Inspiration
Long-running AI coding work rarely fails because an agent cannot generate one more patch. It fails because scope drifts, task windows end, evidence becomes detached from the artifact it proves, retries duplicate side effects, and a green-looking status is mistaken for real completion.
We built LoopSkill to make complex Codex work durable, reviewable, and explicitly finishable.
What it does
LoopSkill is an evidence-bound execution and completion protocol for long-horizon work in the Codex macOS App. It first runs an Intake Gate to decide whether a request should become a Loop at all. Ready requests become a validated Standard or Adaptive Controller Pack with:
- a read-only Controller plus scoped Worker, Reviewer, and Local Verifier roles;
- an MCP State Gateway as the only canonical state writer;
- bounded permissions, budgets, retries, and repair cycles;
- artifact-bound review and validation evidence;
- durable outboxes and report recovery that avoid duplicate dispatches;
- fail-closed transport pause and resume behavior;
- human Decision Cards bound to the exact goal, dispatch, artifact, review surface, and Controller turn; and
- one verifiable completion condition: canonical
FINALIZATION_ACKED.
Small or underspecified requests are routed to direct execution or clarification instead of being forced into orchestration.
How it works
- Intake checks scope, permissions, evidence sources, acceptance criteria, and whether Loop overhead is justified.
- Pack generation produces a self-contained work contract with roles, goals, budgets, repair limits, and stop conditions.
- Execution routes one exact dispatch at a time through a durable outbox.
- Review and verification bind reports and validation files to the current artifact and runtime identity.
- Human decisions are registered and applied through the State Gateway, so stale or replayed choices fail closed.
- Finalization requires final audit, a real heartbeat pause/readback, and canonical acknowledgement. A review PASS alone is not completion.
Built with Codex and GPT-5.6
LoopSkill existed before OpenAI Build Week and was meaningfully extended during the submission period. Codex was the primary engineering environment; GPT-5.6 supported implementation, incident analysis, test design, documentation, independent review, and release hardening.
The Build Week work progressed from the original submission release through public v3.3.3, adding schema-v3 State Gateway workflows, typed MCP runtime transport, exact-artifact report staging, lost-output recovery on the original outbox, transport pause/resume controls, canonical metrics, and the Decision Gateway for bounded visual review choices.
Real operated case
The strongest new evidence is an authorized, real long-horizon Life Blueprint run on the exact installed v3.3.3 build. In that run:
- a bounded visual decision was canonically REGISTERED and APPLIED;
- the final audit passed;
- the registered heartbeat was paused and read back from the App; and
- the canonical state reached
FINALIZATION_ACKED / LOOP_COMPLETE.
The product work stayed local-only: there was no product commit, push, pull request, merge, deployment, or external write. Private project content is not published; only the privacy-safe outcome and protocol evidence are summarized.
Release evidence
The public LoopSkill v3.3.3 release is bound to protected-main SHA 4a46ae95120dad44a1b2e10f42da4293fbaf4c08.
Release evidence includes:
- 657 local tests passing with 80.324457% branch coverage;
- two independent 5,000-case fuzz lanes passing;
- Skill, specification, schema, installer, risk, secret, and source/install zero-drift checks passing;
- an independent review with P0/P1/P2 findings at 0/0/0; and
- exact-tree compatibility CI with 632 canonical tests, 80.091827% branch coverage, both fuzz lanes, Linux/macOS isolated installs, and the final gate passing.
Challenges we ran into
The hardest problems were proving real Codex App task identity, keeping canonical state consistent across retries, recovering reports after lost process output without resending work, preventing old evidence from approving a new artifact, and carrying an explicit human decision across a preview whose loopback port could legitimately change.
The solution was not to invent stronger claims. Unsupported identities and mismatched decisions fail closed with zero canonical side effects. Failed disposable fixtures remain recorded as failures instead of being rewritten as passing evidence.
Accomplishments we're proud of
- A practical Intake Gate that rejects ceremony when a direct task is better.
- Deterministic state transitions with leases, journals, outboxes, and idempotent recovery.
- Review and validation bound to exact artifacts, reports, files, and runtime identity.
- Human steering that cannot silently rewrite canonical history.
- Bounded repair that pauses or stops instead of looping forever.
- A real operated run that reached the same explicit completion contract used by the protocol.
What we learned
Reliable coding agents need protocols, not just stronger prompts. Identity, evidence provenance, side-effect boundaries, human decision integrity, and explicit stop conditions matter as much as model capability.
What's next
Next we want to add a visual pack and evidence inspector, more reusable project templates, stronger App-native receipts when the host exposes them, privacy-safe operated-case metrics, and broader cross-version validation.
Evidence boundary
v3.3.3 is a public, exact-SHA release with strong local, CI, installation, review, fuzz, and real operated-case evidence. We do not claim that one bounded macOS run proves every platform, every Codex version, indefinite unattended operation, or universal production acceptance.
Built With
- codex
Log in or sign up for Devpost to join the conversation.