review-spec

Files (1)

• SKILL.md

SKILL.md (first lines)

<!-- GENERATED by scripts/gen_adapters.py from skills/review-spec/SKILL.md.
Do not edit: run the generator instead. The canonical workflow is
reference/review-spec.md; this file only routes an agent to it. -->
**Before anything else, in Claude Code, check the model** (Codex: skip — pins are Claude model names). Run `python3 .cadence/tools/cadence_config.py --json` and read `model_pins["review-spec"]`, else `models_recommended`. Check again on every resume after a stop: a skill's pin lasts one turn. If one is set and you are not running on it, **say so and stop** — name the model you are on, the one recommended, and how to switch (`/model <name>`, or `"model"` in `.claude/settings.json` to make it stick). Continue only if the author says to.

This is a real gate, not a formality. Cadence cannot switch models: a skill's `model:` pin covers only the turn that invoked it, and this workflow spans many turns because it stops to ask you things. So without this check the pipeline silently runs the first round on one model and everything after on another, and nothing reports it.

Follow the canonical procedure in `.cadence/reference/review-spec.md` exactly.

Read `[paths].specs` and `[paths].principles` from the user's config: `python3 .cadence/tools/cadence_config.py --json`.

**Decide the scope before invoking anyone.** Apply the ordered state table in `reference/spec-pipeline.md` § *The rule* to the latest round **for the gate under review** — round numbers are global across gates, so the highest-numbered round in the directory may belong to the other gate. A gate with zero blockers and a hash matching the artifact is closed: say so and stop, without invoking a reviewer.

Compute hashes with `python3 .cadence/tools/spec_hash.py` and pass them in: `spec` mode on `product.md` (`product_hash: null` on a design-only spec, which has none), `design` mode on `design.md` (Gate 2 only), which covers every file under `design/`. Never compute a hash by hand and never write a placeholder — that is the failure this tool exists to prevent.

Gate 1 → the `product-spec-reviewer` subagent. Gate 2 → `eng-design-reviewer`. The reviewer writes the round file; it never edits the spec.

Report the verdict as the literal first line — `PRODUCT_READY`, `DESIGN_READY` or `NOT_READY`, since callers branch on it — then a Shape B brief, not the reviewer's raw findings.

From arbor-ai-inc/cadence, Apache-2.0.

More from arbor-ai-inc/cadence