use-codex

name

use-codex

description

Use when delegating bounded coding, audit, or review work to codex-cli (OpenAI Codex) for an independent SECOND-OPINION pass that doesn't share Claude's blind spots, or to conserve Claude tokens on heavy content-writing. Codex picker (2026-07-01): gpt-5.5 (frontier/current default), gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark (ultra-fast); gpt-5.3-codex and gpt-5.2 dropped from picker. Spec-driven, diff-first audit pattern — best for peer review of non-trivial diffs (money/auth/webhooks/migrations), bulk audits, multi-file content writes to spec. Skip for small edits or live step-through debugging. Sibling skill of use-cursor.

Use Codex

Core rule

Codex is a bounded peer reviewer + bounded implementation worker. Claude owns the plan, the prompt file, the audit synthesis, the tests, the commits, and the final answer to the user.

The peer-review use case is the killer one: codex doesn't inherit Claude's session context or rationalizations, so its audit catches what Claude misses. Use it BEFORE shipping anything non-trivial that touches money, auth, webhooks, or migrations.

When to use vs when to skip

Use codex-cli	Use Claude Agent tool subagent
Independent peer review of a non-trivial diff	Audits where Claude's tool context (existing reads) is already loaded
"Find bugs the orchestrator missed" pass	Reviews where the deciding rubric is in this codebase's CLAUDE.md only
Multi-file content writes following an explicit spec (alternative to cursor)	Reasoning that needs to feed back into THIS conversation
Bulk audit of N files where each can be reviewed in isolation	Tasks where Claude needs to inspect the result mid-flight
Cross-validation of a fix you're 80% sure is right	Trivial edits — Edit tool is faster
≥3 files of writes; saves Claude output tokens	Anything where 30s of Claude time would beat 5min of codex round-trip

Don't use codex for:

• Small targeted edits (Edit tool is faster)
• Live debugging / step-through reasoning (codex sessions don't yield mid-flight)
• Anything where you need codex's output to immediately influence your next tool call (it's async-friendly, not sync-friendly)
• Tasks that REQUIRE the full Claude conversation context (codex starts cold)

Available models (ChatGPT Plus auth)

These work on codex login --chatgpt accounts. Wrong model name returns {"type":"error","status":400,...,"message":"The 'X' model is not supported when using Codex with a ChatGPT account."} — misleading; same message for "doesn't exist" and "needs API auth". Sanity-test before long-running specs.

Evidence-backed model picker (full benchmark-backed analysis at docs/audits/2026-05-02-cli-model-task-matrix.md; numbers cite vendor blogs + Simon Willison + spectrumailab + MarkTechPost):

Model	Best for	Key benchmarks	Don't use for
gpt-5.5	Frontier review (auth/billing/migrations), reasoning-heavy debugging, terminal/agentic harness work, cybersecurity audits, greenfield architecture	Terminal-Bench 2.0 82.7% (SOTA), GDPval 84.9%, FrontierMath Tier 4 39.6% (~2× Opus 4.7), SWE-Bench Pro 58.6%	Routine work where 5.4 fits — 5.5 doubles in price above 272K input
gpt-5.4	Default everyday coding, multi-file edits, native computer use, 1M-context analysis under 272K	SWE-Bench Pro 57.7%, Terminal-Bench 2.0 75.1%, OSWorld-Verified 75.0%, GPQA Diamond 93.0%	Frontier reasoning beyond 272K (price cliff)
gpt-5.4-mini	Parallel subagent fan-out, OSWorld/screenshot tasks, bulk small-file audits with short context	OSWorld 72.1% (near full 5.4's 75.0%), τ2-bench telecom 93.4%	Long-context retrieval — MRCR v2 8-needle 64K-128K collapses to 47.7% vs 86.0% on full 5.4
gpt-5.3-codex-spark	Ultra-fast coding (slot 4 in the picker): real-time iteration, single-file tweaks, CI/CD codex-exec batches at 1000+ tps. Also the gpt-5.5-exhausted fallback (separate capacity)	Vendor claims Terminal-Bench 77.3% parity with the retired 5.3-codex; independent re-runs measure ~58.4% — 19-point gap	NEVER for hard reasoning. Plus/Team users may see CLI silently route to default — verify with /status, not /model.
gpt-5.3-codex DROPPED from picker (2026-07-01)	(was: long-horizon agent loops, terminal-heavy) — the non-spark coding variant; probe before use (may 400)	Terminal-Bench 2.0 77.3%, SWE-Lancer IC Diamond 81.4%, Cybersecurity CTF 77.6%	New work — gone from picker; use gpt-5.3-codex-spark or gpt-5.4
gpt-5.2 DROPPED from picker (2026-07-01)	(was: multi-day agent loops, input-heavy long-context) — probe before use (may 400)	SWE-Bench Verified 80.0%, AIME 2025 100%, MRCRv2 8-needle 128K-256K 77.0%	New work — gone from picker; superseded by 5.4

Default is gpt-5.5 per ~/.codex/config.toml. Always specify -m explicitly when scripting — don't rely on the default.

Model-name precision: gpt-5.3-spark does NOT exist; the real name is gpt-5.3-codex-spark. The full canonical IDs per OpenAI API docs may also include the alias gpt-5.3 for the codex variant.

Canonical -m model IDs (verbatim from the codex /model picker, 2026-07-01):

1. gpt-5.5 (current/default)      Frontier model for complex coding, research, and real-world work.
2. gpt-5.4                        Strong model for everyday coding.
3. gpt-5.4-mini                   Small, fast, and cost-efficient model for simpler coding tasks.
4. gpt-5.3-codex-spark            Ultra-fast coding model.

These 4 strings are the current picker values. DROPPED from the picker since 2026-05-29 (probe with codex exec -m <id> --skip-git-repo-check - <<<ping before relying — may now return 400 ... model is not supported): gpt-5.3-codex (the non-spark coding variant) and gpt-5.2. Note gpt-5.3-codex-spark moved from slot 5 → slot 4. Any name not routable returns 400 ... model is not supported. (The benchmark-backed when-to-use picker is the table above; this block is the authoritative name list to copy from.)

Fallback chain when gpt-5.5 is maxed out / rate-limited (READ THIS)

gpt-5.5 on ChatGPT Plus auth has a usage cap that exhausts within a heavy session — when it does, codex exec -m gpt-5.5 returns an empty / zero-byte result (no error, just nothing). This silently drops the codex seat from a thinktank or audit. Receipt: 2026-05-29 — codex returned empty across ~5 consecutive dispatches in one session while gemini/droid worked, because the 5.5 cap was exhausted. Detection: the seat's log is empty / has no ^RECOMMENDED: (or no substantive output) while sibling seats succeeded.

When gpt-5.5 is exhausted, do NOT just drop the seat — fall back in order:

1. codex with gpt-5.3-codex-spark — same codex CLI, a different (non-5.5) model that draws on separate capacity: codex exec -m gpt-5.3-codex-spark. Good for second-opinion audits + bounded review; NOT for hard multi-step reasoning (it drifts — see the table). For a frontier-reasoning slot prefer #2.
2. copilot with gpt-5.5 — GitHub Copilot CLI reaches gpt-5.5 through a SEPARATE quota (Copilot Pro/Business, not the ChatGPT Plus cap), so it keeps the frontier 5.5 perspective alive when codex's 5.5 is dead. Use the use-copilot skill: copilot -p "<prompt>" (it routes to a strong default; pin 5.5 if the CLI version supports --model gpt-5.5). This is the right fallback when the task genuinely needs 5.5-class reasoning.
3. Last resort: drop the codex seat and note the missing voice (still get the other seats) — only if both #1 and #2 are unavailable.

In a thinktank dispatch, wire this as: try codex -m gpt-5.5; if its log comes back empty, re-fire that seat as copilot -p (frontier) or codex -m gpt-5.3-codex-spark (fast) rather than synthesizing with a silently missing seat. (Cross-ref: use-copilot skill; apply the same seat-swap rule in any multi-seat council dispatch.)

Workflow

digraph codex_workflow {
    "Identify codex-fittable task" -> "Verify codex auth + model availability";
    "Verify codex auth + model availability" -> "Write task spec file";
    "Write task spec file" -> "Launch codex exec in background";
    "Launch codex exec in background" -> "Continue Claude orchestration";
    "Continue Claude orchestration" -> "Wait for codex exit";
    "Wait for codex exit" -> "Audit output";
    "Audit output" -> "Synthesize into Claude's response";
}

1. Verify codex is available + which model

codex --version             # 0.128.0 or higher
codex login status          # "Logged in using ChatGPT" expected
cat ~/.codex/config.toml | grep "^model"

If codex isn't on PATH, the user installs via npm i -g @openai/codex; do not attempt to install yourself.

2. Write the task spec file in the repo

Save at <repo>/.claude/codex_task_<short-name>.md. Don't put it in .tmp/ or anywhere gitignored — the spec should be discoverable in PRs and post-mortems.

If the task is to write a driver script for a long-running remote job: the spec should require the sub-agent to validate environment prerequisites (dependency versions, credentials, disk) BEFORE the main work starts, so the job fails fast at setup rather than deep into an expensive run. If you maintain a job-setup/verification convention, cite it in the spec.

Spec shape (parallel to use-cursor; codex tolerates more conversational tone):

# codex task: <one-line objective>

You are running as `codex exec` in non-interactive mode against the
`<branch>` branch of `<repo>` at `<absolute path>`. <Constraint sentence —
e.g. "Do not modify any code; produce a markdown audit report.">

## Context (what the orchestrator believes is true)
- Recent shipping history relevant to this task
- F* finding numbers + what they claim to fix
- Tier classifications (CRITICAL / HIGH / MEDIUM / LOW)

## Files to review (focus your attention here)

path/to/file.py # what changed, why it matters ...

## What to look for
(severity-ranked checklist of bug classes)

## Output structure
(exact markdown shape for the audit doc, with headers + tables)

## Constraints
- Do NOT modify code (audit only) OR Do NOT push (just edit)
- Do NOT git add/commit/push
- Use model X end-to-end
- Working dir is <path>
- If file > 500 lines, focus on diff regions via `git log -p -- <file>`

## Done criteria
Print: `[codex-audit] DONE <scope> critical=N high=N medium=N low=N file=<path>`
Then exit. Do NOT continue to interactive mode.

Be explicit. Codex's exec mode is competent but cold-starts every session, so spec MUST contain everything it needs (no "as we discussed earlier"). Inline the orchestrator's existing claims so codex can VALIDATE rather than re-derive.

3. Launch codex exec in the background

ROUTE THROUGH THE HEADLESS WRAPPER (2026-05-13 fix). The wrapper at a local headless wrapper script temporarily renames both Superpowers entrypoints (any auto-loading skill/plugin junctions codex would otherwise pick up) OFF for the lifetime of codex exec, then restores them via try/finally — so the global skill scan can't auto-load using-superpowers/brainstorming/TDD on every headless dispatch and hijack the response. The inline skill-bypass preamble in the spec is now defense-in-depth only; the wrapper is the load-bearing fix. The wrapper also injects --ignore-user-config by default so OAuth MCP servers from ~/.codex/config.toml cannot hang a headless run with rmcp::transport::worker: Auth(TokenRefreshFailed). If the Codex seat truly needs MCP/user config, pass wrapper-only flag --with-user-config and refresh the specific OAuth server first with codex mcp login <server>.

mkdir -p /tmp/codex_logs
cd <workspace>/REPO    # ← REQUIRED on Windows; do NOT use --cd flag

# WSL path → Windows path for powershell.exe
WIN_REPO=$(pwd | sed 's#^/c/#C:/#; s#/mnt/c/#C:/#; s#/#\\#g')

# PREFER the codex dispatch wrapper (v2 2026-05-13) over the raw headless wrapper
# for spec-from-file dispatches from WSL. Three Windows-specific failure modes
# the raw wrapper hits on multi-KB markdown specs:
#
#  1. `< spec.md` stdin-redirect through `nohup powershell.exe -File` does NOT
#     propagate stdin to codex — codex reads zero bytes then exits cleanly,
#     log shows only the prompt echo with no assistant turn.
#  2. The wrapper's `--ignore-user-config` injection (the MCP-hang defense)
#     strips `model_reasoning_effort` from ~/.codex/config.toml — codex runs
#     at reasoning=none and emits zero tool calls. Must explicitly pass
#     `-c 'model_reasoning_effort="high"'`.
#  3. Get-Content -Raw → positional prompt arg makes PowerShell native-exe
#     arg-passing split on whitespace; any markdown hyphenated token (e.g.
#     `log-frequency` in a docstring) gets interpreted as a clap-rs flag.
#
# the codex dispatch wrapper sidesteps all three: Get-Content -Raw -Encoding UTF8,
# mirrors auto-loading skill junction handling inline (mutex + rename + try/finally),
# explicit reasoning_effort, and pipes prompt via `$prompt | & codex -` (the
# `-` is a string literal positional from inside PowerShell, so the wrapper-
# rejection of leading-dash doesn't apply).
#
# Source: `a local codex dispatch wrapper`
WIN_SPEC="<workspace>/REPO/.claude/codex_task_<name>.md"
WIN_REPO="<workspace>/REPO"
nohup powershell.exe -ExecutionPolicy Bypass \
  -File 'a local codex dispatch wrapper' \
  -SpecPath "$WIN_SPEC" \
  -WorkDir "$WIN_REPO" \
  -Model gpt-5.5 \
  -ReasoningEffort high \
  > /tmp/codex_logs/<name>.log 2>&1 &
echo $! > /tmp/codex_logs/<name>.pid

# If you have a one-shot bounded prompt (no spec file), the raw wrapper still
# works because the prompt is the positional arg (no stdin handling needed):
#
#   powershell.exe -ExecutionPolicy Bypass \
#     -File a local codex headless wrapper \
#     -m gpt-5.5 \
#     -c 'model_reasoning_effort="high"' \
#     --skip-git-repo-check \
#     "Your one-line prompt here"

# Sanity-verify the process is alive past startup (catches the silent-die
# failure mode where codex exits in <1s with the os-error-2 bug):
sleep 3
if kill -0 $(cat /tmp/codex_logs/<name>.pid) 2>/dev/null; then
  echo "alive — workdir line in log:"
  grep "^workdir:" /tmp/codex_logs/<name>.log
else
  echo "DIED AT STARTUP — check log:"
  cat /tmp/codex_logs/<name>.log
fi

Flags:

• -m <model> — pin the model explicitly (don't trust default)
• --cd <abs-path> — set working dir; codex defaults to its own launch dir (often a different project!). Gotcha (proven 2026-05-02): when launched via bash+nohup on Windows, --cd C:\\dev\\projects\\REPO (backslash-escaped) causes codex to die at startup with Error: The system cannot find the file specified. (os error 2) and ZERO log output beyond that. Workarounds (in priority order): (a) cd <repo> in the launching shell first and OMIT --cd entirely (codex inherits pwd); (b) use forward slashes --cd <workspace>/REPO; (c) use the cygpath-style --cd <workspace>/... (UNTESTED). The pwd-inheritance approach is the most reliable.
• --skip-git-repo-check — for cases where the spec dir isn't the repo root
• --dangerously-bypass-approvals-and-sandbox — required for unattended background execution; safe ONLY because the spec forbids destructive ops (no commit/push, audit-only). Do NOT pass this for write-mode tasks unless you fully trust the spec
• - < .claude/...md — read prompt from stdin (cleaner than -p inline arg)

The Bash tool's run_in_background: true parameter is an alternative when you want Claude to be notified on completion via the task-notification channel.

Note: codex CLI does NOT need WSL on Windows (unlike cursor-agent). It runs natively from bash, PowerShell, or cmd. Windows path separators in --cd need double-escaped backslashes (C:\\dev\\...) when invoked from bash.

4. Continue Claude orchestration

Codex sessions take 5-15 min for typical multi-file audits with gpt-5.5, or 2-5 min with gpt-5.3-codex-spark. Use that time for parallel Agent-tool subagent dispatches, your own focused work, or other codex jobs on different scopes. Never tail the codex log via cat /tmp/codex_logs/*.log in a hot loop — wastes tokens. Check it once at the end.

For long-horizon waits (>5 min), use the ScheduleWakeup tool to set a wakeup so you're not blocked.

5. Wait for codex exit, then audit the audit

while kill -0 $(cat /tmp/codex_logs/<name>.pid) 2>/dev/null; do
  sleep 30
done
echo "exited"
tail -30 /tmp/codex_logs/<name>.log

Then:

1. Read the audit doc codex produced (NOT the log — the log is noisy token-stream output; the audit doc is the structured deliverable).
2. Diff the audit's findings against your existing understanding.
   • Findings codex agrees with: validation; mark in your tracker.
   • Findings codex flags as missed: triage by severity; the highest-severity ones may require Round N+1 fixes.
   • Findings codex calls WRONG: investigate. Codex doesn't have your full context, so it may misread; look at the specific file:line it cites.
3. Synthesize into your response to the user. Don't dump the codex doc verbatim — extract the verdict + the actionable items.
4. If you trust the codex finding and it requires a fix: dispatch a Claude Agent subagent (or write the fix directly), reference the codex audit doc in the commit message ("found by codex audit docs/audits/2026-05-02-codex-backend-audit.md").

6. Run-many pattern (audit-by-fanout)

For large audits, run multiple codex jobs in parallel with DIFFERENT scopes (not the same scope twice — that's redundant). Example split across a typical backend/web/docs audit fan-out:

• Job A (frontier model): backend audit — API + DB + cron + startup checks
• Job B (fast model): web audit — UI + types + hydration risks
• Job C (mid-tier model): docs audit — claim vs code drift in CLAUDE.md / skills

Each writes to docs/audits/<date>-codex-<scope>-audit.md. Diff against each other for cross-validation; if Job A and Job B both flag the same boundary issue (e.g., API returns X but web expects Y), it's a real bug.

Spec shape — required fields

Field	Why
Working dir as absolute path	Codex defaults to its own launch dir (often another project)
Model explicit	Default may have shifted; -m forces it
Branch + repo path	Stops codex from wandering into the wrong worktree
Files to review (read-only or write-target)	Bounds blast radius
Orchestrator's existing claims	So codex can validate, not re-derive
Severity rubric (CRITICAL/HIGH/MEDIUM/LOW with examples)	Codex's calibration drifts without anchors
Output file path + structure	The deliverable, not the log
"Do NOT git commit/push/add"	Bounds the destructive blast radius
"Do NOT switch model mid-session"	Some models silently switch; pin it
"Do NOT continue to interactive mode"	Forces clean exit
Done-criteria line	Audit anchor for the orchestrator

Guardrails

• NEVER write the literal e-v-a-l( (the four-letter identifier immediately followed by an open paren) into a codex spec file — even inside a "do NOT do this" instruction. The PreToolUse security_reminder_hook matches the contiguous 6-character substring and BLOCKS the Write tool, regardless of context (Python file, Markdown spec, plan doc, JSON, HEREDOC). Recursive trap: documenting the rule triggers the rule. Safe substitutions when telling codex what to avoid:

  • "do not use the model's e-val method (with parens) — use model.train(False) instead"
  • Quote the method NAME without parens: model.eval — and describe the call verbally as "calling that method".
  • Or embed the runnable substitute directly in a code fence and skip the negative form entirely:

    model.train(False)  # equivalent to the e-val method call
    

  • For ANY file the orchestrator writes (codex spec, plan doc, skill update, memory file, this very SKILL.md), grep the draft for the 6-char literal before submitting Write. Cost: 5 seconds. Cost of a hook block: one wasted Write turn + a context-bloating error message in transcript. Receipts: 2026-05-09 (twice — script then plan doc), 2026-05-10 (codex spec writing about the trap itself). Keep a note of this trap wherever your own project tracks recurring gotchas.

• Never paste raw markdown into bash -lc "$(cat ...)". Shell expansion can execute prompt text or corrupt the task. Use stdin-from-file (< spec.md) or - argument and let codex read it.

• No backticks in the launch command's prompt string. They get evaluated by the host shell. Reference the spec file by path.

• Process pollution check after a bad launch: ps -ef | grep -E "codex" | grep -v grep

• Do not pass HF_TOKEN, OPENAI_API_KEY, or other secrets in the prompt. Codex reads them from ~/.codex/config.toml or .env.local; the spec must say so without inlining the value.

• Codex auth gates which models work. A spec asking for gpt-5.3-codex-spark on a ChatGPT account will fail with a misleading error. If you're not sure which auth the user has, ask before dispatching, OR test with a 2-token sanity ping first through the wrapper: powershell.exe -ExecutionPolicy Bypass -File a local codex headless wrapper -m <model> --ephemeral "say only OK".

• Model-name typo silently kills the session. Always sanity-test the model string before launching a long-running spec.

• Working-directory mismatch silently kills the session. Codex's --cd flag is mandatory for non-default workspaces. The user's status pane will show codex's actual directory; verify it matches your spec's intent.

• --dangerously-bypass-approvals-and-sandbox is a safety override. Use ONLY when the spec is audit-only or has explicit no-write/no-commit constraints. Never combine with a write-capable spec without also enabling Claude-side audit of the diff before any commit.

• Codex's success report is optimistic. It says "DONE" even when the output file is incomplete. Always verify the deliverable file exists, has content, and follows the spec's structure before declaring success.

• Codex git commit -m "..." on Windows silently fails for multi-line messages with backticks, single quotes, or embedded newlines. When headless codex runs on Windows via the codex headless wrapper, its git commit -m "..." invocations route through pwsh.exe -Command '...'. PowerShell's quote-parser corrupts the message string, the shell exits 0 (silent success!), and git commits nothing. Codex's DONE report says "commits=N" but git log shows zero new entries. Files are in the working tree as untracked.

  Receipt: one codex job wrote 4 files + 1 test file cleanly. Its pwsh.exe -Command 'git commit -m "..." -m "..."' with embedded \nand single-quoted inner content failed silently. Detected on audit; manuallygit add+git commit` recovered the work. Cost: 5 min audit time, zero rework.

  The fix. In codex specs that include commit steps, always include:

  ## Commit instructions
  
  Use `git commit -F <message-file>` (write the message to a tmp file
  first). Do NOT use `git commit -m "..."` with an inline multi-line
  message — on Windows this routes through PowerShell whose quote-parser
  silently corrupts embedded newlines, backticks, and single quotes.
  
  After every `git commit`, run `git log -1 --oneline` and verify the
  expected subject is present. If absent, re-stage and retry with the
  `-F` form. Do NOT print "DONE" until `git log` confirms the commit.
  

  Orchestrator-side mitigation: after every codex DONE, run git status --porcelain | grep "^??" (untracked files = silent commit failure) AND git log --oneline -3 (commit subjects match expected). Catches the bug in 2 seconds. If untracked files surfaced by codex's spec exist, manual git add + commit is the recovery — the work itself is intact.

When to escalate / abort

• Codex exits non-zero or prints BLOCKER → read log tail (last 30 lines), diagnose, possibly rewrite the spec + relaunch
• Codex's audit doc claims something contradicting an obvious code reality → spot-check the cited file:line; if codex is wrong, document your counter-finding and DON'T relaunch
• Codex stalls past expected wall (>20 min for typical multi-file audits at gpt-5.5; >10 min for spark) → kill $(cat /tmp/codex_logs/*.pid), inspect log, decide whether to retry with a tighter spec
• Codex produces an audit doc with NO findings on a surface you KNOW has bugs → spec was too narrow; widen the file scope and relaunch

Windows CreateProcess refresh failure — orchestrator takeover (2026-05-25)

Symptom: codex returns Rejected("Failed to create unified exec process: spawn setup refresh") immediately, without reading any files or producing output.

Receipt: in one session, two separate codex dispatches on the same repo both failed with this error within 30 seconds. Retrying with the same spec failed again. The failure is a Windows-specific issue in the codex CLI's process-spawn layer; it is NOT a spec error, NOT a model error, and NOT a network issue.

Failure identification: if Rejected("Failed to create unified exec process: ...") appears 2+ times in succession on distinct specs, it is a codex process-spawn state issue, not a spec problem.

What to do (orchestrator takeover):

1. Do NOT rewrite the spec or retry codex.
2. Verify the files codex was supposed to audit via git status + targeted Grep.
3. Run the tests and lint gate yourself:

   black --check --line-length 100 api/app/ api/tests/
   ruff check api/app/ api/tests/
   PYTHONPATH=. pytest api/tests/ -q ...
   

4. Commit when tests + lint pass. Codex's review pass is skipped for this slice; document in the commit message: (codex spawn-failure, orchestrator-verified).
5. If a second-opinion review is load-bearing (money/auth/migration), fall back to a gemini --approval-mode plan pass (dead post-2026-06-18 -> use agy/droid) or a Claude general-purpose Agent audit subagent.

Recovery attempts that do NOT help: rewriting the spec, changing model, changing --cd, re-running codex login, killing + restarting the terminal. Only a full workstation restart or waiting ~5 minutes resets the spawn state.

If you keep a durable-memory file for recurring tool gotchas, log this one there for next time.

Real-world impact (backend/web/docs audit fan-out)

• Two-model parallel dispatch worked: a frontier model on a backend audit surface + a fast model on web, in parallel. Total wall time: ~15 min vs estimated 30+ min if sequential. Both wrote to separate audit files; cross-validation by diff caught one boundary issue (API field name vs web expectation) that neither model would have caught alone.
• Spec precision pays off: the spec that inlined the orchestrator's existing findings verbatim ("finding X registers 5 phantom tools at lines Y/Z, validate these registrations exist") got crisp validation findings. The spec that said "review recent backend changes" got a generic "looks reasonable" audit with no actionable items. Always inline the existing claims.
• Working-dir gotcha bit us once: codex defaulted to a sibling project's directory and produced an audit on the wrong codebase before the orchestrator caught it. --cd <abs-path> is mandatory.

Red flags — STOP and rewrite the spec

• "Audit the codebase" — too open-ended; codex will produce a review of whatever file it reads first
• "Find bugs" — needs explicit severity rubric and bug-class examples
• "Validate the fix" — needs the specific file:line + claim of what the fix does
• "Suggest improvements" — codex will rewrite working code
• Multi-step plans without explicit sequencing → split into N specs

Cross-references

• Sibling skills (the rest of the delegation matrix):
  • use-cursor (~/.claude/skills/use-cursor/SKILL.md) — bounded WRITE work in WSL on auto-model. Complementary (cursor implements, codex audits).
  • use-gemini (~/.claude/skills/use-gemini/SKILL.md) — third-opinion audit via Google Gemini CLI. Pair codex+gemini for cross-vendor cross-validation on high-stakes diffs.
  • use-claude-in-chrome (~/.claude/skills/use-claude-in-chrome/SKILL.md) — drives the user's REAL Chrome session (cookies/OAuth/MFA inherited). When codex needs to inspect or interact with an authenticated SaaS dashboard, dispatch via this skill rather than codex's own chrome-devtools-mcp (which is fresh-headless and lacks the user's session). Codex can invoke mcp__claude-in-chrome__* tools if the orchestrator loads them via ToolSearch first.
• For research/audit work that needs to feed back into the conversation, prefer the Claude Agent tool (subagent_type: feature-dev:code-reviewer, feature-dev:code-architect, general-purpose with mcp__Exa__*). Agent subagents return their report inline; codex commits to a file and you audit the file.
• Codex MCP server connectors live in ~/.codex/config.toml [mcp_servers]. Currently registered (2026-05-02): context7, github (needs GITHUB_TOKEN), chrome-devtools, exa, ref. These are AVAILABLE to codex sessions automatically — useful for specs that need exa search or chrome-devtools inspection during the audit.