skills/pr-review: clarify AGENTS.md rule updates and remove assumptions of /learn on Claude Code / Sonnet review lanes

deno.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://raw.githubusercontent.com/denoland/deno/main/cli/schemas/config-file.v1.json",
   "name": "@webjam/tools",
-  "version": "1.32.32",
+  "version": "1.32.33",
   "exports": "./src/uptime/cron.ts",
   "tasks": {
     "sheet-music:generate": "deno run --allow-read --allow-write src/sheet-music/generate_docx.ts",

scripts/handle-agy-tasks.sh

@@ -554,11 +554,11 @@ Rules:
   repos use "npm run test:lint" / "npm run test:unit").
 - Do not switch branches and do not add new dependencies.
 - $LOCAL_TESTING_INSTRUCTIONS
-- Before opening the PR (web-jam-tools#239): run the \`/learn\` slash command in
-  this session, then commit whatever changes it makes to AGENTS.md onto THIS
-  SAME branch as its own commit (e.g. "chore: fold /learn updates into
-  AGENTS.md") — do not open a separate PR or defer it to a follow-up step. If
-  \`/learn\` makes no changes, that's fine; just don't skip running it.
+- Before opening the PR (web-jam-tools#239, web-jam-tools#867): on Antigravity (agy) sessions, run
+  the \`/learn\` slash command in this session if available, then commit whatever changes it makes
+  to AGENTS.md onto THIS SAME branch as its own commit (e.g. "chore: fold /learn updates into
+  AGENTS.md") — do not open a separate PR or defer it to a follow-up step. If \`/learn\` makes no
+  changes or is running on a non-Antigravity surface (e.g. Claude Code), that's fine; just continue.
 - When lint and tests are green, finish by opening a draft PR — run:
     ~/WebJamApps/web-jam-tools/scripts/create-draft-pr.sh --author "agy — <the model you are running as>" \\
       --summary-file /tmp/pr-summary.md --test-plan-file /tmp/pr-test-plan.md --test-evidence-file /tmp/pr-test-evidence.md

skills/pr-review/SKILL.md

@@ -141,14 +141,18 @@ Review the PR diff, description, mergeability, and non-CircleCI status checks (S
 4. **Issue Acceptance Criteria & Scope**:
    - Does the diff fulfill all requirements and acceptance criteria stated in the linked issue?
    - Is the PR tightly scoped to the issue task? Are there any out-of-scope files, unintended refactors, or stray code additions?
-   - **Exception — `AGENTS.md` updates are never a scope violation.** A `/learn`-sourced change to
-     `AGENTS.md` (an agent folding in a lesson or guardrail it picked up while doing the PR's actual
-     work) is standing cross-cutting instruction content, not scope creep — it has no natural issue
-     of its own to belong to, and routing every such tweak through a separate PR would suppress the
-     mechanism by which agents correct their own guidance. Do not raise a Must Fix (or any) finding
-     over an `AGENTS.md` diff being unrelated to the PR's linked issue. Still review the content on
-     its own merits — it must not contradict or duplicate existing `AGENTS.md`/`docs/cross-ai-rules.md`
-     text (flag that as a normal correctness finding if it does).
+    - **Exception — `AGENTS.md` and `docs/cross-ai-rules.md` updates are never a scope violation.** Rule,
+      lesson, and guardrail updates to `AGENTS.md` or `docs/cross-ai-rules.md` (whether learned
+      interactively during task work or sourced from `/learn` on Antigravity sessions) are standing
+      cross-cutting instruction content, not scope creep — they have no natural issue of their own to
+      belong to, and routing every such tweak through a separate PR would suppress the mechanism by
+      which agents correct their own guidance. Note: `/learn` is a Google Antigravity-native command;
+      Claude Code (Sonnet) sessions do not have `/learn` and update memory/rules directly. Do not
+      raise a Must Fix (or any) finding over an `AGENTS.md` / `docs/cross-ai-rules.md` diff being
+      unrelated to the PR's linked issue, and never expect or flag a missing `/learn` invocation in
+      reviews. Still review the instruction content on its own merits — it must be purely additive and
+      must not contradict or duplicate existing `AGENTS.md`/`docs/cross-ai-rules.md` text (flag that
+      as a normal correctness finding if it does).
 
 5. **Single Semver Version Bump per PR**:
    - Check `package.json` (or `deno.json` for `web-jam-tools`).

test/skills_validation.test.ts

@@ -292,6 +292,24 @@ Deno.test("skills/pr-review/SKILL.md uses ### 🟡 Suggestions heading and not A
   );
 });
 
+Deno.test("skills/pr-review/SKILL.md clarifies AGENTS.md scope exemption and Antigravity-specific /learn command (#867)", async () => {
+  const prReviewPath = `${SKILLS_DIR}pr-review/SKILL.md`;
+  const text = await Deno.readTextFile(prReviewPath);
+
+  assertStringIncludes(
+    text,
+    "**Exception — `AGENTS.md` and `docs/cross-ai-rules.md` updates are never a scope violation.**",
+  );
+  assertStringIncludes(
+    text,
+    "`/learn` is a Google Antigravity-native command",
+  );
+  assertStringIncludes(
+    text,
+    "never expect or flag a missing `/learn` invocation",
+  );
+});
+
 Deno.test("skills/design-issue/SKILL.md contains absolute standing rule that skill never dispatches", async () => {
   const designIssuePath = `${SKILLS_DIR}design-issue/SKILL.md`;
   const text = await Deno.readTextFile(designIssuePath);