kpi-healerlisted
Install: claude install-skill Shadysmetools/claude-skills
# KPI Healer
"Self-healing workers checking every KPI issue, verifying it and updating it, so the KPIs get
smarter every day." The healer never trusts yesterday's status: every run re-proves each issue
from live data, then either closes it with evidence, escalates it, or ships a fix PR.
## Configuration (`KPI_HOME`, default `./kpi`)
| Item | What it is |
|---|---|
| `scorecards/<date>.json` | Today's KPI scorecard: `[{ "id", "category", "value", "target", "status": "green|amber|red" }]`. If missing, run your scorer command (`SCORER_CMD`) first. |
| Issue store | One row per issue: `id, kind, category, status (open/acknowledged/regressed/auto_resolved/resolved), evidence, last_seen_at, ticket, pr`. Any table or a `issues.jsonl` file works. If it does not exist yet, build the list from the scorecard and say so. |
| `KPI-OWNERS.md` | Category -> owner agent/role (e.g. "Ingestion -> backend agent"). |
| `KPI-PLAN.md` | Targets and dates per KPI (context only). |
| Tracker | GitHub repo/project (or any tracker with search + create). |
Credentials come from your environment or a secrets file; never print them.
## Who runs it
Run as a tech-lead agent. It dispatches the category **owner agents** from `KPI-OWNERS.md`, but only
for categories with red KPIs or new/regressed issues, max 3 in parallel. Each owner runs steps 1-5
for its own KPIs and writes `owners/<category>/<date>.md`; the tech lead consolidates into
`healer/<date>.md`.
## Loop (per run)
0. **Optional pre-checks.**