AI PR reviewer that knows framework conventions (runs locally + in CI)

[MUKE-coder]

Vision

Code review is the slowest step in most teams. Grit's strict conventions mean an AI reviewer can be specific instead of generic — "this should use SearchableSelect", "this handler is missing RequirePermission", "this list endpoint isn't returning the standard envelope".

Proposal

`grit review` (locally) and a GitHub Action (in CI):

• Reads the diff
• Loads the framework conventions from CLAUDE.md / AGENTS.md (grit init writes a CLAUDE.md / AGENTS.md template that codifies framework conventions #35)
• Calls Claude with a prompt that knows specifically how Grit is supposed to be used
• Posts inline review comments per file with severity (info / warn / block)
• Blocks merge if convention violations are found at `block` level (configurable)
• Runs in <60 seconds for typical PRs

Differentiator vs. generic AI review: Grit's reviewer has a baseline of "how this codebase works" so it doesn't waste comments on subjective style — it focuses on framework drift.

Why this is differentiating

Cursor, Continue, Claude Code all do generic AI review. None know framework-specific rules. A framework-aware reviewer is sharper, has fewer false positives, and is shipped with the framework so every team gets it on day zero.

Acceptance

• Opening a PR that uses native `` for an FK field gets a review comment pointing to the SearchableSelect rule A handler missing `RequirePermission` gets blocked The reviewer learns when the team configures exceptions (e.g. "this app doesn't use the matrix editor")