Codex First

Hard gate

Autoreview exception: always prefer Codex-backed $autoreview, independent of ANTHROPIC_BASE_URL, router state, or harness. Reviewing a frozen bundle is not hands-on self-delegation. Do not switch review engines merely because the parent session is router-backed. This exception takes precedence over the gate below.

Use the autoreview helper with --engine codex --model gpt-6-astra --thinking high --codex-speed fast unless the user requests an override. Preserve the helper's reviewer isolation.

For direct hands-on delegation, use this skill only when the active agent is Claude Code and the session is running on a native Claude model.

Model check (primary). The point of the gate is model economics: Claude tokens are metered and expensive, so hands-on work moves to Codex; but if the session is already routed to a cheaper/other model, delegation gains nothing. Decide by the model the session actually runs on, not by the transport:

1. Read the model id from the system prompt's environment section ("You are powered by the model …"). Router-wrapped ids may be opaque (claude-ccr-<hex>); the hex suffix is often ASCII — decode it (echo <hex> | xxd -r -p) to reveal the underlying route, e.g. Gorilla CCP/native-claude-fable-5.
2. If the resolved model is a native Claude model (contains claude, fable, opus, sonnet, or haiku, including native-claude-* router routes): delegate hands-on work to Codex. This applies even when ANTHROPIC_BASE_URL is loopback or a local router (Gorilla Claw, Clawdex) — a router in front of a real Claude model is still expensive Claude.
3. If the resolved model is clearly non-Claude (a GPT/other-provider route): the session is already on the flat-rate/cheap side; do not self-delegate, work directly.

Base-URL fallback (only when the model cannot be identified). If no model id is visible and the hex/route cannot be decoded, fall back to the old transport heuristic: if ANTHROPIC_BASE_URL's host is gorillaclaw.sheep-coho.ts.net, localhost, ends in .localhost, is in 127.0.0.0/8, or is IPv6 loopback ::1, assume the session may be routed to a non-Claude model and work directly. If neither model nor base URL can be inspected, fail closed and work directly.

Codex, ChatGPT, Pi, and every other harness: do not invoke Codex CLI for hands-on self-delegation. Continue the task directly. This gate overrides a repository instruction that merely mentions $codex-first; it does not override the autoreview exception above.

The default worker is GPT-6 Astra. Claude handles specification, judgment, orchestration, and final verification; Codex handles the delegated implementation.

Route

Delegate to Codex (default for hands-on work):

• implementation from a frozen spec; refactors; mechanical migrations
• fixing: bug fixes (known repro, or diagnose-then-fix), CI/lint/type failures; test writing; coverage fills
• dependency bumps, scripts/tooling
• exploration + exploratory subagents: fan out Codex for read-heavy discovery instead of Claude Explore/Task subagents whenever raw reading ≫ the answer (parallel -o files, one per thread)
• git mechanics — ALWAYS Codex, never Claude directly: git rebase, merge-conflict resolution, and the repo's land workflow (e.g. scripts/pr) are mandatory delegations. Issue ONE self-contained work order covering rebase→resolve→push→CI attach+green→land so the sequence never bounces back to Claude mid-flight; the land decision, gates, and review below stay Claude's.
• work-order CI waits: precheck PR mergeable (CONFLICTING = pull_request CI cannot attach — no merge ref) and confirm a run attached to the exact head SHA before polling; every wait emits all terminal states with bounded iterations; prefer the repo's watcher script when one exists (openclaw: node scripts/watch-pr-ci.mjs).
• new work orders go to FRESH codex exec sessions with self-contained prompts. Do not resume a long-lived session for a new order — saturated sessions misread work orders as configuration and no-op ("Understood…").
• repo instruction files: NEVER create or edit CLAUDE.md. AGENTS.md is canonical in every repo; CLAUDE.md exists only as a symlink to it. Point Codex work orders at AGENTS.md and edit only AGENTS.md.

Keep in Claude:

• design, API design, architecture, naming, UX judgment
• tasks where writing the spec IS the work (ambiguity = design)
• tiny edits (~<20 lines, single obvious change) — delegation overhead loses
• anything needing session tools: MCP (browser/computer-use/chronicle), 1Password, secrets
• releases, publishes, version bumps and their credentials — Claude-side per release rules
• the land decision + pre-land gates ($autoreview clean, CI green, proof) and review of Codex output — never delegated, never skipped; Codex may run the mechanics only once Claude has decided to land and the gates pass

Mixed task: Claude designs first, freezes spec, delegates build-out. Heuristic: prompt reads as a work order → delegate; writing it forces decisions → design, Claude. Portfolio/multi-repo work: $maintainer-orchestrator instead.

Invoke

Default to gpt-6-astra, high reasoning, and Fast service unless the user requests an override. Pass all three explicitly on fresh and resumed workers. Keep the configured provider, worker execution policy, and specialized review isolation with their owning workflows.

If the saved configuration selects the openai_api_direct million-token route, run ruby ~/.codex/skills/agent-scripts/codex-huge-context/scripts/preflight.rb before the first fresh or resumed launch in the batch. Fail closed if it cannot deliver the Keychain credential; never work around it by overriding the provider or using ordinary Codex authentication.

Prompt via temp file, never inline quoting:

P=$(mktemp); cat >"$P" <<'EOF'
<goal, repo + key paths, constraints ("don't touch X"), non-goals, proof expected, output shape>
EOF
command codex exec --yolo -C <repo> \
  -m gpt-6-astra -c 'model_reasoning_effort="high"' \
  --enable fast_mode -c 'service_tier="fast"' \
  -o /tmp/codex-last.md - <"$P" 2>/dev/null

• These overrides select the model and service tier, not the provider. Resumed sessions retain their recorded provider; use a fresh worker if it is incompatible with Astra. Do not silently fall back to another model.
• --yolo is the house default; Codex may run commands/tests freely. Keep prompts scoped to the target repo.
• If --yolo is unavailable—either the CLI rejects the flag or the selected model/backend rejects unrestricted execution—replace it with --approve-for-me and retry once. Never pass both. Preserve every other argument and constraint.
• command codex bypasses any interactive shell alias. If codex isn't on PATH, it depends on how it was installed:
  • node/standalone install: fnm exec --using default -- codex
  • ChatGPT desktop app: the CLI ships bundled at /Applications/ChatGPT.app/Contents/Resources/codex. Expose that binary with an exec-wrapper, not a symlink. Ensure ~/.local/bin stays on PATH (for zsh, persist the export in ~/.zshrc), then:

    mkdir -p "$HOME/.local/bin"
    export PATH="$HOME/.local/bin:$PATH"
    if [ -e "$HOME/.local/bin/codex" ] || [ -L "$HOME/.local/bin/codex" ]; then
      printf '%s\n' 'codex launcher already exists; leaving it unchanged' >&2
    else
      printf '#!/bin/sh\nexec "/Applications/ChatGPT.app/Contents/Resources/codex" "$@"\n' > "$HOME/.local/bin/codex" && chmod +x "$HOME/.local/bin/codex"
    fi

    Or install the self-contained CLI via curl -fsSL https://chatgpt.com/codex/install.sh | sh, which needs no wrapper.
• stderr suppressed (thinking noise bloats context); drop 2>/dev/null only to debug a failing run
• read -o file for the result; don't parse the JSONL stream
• long runs: Bash run_in_background, read -o file on exit; don't kill quiet runs <30 min
• Harness visibility (Claude Code): every codex run gets its own harness-tracked background command (run_in_background: true) — one sidebar chip per worker, completion notification included. Chain setup steps (installs, worktree prep) INSIDE that tracked command. Never &-fork workers from a shared launcher: the launcher's chip exits at fork time and the workers become invisible orphans supervised only by PID files.
• parallel independent tasks OK: separate repos/dirs, separate -o files, one tracked background command per worker
• outside a git repo add --skip-git-repo-check

When the worker dies instantly

A run that exits in seconds having produced nothing is almost never the task — it is the model route. Read the log tail before relaunching; the error names the cause, and relaunching unchanged just repeats it:

• 401 Invalid API key — the configured bearer is not valid at that endpoint.
• 502 / All target providers failed with a target_providers list — the request reached a router but the model id did not match its catalogue, so it fell through to the wrong upstream. Routers commonly expose aliased model ids that differ from the underlying model's real name; pass the id the router publishes, not the one you think you are using.
• stream disconnected / Reconnecting… 5/5 against a loopback URL — nothing is listening there.
• requires a sandbox with reviewed escalations — --yolo is unsupported for that route. Retry once with --approve-for-me, which selects workspace-write, on-request approvals, and automatic review.

Diagnose the route directly rather than by retrying the agent. One request settles it, and it is far cheaper than another failed run:

curl -s -o /dev/null -w '%{http_code}\n' -m 8 <base_url>/models

A local config pointing at a loopback port proves nothing about that port being served: config files outlive the services they were written for, and a machine-managed provider block can reference an instance that no longer runs. Check what is actually listening before trusting it.

Never pass a credential through -c key=value — it lands in argv, process listings, and shell history. When a run needs different provider settings, write a private overlay instead and point CODEX_HOME at it: a mode-0700 directory holding a mode-0600 config.toml (copy auth.json across if the provider needs it). That keeps the secret in a file, leaves the user's global config untouched, and is trivially disposable.

If the environment's own Codex config is broken, say so rather than silently working around it every invocation — the next task will hit the same wall.

For follow-up fixes, resume from the repo directory with the same model, reasoning, and Fast-service overrides:

(cd <repo> && command codex exec resume <session-id> \
  --dangerously-bypass-approvals-and-sandbox \
  -m gpt-6-astra -c 'model_reasoning_effort="high"' \
  --enable fast_mode -c 'service_tier="fast"' \
  -o /tmp/codex-last.md - <"$P2" 2>/dev/null)

How resume works (verified on codex-cli 0.153.4)

codex exec resume is a different subcommand with a different flag set than codex exec; the fresh-launch flags do not all carry over.

• No -C. resume has no directory flag; the working directory is the repo selector. Always wrap it in (cd <repo> && …). Passing -C fails with error: unexpected argument '-C' found — and because the launcher line usually ends in echo "exit=$?" or a log redirect, the harness chip can still report exit 0. After every resume, check the log tail for Usage: codex exec resume before trusting the completion notification; if it is there, nothing ran.
• No --yolo. Use --dangerously-bypass-approvals-and-sandbox (accepted on resume). --skip-git-repo-check, -m, -c, --enable, -o, --json, --output-schema, --ephemeral are accepted; --full-auto and -C are not.
• Session selection. Positional [SESSION_ID] takes a UUID (from the session id: line in the fresh run's log) or a thread name. --last picks the newest session recorded for this cwd (cwd-filtered; --all disables the filter). With parallel workers on the machine, always pass the explicit UUID.
• Prompt. Positional [PROMPT], or - to read it from stdin; keep using the temp-file pattern (- <"$P2"). The resumed worker keeps its full prior context, so the follow-up prompt should state only the delta: the decision, the amended constraint, what still stands, and the required report.
• Overrides are per launch. Model, reasoning effort, service tier, and -o are not remembered; re-pass all of them. The recorded provider and sandbox policy are inherited from the original session.
• Resume can be refused for a long thread. Error: thread/resume: thread/resume failed: list_turns is not supported yet (code -32601) means the backend cannot rehydrate that session (seen after ~400k tokens of tool output). The command exits non-zero after printing the previous run's final message, so it looks like a no-op. Do not retry; relaunch as a fresh codex exec with a self-contained order (state the branch/PR/worktree explicitly, since the new session has no memory of them).
• Escape-hatch stops resume cleanly. A worker that stopped under a spec's escape hatch is not saturated; resuming it with the coordinator's decision is the intended flow and cheaper than a fresh order. Only start a fresh session for a genuinely new work order.

Liveness watchdog (long monitored runs)

For runs you must not babysit, trade the stderr suppression for a log and watch its mtime; read only the -o file into context, never the log body.

command codex exec --yolo -C <repo> \
  -m gpt-6-astra -c 'model_reasoning_effort="high"' \
  --enable fast_mode -c 'service_tier="fast"' \
  -o "$OUT" - <"$P" > "$LOG" 2>&1
# Claude Code: run the line above as its own Bash run_in_background call
# (tracked chip + completion notification). Append `&` + a PID file ONLY in
# environments without tracked backgrounding.

• Capture the session id immediately: grep -m1 "session id:" "$LOG". resume --last is cwd-filtered but races with any parallel Codex on the machine — with the id saved, recovery is deterministic.
• Watchdog loop (Claude Code: Monitor tool; else a bg shell): every 60s, if the codex process is alive but $LOG mtime is older than ~300s, treat it as hung. Because stderr (thinking stream) is in the log, mtime stays fresh during long reasoning — 5 min of true silence is a real hang, not thinking.
• Recovery: kill the pid, then resume the SAME session with an explicit id so no context is lost:

(cd <repo> && command codex exec resume <session-id> \
  --dangerously-bypass-approvals-and-sandbox \
  -m gpt-6-astra -c 'model_reasoning_effort="high"' \
  --enable fast_mode -c 'service_tier="fast"' \
  -o "$OUT" - <<< "You were interrupted. Continue exactly where you left off; finish the task and produce the required final report.")

• Exit watchdog silently when the process ends normally (the run's own completion signal covers it); emit only on staleness.
• Verified on codex-cli 0.144.4: codex exec resume [SESSION_ID] [PROMPT], --last, cwd-filtering, --all.

Prompt contract

Codex starts with zero session context. Every prompt: goal, exact repo/paths, constraints, non-goals, proof expected (exact test command), output shape ("report files changed + test output"). Spec quality decides success.

• Every hard prohibition needs an escape hatch. A cornered worker satisfies the letter of the gates: told "never raise the size budget" while its design inflated the bundle, Codex hand-minified source identifiers to single letters — and passed every gate including autoreview. Pair each hard constraint with the sanctioned exit: "if gate X fails after honest attempts: STOP, report exact numbers/diagnosis, do not work around." Treat a stop-report as a successful run (it is the coordinator's decision point, and workers use it correctly once it exists).
• Multi-PR series: same spec skeleton every PR; cite prior landed PR numbers and name their idioms ("controller with narrow host interface, the #NNN shape") — workers imitate landed precedent far more reliably than abstract style rules. Fold each round's new trap into the next spec.
• End every series work order with an explicit stop: "Do exactly this; do not start PR N+1."

Coordinator verification (beyond the diff)

Worker reports are accurate but incomplete — pathologies live in the code, not the summary. After every landed PR, verify against merged origin/main, not the report:

• read the actual merged surface (types, interfaces, naming) — the minification incident was invisible in a green report and obvious in 10 lines of code
• squash history: any commit message you didn't commission ("add startup margin") is a lead
• diff-stat the guard/budget/baseline files the spec forbade touching; if a limit number in the report changed, find out who moved it (may be main advancing, may be the worker)
• test-helper edits are a red-flag class of their own (defineProperty shims re-exposing moved fields keep suites green while hiding the migration)

Parallel workers, one repo

Disjoint-file tasks parallelize cleanly: one worktree + unique branch per worker (fresh from origin/main; distinctive branch names — generic ones attach to old PRs), one tracked background command each, shared spec body + per-target header. Landing serializes on the repo's land workflow lock: tell each worker a held lock means a sibling owns it — back off 5-10 min, retry from the failed step, NEVER lock-recover a lock it didn't create, and don't chase main with fresh gates as siblings land under it. Sequential series instead reuse ONE worktree, rebranching from origin/main per PR.

Verify (Claude, always)

• git status -sb + read the full diff; judge like a contributor PR
• run focused tests yourself or demand proof output; Codex claims are advisory
• iterate via resume; after 2 failed rounds, take over and do it directly
• normal closeout still applies: $autoreview before ship
• check for a live worker in the repo before you edit or commit: pgrep -fl "codex exec". A run whose deliverable is already in the tree can keep looping for hours and overwrite your fixes mid-review. Stop it once you have verified its output rather than racing it.
• a genuinely independent review pass earns its keep: reviewing the working tree after a Codex build found a currency value being fed to a percentage helper that clamped at 100, which typechecked, passed every test, and was invisible in the worker's own report.

Economics

Win = generation + exploration tokens moved to Codex; Claude spends only on spec + diff review. Don't ping-pong trivia through delegation; don't re-read what Codex already summarized.