Codex First

Launch flags — read first, copy verbatim

Default every Codex worker, fresh or resumed, to GPT-6.1 Sol, high reasoning, Ultrafast service tier (service_tier="ultrafast"). Pass all three settings unless the user requests an override.

# fresh
codex exec --yolo -C "$WT" \
  -m gpt-6.1-sol -c 'model_reasoning_effort="high"' \
  --enable fast_mode -c 'service_tier="ultrafast"' \
  -c 'model_auto_compact_token_limit=260000' \
  -o "$OUT" - < "$ORDER"
# resume: pass the same flags again; resume does not inherit them
codex exec resume "$SID" --dangerously-bypass-approvals-and-sandbox \
  -m gpt-6.1-sol -c 'model_reasoning_effort="high"' \
  --enable fast_mode -c 'service_tier="ultrafast"' \
  -c 'model_auto_compact_token_limit=260000' -o "$OUT" -

• Ultrafast has a context ceiling (measured 2026-10-13, direct Responses probes, gpt-6.1-sol + gpt-6-astra): served ultrafast up to ~282k prompt tokens, flaky at 288–294k, fast at ≥300k — silently, HTTP 200. Config default compacts at 400k, so long workers drift to Fast; the 260k compact cap keeps them Ultrafast. Rollouts record only the requested tier; prove the served tier with a direct probe (response.service_tier).
• Context: workers get the base 500k window. Add -p sol61-1m (or -p astra-1m) only when the order needs a whole-repo context; see $codex-huge-context. Those profiles forfeit Ultrafast above ~280k.
• Copy all model/reasoning/tier flags into wrappers; override only on user request. Verify launch arguments (ultrafast default; requested fast is valid).
• For requested Fast reviews, use --codex-speed fast.

Security audits and cyber-classifier refusals → Daybreak Red (Peter, 2026-10-15/18)

Security audit lanes (threat models, injection/secret-exposure audits, permission inventories, security review of hardening diffs) launch on Daybreak Red from the start (Peter 2026-10-18: "use daylight red model for security audits"). Implementation workers stay on Sol.

Worker stops with This content was flagged for possible cybersecurity risk → relaunch or resume the same order on Daybreak Red (Peter: "use daybreak red if you hit a blocker with regular models"). Don't rephrase to dodge the classifier or hop between regular models.

• Direct API only: the ChatGPT-account route returns 400 not supported when using Codex with a ChatGPT account. Add -c 'model_provider="openai_api_direct"' (API key via the provider's auth command in ~/.codex/config.toml).
• Slug: the model_catalog_json entry whose display_name is Daybreak Red. Keep it out of public artifacts (commits, PRs, issues, logs) per public model naming.
• No Ultrafast: drop --enable fast_mode / service_tier; keep model_reasoning_effort="high".
• Verified 2026-10-15 with a one-line codex exec probe on the direct provider.

Capacity fallback (Peter, 2026-10-12)

Org caps gpt-6.1-sol at 40M tokens/min. Worker or resume fails with Rate limit reached for gpt-6.1-sol ... tokens per min (TPM), Selected model is at capacity, or overloaded → resume that same session on GPT-6 Astra, high, Ultrafast immediately (Peter, 2026-10-12: "use astra if sol has rate limits"; Astra serves ultrafast, verified 2026-10-13); no waiting loop:

codex exec resume "$SID" --dangerously-bypass-approvals-and-sandbox \
  -m gpt-6-astra -c 'model_reasoning_effort="high"' \
  --enable fast_mode -c 'service_tier="ultrafast"' \
  -c 'model_auto_compact_token_limit=260000' -o "$OUT" -

• Resume onto Astra keeps the thread (verified 2026-10-12: Sol-created session resumed on Astra and replied). Fresh orders at capacity: same flags with codex exec --yolo -C "$WT"; a fresh launch that dies on capacity already has a session id in stderr (session id: …) — resume that, don't relaunch.
• Back off only when Sol and Astra are both limited; then alternate with growing sleeps.
• Wrapper retry triggers must match all three strings (rate limit|at capacity|overloaded); matching only "rate limit" made capacity failures exit instead of falling back.
• Autoreview at capacity: --model gpt-6-astra --thinking high --codex-speed fast. If autoreview refuses because its provider/route doesn't match operator config, don't change config — review with a fresh-context subagent (diff + brief) and say so in the PR.
• New workers still start on Sol 6.1 Ultrafast; fall back per worker, only on the capacity error.
• Restarts and recoveries: ~/.codex/skills/agent-scripts/codex-first/scripts/codex-resume does the fallback, busy check, safe stop, and classified exits (resume and watchdog).

Wrapper and process hygiene (2026-10-12/13)

• Never edit a running wrapper script in place: bash reads scripts incrementally, so live runs execute the new bytes at their old offset. Write x.new, bash -n, chmod +x, then mv (rename keeps the old inode for running copies; forgetting chmod gives exit 126).
• Never run codex exec under set -e in a wrapper: a rate-limit death exits non-zero and ends the script before its fallback runs (2026-10-13, a 27-worker fleet). Capture the code (… && rc=0 || rc=$?) and classify it, or hand the session to codex-resume.
• Killing a wrapper orphans its codex exec children; the next resume then fails with thread … already has an active writer. Stop and resume with codex-resume --stop "$SID": it finds the worker through its open rollout file (fresh launches don't carry $SID in argv, so pgrep -f misses them) and refuses while scripts/pr or git runs beneath it. A leftover ~/.codex/thread-writer-locks/$SID.lock file without a live holder is harmless.
• To redirect a running worker (new design/priority), codex-resume --stop -p "$ORDER_FILE" "$SID" (or --message "…"); its worktree changes and context survive.
• To retire a worker for good, do not use codex-resume: --stop always resumes afterward (2026-10-18, a cancelled worker restarted in its deleted worktree). Kill the native codex binary (pgrep -P <node-shim-pid>), then the node shim and its wrapper.

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 (defaults: 6.1 Sol, medium, Ultrafast) unless the user requests an override. It tries the ChatGPT subscription first and, on an auth failure, retries once on the configured direct-API provider; pass --codex-config 'model_provider="openai_api_direct"' only to skip the subscription attempt. 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-6.1 Sol on the Ultrafast tier. 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). Never gh run watch or gh pr checks --watch in work orders: gh run watch polls every 3 s by default (~1,200 calls/hour), and two such watchers drained a shared 5,000/hour writer token twice on 2026-10-15. Poll one exact run with gh run view <id> --json status,conclusion at most every ~10 minutes, or let the repo's landing tool own the wait.
• 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

• Run timeout 8 git --version; for an Xcode-beta hang, set DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer in the session and every order.
• For openai_api_direct, run ruby ~/.codex/skills/agent-scripts/codex-huge-context/scripts/preflight.rb before the batch's first fresh/resumed launch. If Keychain delivery fails, stop; never override the provider or use ordinary Codex auth.
• Keep the configured provider; resume inherits it. If incompatible with 6.1 Sol, start fresh; never silently change models.
• Use repo-scoped --yolo, temp-file prompts (never inline quoting), and -o results (not JSONL).
• 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
• For Fast/catalog checks, CLI setup, or execution details: read launch details.
• On failure or provider changes: troubleshooting; never put credentials in -c key=value.
• Before resuming/monitoring: resume and watchdog. Resume and restart through scripts/codex-resume (--dry-run first when unsure).

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.

• Name the domain skill. When the repo has a skill for the area being changed (e.g. openclaw .agents/skills/solid/SKILL.md for Control UI/Solid work), the work order says to read it, plus the references it routes to, before the first edit. Codex auto-loads only the AGENTS.md chain from its working directory, so a skill linked from a nested AGENTS.md is easy to miss.

• 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."

• Work-order proof: PR CI is the default broad proof; open a ready PR once the build and focused local tests/typecheck/lint/format pass, without waiting for pre-PR broad suites; use Crabbox/Testbox only for gaps CI/local cannot cover and only if granted immediately (queued → local; no Testbox → AWS → Hetzner ladder); fix real failures, read CI's pre-existing/new classification, and never rerun/re-push just for green. Keep untrusted code secretless, live credentials authorized, and Gateway/state isolated.

Before reviewing/landing or coordinating workers, read coordination; check for live workers before editing/committing.