Codex Delegation

Codex Delegation

This skill is the policy and the loop for delegating work to Codex without a /codex:* command. Presenting output stays governed by the codex-result-handling skill; prompt shape stays governed by the gpt-5-4-prompting skill. Do not duplicate either — defer to them.

When to delegate — and when not

• Delegate when an independent deep pass adds real value: a second opinion on a risky change, security or architecture analysis, an adversarial review of a substantive diff, or an implementation that can run in parallel while the session continues on something else.
• Keep trivial work local. A rename, a one-file edit, a question Claude can answer directly — spawning a background agent there is overhead, not help.
• Bench evidence, 2026-08-08 (docs/bench-pilot-2026-08-08.md): on the pilot's one comparable cell — a well-localized single-file bugfix with a reproducible symptom — a Codex-delegation arm matched solo Claude's fix rate exactly (3/3 both arms) while costing ~30% more wall time. n=3, one task shape: a direction, not a law. For fixes of that measured shape, default to solo unless the change carries a genuine risk or second-opinion dimension; broader bugfix classes keep the guidance above unchanged until more cells replicate.
• Never delegate when the user explicitly asked Claude to do the work personally.
• Run one delegated job of a class at a time (one task, one review). Before firing a new one, run status --json and confirm nothing of the same class is still active.

Disclosure

• Announce every delegation in one short line when starting it — what is being delegated and that it is going to Codex. Never silently spawn CLI work.

The loop: fire, await, collect

Fire — start the job in the background:

node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-companion.mjs" task --background [--write] [--profile <deep|fast>] [--effort <tier>] "<prompt>"

• The output line started in the background as <jobId> carries the job id. Record it.
• Add --write only when Codex is supposed to edit files; leave it off for analysis and second opinions.
• --profile <deep|fast> is accepted on task (and its rescue forwarder) only; review and adversarial-review reject it.
• This fire → await → collect loop is for task only. Reviews detach differently — see below.

Await — issue the wait as a background Bash task so the harness notifies on completion instead of blocking the session:

Bash({
  command: `node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-companion.mjs" status <jobId> --wait --timeout-ms 1800000 --json`,
  description: "Await Codex job",
  run_in_background: true
})

• On wake, parse the JSON. If waitTimedOut is true and the job is still active, re-issue the same wait instead of assuming failure — a timeout is not a verdict.

Collect — fetch the stored output once the job is terminal:

node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-companion.mjs" result <jobId> --json

• Present it per the codex-result-handling skill — including its rule that review findings are never auto-applied: present them, stop, and ask which ones to fix.

Reviews: one background step

review and adversarial-review have no companion-side enqueue: the companion parses --background but always runs the review in the foreground, and no started in the background as <jobId> line is printed. Delegate a review by running the whole command as a background Bash task — the wake IS the collect step, carrying the rendered review on stdout:

Bash({
  command: `node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-companion.mjs" review [--base <ref>]`,
  description: "Codex review in background",
  run_in_background: true
})

• adversarial-review [--base <ref>] [focus text] detaches the same way.
• The run still records a tracked job, so status and result work on it afterwards.

Brief shape

• Prefer a design-to-verdict brief for analysis delegations — hand Codex a concrete design and ask for numbered verdicts and corrections — over an open-ended "design everything" ask. When a delegated run goes likely-dead or hangs, refine by narrowing the ask (and dropping to --effort high for verdict checks), never by re-sending the same brief. Ledger evidence, 2026-08-07: two open-ended xhigh analysis briefs died or hung; both narrowed re-delegations returned approve-with-corrections and merged.

Routing

• Deep analysis, security review, architecture validation → task or adversarial-review; raise --effort when a verdict needs more depth.
• Leave model and effort unset by default — Codex chooses its own defaults.

Failure honesty

• If the companion errors, or Codex was never actually invoked, report that plainly and stop. Never substitute Claude-authored output as the delegate's.