| name | code-review |
| description | Parallel multi-agent review for Laboratory changes. Prioritizes correctness,
scientific validity, provenance, tests, and harness fit. Trigger:
/code-review, /review, /critique.
|
| argument-hint | [branch|diff|files] |
/code-review
Review the diff as a Laboratory artifact. Findings lead; summaries are
secondary. Block on bugs, invalid scientific claims, missing tests, broken
simulate/live separation, missing trace/provenance, and runtime paths that were
not exercised.
Scope
Default scope is git diff <base>...HEAD, where base is master unless the
user names another base. Include uncommitted changes when the user asks for a
working-tree review.
Classify touched surfaces:
- experiment design or manifest
- harness runner
- analysis/statistics
- data artifact
- report/publication artifact
- site/content
- Spellbook harness/skill/agent surface
- repository infra
Review Lanes
Use fresh reviewers for substantive reviews:
- correctness and runtime behavior
- scientific methodology and provenance
- simplicity and harness fit
Use Thinktank or cross-harness review when the diff is broad, high-risk, or
touches experiment conclusions. Provider output is evidence, not authority; the
lead synthesizes and verifies.
Blocking Checks
- New or changed executable paths were directly exercised: harness CLI,
analyzer, Make target, script entrypoint, site build, or wrapper.
- Harnesses preserve deterministic
--simulate behavior and require explicit
--live for API calls.
- Live paths have provider preflight, budget controls, model/config logging,
and trace artifacts.
- Analysis refuses simulated or trace-missing data for publishable findings.
- Reports cite live/provenance-backed data and do not overclaim.
- Tests cover new deterministic logic.
- Spellbook harness changes preserve shared
.agents roots and bridge links.
Verdicts
ship: no blocking findings and required executable paths were exercised.conditional: only clearly bounded follow-up remains and the user can accept
the risk.dont-ship: any correctness, scientific-validity, provenance, safety, or
unverified-runtime blocker remains.
Output
Return findings first, ordered by severity with file/line references. Then list
open questions, commands or artifacts reviewed, and the verdict.
