.agent/tasks/<TASK_ID>/, installs project-scoped Codex and Claude subagents, updates repo guidance, and drives a strict loop:
What it solves
Three problems make non-trivial coding tasks hard to track and audit:- Proof lives outside the repo. Results, test output, and screenshots disappear when a session ends.
- Implementation and verification are mixed. The same agent that writes code also declares it correct, with no independent check.
- Task state is opaque. Picking up a half-finished task requires reconstructing context from scratch.
What it creates
Runninginit creates the following structure inside the target repository:
AGENTS.md- The repo’s Claude guide file:
CLAUDE.mdor.claude/CLAUDE.md
Compatibility
| Requirement | Details |
|---|---|
| Python | 3.10 or later (for bundled scripts) |
| Codex | Project skill at .agents/skills/repo-task-proof-loop/ |
| Claude Code | Project skill at .claude/skills/repo-task-proof-loop/ |
Quick Start
Go from zero to a completed verified task as fast as possible.
Workflow
Understand the full spec-freeze → build → evidence → verify → fix loop.