.
How to delegate
| Situation | Route |
|---|
| Quick, bounded task | /codex:rescue (foreground) |
| Big, open-ended, or long-running | codex-rescue subagent (already model: sonnet, Bash-only — passes agent-model-guard) or --background; poll /codex:status, fetch /codex:result |
| Continue prior Codex work | say "resume"/"continue" (maps to --resume-last) |
| Routine pre-commit review or second-opinion diagnosis | /codex:review — default here before spending Claude subagent tokens; reserve /code-review or /adversarial-review for high-stakes diffs |
Task prompt shape (keep it focused)
Goal: [one sentence]
Files: [paths]
Constraints: [style, don't touch X]
Done when: [verification command + expected result]
One task per delegation. Split unrelated work into separate rescues.
Workspace root gotcha: Codex's sandbox root is the session's cwd, not the task's target repo. If the target repo differs from cwd, the codex-rescue subagent and /codex:rescue will spawn a task that can't write ("patch rejected: writing outside of the project") — Codex burns the whole run discovering this. Invoke the companion runtime directly instead:
node "<plugin>/scripts/codex-companion.mjs" task --prompt-file <spec.md> --cwd "<target-repo>" --write --background
Long specs go in a --prompt-file (avoids shell quoting). Job state is keyed to the --cwd value — pass the SAME --cwd to status/result/cancel or the job appears not to exist. Errors print to stderr; don't 2>/dev/null status calls.
After Codex finishes — never accept blindly
git diff — read what actually changed.- Run the "Done when" verification command yourself and read the output.
- Wrong or sloppy → re-delegate with specific corrections (use resume); small issues → fix inline.
When NOT to delegate
- Trivial edits (faster inline than writing the delegation prompt).
- Architecture or product decisions — Claude's job.
- Anything under CLAUDE.md Hard Stops (deploy, migrations, auth/billing config, external comms).
- Tasks needing conversation context Codex doesn't have — inline the needed context into the task prompt or keep it in Claude.
