architrave-review

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

Prefer the host-native reviewer. When the host provides one (GitHub Copilot
`rubber-duck`, or `code-review` for diffs), run the semantic review through it
instead of launching `adversarial-judge`, so there's never a double review.
Pass the frozen review subject, acceptance criteria, `gates/rubric.md`, and
deterministic evidence, and require one PASS/REVISE/FAIL verdict with findings
by severity and uncovered requested criteria. Use the rubric's compact output;
do not repeat the contract or score irrelevant dimensions.

`adversarial-judge` (`architrave_judge` in Codex) stays the rubric source and
the fallback when no native reviewer exists. The Codex role is advisory and
inherits parent skills, MCP, and permission policy.

For an adopted canonical Run, record its result with `gate-record --type semantic --reviewer
host-native|architrave-judge --family <model family> --effort <requested>:<mapping>`.
For ordinary direct work, return the review in conversation; do not create a
Run, receipt or report solely to record a review. This does not waive required
source-bound admission for an existing Run or authorized external operation.
For R3/R4 request `effort: high`, mapped only to what the host exposes (auto
tier or `reasoning_effort`), else none; never name a model. One independent
review suffices. Only when `review.crossFamily` is true (building Architrave
itself) does R3/R4 need two different families: get the second through a model
override on the same native reviewer when the host allows it, otherwise through
`adversarial-judge`. Never run two reviews of the same family; the runtime
rejects a duplicate same-family R3/R4 PASS (`DUPLICATE_REVIEW_FAMILY`).

From dragoshont/architrave, MIT.

More from dragoshont/architrave