crew-consult

Overview

Independent, multi-model review of a design or plan document. Report-only — it prints a report and never edits the doc or triggers a build; acting on the verdict is the user's separate decision. The review is produced by one script (crew.mjs): run the script. Do not perform the review yourself or spawn your own reviewer subagents. Your job is to run it and surface its report faithfully.

Usage

Require a file: /crew-consult <file> [--item <label>] [--read family[:tune] ...] [--judge family[:model]] [--timeout <sec>]

Resolve the script

The script lives at <this skill's base directory>/../../scripts/crew.mjs. Resolve it from the base directory the harness gives you for this skill — never guess or hardcode another location.

Run it

Echo the exact command, then run it from the current repo root:

node <resolved-path>/crew.mjs consult <file> [flags…]

Families: claude · openai (codex CLI) · google (agy CLI). With no flags the script detects installed CLIs and prefers a cross-family crew automatically.

Tiers are fast / balanced / extra (aliases cheap / standard / most-capable); reviewers get balanced at minimum. For codex, read codex debug models, choose a model and effort by capability, and pass --read openai:<model>@<effort>; never copy a model id from a table or an old session.

After it finishes

• Surface compactly: the anchor (the doc's sha256 digest), the CREW composition line, each read's verdict and findings, and the judge's verdict with its discarded/systemic lists.
• Preserve the report-only boundary: do NOT act on the verdict — no edits to the doc, no follow-up build. Acting is the user's separate decision.
• On exit 1/2, relay the stderr message verbatim (it names the missing CLI, bad file, or usage problem) AND surface any report that printed on stdout — an exit 1 with all reads skipped still prints a useful report there.
• The run appends telemetry to ~/.preflight/modelcalls.jsonl and writes nothing into the reviewed repo.