# Evaluator Profile

This file defines the `tikv-slow` evaluator profile used to evolve
`tikv-slow-diagnosis`.

## Goal

Use real incidents to review whether `tikv-slow-diagnosis` would:
- route into the right diagnosis entry
- choose the right path and scope
- stop at the earliest proven bottleneck
- separate repeated spikes into distinct review windows when the case clearly
  contains multiple duration rises
- avoid overclaiming when evidence is incomplete

This evaluator produces review output and improvement candidates only. It does
not directly modify `tikv-slow-diagnosis`.

## Default Scope

- time range: last `30d`
- sources: `Pingkai ticket + Jira GTOC`
- owner filter: `江红梅`
- sample size: `5`
- require a Clinic link in the initial ticket body or Jira description
- exclude Grafana-only cases by default
- do not silently expand discovery to `TCOC`

## Replay Boundary

Allowed replay inputs:
- ticket summary
- ticket description
- cluster id if present in the initial ticket body
- Clinic links if present in the initial ticket body
- Grafana links if present in the initial ticket body
- incident time window if present in the initial ticket body

Disallowed replay inputs:
- all comments
- owner-side conclusions
- later-added evidence that appears only in comments
- retrospective or manually curated RCA summaries

## Entry Resolution

Supported evaluator entry modes:
- `explicit_case_set`
- `default_benchmark`
- `custom_scope_benchmark`

If the user asks for one live incident only, use `tikv-slow-diagnosis` or
`oncall-toolkit analyze` instead of this evaluator path.

## Multi-Window Review

When one incident contains multiple spikes inside the same case, the evaluator
may record optional `analysis_windows` and `window_reviews`.

Use multi-window review when:
- the ticket or owner-side evidence explicitly mentions multiple windows
- Grafana or Clinic evidence shows repeated spikes that should not be collapsed
  into one blended narrative
- the same mechanism repeats across multiple rises and the maintainer wants to
  preserve that pattern for future skill edits

Do not force multi-window output when the case has only one clear window or the
available evidence cannot distinguish the windows reliably.