Skill: Pattern Mine

Read-only history mining for trigger candidates. Does not design automation YAML. Query past state via grafana-ha-history; resolve live entity IDs via HA MCP when needed.

Steps

1. Normalize the brief. Turn narrative and/or structured events into an ordered anchor list: approximate time, what happened, optional entity hints, and the search window. Done when every anchor has a time (or relative order) and a plain-language label.

2. Seed-episode forensics. For the primary window, query state changes for hinted entities and obvious neighbors (same area, related domains: motion, presence, lock, door, alarm, light). Align timestamps to the anchors. Done when each anchor has supporting or contradicting sensor evidence cited.

3. Neighborhood expand. From entities that moved near anchors, discover related IDs (SHOW TAG VALUES, area packages, HA MCP). Drop noise (always-on chatter, unrelated rooms). Done when the candidate set is listed with why each was included.

4. Tier confidence.

   • Goal is episode understanding only → stop; label candidates episode-only.
   • Goal asks for reliable triggers → continue to step 5. Done when every candidate has a confidence tier.

5. Multi-occurrence pass (reliable only). Find comparable episodes in the parent window (default: last 14 days, need ≥5 matches unless the Bar says otherwise). For each candidate, count hits / misses / false positives. Cite InfluxQL (or HA history calls) used. If history is too thin, label insufficient-history — not reliable. Done when each reliable claim meets the Bar or is explicitly demoted.

6. Emit the Artifact. Ranked candidates, rejected/noisy signals, evidence, confidence tier. Hand to pattern-reviewer (do not self-grade against the Bar).

Default Bar

A candidate may be labeled reliable only if all hold (parent may tighten):

#	Requirement
1	Correlates with target anchors in the seed episode (cited timestamps)
2	Holds across ≥5 comparable episodes in the last 14 days (or parent N/window)
3	False-positive rate in those episodes is stated with evidence, or history is flagged insufficient
4	Every claim cites the query/tool used — no vibes

Return

Pattern mine — <slice or brief title>

Confidence: episode-only | reliable | mixed

Candidates (ranked):
- entity / event shape — why — evidence (times + query) — tier

Rejected / noisy:
- entity — why dropped — evidence

Gaps for reviewer:
- missing history, ambiguous anchors, open entity questions