Delegate bounded coding tasks to OpenAI Codex and review diffs

Lets an orchestrating agent hand a bounded coding task to the OpenAI Codex CLI as a separate implementer, then review what it produced and land it itself. Dispatch runs through the bundled relay, which wraps codex exec --json, defaults to a write-capable workspace-write sandbox (with read-only and an explicitly human-authorized danger-full-access also available via --sandbox <mode>), and writes a structured result.json to a temp directory so the repo under review stays clean. The relay never commits, because whether Codex's own sandbox can reliably write .git varies by version, OS, and execution path - the deliberate contract is Codex edits, the orchestrator commits. Rework resumes the exact thread via --session <threadId> taken from the prior result.json (preferred, since --resume-last targets the globally most-recent Codex session, which an unrelated run can steal). --clean-env strips inherited environment variables down to runtime basics before both the version preflight and dispatch, with --keep-env <name> re-adding any environment-backed auth, provider credential, proxy, or MCP variable the run actually needs - explicitly not a security boundary, since Codex can still reach files and other same-user secrets through the sandbox and OS.

Because multiple Codex installs are common (a current npm/nvm copy shadowed by a stale Homebrew one), the skill has the orchestrator check command -v codex and codex --version before dispatch, and the relay separately records the binary version it actually ran into result.json for after-the-fact diagnosis. The relay also doubles as a clean way to get an adversarial second opinion with zero write risk: dispatch --read-only with a brief listing each contested point and both positions, and ask Codex to defend or concede each in its final message - Codex's sandbox enforces the read-only guarantee, unlike some other implementers whose read-only mode is only best-effort and flagged after the fact.

The skill is explicitly complementary to the official openai-codex Claude Code plugin rather than a replacement: the plugin's codex:codex-rescue agent is a one-way forwarder that hands a task to Codex and returns the output without polling, reviewing, or committing, and its review command runs the inverse direction (Codex reviews your work); this skill is the orchestration loop in the other direction - the orchestrator drives Codex to implement, then reviews and lands each result itself.

Use it to delegate a bounded coding task to a separate codex implementer and then review its diff yourself, when the user explicitly asked for delegation to this implementer. Do NOT use it when: the task is small enough to just do inline and delegation overhead isn't worth it; the codex CLI is not installed or not authenticated (codex login); or you want to write the code yourself, or only need a review - in that case use Codex's own review command instead of this orchestration loop.

The brief is a compact, block-structured document (<task>, <verification_loop> naming the project's real gate commands, <action_safety> forbidding git add/git commit, <structured_output_contract>), since Codex sees only the text sent to it plus whatever it reads from the working tree (it does auto-load a repo's own AGENTS.md). Dispatch:

node "<skill-dir>/scripts/relay.mjs" --brief brief.txt --cd /path/to/repo

Output is result.json (schema delegate-relay.result.v1) with status (completed/failed/timeout/aborted/codex_unavailable), codexVersion (the binary that actually ran), threadId (feed to a later --session <id>), finalMessage, and touchedFiles (git status --porcelain, the review starting point) - plus raw artifacts briefPath, eventsPath, and finalPath. Codex's JSON event stream reports a file change only as its path and kind, not the diff contents, so recovering lost work from events.jsonl alone is not possible if the working tree itself is damaged.

Wraps the OpenAI Codex CLI (npm i -g @openai/codex, authenticated with codex login), passing --model and --effort (as Codex's own -c model_reasoning_effort=<level>) through to the underlying model, and can apply a fleet lane's dials from a companion delegate-setup configuration via --lane. Complements, but does not require, the official openai-codex Claude Code plugin, whose own CLI is an optional alternative dispatch backend with background-job tracking (queued/running states via a broker process) versus the bundled relay's simpler blocking model.

Orchestrating agents - designed for and run on Claude Code, with other orchestrators (OpenCode, etc.) treated as designed-for but not yet proven - that need to delegate a scoped coding task, or a read-only adversarial second opinion, to OpenAI Codex while keeping review-and-commit authority. Adapted from the community amElnagdy/delegate-skills repository under the MIT license; this catalog entry is a docs-only import, since the executable scripts/relay.mjs is not bundled here.