Ariadne Loop: L1 loop contracts for coding agents

[zhangzeyu99-web]

I am building Ariadne Loop, a local-first workbench for turning rough coding-agent context into loop contracts, agent packets, verifier gates, and handoff reports.

Pattern picked

I started with an L1/report-first loop rather than an unattended loop. The closest fit is a mix of Loop Design Checklist plus PR Babysitter-style verifier gates: read current state, take one bounded action, verify, then decide whether to continue, stop, rollback, or ask.

First loop command

/loop report-first
Inspect the current repo and public page. Produce an agent packet with scope, verifier gates, stop rules, rollback criteria, and a handoff report. Do not make code changes unless the packet is accepted.

One surprise

The hard part was not generating a plan. The hard part was keeping the agent from expanding scope after every inspection pass. The useful primitive became the explicit continue | stop | rollback | ask decision at the end of each pass.

Ariadne publishes an agent-facing index here: https://zhangzeyu99-web.github.io/ariadne-loop/llms.txt

[zhangzeyu99-web]

Quick update after the latest iteration:

Ariadne Loop now generates a full Loop Run Kit instead of only a loop spec.

ariadne-loop quickstart --output .ariadne/quickstart writes:

• agent-packet.md for the next coding-agent turn,
• PROGRESS.md as externalized memory between turns,
• RUNBOOK.md for what to read, run, and report,
• reports.jsonl for structured turn reports,
• loop.json as the machine-readable loop contract,
• decision.json for the supervision decision.

The loop cycle is now inspect -> act -> verify -> persist -> decide, which maps more directly to the Orange Book framing around externalized memory, independent verification, verification debt, comprehension rot, token blowout, and human gates.

I also added case packs for desktop app development, GitHub Pages publishing, bug repair, docs refactor, and OpenClaw scheduled work, plus an ariadne-loop audit --dir <run-kit-dir> command to check whether a handoff is complete before giving it to another agent.

Repo: https://github.com/zhangzeyu99-web/ariadne-loop

I would still value feedback from people working on loop engineering: is this run-kit shape too heavy, too light, or missing a key artifact for long-running coding-agent loops?