[Cogitations] Bridge 27 eval JSON files to automated pytest assertions

[zircote]

Quality Gate Reference

• Checklist Item(s): TDD-001 (Test Coverage Baseline), TDD-003 (Test Pyramid Balance)
• Current Score: TDD domain 11/100 — Tier 1 blocker
• Impact: High — leverages existing eval investment
• Effort: High

Description

The project has 27 JSON eval files across evals/ and skills/*/evals/ that define structured behavioral expectations (prompts, expected tool calls, assertions). These are currently declarative specifications evaluated manually or by AI — not connected to any automated test runner.

Implementation Plan

1. Write a pytest plugin or parametrized test loader that reads eval JSON files
2. For each eval entry, extract the expectations or assertions arrays
3. Map deterministic expectations to pytest assertions:
   • Skill trigger detection → verify correct skill matches prompt
   • Expected tool calls → validate against known tool schemas
   • Expected state changes → check output structure
4. Use @pytest.mark.parametrize with ids from eval name fields
5. Non-deterministic expectations (LLM output quality) → mark as pytest.mark.skip with reason, or use similarity thresholds

Benefits

• Converts existing eval investment into CI-enforceable tests
• Fills the top of the test pyramid (acceptance/behavioral specs)
• Combined with unit tests from issue chore: quality improvement 53.9 → 81.5 (tier 2) #1, creates proper pyramid shape

Acceptance Criteria

• TDD-001 score improves
• TDD-003 (test pyramid) improves from 0.05
• Eval failures detected in CI before merge
• No regression in other domain scores

References

• Promptfoo — eval framework patterns
• DeepEval — pytest-integrated eval framework
• Existing eval files: evals/test-architect-evals.json, skills/refactor/evals/refactor-evals.json

---

Generated by Cogitations /cog-discover