pilastdigit/trip-workflow/codex-ask

Codex Ask

Free-form second opinion from Codex CLI on any matter — architecture decisions, debugging hypotheses, research conclusions, trade-off calls — not just plans and diffs. Codex answers from inside the repository (read-only sandbox), so its opinion is grounded in the actual code, not in whatever excerpt happened to be quoted.

Advisory, not authoritative. Unlike codex-plan-review / codex-code-review, there are no verdict tags and nothing is gated on the answer: treat the response as one input to your judgment, exactly like a colleague's opinion. Agreement is weak evidence; disagreement is a strong signal that something deserves the user's attention.

State persisted per topic label under .claude/skills/codex-ask/state/ — follow-ups resume the same thread, enabling multi-round discussion. Invoke this skill's own wrapper scripts under .claude/skills/codex-ask/scripts/ — they pin the correct STATE_DIR and delegate to the shared codex-plan-review implementation, so there is no export to remember (a forgotten export would silently operate on the plan-review thread for the same label).

Arguments

• <topic-label> — short kebab-case label for the discussion (becomes the state key), e.g. orchestrator-choice, flaky-auth-test. Auto: start if no thread, resume if one exists.
• <question> — the actual question, passed as trailing text. Include your own draft position when you have one ("Here is my recommendation: … Red-team it").
• reset <topic-label> / show <topic-label> — drop state / display last answer.

Execution

1. Start a discussion:

``bash bash .claude/skills/codex-ask/scripts/start.sh \ --prompt-file .claude/skills/codex-ask/prompts/ask.tpl \ <topic-label> "<question — include your draft position and ask for disagreement>" ``

1. Follow up in the same thread (counterpoints, new evidence):

``bash bash .claude/skills/codex-ask/scripts/resume.sh \ --prompt-file .claude/skills/codex-ask/prompts/followup.tpl \ <topic-label> "<follow-up or counterpoint>" ``

1. Reset: bash .claude/skills/codex-ask/scripts/reset.sh <topic-label> — Show: bash .claude/skills/codex-ask/scripts/show.sh <topic-label>

When to use

• Second opinion on an architecture/design decision before it hardens (e.g., at the end of a research session, before writing the plan).
• Root-cause help when genuinely stuck on a bug — fresh eyes, different blind spots.
• "Red-team this conclusion" on a memo or recommendation you are about to present.

When NOT to use

• Questions that need the user's preference or judgment — ask the user, not Codex.
• Trivial lookups or anything settled by reading the code yourself — every ask costs a Codex run.
• As a gate: never block or approve work based on the answer; that is what the review skills with verdict tags are for.

Notes

• Run Codex calls in a background shell. Invoke start.sh / resume.sh via the Bash tool with run_in_background: true — never as a foreground/inline command. Codex runs at xhigh effort routinely outlast the foreground command timeout; the background task notifies on completion, then read its output. reset.sh / show.sh are instant and fine in the foreground.
• Set `CODEX_TIMEOUT=1800` (30 min) when invoking start.sh / resume.sh — a generous circuit breaker against hung runs (script default 0 = no timeout); fails through the normal error path with a "timed out" message on expiry.
• Read-only sandbox — Codex can read the repo but change nothing.
• Model/effort/tier defaults live in codex-plan-review/scripts/_common.sh, derived from STATE_DIR: implementation → gpt-6-luna at high effort on the fast service tier; plan review → gpt-6-astra at xhigh on standard routing; code review → gpt-6-sol at xhigh on the fast tier; codex-ask → gpt-6-astra at xhigh on standard routing. Adjust that one file to your preferred models, or override per run via CODEX_MODEL / CODEX_EFFORT / CODEX_TIER env vars; the scripts echo the effective values.
• Surface Codex's answer to the user verbatim when it disagrees with your position — the disagreement itself is the valuable output.