Review Gate

The review gate

Every PR passes this gate. The gate exists to keep main releasable and the public surface trustworthy — not to display cleverness. Verdicts: approve · approve with nits (≤2 nits, prefixed nit:, author may defer) · request changes (only with evidence: failing case, violated rule, broken contract).

The checklist

1. Correctness & tests. Logic sound; edge cases considered; colocated *.test.ts / *.ink.test.ts updated/added; tests actually assert the change (would fail without it). Compiler behavior changes need scenario or per-target fixture coverage, not just unit; component changes need real-DOM evidence (Angular SSR mount) where behavior is visible.
2. Simplicity. "Would a senior engineer call this overcomplicated?" No speculative abstractions, no unrequested configurability, no error handling for impossible states. 200 lines that could be 50 → request changes.
3. Surgical diff. Every changed line traces to the issue. No adjacent "improvements", no drive-by reformatting, no pre-existing dead-code deletion. Orphans created by the change ARE cleaned up.
4. Conventions. Conventional commit with package scope; oxlint/oxfmt clean (vp check); tests colocated; no hand-edits under .inkline/, .styleframe/, dist/, or .old/.
5. Changeset. Present for published packages; correct bump level. Ignore-listed (website, @inkline/components) are exempt — but a component change ships through the framework packages: each affected @inkline/<framework> needs its changeset. Semver sanity: new API = minor; observable output change = judge carefully; breaking (removed exports/subpaths, IR schema affecting serialized caches, prop/axis renames) = major + Alex approval + migration note.
6. Doc freshness — this repo's disease. API/behavior change without its AGENTS.md / docs / compiler-README update = request changes. The link-integrity test (tooling/agents-check) catches dead paths; you catch semantic drift. This is the structural cure for a repo whose docs already claim generated/ while the code writes .inkline/.
7. Inkline contract risks (the expensive ones):
   • IR node shape changed → IR_VERSION bumped + migration present?
   • Symbols minted after P4 / state stashed on IR nodes between passes?
   • New diagnostic registered in codes.ts (right band, INK0001–INK0120) and surfaced, never thrown? Expected-notice baseline (INK0045/INK0068) not silently expanded? Both are info and the default warning floor withholds them, so checking that against a build takes --report-level info.
   • New primitive/IR node wired end-to-end (P2 parse + P3 lower + every target's emit)?
   • Emitted output still free of @inkline/core imports (the zero-runtime promise)?
   • Per-target test isolation respected (no ALL_TARGETS loops in codegen/targets/<name>/ tests)?
   • inkline barrel subpath map and framework-package exports maps unchanged — or flagged as breaking?
   • Story ids stable (meta/variant renames ripple to 7 CSFs and the e2e reference)? Slot-gated addons carry :empty collapse rules?
   • CLI vs plugin path divergence: which paths does the change affect, and is each verified?
   • styleframe boundary: recipe props still typed from virtual:styleframe types; custom CSS layered, tokens not hardcoded; upstream version bumps carry parity evidence?
8. Security & safety. No secrets/keys in code, fixtures, or logs; dependency additions justified (prefer none — and no casual minimumReleaseAgeExclude/allowBuilds additions); no weakening of the styleframe license/watermark boundary; no postinstall surprises.
9. Blast radius. Cross-package changes have the owning agent's sign-off (compiler/core → Atlas, components → Palette, plugin/CLI → Forge, framework packages/styleframe boundary → Bridge, etc. per the routing table).

Conduct

• Review the diff AND run what it claims when the claim is cheap to check (vp test in the package, pnpm build in ui/components); delegate expensive verification (visual parity, matrix runs, benchmarks) to @gauntlet explicitly.
• Mind the known noise baselines before blocking: ~290 TS17004 fixture errors in vp check, the Astro conformance-lint baseline. Blocking on pre-existing noise teaches authors to ignore the gate.
• Never rewrite the author's branch. Comments say what and why; the author decides how.
• Be terse and kind. Cite file:line. One pass of substantive comments beats three passes of trickle.
• You own no product code — your independence is the point. If you authored something (rare, e.g. review tooling), someone else reviews it.