Repo Task Proof Loop 🧩 Skill Open source

What it does

Repo Task Proof Loop is a workflow skill for non-trivial coding tasks that keeps the proof of work inside the repository. When you start a task it creates a folder under .agent/tasks/<TASK_ID>/, installs project-scoped subagents for Claude Code and Codex, and adds managed blocks to AGENTS.md and CLAUDE.md. The agent then follows a fixed loop: freeze a spec, build, collect evidence, run a fresh verification, apply a minimal fix and verify again. Separating the builder from a verifier that starts with clean context is the core idea, and the saved files make a task easy to resume or audit later.

What is inside

• SKILL.md with the loop definition and reference files for commands, subagent roles and platform differences
• Four generated subagents for each host: spec freezer, builder, verifier and fixer, as Markdown for Claude Code and TOML for Codex
• Task artifacts: spec.md, evidence files with raw build, test and lint logs, screenshots, a verdict.json and a problems list
• scripts/task_loop.py with init, validate and status commands, and a packaging smoke test
• For Codex, an optional bounded fan-out to explorer or worker children when a task splits cleanly

Works with

Claude Code and Codex. You install the skill as a project skill in each tool's skills folder, or keep one copy and sync it.

How to install

Copy the skill folder into .claude/skills/repo-task-proof-loop/ or .agents/skills/repo-task-proof-loop/, then either mention it in a prompt or initialise a task directly:

python3 "$SKILL_PATH/scripts/task_loop.py" init \
  --task-id feature-auth-hardening \
  --task-file docs/tasks/auth-hardening.md

Maintenance and safety

The project is Apache-2.0 licensed. Its last update was in spring 2026, so check it against current host versions. There is no one-line installer, so setup is manual. The script writes into your repo, including guide files and agent definitions, so review the diff it creates. The README credits a research paper as inspiration, which is a design influence rather than evidence that the skill works better.

Who should use it

Developers using Claude Code or Codex on larger changes who want a written spec, stored evidence and an independent verification pass for each task, and who are happy to commit or ignore a .agent/ folder in their repositories.