Prompt file imported from Caetanogp/SISTEMA_GTM_REVENUE_AI_ENG (
.codex/prompts/autonomous-loop.md). Copyright stays with the author.
autonomous-loop
Run unattended development work against a pre-approved queue while the user is away. Use when asked to set up, start, watch or stop a "night shift" / autonomous loop run.
Read docs/playbooks/autonomous-loop.md and follow it. That playbook is shared with Claude Code
(.claude/skills/autonomous-loop/SKILL.md) — edit the playbook, never the wrapper.
The queue it works from lives at .handoff/AUTONOMOUS_QUEUE.md; the judge of "done" is
scripts/autonomous_gate.py. Never trust a self-reported "finished" — trust the gate's exit code.
Codex runs are controlled by scripts/codex_loop_supervisor.py. The normal executor handles one
item per process. A technical design gap returns the structured architecture_required outcome
before any implementation; the supervisor then invokes a read-only strong-model architect and a
separate reviewer. HUMAN_REQUIRED remains a hard stop. Never edit the queue, supervisor, gate,
policy, or external authorization artifact from an executor turn.
Project rules that always apply: AGENTS.md, specifically the "Autonomous loop mode" section.