fire

name	fire
description	Delegates a well-specified implementation task to Codex CLI (or, via --with, Claude Sonnet 5 or opt-in GLM-5.2) in the background. Use when the user asks to hand work to Codex, or for substantial spec-able work - features, refactors, migrations, boilerplate; offer first unless the routing policy is autonomous. Not for small fixes or ambiguous design; never fire silently.

Fire - hand the ticket to the sous-chef

You are the head chef. Codex is your sous-chef: a strong implementer with no memory of this conversation, executing exactly one written ticket per run. Two rules fall out of that: everything the worker needs goes on the ticket, and everything the run produces lands on disk - the job dir, not this conversation, is the record of the run.

When to fire vs. cook it yourself

Fire when ALL of these hold:

• The task is implementation, not design - you already know what the end state looks like.
• It spans multiple files or is mechanical bulk work (refactors, renames, migrations, boilerplate, test scaffolding).
• You can state "done" as checkable criteria (tests pass, command output, types compile).

Cook it yourself when ANY of these hold:

• One-file or few-line surgical fix - the delegation round trip costs more than doing it.
• The approach is still ambiguous - resolve design questions first, then fire.
• The task depends on conversation context that can't be written into a ticket.

Delegation sends code to another vendor and spends their quota: if the user didn't explicitly ask for it, propose it in one line rather than firing silently. Exception - an autonomous routing policy in the user's CLAUDE.md pre-authorizes the delegation; the one-line announcement then replaces the proposal: announce and fire.

Preflight (all deterministic, run before writing the ticket)

1. Git repo with at least one commit - git rev-parse HEAD succeeds. Codex refuses non-repos by default, and diff review needs a baseline. If not: tell the user to git init / make an initial commit first.
2. Profile exists - test -f ~/.codex/sous-chef.config.toml. This check is load-bearing: Codex silently ignores a missing profile (exit 0, runs under the user's own defaults - possibly no sandbox at all). If missing: stop and offer /sous-chef:mise.
3. Job directory - mint one per fire: JOB=$(mktemp -d "$SCRATCHPAD/fire-XXXXXX") ($SCRATCHPAD = your session scratchpad directory; substitute its absolute path). Never share ticket/result/log paths between jobs - fixed paths let concurrent or sequential runs clobber each other and serve a stale result as a fresh success.
4. Snapshot the tree - if git status --porcelain is non-empty, warn the user their uncommitted changes will share the tree with Codex's edits (suggest committing/stashing first), and either way save the baseline: git diff > "$JOB/pre-fire.patch"; git status --short > "$JOB/pre-fire.status". At plating you review Codex's delta against this baseline, not the raw diff.

Writing the ticket

Write the ticket to $JOB/ticket.md using the template in references/ticket-template.md. The contract is XML-block-structured because Codex follows explicit contracts far better than prose requests:

• <task> - the concrete job, with enough repo context to orient a fresh agent.
• <done_when> - checkable success criteria. Declarative beats imperative: give success criteria, not step-by-step instructions.
• <files> - files/dirs to create or modify, and files/dirs NOT to touch. Without this, Codex will make assumptions.
• <interfaces> - exact signatures, types, or API shapes other code depends on.
• <constraints> - hard rules (no new dependencies, match existing patterns, etc.).
• <verification> - exact commands Codex must run before finishing (test, typecheck, lint).
• <output_contract> - what the final message must contain: what changed, what was verified with actual command output, what was left undone and why.
• <follow_through> - the default when routine questions come up: keep going, make the reasonable choice, record it; stop only on hard blockers.
• <action_safety> - stay narrow; no unrelated refactors; no deleting code it doesn't understand.

Repo-level standards (build commands, conventions, do-not-touch areas) belong in the repo's AGENTS.md, which Codex reads automatically on every run - don't duplicate them on the ticket. Run /sous-chef:mise once per repo to set that up.

Choosing the worker - --with

The arguments may begin with --with <worker>; strip it before treating the rest as the task description. Workers:

--with	Worker	Route
(absent) / codex	GPT-5.5 via Codex CLI	the default invocation below
sonnet	Claude Sonnet 5, user's own subscription	references/glm-routes.md Route C
glm	GLM-5.2	references/glm-routes.md Route A or B, whichever is installed

Loose phrases ("fire with sonnet", "use GLM for this") mean the same thing - --with is just the unambiguous spelling, immune to task text that happens to mention a model name. The ticket, job dir, and plating are identical for every worker; only the invocation changes. Preflight differs per worker: step 2's Codex-profile stop applies to the Codex route only - Route C's preflight is just command -v claude.

Firing

Run from the repo root (workspace-write scopes writes to the working directory), in the background - never in the foreground, where the Bash timeout ceiling kills long runs. Tool-level backgrounding is the only backgrounding: the command itself must NOT contain &, nohup, or disown, or Claude Code will track a wrapper that exits immediately, fire a false completion notification, and leave the real worker orphaned.

Bash (run_in_background: true), cwd = repo root:
env -u CODEX_API_KEY -u CODEX_ACCESS_TOKEN codex exec --profile sous-chef \
  --output-last-message "$JOB/result.md" \
  - < "$JOB/ticket.md" > "$JOB/job.log" 2>&1

Notes on the invocation:

• --profile sous-chef loads ~/.codex/sous-chef.config.toml (workspace-write sandbox, approvals never - it never pauses for input that will never arrive). Model and reasoning effort deliberately fall through to the user's ~/.codex/config.toml defaults.
• env -u CODEX_API_KEY -u CODEX_ACCESS_TOKEN pins the run to the user's codex login (ChatGPT subscription) auth - those two are the only env vars that override it in codex exec, and if either is set the run silently bills per-token instead. (OPENAI_API_KEY is NOT read for auth by current Codex, and unsetting it would break custom providers that use it as their env_key.)
• Prompt goes via stdin (- <) to avoid shell-quoting damage to the ticket.

Then tell the user, in one or two lines: what was delegated and to which model (read model from ~/.codex/config.toml - don't assert a model you didn't check), that it typically takes 5–20+ minutes at high reasoning effort, a paste-ready tail -f "$JOB/job.log" (absolute path) to watch it cook - warning that stray MCP transport noise early in the log is usually harmless, not the run failing - the ticket at $JOB/ticket.md for what was ordered, and that they can cancel anytime. Offer progress ticks (below) as a clause they can opt into by replying, not a blocking question.

To route the ticket to GLM-5.2 (user opt-in) or to Claude Sonnet 5 on the user's own subscription (no extra key - the natural fallback when Codex hits its usage limit mid-serve), see references/glm-routes.md - same ticket, different worker invocation.

While it cooks

Do NOT poll - polling loops against a running Codex job are the documented way to incinerate quota while producing nothing. Work on something else or end your turn; the backgrounded job re-invokes you when it exits. If you must watch for a condition, arm a single Monitor with an until-loop on $JOB/job.log matching terminal states (completion AND error signatures like ERROR:, stream disconnected), not a poll loop.

A long run need not be a silent one. If the user opted into progress ticks (or serve turned them on), arm a self-paced wakeup loop - /loop with no interval, or whatever wakeup scheduler the harness offers - with the absolute $JOB/job.log path in the armed prompt: a tick must not depend on conversation memory to find its log. Each tick, read the tail of the log and report ONE distilled line ("three files edited, tests running now"), then re-arm, pacing ticks a few minutes apart (under five, so each wakeup lands on a warm prompt cache). A tick that finds a terminal state in the log says nothing and does not re-arm - completion re-invokes you regardless, and reporting the outcome, like plating, belongs to that turn. This is not the forbidden polling: it is bounded by the run, disarms itself, reads a local file, and never queries the worker. No wakeup facility? The tail -f handoff stands alone.

If the user cancels: kill the background task, then run git status + git diff - the workspace-write worker may have left a half-applied change. Show the user what's there and let them decide keep or revert.

Plating - when the job exits

1. Check the outcome before trusting the plate. If the job exited non-zero, or $JOB/result.md is missing or empty, the run failed - read the tail of $JOB/job.log, show the user the error verbatim, and offer one rerun or taking over yourself. Two errors worth naming for the user: "You've hit your usage limit" means wait for the plan's 5-hour window to reset (or escalate plans) - or offer to continue now with --with sonnet (Route C); a persistent 401 means their codex login needs redoing. Never present a missing result as a clean outcome. (MCP transport errors near the top of the log are usually harmless noise from the user's Codex-side MCP servers - the real signal is the last lines.)
2. Glance at the log's opening banner: its sandbox: line is ground truth for what actually ran. If it isn't workspace-write, say so.
3. Read $JOB/result.md, then compare the post-baseline changed file set (git status/git diff minus $JOB/pre-fire.*) to the ticket's <files> Touch list. Outside-list paths are unresolved until classified: paths confirmed as another session's concurrent edits must be named with the warning concurrent edit detected - these changes are NOT part of this run's review and excluded from the worker-attributed delta, while paths that are the worker's own out-of-scope changes must be reverted or explicitly flagged to the user before the run can be accepted. This is a path-level check: it cannot catch a concurrent session editing a file that is on the Touch list - those edits merge into the same file's diff and only the line-by-line read in step 4 will separate them, so treat a Touch-listed file that changed more than the ticket asked as suspect too. Then review Codex's actual delta against $JOB/pre-fire.patch - don't attribute the user's own WIP to Codex.
4. Review the diff carefully, line by line. Codex is a competent implementer that makes wrong assumptions without checking - that is exactly the failure mode you're here to catch.
5. Run the <verification> commands yourself. Codex's claims are not evidence; command output is.
6. Then either:
   • Accept - summarize what shipped and what you verified, plus the token usage from the log's closing summary when the log carries one (the tokens used block near the end of job.log; Claude-worker routes emit none - say token usage is unavailable) - quota spend is otherwise invisible to the user. Then add the job to the running tab: append one line to ~/.sous-chef/ledger.jsonl (mkdir -p ~/.sous-chef first) of the form {"ts":"<UTC ISO-8601>","repo":"<repo basename>","skill":"fire","model":"<model from the log banner>","tokens":<total from the closing summary>}. If the log carries no token summary, skip the ledger line - never invent a number.
   • Send ONE delta: a fresh fire (new $JOB, short ticket that states what the previous run got wrong, quotes the failing output, and scopes the fix). Do NOT use codex exec resume - resumed sessions rebuild config from the user's defaults, silently dropping the sandbox, and --last may grab a different session entirely. Fresh run + state on disk is the reliable path.

Cap follow-ups at two rounds. If it's still not right after two deltas, take over and finish it yourself - further debate has diminishing returns.

If Codex is unavailable

If codex is missing, unauthenticated, or the profile doesn't exist (preflight step 2), say so and offer to run /sous-chef:mise - don't silently implement the task yourself without telling the user the delegation failed.