@@ -21,34 +21,73 @@ Run a comprehensive review using multiple specialist agents in parallel.
2121
22222 . ** Collect results** from all 4 agents.
2323
24- 3 . ** Produce consolidated summary:**
24+ 3 . ** Produce consolidated summary** (see format below).
25+
26+ ## Agent Prompt Rules
27+
28+ ** CRITICAL: Include these rules in EVERY agent prompt to prevent noise.**
29+
30+ Each agent prompt MUST include the following section verbatim:
31+
32+ ```
33+ ## What to Report
34+
35+ Only report issues that meet ALL of these criteria:
36+ 1. **Demonstrably wrong** — you can show it produces incorrect behavior, crashes, data loss, or security vulnerability
37+ 2. **Reproducible** — you can describe concrete steps or inputs that trigger the issue
38+ 3. **Not a trade-off** — the current approach was not an intentional design choice with documented rationale
39+
40+ ## What NOT to Report
41+
42+ Do NOT report any of the following. These are explicitly banned:
43+ - "Consider adding..." / "Could be improved by..." — speculative improvements
44+ - Performance optimizations for code that handles < 10,000 items (premature optimization)
45+ - Missing tests for private/internal functions — only flag missing tests for PUBLIC API
46+ - "This could break if..." with hypothetical future scenarios that don't exist today
47+ - Code style preferences already handled by the project's linter (Biome)
48+ - Suggesting abstractions, wrappers, or classes for code that works fine as-is
49+ - Duplicated findings from previous review rounds that were evaluated and accepted
50+ - "Add a comment explaining..." — if the code is clear, it doesn't need a comment
51+ - Alternative approaches that are equally valid but not better ("you could also do X")
52+ - Module-level state that is intentional and correctly managed (e.g., caches, singletons)
53+ - Missing error handling for errors that cannot occur in practice
54+
55+ ## Severity Calibration
56+
57+ - **CRITICAL**: Currently broken in production. Data loss, security hole, crash. Must have reproduction steps.
58+ - **MAJOR**: Demonstrable bug or correctness issue with concrete example. NOT "this might cause problems someday."
59+ - **MINOR**: Real code smell with a clear fix, NOT a suggestion or preference.
60+
61+ ## Hard Limits
62+
63+ - Maximum 3 MAJOR issues per agent. If you find more, keep only the top 3 by impact.
64+ - Maximum 5 MINOR issues per agent. If you find more, keep only the top 5.
65+ - If you find 0 issues, return APPROVE with "No issues found." Do NOT invent issues to fill the template.
66+ - It is BETTER to return APPROVE than to pad findings with noise.
67+ ```
68+
69+ ## Consolidated Summary Format
2570
2671```
2772## Critical Review Summary
2873
2974### Architecture: [ALIGNED / DRIFT DETECTED]
30-
31- [Key findings from ts-architect]
75+ [Only genuine architectural misalignment]
3276
3377### Code Quality: [APPROVE / REQUEST CHANGES]
34-
35- [Key findings from ts-code-reviewer]
78+ [Only demonstrable bugs or correctness issues]
3679
3780### Test Coverage: [SUFFICIENT / GAPS FOUND]
38-
39- [Key findings from ts-qa-developer]
81+ [Only missing tests for public API with real risk]
4082
4183### Search Quality: [ACCEPTABLE / NEEDS TUNING]
42-
43- [Key findings from search quality check]
84+ [Only measurable search quality problems]
4485
4586### Blockers
46-
47- [List any BLOCKER-severity issues]
87+ [List any CRITICAL-severity issues, or "None"]
4888
4989### Action Items
50-
51- [Prioritized list of changes needed]
90+ [Deduplicated, prioritized list — max 5 items total across all agents]
5291```
5392
5493## Success Criteria
0 commit comments