tools: claude: use new review prompt to catch more issues

[pzmarzly]

Seems like telling AI to do more work makes it do more work. (And costs more money.) Shocked Pikachu.

Tested on #489 using my local testing scripts (pzmarzly/claude-pr-review#test/):

#	Issue	Upstream bot	New bot
1	Path traversal: unvalidated cache_dir POST field flows into shutil.rmtree (bfoptimize-web:119)	must-fix	suggestion
2	git reset --hard silently destroys uncommitted work (bfoptimize:542)	suggestion	—
3	bypassPermissions gives Claude agent unrestricted Bash (bfoptimize:259)	suggestion	suggestion
4	NO_TESTS option description is stale, now also disables benchmarks (CMakeLists.txt:33)	suggestion	—
5	NO_TESTS doc description should mention benchmarks (build.rst:98)	suggestion	—
6	No validation on model/effort string fields (bfoptimize-web:48)	suggestion	—
7	daemon: cgen: commit prefix is stale — daemon was removed from project (bfoptimize:249)	—	must-fix (5/5 reviewers)
8	-DNO_TESTS=1 contradicts make unit e2e integration — test targets won't exist (bfoptimize:246)	—	must-fix (5/5 reviewers)
9	results.html.j2 deleted instead of moved — bfbencher history --report-path broken at runtime (bfbencher:39)	—	must-fix (4/5 reviewers)
10	DELETE /history crashes with FileNotFoundError if cache dir doesn't exist yet (bfoptimize-web:180)	—	suggestion
11	--fail-on-benchmark-error dead code for history subcommand (bfbencher:1298)	—	suggestion (3/5 reviewers)
12	CI build job may fail — find_package(benchmark REQUIRED) now runs without benchmark dep (ci.yaml:85)	—	suggestion
13	git commit -am stages all tracked files, not just CGEN_DIR/ (bfoptimize:249)	—	suggestion
14	No subprocess output captured on benchmark failure (bfoptimize:238)	—	nit

And checked if workflow still works in: pzmarzly/claude-pr-review#16

[github-actions]

Claude review of PR #490 (4a027ab)

Nits

• Stale "subagent" terminology — .github/workflows/claude-pr-review.yaml:210 — One unchanged context line still says "not for subagent output" while all modified lines use "agent"/"reviewer"

---

Clean PR overall. YAML syntax is correct, --effort max and Agent tool additions are properly placed, the prompt instructions are clear and internally consistent, and the commit message follows project conventions (tools: claude:, imperative mood, under 72 chars). The read-only security posture of the review job is preserved since sub-agents inherit the same allowedTools restriction.

Workflow run

[qdeslandes]

Thanks! Looking forward to try this out!