Codex Dispatch

Use this skill to call OpenAI Codex CLI directly from Claude Code while preserving the orchestrator's role boundaries and verification discipline.

Core Rule

All dispatches must go through:

python3 ~/.claude/skills/codex-dispatch/scripts/codex_dispatch_role.py --task <task-file>

The script always calls codex exec with:

--dangerously-bypass-approvals-and-sandbox

Safety therefore comes from the task packet, wrapper checks, git status snapshots, write-scope validation, and orchestrator review of the result artifacts.

Modes

Pick exactly one mode per dispatch:

Verifier, reviewer, and synthesizer modes are read-only by contract. The wrapper will report a policy violation if the working tree changes.

Task Packet

Create a task packet Markdown file. Use this minimum shape:

MODE: verifier
WORKDIR: /path/to/your/project
OBJECTIVE: Verify that the targeted test and build checks work.

WRITE SCOPE:
- none

NON-GOALS:
- Do not modify files.
- Do not install packages.

VERIFICATION:
- pytest tests/ -v
- pytest tests/ -k "specific_test" -v

DELIVERABLE:
- Return whether the claim is supported and cite command evidence.

For full field guidance, read references/task-packet.md.

Worker laziness (LAZY)

worker dispatches inject a "lazy senior dev" constraint by default: a YAGNI / stdlib-first / minimal-diff ladder that biases Codex toward smaller, self-verifying patches. It appears verbatim in the generated prompt.md, so it is auditable. Disable per-dispatch with LAZY: off. Read-only modes (verifier, reviewer, synthesizer) never receive it, since its code-first terseness would be noise there. Grounded in the ponytail A/B (2026-06-15): the ladder cut worker implementation logic ~38% and forced a runnable check the orchestrator can rerun for command-tier evidence.

Run

python3 ~/.claude/skills/codex-dispatch/scripts/codex_dispatch_role.py \
  --task /path/to/task.md

Useful options:

--dry-run                 # build prompt and artifacts without calling Codex
--output-dir <dir>        # default: <workdir>/.codex-dispatch/runs
--allow-dirty-overlap     # allow worker mode to touch already-dirty scoped files
--stall-retries <n>       # fresh retries on a zero-stdout stream stall (default 2)

Intermittent Codex stream stalls kill a run before any stdout (empty events.jsonl); the wrapper auto-retries that exact signature up to --stall-retries times (stall_retries_used lands in policy.json). Real output followed by a stall, or a non-zero exit, is not retried.

The script writes:

.codex-dispatch/runs/<run-id>/
├── task.md
├── prompt.md
├── events.jsonl
├── result.json
├── result.md
├── pre-status.txt
├── post-status.txt
├── pre-diff-stat.txt
├── post-diff-stat.txt
└── policy.json

Completion

After a run, read result.md and policy.json. For worker mode, also inspect the git diff. Do not accept Codex's own success claim without fresh evidence.

If policy.json reports policy_violation: true, stop and surface that before continuing.

Evidence tiers (採信順序 ci > command > manual)

Rank Codex's completion claim by how it was produced, not by what it asserts:

• manual — Codex self-reports "tests pass" / "done" in result.md. Records intent only; never accept as passing.
• command — the orchestrator (not Codex) reruns the verification command and sees exit 0 + real output. This is the minimum bar to accept a worker result.
• ci — external CI / GitHub Actions artifact. Highest trust.

Rule: a worker completion may not be accepted below command tier.

Fingerprint-disabled = evidence downgrade

policy.json skips the working-tree fingerprint check when dirty_file_count_before exceeds the limit (200), reporting fingerprint_enabled: false / Fingerprint check skipped. In a chronically-dirty repo (e.g. For_Claude ~1850 dirty) this layer is silently off, so out-of-scope edits to already-dirty files are not caught. Treat fingerprint_enabled: false as a downgrade signal: the manual git-diff review is now mandatory, not optional. For important worker write scopes, consider dispatching in a clean worktree — see /Users/fredchu/Documents/For_Claude/company/_shared/references/cx-worktree-worker-dispatch-runbook-2026-06-14.md.

References

• Task packet format: references/task-packet.md
• Mode policy: references/mode-policy.md
• Orchestrator integration notes: references/orchestrator-integration.md