Bounded, never infinite
Every unattended run declares a machine-checkable done condition, a hard cap, a cost or time budget, a no-progress rule, and escalation triggers before it starts.
▸ system architecture
Autonomy with an instrument panel.
The collection treats delivery as a bounded state machine: every phase has evidence, every loop has a stop set, and every long run has a durable way back in.
RUN CONTRACT / 06
01Bounded, never infinite
02Fails closed
03Verifies before acting
04Durable and resumable
05Escalates, never guesses
06Measures, never assumes
▸ lifecycle
Eight phases. One directional pass.
Each phase consumes grounded artifacts from the phase before it. Optional phases are explicitly skipped; required red gates block everything downstream.
01 · Define
/auto-spectestable spec
02 · Plan/auto-planvalidated DAG
03 · Build/auto-buildtask commits
04 · Clean/auto-simplifysimpler code
05 · Prove/auto-testmeaningful tests
06 · Gate/auto-reviewverified findings
07 · Measure/auto-performanceproven gains
08 · Ship/auto-shiprelease-ready
↻ close the loop auto-learn runs from the checkpoint; auto-map refreshes project context after a real completed run.
▸ autonomy contract
Six rules every skill inherits.
The implementation differs by platform. These guarantees do not.
Fails closed
A gate that did not run is not clean. Exhausted is not converged. Red evidence remains red all the way to the final report.
Verifies before acting
Findings that drive edits are attacked by independent skeptics. Rejected and unverified claims remain visible instead of quietly becoming work.
Durable and resumable
Long work records a checkpoint and resumes by skipping units already verified done. The checkpoint, not a session transcript, is the durable truth.
Escalates, never guesses
Ambiguity, irreversible actions, repeated stalls, and authority changes stop at a human boundary. One approval never becomes blanket permission.
Measures, never assumes
Tests, validators, benchmarks, diffs, and reachability decide whether work landed. Agent assertions are useful context, never proof.
▸ enforcement boundary
Be exact about what is mechanical.
Static guards block known dangerous tool shapes. Mutation checks, validators, benchmarks, checkpoints, and adversarial review cover the claims a hook cannot decide safely.
| Mechanism | Enforces | Evidence |
|---|---|---|
guard-test-integrity | Common static suite-gaming forms added to test files: .only, .skip, xit, and suppressions. | scripts/test-guards.sh |
guard-git-hygiene | Bulk staging, destructive reset or clean, commit -a, and unsafe force-push forms during a live build. | scripts/test-guards.sh |
guard-ship-irreversibles | Unapproved forced updates and remote ref deletion while ship preparation is active. | scripts/test-guards.sh |
checkpoint.mjs | Reinitializing a live run, demoting done work, and finalizing an incomplete checkpoint as done. | scripts/test-checkpoint.sh |
▸ durable truth
Resume from evidence, not memory.
Checkpoint
Locked, atomic tooling records unit state and evidence. Observability failures stay visible instead of being promoted into clean state.
Skip done
Resume skips only work that the durable state and repository evidence still support. Stale done records block rather than disappear.
Learn carefully
Candidate lessons must be true, general, and actionable before they are routed into provider-native project context.
Keep authority narrow
Plan approval governs the autonomous stretch. Ambiguity and irreversible actions still stop for a fresh user decision.