Governed autonomy loop

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	the reconcile_skip reason classification (which reason to record)

The rule: precision work with one right answer per input lives in tested code; the model's intelligence builds the constraint that then constrains the model. The scripts are pinned by 65 unit/integration tests + an E2E smoke.

The invariant spine vs the swappable adapters

Keep the spine verbatim (references/invariant-spine.md): the loop shape (outer A–E + inner F), the ten binding invariants, the metacognitive gate, the irreversible-act delegation, resumability + rot-reseed, comment-first attribution, the trust boundary. Fill four adapters (references/adapters.md):

Adapter	What you provide	Reference
Tracker	list/read/comment/transition/read-answer + the write-surface denylist	linear-server MCP + templates/denylist.linear.json
Irreversible-act + enforcement	the act (merge/deploy/publish) + the 3-layer gate	PR-merge + P20 + .control/policy.yaml + merge-gate.yml
Runtime	what pokes tick.sh on a cheap interval	launchd / systemd / cron / k8s CronJob
Partition	the disjoint label discriminator (N governors, one queue)	agent-ok vs life-agent-ok

Procedure — instantiate a new loop

1. Scaffold the instance dir: copy scripts/ verbatim; copy templates/runner-prompt.template.md and fill every {{SLOT}} for your adapters; copy + edit templates/config.env.template (starts DRY_RUN=1).
2. Author the tracker adapter: point the runner-prompt at your queue's tool surface; write templates/denylist.<queue>.json (full write surface + both denylists) and run python3 scripts/denylist_check.py <that file> — it must PASS.
3. Author the enforcement adapter + ship the server-side net: follow references/server-side-net.md — copy templates/merge-gate.yml, prove it green on one PR, THEN require it in a branch ruleset with no admin bypass.
4. Wire the runtime: fill templates/launchd.plist.template or templates/systemd/*.template; set GAL_* env; enable it.
5. Validate before arming (all offline, zero side effects): python3 -m pytest tests/ · bash tests/smoke.sh · DRY_RUN=1 FORCE=1 GAL_CLAUDE_BIN=echo bash scripts/tick.sh → assert a well-formed digest + runner_exit rc=0.
6. Arm deliberately: review at least one DRY digest, then flip DRY_RUN=0. The first live fire is an operator decision, never an accident (DRY fails toward observation; the kill switch fails closed).
7. Dogfood the generalization: instantiate a SECOND loop (a different queue or partition) — proving the spine/adapter factoring holds is the real test that you skillified the pattern, not re-described the instance.
8. Keep it learning (once a loop is running): periodically mine_loop_log.py taxonomy <loop-log> — it grounds the controller in the loop's observed behavior and exits 3 on drift (a reason the loop emits that the skill's contract does not yet know), the signal to update loop_state + Step A. The reference loops mine to ✓ no drift across 1,836 live decisions. See references/live-telemetry.md.

Key parameters (config.env)

DISPATCH_ENABLED (kill switch, exactly 1) · DRY_RUN (live = exactly 0) · WIP_CAP · LABEL (the trust boundary — applying it authorizes code execution) · FIRE_INTERVAL_HOURS · RESUME_ENABLED · RESEED_TURN_CAP (context-rot bound) · RESEED_MAX_GENERATIONS (runaway guard). Full inline docs in the template.

Anti-rationalization

Excuse	Reality
"The governor can just merge when it's green."	Then it is a bot, not a controller. The governor NEVER performs the irreversible act (invariant 4) — an isolated arc does, only behind the independent enforcement pipeline.
"The prose controller is enough; skip the scripts."	The scheduler + reseed gate + fail-closed config are ARITHMETIC. Latent space doing them is the bug — it silently drops a safety block. They are tested code.
"I'll require the individual CI jobs."	Path-filtered CI skips on docs-only PRs and blocks them forever. Require the always-runs aggregate gate instead (references/server-side-net.md).
"Admin bypass is fine, I'm the admin."	The arc runs as an admin identity. Admin-bypass = no gate for it. No bypass.
"One loop proves it."	One loop proves the instance. Instantiate a second to prove the pattern generalized.

Validation (skill self-test)

• Deterministic core: python3 -m pytest tests/ -q (65 tests) + bash tests/smoke.sh (E2E DRY tick, zero side effects).
• Gate: python3 <skillify>/scripts/skillify_check.py <this dir> --run-tests exits 0 (SKILL.md contract + syntax-valid scripts + real unit tests).

References

• references/invariant-spine.md — the fixed control-loop core + the 10 invariants.
• references/adapters.md — the four adapter contracts + the optional Step-F adapters.
• references/server-side-net.md — the mandatory checklist for autonomous acts.
• references/scenarios.md — the gradeable latent-decision scenarios (the run-41 defer / reseed-exhaust / merge_not_authorized cases + live-grounded ones).
• references/live-telemetry.md — what the reference loops actually do (mined from 1,836 live decisions), the drift check, and how to keep the skill learning.
• Reference implementation: broomva scripts/ticket-dispatch/ (BRO-1740 + BRO-1833). Composes bstack primitives P3/P4/P5/P9/P12/P19/P20; it is the concrete realization of P19's external-trigger + across-session cube cell.