← ClaudeAtlas

opascope-loop-builderlisted

Prepare a long unattended agent run with a spending limit, a stopping condition you can check, and progress saved to files. Use when asked for unattended work, a resumable loop, or to keep Claude Code or Codex working until a verified outcome. Builds and validates the loop; launches only when execution is requested.
opascope/skills · ★ 0 · AI & Automation · score 75
Install: claude install-skill opascope/skills
# Build an unattended Claude Code or Codex loop Read [shared.md](shared.md) first. Keep the loop in the shared task directory. Read [references/contract.md](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 ch