# tikv-slow evaluator

This directory owns the maintainer-facing evaluator definition layer for
`tikv-slow-diagnosis`.

Purpose:
- keep the evaluator close to the skill it evolves
- keep runtime diagnosis guidance separate from benchmark and replay policy
- let `oncall-toolkit` remain the execution entry point without owning the
  `tikv-slow` evaluator's domain policy

Boundary:
- `tikv-slow-diagnosis/SKILL.md` and `references/*.md` are runtime diagnosis
  knowledge
- `tikv-slow-diagnosis/evaluation/*.md` are maintainer and evaluator policy
- `oncall-toolkit` should execute `evaluate tikv-slow`, but should consume the
  contracts defined here instead of being the long-term source of truth

Recommended ownership split:
- this directory owns:
  - evaluation scope and replay boundary
  - profile-specific discovery and exclusion policy
  - case schema
  - mismatch taxonomy
  - comparison rubric
  - improvement promotion policy
- `oncall-toolkit` owns:
  - candidate discovery
  - Jira and Pingkai retrieval
  - bundle assembly
  - command-line execution and reporting

Migration direction:
1. move evaluator policy here first
2. keep `oncall-toolkit evaluate tikv-slow` as the stable CLI entry
3. later migrate Python evaluator contracts and workflows to a skill-owned
   companion implementation if needed

Files in this directory:
- `PROFILE.md`: profile scope, defaults, and replay boundary
- `discovery.py`: profile-specific focus matching, case shaping, and exclusion rules
- `CASE_SCHEMA.md`: canonical case and review object schema
- `WINDOW_REVIEW_POLICY.md`: optional repeated-spike and multi-window review policy
- `MISMATCH_TYPES.md`: allowed mismatch taxonomy and interpretation
- `PROMOTION_POLICY.md`: how reviewed cases may influence future skill edits