governed-autonomy-looplisted
Install: claude install-skill broomva/skills
# governed-autonomy-loop
> Verbal incantation ("be autonomous", "merge when green") is **open-loop** control.
> A governed autonomy loop is the **closed-loop** version — a controller whose
> verification signals are causally independent of the agent (`h ⟂ U`), so
> "autonomous" becomes a machine-checkable behavior instead of a hope.
This skill packages a *demonstrated controller*, not a described one: the reference
loop ran in production, hit two real runtime failures, degraded gracefully without
paging a human, and self-healed to close its arc — all observable in
`loop-log.jsonl` (broomva BRO-1740 + BRO-1833, first fully-autonomous merge
2026-07-10). It generalizes that instance into a recipe.
## The latent-vs-deterministic split (the load-bearing idea)
| Deterministic → tested code (`scripts/`) | Latent → prose (`templates/runner-prompt.template.md`) |
|---|---|
| `tick.sh` — durable scheduler: fire-timing, flock, run-counter, quiet hours, mode decision, watchdog, child-guard | reconcile classification (merged? dead? stale?) |
| `loop_state.py` — in-flight fold, **reseed gate arithmetic**, busy-guard, arc-status contract | dispatch selection (which candidate, trust-bounded) |
| `validate_config.py` — fail-closed kill switch, DRY_RUN, num_or, partition guard | the govern/escalate judgment |
| `denylist_check.py` — tracker write-surface coverage | the adversarial (P20) review verdict |
| `mine_loop_log.py` — taxonomy + fixtures + drift-check from a running loop's log | th