chronicle-mcp: Skill for Claude Code

Skill TDD

Test and benchmark skills (SKILL.md), subdocs, and policies (AGENTS.md) using red-green-regression cycles, 4-axis coverage, single-call parallel dispatch, and hypothesis ledgers.

Directives

1. Red Baseline First: MUST run a baseline subagent without the skill on negative fixtures to verify failure (Red) before testing candidate versions.
2. Double-Sided Fixtures: MUST test both a Negative Fixture (agent must block/refactor) and a Positive Fixture (agent must clear/approve) to prevent false-positive hyper-critique.
3. 4-Axis Coverage: Test suites MUST cover Directives, Decision Branches, Anti-Patterns, and Domain Diversity (minimum 2–3 distinct domains) per REFERENCE.md.
4. Test Matrix Budget:
   • Fast Dev Loop: $N_{\text{tests}} \ge 2$, $N_{\text{runs}} \in [1, 3]$ per test.
   • Graduation Gate: $N_{\text{tests}} \ge 3\text{--}4$ diverse domain fixtures, $N_{\text{runs}} \ge 5\text{--}10$ with fresh 0-prior-context subagents.
   • Graduation Criteria: $\Delta_{\text{Suite}} \ge +60%$, 100% positive clearance, and 0 cheat flags.
5. Single-Call Parallel Matrix Dispatch: MUST launch all independent test trials (baseline and treated runs across all fixtures) in a single concurrent batch via invoke_subagent(Subagents=[...]). NEVER dispatch trials sequentially across multiple turns. Tag each subagent clearly: Role: "[Baseline/Treated] | <Fixture> | Run <K>".
6. Spec Decoupling & Anti-Snooping Audit:
   • TEST_SPEC.md and HYPOTHESIS.md MUST reside in .scratch/<skill>-versions/specs/ and NEVER inside fixture code directories (.scratch/fixtures/<name>/).
   • Evaluators MUST inspect subagent tool calls in transcripts. If a baseline subagent reads repository SKILL.md files or ANY subagent reads TEST_SPEC.md/HYPOTHESIS.md, abort immediately with:
     Overfitting cheat detected: Subagent snooped <path>.
7. Immutable Prompt Contract: MUST copy prompts verbatim from TEST_SPEC.md. NEVER inject hints, line numbers, or leading context.
8. Zero Prior Context: Spawn every test run using a fresh subagent with 0 prior context. NEVER reuse subagent conversation threads.
9. Sandbox Delta Versioning:
   • NEVER edit production skills/policies directly on test failure.
   • Create a sandbox (.scratch/<skill>-versions/) with candidate files (SKILL.v1.md) and an append-only HYPOTHESIS.md (NEVER delete entries).
   • Point test subagents to candidate versions. Graduate to production only after meeting the $\Delta_{\text{Suite}}$ threshold.
10. Anti-Overfitting Cheats:
    • SKILL.md and REFERENCE.md MUST express domain-agnostic rules. NEVER hardcode test fixture entities, class names, or mock values.
    • Evaluators detecting leaked words or prompt hints MUST abort immediately with:
      Overfitting cheat detected: <reason>.
11. Zero-Delta Anomaly: If 100% of runs yield Double-GREEN ($\Delta = 0$), classify as a Zero-Delta Failure (harden test fixture or prune redundant skill rules).
12. Seam-Only Assertions: Verify behavior at public seams (verdict, scorecard rows, tool call constraints, stop boundary). NEVER assert on private reasoning traces.