ci: enhance claude code review and clean up enhance-claude-docs skill

[ninja-shreyash]

Why

Three pain points across our review/docs automation:

1. Bombardment in claude-code-review.yml — every push re-runs a full review with no awareness of prior comments, re-posting the same inline comments even after fixes/resolves/declines.
2. No way to opt out of the review workflow on WIP or trivial PRs.
3. enhance-claude-docs skill bakes provenance into doc files — (Source: PR #184) and @swati354-style references end up permanently inside agent_docs/*.md (see docs: update Claude docs from PR review analysis #360), creating cross-link noise and burning contributor handles into long-lived docs.

What this PR does

1. Resolved-thread dedupe (claude-code-review.yml)

Fetches existing review threads via GraphQL before reviewing, then gates every inline-comment posting:

Prior thread	Original was	Code now	Action
Unresolved	(any)	(any)	Skip
Resolved	Actionable ask	Fix applied	Skip
Resolved	Actionable ask	Fix NOT applied	Unresolve the thread (no new comment)
Resolved	Author replied "won't fix" / "intentional"	(any)	Skip
Resolved	Nit / FYI	(any)	Skip

Decline detection reads thread reply text — no labels or syntax needed.

2. Opt-out label (claude-code-review.yml)

if: ${{ !contains(github.event.pull_request.labels.*.name, 'skip-claude-review') }}

Apply skip-claude-review (label created in this repo) → workflow skipped on subsequent commits. Remove → re-enabled. Symmetric.

3. Silence summary on no-op runs (claude-code-review.yml)

The current workflow posts zero plugin summary comments on most PRs (e.g., 8 commits on #358 → 0 plugin summaries). Without explicit guidance, the new dedupe prompt would have started posting filler summaries like "No new issues" on runs where nothing actually happened.

STEP 5 of the prompt now gates the summary: post only if the run produced an action (a new inline comment or an unresolved thread). Otherwise stay silent. Preserves current "no chat on no-op" behavior while keeping audit value when state actually changed.

4. Content rules for enhance-claude-docs skill

Step 6 of .claude/skills/enhance-claude-docs/SKILL.md now forbids #N refs, @username mentions, and verbatim quoting inside the doc files. Provenance stays in the PR body and terminal output.

Testing — review workflow

Validated end-to-end on ninja-shreyash/cherrypick-test#20 across 8 scenarios. All passed.

#	Scenario	Validates	Result
1	Baseline — open PR with 3 deliberate issues	Initial review posts findings	3 inline threads at lines 3, 4, 8 ✅
2	Idle re-run (empty commit)	No duplicate inline comments	Still 3 threads, 1 comment each ✅
3	Resolve thread without fix → push	Bot unresolves the thread	Thread flipped resolved → unresolved, 0 new comments ✅
4	Resolve + apply fix → push	Bot leaves it resolved	Stayed resolved (line 8 → 12 drift tracked correctly), 0 new comments ✅
5	Reply "won't fix", resolve → push	Bot respects decline	Stayed resolved, 0 new comments ✅
6	(Emergent) Bot finds new issue near existing thread	New finding on different topic posts; bot references existing thread	Posted on line 5 about a logic bug, body referenced line-4 thread ✅
7a	Add skip-claude-review label → push	Opt-out skips workflow	Run completed in 1s with status skipped, no compute ✅
7b	Remove label → push	Workflow resumes normally	Ran successfully, dedupe respected ✅
8	No-op push after silence-on-no-op clause added	Bot posts no filler summary	Same threads, comment count unchanged, 0 new summaries posted ✅

Across 4 follow-up pushes (scenarios 2–5), zero duplicate inline comments were posted. Scenario 8 confirms the silence-on-no-op clause from change #3 fully suppresses filler summaries — same conditions previously produced a ## Code Review — No New Issues summary, now produces nothing.

Test PR: ninja-shreyash/cherrypick-test#20

Testing — docs skill change

Prompt-level change. Exercised by the next scheduled claude-docs-update.yml run (Mondays 9am UTC) or manual workflow_dispatch. Next run's docs PR should have no #N or @username strings inside modified agent_docs/*.md files.

Risk

• CI / skill files only — no SDK code touched.
• Backward-compatible: PRs without the opt-out label behave as before, plus dedupe.
• If the GraphQL query fails, falls back to prior full-review behavior. Worst case = one run of duplicate comments, not a broken workflow.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud