dispatching-to-codex
/dispatching-to-codex
Use when the user explicitly asks to hand a task to Codex — "use codex", "delegate this", "send it to codex", "交给codex", "派给codex", or the /delegate command. Runs the full dispatch lifecycle with frozen acceptance criteria and a binding independent judge. NEVER auto-dispatch — delegation happens only on the user's word. If a task looks Codex-shaped (bulk mechanical edits, wide read-and-report analysis, test scaffolding), you may offer the option in ONE line; if the user doesn't take it, drop it.
- Author
- ccplugins ↗
- Repository
- awesome-claude-code-plugins ↗
- Licence
- Apache-2.0
- Last repository change
- 12.08.2026
Assessment and security
- Files in this package
- 2
- Scripts
- 1
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 12.08.2026, without references to retired models or APIs.
Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.
- 1 scripts in this package. Review them before installing, as Claude may execute them.
Scripts: preflight.sh
What it does and when to use it
Use when the user explicitly asks to hand a task to Codex — "use codex", "delegate this", "send it to codex", "交给codex", "派给codex", or the /delegate command. Runs the full dispatch lifecycle with frozen acceptance criteria and a binding independent judge. NEVER auto-dispatch — delegation happens only on the user's word. If a task looks Codex-shaped (bulk mechanical edits, wide read-and-report analysis, test scaffolding), you may offer the option in ONE line; if the user doesn't take it, drop it.
Triggers from the description
Original author description
Use when the user explicitly asks to hand a task to Codex — "use codex", "delegate this", "send it to codex", "交给codex", "派给codex", or the /delegate command. Runs the full dispatch lifecycle with frozen acceptance criteria and a binding independent judge. NEVER auto-dispatch — delegation happens only on the user's word. If a task looks Codex-shaped (bulk mechanical edits, wide read-and-report analysis, test scaffolding), you may offer the option in ONE line; if the user doesn't take it, drop it.
How it works
- Iron Laws (no exceptions)
- Is it worth dispatching? (say so BEFORE starting)
- Lifecycle
- Red Flags — excuses that mean STOP
SKILL.md preview
---
name: dispatching-to-codex
description: Use when the user explicitly asks to hand a task to Codex — "use codex", "delegate this", "send it to codex", "交给codex", "派给codex", or the /delegate command. Runs the full dispatch lifecycle with frozen acceptance criteria and a binding independent judge. NEVER auto-dispatch — delegation happens only on the user's word. If a task looks Codex-shaped (bulk mechanical edits, wide read-and-report analysis, test scaffolding), you may offer the option in ONE line; if the user doesn't take it, drop it.
---
# Dispatching to Codex
Claude is the brain; Codex is the hands. You plan, freeze the bar, and judge. Codex executes on its own quota, in its own process, so your context stays small and your tokens go to judgment — not grunt work.
## Iron Laws (no exceptions)
1. **NO DISPATCH WITHOUT FROZEN ACCEPTANCE CRITERIA COMMITTED TO GIT.** If you cannot write a concrete pass/fail check, the task is not ready to delegate — do it yourself.
2. **NO ACCEPTANCE WITHOUT A FRESH-CONTEXT JUDGE VERDICT.** A FAIL cannot be overridden — not by you, not by "the diff looks fine".
3. **MAX 2 RETRIES, THEN LOUD TAKEOVER.** Every failure path ends with you doing the work and telling the user delegation failed. Silent success is forbidden.
## Is it worth dispatching? (say so BEFORE starting)
Dispatch overhead is real: branch + freeze + judge ≈ minutes of wall-clock and some of both quotas. If you could finish the task faster than the overhead, tell the user plainly: *"This is faster for me to do directly — still want Codex on it?"* Then respect their answer.
Good dispatches: bulk/repetitive edits, migrations, mechanical test-writing against a clear spec, wide "read everything, report back" analysis, output-heavy runs.
Bad dispatches: design work, subtle debugging, anything ambiguous, tiny tasks.
## Lifecycle
0. **Preflight.** Run `${CLAUDE_SKILL_DIR}/preflight.sh`. Not ready → one-line explanation (e.g. "run `codex login` once and I can dispatch"), then do the work yourself. Never nag twice in a session.
**Analysis-only shortcut:** if the dispatch edits NO files (wide read-and-report, investigation), skip ISOLATE/FREEZE/JUDGE — run `codex exec --sandbox read-only` and review the report yourself. The gate exists for merges, not for reading. Everything below is the write path.
1. **Isolate.** `git switch -c codex/<slug>` — throwaway branch; main is never touched.
2. **Freeze.** Write `.codex/ACCEPTANCE.md`: one-line GOAL + the exact commands that must pass (exit codes, not vibes). Commit it BEFORE Codex exists. See skills/writing-codex-briefs for the format.
3. **Brief + dispatch.** Self-contained brief (Codex sees zero chat history). Stock config:
`codex exec -m gpt-5.5 -c model_reasoning_effort="xhigh" --sandbox workspace-write "<brief>" 2>/dev/null`
Model/effort overrides, in priority order: `PRAETOR_MODEL` / `PRAETOR_EFFORT` env vars (use verbatim if set) → preflight reported `config=custom` (active relay provider: DROP the `-m`/`-c` flags, keep `--sandbox`, respect their config) → stock default above. Never `danger-full-access`. Always a hard timeout. Big output → `-o <file>`, never into your context.
4. **Judge.** Spawn a FRESH subagent with `agents/codex-judge.md` + repo path + branch name. Codex's work is UNCOMMITTED — the judge reviews the working tree (`git diff HEAD`), runs every frozen check, checks the bar wasn't moved. Returns PASS or FAIL + evidence.
5. **Resolve.** PASS → you commit (Codex never commits), report in plain language, append the ledger. FAIL → re-brief with `codex exec resume --last "<fix>"` (≤2 retries) or take over loudly.
6. **Cleanup.** Merge or delete the branch. Append one line to `.codex/ledger.jsonl`: `{task, model, verdict_history, wall_seconds, dispatched_at}`.
**Kill switch:** a `STOP` file in the repo root halts all dispatching — check before every dispatch.
## Red Flags — excuses that mean STOP
| Excuse | Reality |
|---|---|
| "The diff looks obviously fine, skip the judge" | Obvious diffs are where silent regressions hide. Law 2. |
| "No time to write acceptance criteria" | Then there's no way to know Codex succeeded. Law 1. |
| "One more retry will fix it" | Retry 3+ burns both quotas for a coin flip. Law 3. |
| "Codex said tests pass" | Codex reporting ≠ judge verifying. Only exit codes count. |
| "It's just a small task, skip the branch" | Small tasks on main are how main breaks. Isolate or don't dispatch. |
| "User probably wants this delegated" | Probably ≠ said so. Offer in one line, then drop it. |
Files in this package (2)
- SKILL.md 5 KB
- preflight.sh 2 KB
Versions
The database has no other skills named dispatching-to-codex.
Database snapshot: 13.11.2026 · package version 9bc5103b13e9 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100writing-codex-briefs
Use when composing the task brief and acceptance criteria for a Codex dispatch (invoked from dispatching-to-codex). Covers the self-contained brief format and how to write acceptance checks that actually protect you — weak checks are the sneakiest way a dispatch goes wrong. - 100codex-rescue
Delegate an implementation task to Codex, then Claude reviews the result. Use when asked \"codex rescue\", \"delegate to codex\", \"have codex do it\", or wants Codex to implement or fix something. - 96codex
用户输入 /codex <任务> 时,通过 peer 调度 Codex 执行该任务并汇报结果。Codex 可读取并写入当前工作目录。 - 100codex-delegate
把长代码库任务委托给本机 Codex CLI 后台执行,完成时自动通知当前会话,并可续聊同一个 Codex 会话。当用户说用 codex skill、codexskill、codex delegate、委托 codex、后台 codex、让 codex 去做、继续那个 codex 会话时使用。 - 84codex-delegate
Offer to hand heavy work to Codex (GPT-5) as a subagent so it runs on the ChatGPT plan quota instead of burning Claude context and tokens. Use at the START of any task that is large, repetitive, or asset-producing — bulk refactors across many files, generating boilerplate or… - 90codex-delegate
Orchestrate an implementation task where Claude plans and coordinates while OpenAI Codex performs the actual code edits; plans are cross-reviewed with Codex at its highest reasoning effort before implementation. Use when the user asks to "use codex", "use gpt", "use gpt-5.5",…