consult-codex

Consult Codex

Claude-orchestrated three-phase skill: subprocess Phase 1 → direct MCP call Phase 2 → subprocess Phase 3 → subprocess Phase 4 verdict. MCP tools like mcp__codex__codex are only callable from Claude's own context, so the skill is NOT a single script that runs end-to-end — it is an orchestration guide Claude follows.

When to invoke

Proactive (automatic, from other lifecycle skills):

• mode: review-design — after writing a design doc, before verification (integrated into feature-flow:design-document)
• mode: review-plan — after verify-plan-criteria mechanical check passes (deferred — follow-up plan)
• mode: review-code — before Harden-PR step, after all tests pass (deferred)

Reactive (manual or auto-suggested):

• mode: stuck — user typed stuck: or a signal-collector hook emitted a stuck suggestion (deferred — follow-up plan)

If invoking mode: stuck, consider calling advisor() first for a fast same-family check before spending a codex call.

Orchestration — follow these phases in order

Phase 1 — consult.js start

Run the start subprocess to check preconditions and build the brief.

FEATURE_FLOW_SESSION_ID=<session-id> \
FEATURE_FLOW_FEATURE=<feature-name> \
FEATURE_FLOW_WORKTREE=<abs-path-to-worktree> \
node skills/consult-codex/scripts/consult.js start --mode <mode> [--signal-key <key>]

Parse the JSON on stdout. Possible statuses:

• "disabled" → stop. Codex is disabled. Report the message to the user. Skip the rest of this skill.
• "skipped" → stop. A precondition (budget, escape-hatch, model-unresolvable) rejected this call. Report the reason.
• "ready" → proceed to Phase 2 with the returned { brief, model, timeout_ms, worktree, mode, signal_key }.
• "error" → stop. The CLI rejected the call (e.g., unknown mode). Surface the message to the user; do not proceed.

Optional: call advisor() to sanity-check the brief for missing context before invoking codex. Cheap same-family review prevents wasting a codex call on an ambiguous brief.