| name | codex-implementation |
| description | Delegate a bounded, well-specified coding task to GPT-5.5 via the Codex CLI (codex exec) and verify the result. Use for bulk or mechanical work routed to gpt-5.5 per the model policy — clear-spec implementation, migrations, repetitive refactors, test backfills, data transforms — or when the user says "ask codex", "let codex build/fix this", "delegate to gpt-5.5", or "codex exec". Not for taste-critical UI, copy, or public API design. |
Codex Implementation
Delegate implementation to GPT-5.5 through codex exec. You own the outcome: you write the brief, Codex writes the code, you verify, judge, and report.
1. Delegate or not
Delegate when all three hold:
- Clear spec — a skilled dev could start without follow-up questions. If not, tighten the spec first (files, constraints, acceptance criteria, gate command). Never forward vagueness.
- Bounded — you can name the files or subsystem it touches.
- Taste is not the bottleneck — gpt-5.5 codes fast but has taste 5. UI, copy, and public API design need taste ≥ 7: do those yourself, or hand Codex a spec so exact that no taste decisions remain.
For multi-slice projects with human gates, use the architect-loop skill; this skill is the single-task primitive.
2. Dispatch contract
RUN_DIR="$(mktemp -d "${TMPDIR:-/tmp}/codex-run.XXXXXX")"
codex exec \
-s workspace-write \
-C "$REPO_ROOT" \
--skip-git-repo-check \
-o "/result.md" \
-c model_reasoning_effort= \
</dev/null