Codex Delegate

Codex Delegate Skill

Claude is the supervisor. Codex CLI runs the mechanical work. Claude plans, constrains scope, reviews the diff, and verifies outcomes.

Why use this instead of raw codex exec

This wrapper is not cosmetic. It exists because every shipping task that bypasses it loses one of three things that have caused real incidents:

Token-saving ratios (directional dogfood data — measured treatment, ESTIMATED control)

From a 6-round mixed-workload session (awesome-agentic-ai-zh 2026-05-14, see agent-collab-skills/docs/measured-benefits.md). Honest methodology note (per that doc's own words): the delegated arm was measured (bytes/3 proxy), the Claude-inline control arm was "estimated, not measured, because we didn't run the control condition" — treat every ratio below as directional, not a precise multiplier. The 17-22× row was additionally produced on the since-retired gemini lane (R4 of that session) and has not been re-measured on Codex.

The wrapper helps for token-heavy mechanical work. It does not help for judgment-heavy work — be honest about which one your task is.

Anti-patterns this skill prevents

• F11: Codex over-applies a sweep rule to documentation describing that rule. Prevented by agent-task-splitter brief that explicitly lists in-scope files.
• F14: Operator (Claude) defaults to Bash("codex exec ...") despite the rule saying use Skill("codex-delegate"). Prevented by a PreToolUse hook on the operator side (template below).
• Codex hang: 25-minute zero-byte run on 2026-05-14 because raw codex exec was invoked without </dev/null. Prevented by wrapper's mandatory stdin closure.

CLAUDE.md snippet to enforce routing

Drop this into your ~/.claude/CLAUDE.md (or repo-level CLAUDE.md) to make the rule operational rather than aspirational:

## Codex routing rule (enforced)

For any task that is mechanical, multi-file, has a clear pattern, or
batch-edits ≥ 3 files (canonical tripwire: fable-style-project-harness/docs/codex-delegation-policy.md — single source; older ≥5 statements are superseded): invoke `Skill("codex-delegate", args="brief=...")`.
Do NOT use `Bash("codex exec ...")` for shipping tasks — it loses stdin
closure (codex hangs), the `.result.json` contract (no structured status),
and the brief template (Codex drifts F11/F12).

Optional mechanical enforcement: a PreToolUse hook on Bash that nudges raw
`codex exec` toward `Skill("codex-delegate")`. Reference implementation:
https://github.com/WenyuChiou/dotfiles-claude/blob/main/hooks/check_codex_skill_routing.py

→ The hook is non-blocking (warns only) so you can still run codex --version / codex exec resume / codex exec review without friction. Replace return 0 with return 2 in the hook to make it block instead of warn.

Hard rules

• Before invoking the wrapper, run codex --version via Bash. If it fails, stop and tell the user to npm install -g @openai/codex.
• When calling codex exec directly (not through the wrapper), close stdin explicitly (</dev/null); the wrapper handles this internally.
• Wrapper run leaves machine-readable status at <log-file>.result.json. Acceptance is Claude's job, not the wrapper's.
• Do not wrap the wrapper in Start-Process. Call it inline so file writes persist.

Brief discipline

The wrapper now enforces on-disk brief traceability (added 2026-05-15 after the research-hub v0.89.1 post-release audit caught 2 orphan results from inline --prompt dispatches with no brief on disk).

• Brief MUST be on disk before invoking the wrapper. Filename pattern: .ai/codex_task_<NN_chronological_or_version>_<short_slug>.md (e.g., .ai/codex_task_v090_paginated_hub.md). The brief is the audit record — if it's not on disk, post-hoc reconstruction is guesswork.
• Pass --brief-file <path> (preferred) instead of an inline --prompt. The wrapper auto-derives the prompt as Read <brief> and execute all instructions inside. and places results next to the brief at <brief_stem>.txt and <brief_stem>.txt.result.json — no more stem drift where briefs live in .ai/ flat and results land in .ai/2026/05/13/.
• Guard fires on inline --prompt > 500 chars with no brief reference. Wrapper exits 2 with a fix-it message. Short inline prompts (e.g., --prompt 'Read .ai/codex_task_v090.md and execute') still work — the guard only refuses long inline prompts that ship instructions Codex cannot trace back to a file.
• Escape hatch: CODEX_DELEGATE_ALLOW_INLINE=1 (env var) bypasses the guard. Use this for one-off shell debugging only, never for shipping tasks.

When to delegate

Mechanical → codex · Reasoning / long-context synthesis → claude (chunked; Gemini lane DEPRECATED 2026-06-18).

Full routing table and good/bad examples: references/delegation-targets.md.

Workflow

1. Brief: write .ai/codex_task_<name>.md with Context / Goal / Constraints / Acceptance. Template: references/task-template.md. For a judgment-sensitive task (debugging, write-capable change, review, research) — where an unsupported guess or a half-finished fix would hurt — also add the XML prompt blocks from references/codex-prompt-blocks.md to the Goal/Constraints. A pure mechanical sweep does not need them. If the brief was already written by agent-task-splitter at .ai/codex_task_<NNN>_<slug>.md, read .coord/plan.yml for round context first.

2. Run: from Claude Code Bash, invoke the wrapper from its install location:

   bash ~/.claude/skills/codex-delegate/scripts/run_codex.sh \
     --brief-file .ai/codex_task_<name>.md \
     --repo "$PWD"
   

   With --brief-file the wrapper auto-derives the prompt and places results next to the brief at .ai/codex_task_<name>.txt and .ai/codex_task_<name>.txt.result.json (canonical location — eliminates the stem-drift class where briefs and results live in different subdirs). Override with --log-file only when you have a reason to. --repo defaults to the caller's $PWD; pass it explicitly to be defensive about the working directory at invocation. On non-Claude-Code agentskills.io hosts, substitute the host's skills directory (e.g. ~/.hermes/skills/<category>/codex-delegate/scripts/run_codex.sh). PowerShell variant + env vars: references/wrapper.md.

3. Read status: cat .ai/codex_task_<name>.txt.result.json.

   • success → diff still needs review.
   • fallback → Codex quota hit; Claude must take over.
   • error → wrapper failed; check <log>.error.

4. Accept: read the diff, confirm scope, run the verification commands listed in the task file. Reject if Codex drifted. Extended checklist: references/review-checklist.md.

Output contract

.result.json includes at minimum: status (success|fallback|error), delegate ("codex"), model, log_file, output_file, summary, risks, files_changed, tests_run, timestamp_utc. Full schema and status semantics: references/output-contract.md.

Compatibility

• Tested with @openai/codex 0.128.0–0.144.1 (May–July 2026). Should work with any version that accepts codex exec --sandbox workspace-write.
• Default model: gpt-5.5 (bumped from gpt-5.4 on 2026-05-14 per operator preference; override via --model or -Model). gpt-5.4 remains available and is ~3× cheaper per token if cost-sensitive — trade-offs and an A/B-test recipe live in references/model-selection.md. Other models on your CLI: see codex models.
• Wrapper calls codex exec --sandbox workspace-write -C <repo> -m <model>. The older --full-auto flag is deprecated in 0.128+ and was replaced.
• codex exec runs in non-interactive mode and auto-approves (no --ask-for-approval flag exists on exec; that flag is top-level only).
• Direct codex exec calls must close stdin (</dev/null) to avoid the historical hang (issue #20919).
• PowerShell wrapper requires $ErrorActionPreference to NOT be Stop so stderr writes (warnings, banners) don't trip the catch block.

Harness-aware delegation (target repo carries the fable-method-harness pointer)

If the target repo's AGENTS.md carries the conditional pointer generated by the fable-method-harness (python scripts/setup_harness.py --print-wiring, run in the harness repo), do NOT re-teach harness discipline in the brief — the pointer's conditional block decides activation per task (measured on Codex: 4/4 trigger recall, 0/2 routine over-trigger, 1/1 rollback; proof-of-instrument, not a token-superiority A/B — see that harness's docs/runtime_activation_telemetry_2026_07_15.md). A brief that restates the discipline pays the context twice. For bidirectional multi-round work (ai-review experiments, iterative implementation), use references/handoff-protocol.md — files, not clipboard.

See also

• references/handoff-protocol.md — bidirectional Fable↔Codex rounds via .ai/handoff/ files (kills copy-paste); harness-aware brief rule; codex exec --json usage telemetry
• references/delegation-targets.md — when to use vs avoid
• references/wrapper.md — full wrapper invocation, env vars, Windows runner notes
• references/task-template.md — task brief template
• references/codex-prompt-blocks.md — XML prompt blocks + recipes + anti-patterns for judgment-sensitive briefs
• references/output-contract.md — full .result.json schema, status semantics, .fallback_claude quota sentinel
• references/review-checklist.md — extended acceptance gate
• references/patterns.md — five single-task delegation shapes (context file, parallel, resume, structured output, review mode)
• references/multi-agent.md — leaf role in router/leaves architecture; when to route through research-hub-multi-ai or agent-task-splitter
• references/examples.md — concrete invocation examples on codex-cli 0.128.0+ syntax
• references/model-selection.md — gpt-5.4 vs gpt-5.5 trade-offs (A/B-tested) + when to override the default