Build an unattended Claude Code or Codex loop

Read shared.md first. Keep the loop in the shared task directory. Read references/contract.md before writing its files.

The portable core is an outcome, immutable criteria, durable context and independently runnable proofs. The launcher is a small foreground Python process. Neither runtime needs native goal commands or model aliases.

Build

1. Identify the intended outcome and what would prove it. Use prior answers and inspected files first. Ask only for material missing choices. A focused job gets one item; a longer job gets a dependency-ordered checklist. Both use durable artifacts so a fresh runtime call has enough context.
2. Read project instructions at every scope the work will touch. Put relevant complete constraint sections and the user's authority limits in SAFETY.md, with source paths. Include nested instructions in any worker's scope. Do not copy unrelated private context into a shareable loop.
3. Fill CONTEXT.md with the objective, input paths, reference facts and settled decisions. Use stable headings. Fill loop.json with immutable item IDs, descriptions, dependencies, proof argument arrays and a final behavioral proof. Include verifier script paths for fingerprinting and output paths for progress detection. No proof may rely only on the worker's done claim.
4. Write CHECKLIST.md with pending items matching the contract, NOTES.md with a checkpoint and decisions section, BLOCKERS.md with no unresolved entries unless actually blocked, and init.md with an explicit cold-read order.
5. Run the checks once to establish the baseline. Expected task failures are allowed; missing tools, malformed proofs or nonexistent inputs are not a valid baseline. Read proof code before running it. It runs with the user's privileges, so only trusted checks are eligible.
6. Run python3 "KIT" loop validate "/full/task/path", then loop seal with that path. These validate structure and fingerprint criteria; they do not prove the task is complete. Never clobber an existing sealed/live loop. Reuse it to resume; create a new task for a reviewed change in scope or criteria.

Launch and resume

If asked only to build a loop, deliver the artifact path, baseline findings and the concrete launch command, then stop. If the user also asked to run it, use the appropriate command, within the already authorized scope:

python3 "KIT" loop run "/full/task/path" --runtime claude --steps 10 --seconds 300
python3 "KIT" loop run "/full/task/path" --runtime codex --steps 10 --seconds 300

Replace KIT and the task path with actual resolved paths. Select the installed runtime requested by the user; leave their model selection intact. The runner starts fresh calls that read artifacts before each item, so it also resumes after context loss or a runtime switch. Inspect checkpoints and blockers before a subsequent explicit run. Clear a blocker only when its unblock condition is verified, preserving its history in NOTES.md.

Completion discipline

The worker updates deterministic item status only after showing actual proof. It never changes success criteria or verifier files. An independent reviewer must handle a judgment clause; if unavailable, record the missing review as a blocker, not a self-issued PASS. A reviewed scope amendment becomes a new sealed task linked to the old one, so removed requirements remain visible.

The runner checks all item proofs and the final proof itself. It detects criterion drift and stops for blockers, runtime failure, timeout, unchanged progress or budget exhaustion. None of those is completion. Exit 0 alone means all proofs passed. Read the completion artifact and inspect the resulting work before reporting success; a poor proof can still certify the wrong outcome.

Explain these limits: the launcher is not an OS sandbox or a scheduler; no automatic wakeup occurs. Runtime permissions may block an unattended action, and that is an actionable blocker. Do not bypass them to keep a loop alive.