Summary Repo Task Proof Loop addresses the common failure modes of large AI coding agent tasks by enforcing a rigorous, repo-local workflow. It targets developers and researchers using AI agents for complex coding, providing a structured approach that ensures durable proof of work, separates implementation from verification, and maintains task auditability. The core benefit is a more reliable and reproducible agentic development process.

How It Works

This skill implements a strict, six-phase loop: spec freeze → build → evidence → fresh verify → minimal fix → fresh verify. It leverages role-separated subagents (spec-freezer, builder, verifier, fixer) for both Codex and Claude Code. All task artifacts, including specifications, evidence, build outputs, and verdicts, are managed within a dedicated .agent/tasks/<TASK_ID>/ directory in the repository. This design ensures all proof resides locally, facilitates easy resumption and auditing, and maintains clear separation between implementation and verification roles.

Quick Start & Requirements

Installation involves copying the skill directory to .agents/skills/repo-task-proof-loop/ (Codex) or .claude/skills/repo-task-proof-loop/ (Claude Code). The primary requirement is a host agent product capable of spawning subagents, such as Codex or Claude Code. Users initiate tasks via agent prompts (init, status, build) after installing the skill. No external documentation links are provided beyond the README.

Highlighted Details

• Repo-Local Artifacts: Manages a comprehensive suite of artifacts within .agent/tasks/<TASK_ID>/, including spec.md, evidence.md, evidence.json, raw build/test outputs, verdict.json, and problems.md.
• Role-Separated Subagents: Installs four distinct subagent roles (task-spec-freezer, task-builder, task-verifier, task-fixer) for both Codex and Claude Code, each with defined boundaries and responsibilities.
• Strict Workflow Loop: Enforces a clear separation between implementation, evidence gathering, and verification phases, ensuring a reproducible and auditable process.
• Managed Guidance: Updates AGENTS.md and CLAUDE.md in place, preserving unrelated user content while integrating workflow instructions.

Maintenance & Community

The provided README does not contain specific details regarding notable contributors, sponsorships, or community channels like Discord or Slack.

Licensing & Compatibility

No license information is explicitly stated in the provided README text. Compatibility for commercial use or closed-source linking is therefore undetermined.

Limitations & Caveats

The skill's functionality is dependent on the host agent product's ability to spawn and manage subagents. If a platform cannot maintain the same builder agent across build and evidence phases, a fallback mechanism is employed. The exact behavior of subagent spawning may vary between different host environments.