Use latest Cursor model for PR review automation

.cursor/commands/read-branch.md

@@ -0,0 +1,57 @@
+# Read branch / PR review
+
+This file is the system prompt for the Flatbread PR Review Cursor
+automation (`/read-branch`). Execute it against the current pull request.
+
+## Models
+
+Use the **latest Cursor Grok** model for architecture, proof-journal, and
+correctness work. Do not pin an old Grok version.
+
+1. Do not pass `cursor-grok-4.5-high`, `cursor-grok-4.5-high-fast`, or any
+   older `cursor-grok-*` slug to Task subagents.
+2. Prefer omitting Task `model`, or pass `inherit`, so children match the
+   parent run. The automation parent should already be the latest Cursor
+   Grok.
+3. If the Task tool requires an explicit slug, pick the newest
+   `cursor-grok-*` entry in that tool's allow-list. Do not invent a slug
+   that is not listed.
+4. Keep the simplify/quality domain on Composer: `composer-2.5`, or the
+   newest `composer-*` slug in the allow-list.
+5. Do not write a versioned Grok slug into automation memory. Memory may
+   say "latest Cursor Grok". It must not say `cursor-grok-4.5-high`.
+
+When the review footer names models, name the slugs you actually used.
+
+## Fan-out
+
+Spawn one subagent per domain:
+
+1. architecture
+2. opportunities to simplify and improve code quality (including
+   complexity)
+3. Proof journal quality (`/proof`). If the PR contains a significant
+   event, a Proof record should likely accompany it.
+4. code correctness and real-world edge-case coverage
+
+Skip the run when the diff is dependency-only (lockfiles / `package.json`
+only). Continue if any non-dependency source or docs remain.
+
+If Oven (`pnpm exec oven`) is missing, use the Task tool as the fallback.
+The parent writes `/tmp/review-judge-final.md`.
+
+## Synchronize
+
+On a `synchronize` event:
+
+1. Classify prior automation threads as addressed or still open.
+2. Resolve prior threads, then `cleanup_previous`.
+3. If no threads are open, skip resolve.
+
+Do not bump Unreleased CHANGELOG notes for packages with no `src` in the
+PR.
+
+## Output
+
+Lead with a review verdict. Then list prior notes closed, domain verdicts,
+and the models actually used.

.cursor/rules/cursor-grok-latest.mdc

@@ -0,0 +1,20 @@
+---
+description: PR review and Grok Task subagents use the latest Cursor Grok model, never a pinned 4.5 slug.
+alwaysApply: true
+---
+
+# Latest Cursor Grok for review subagents
+
+When this run is a pull-request review, or when you spawn Task subagents
+for architecture, proof-journal, or correctness:
+
+1. Use the **latest Cursor Grok** model. Prefer omitting Task `model`, or
+   pass `inherit`, so children match the parent.
+2. If the Task tool requires an explicit slug, pick the newest
+   `cursor-grok-*` entry in that tool's allow-list.
+3. Do **not** pass `cursor-grok-4.5-high` or `cursor-grok-4.5-high-fast`.
+4. Do **not** write a versioned Grok slug into automation memory.
+
+The Flatbread PR Review automation executes
+`.cursor/commands/read-branch.md` as `/read-branch`. That file is the
+source of truth for this routing. Quality/simplify work stays on Composer.

AGENTS.md

@@ -59,6 +59,15 @@ The repo uses Mergify stacks for PR management. The `mergify-cli` is installed v
 - The commit-msg hook (`.husky/commit-msg`) auto-appends a `Change-Id` trailer for stack tracking.
 - See `.agents/skills/mergify-stack/SKILL.md` for the full workflow.
 
+### PR review models
+
+The Flatbread PR Review automation runs `.cursor/commands/read-branch.md`
+as `/read-branch`. Architecture, proof-journal, and correctness subagents
+use the **latest Cursor Grok** model. Do not pin `cursor-grok-4.5-high`.
+Prefer `inherit` (or omit Task `model`) so children match the parent.
+Quality/simplify stays on Composer. See
+`.cursor/rules/cursor-grok-latest.mdc`.
+
 ### Gotchas
 
 - **Native build scripts are approved in `pnpm-workspace.yaml`.** The `onlyBuiltDependencies` list allows esbuild, sharp, @swc/core, etc. to run their postinstall scripts automatically during `pnpm install`.

scripts/cursor-grok-latest.test.js

@@ -0,0 +1,34 @@
+import test from 'ava';
+import { readFileSync } from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const root = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..');
+
+function readRepo(rel) {
+  return readFileSync(path.join(root, rel), 'utf8');
+}
+
+const command = readRepo('.cursor/commands/read-branch.md');
+const rule = readRepo('.cursor/rules/cursor-grok-latest.mdc');
+const agents = readRepo('AGENTS.md');
+
+test('/read-branch tells review subagents to use the latest Cursor Grok', (t) => {
+  t.true(command.includes('latest Cursor Grok'));
+  t.true(command.includes('inherit'));
+  t.true(command.includes('Do not pass `cursor-grok-4.5-high`'));
+  t.false(command.includes('models cursor-grok-4.5-high'));
+});
+
+test('always-on rule forbids pinning Grok 4.5 on review Task children', (t) => {
+  t.true(rule.includes('alwaysApply: true'));
+  t.true(rule.includes('latest Cursor Grok'));
+  t.true(rule.includes('cursor-grok-4.5-high'));
+  t.true(rule.includes('Do **not** pass `cursor-grok-4.5-high`'));
+});
+
+test('AGENTS.md points PR review routing at the in-repo command and rule', (t) => {
+  t.true(agents.includes('.cursor/commands/read-branch.md'));
+  t.true(agents.includes('.cursor/rules/cursor-grok-latest.mdc'));
+  t.true(agents.includes('latest Cursor Grok'));
+});