Agent Task Contract

English | 简体中文

agent-task-contract is a skill for coding agents that keeps repository work scoped, verifiable, and ready for review.

It gives an agent a lightweight workflow for:

• defining the task before editing
• keeping non-goals and scope boundaries visible
• choosing verification that matches the risk of the change
• recording useful failure context instead of blindly retrying commands
• handing off changes with clear test evidence and PR notes

Repository Layout

agent-task-contract/
  SKILL.md
  agents/openai.yaml
  references/
  scripts/inspect-repo.ts
scripts/
  validate-skill.ts
.github/
  workflows/validate.yml
  PULL_REQUEST_TEMPLATE.md

Install

Copy or symlink the agent-task-contract folder into the skill directory used by your agent runtime.

For Codex-style local skills:

Copy-Item -Recurse .\agent-task-contract $env:USERPROFILE\.codex\skills\

For Claude Code, OpenClaw, or other agent tools, keep SKILL.md, references/, and scripts/ together and place the folder where that tool loads skills or plugins.

Examples

• Bug fix with verification evidence
• Scoped feature change
• PR-ready handoff
• Browse all examples

Validate

Install the TypeScript toolchain first:

npm install

npm run validate

Inspect a repository before using the skill:

npm run inspect

The helper scripts are written in TypeScript and compiled into dist/ by npm run build.

Example Prompts

Use $agent-task-contract to fix this failing test and show the verification evidence.

Use $agent-task-contract to prepare this repo change for a PR-ready handoff.

Contributing

Keep the skill folder focused on runtime instructions. Put detailed checklists in references/, deterministic helpers in scripts/, and repository maintenance files at the project root.

Before opening a PR, run:

npm run validate