insight-harvesterlisted

# Insight Harvester Turn raw observations into configuration improvements with evidence. ## Inputs - `memory/learnings/*.md` (raw, per-session) and `memory/learnings/CONSOLIDATED.md` - Pasted output of the built-in `/insights` command, when available - Recent `docs/specs/*/` story files with `## Findings` walls ## Method 1. **Cluster, do not enumerate.** Group observations by underlying cause, not surface symptom. "Worker touched out-of-scope file" three times across different specs is ONE pattern. 2. **Demand evidence.** Every pattern cites at least 2 occurrences (file + date). Single events are noted, not acted on. 3. **Map pattern -> smallest fix.** Prefer, in order: a path-scoped rule in `.claude/rules/` > a line in an agent prompt > a CLAUDE.md constitution change (most expensive context-wise - last resort). 4. **Write fixes as diffs**, not advice. "Add to `.claude/rules/api.md`: 'All new endpoints require an authz check; see wiki/auth-invariants'" - copy-pasteable. 5. **Specificity test:** a rule that would equally apply to any random repo is too generic to add. Project-specific or nothing. ## Output A ranked table: pattern | evidence (n, links) | proposed diff | expected effect | rollback note.