youdede toolkit: skills, hooks and subagents for coding agents
Every artifact states what it refuses to do, and ships the tests or behavioral cases that check it. Install one at a time. Read every line before you do.
The problem
Your agent just ran git push --force on main. Or committed a .env. Or told you the migration was safe after reading one error line. Or answered a question about a database it never connected to.
You cannot review every action it takes. So the rules have to live where the agent is, not in your head.
What this is
A catalog of 319 things you can install into a coding agent, one command each:
- 170 skills: procedures an agent loads when a task matches. Each one states what it does, what evidence it requires, and what it refuses to do.
- 64 hooks: deterministic checks on an agent's tool calls and session events. Most inspect a command or file write before it runs and block it; a few only observe and warn. No model in the decision path.
commit-guard scans staged blobs for secrets. protected-branch-guard refuses a force push to main.
- 62 subagents: specialists with a deliberately narrow brief, no authority to edit or approve anything, and a required section for what they could not verify.
- 13 packs: the above composed into versioned outcomes, with explicit member-to-evidence mappings and behavioral evaluations. The workflow packs include
plan-and-verify-loop, session-continuity, daily-driver, mcp-baseline, remote-review-evidence, and browser-verification.
Every artifact is MIT-licensed plain text, inspectable before install.
Try one in thirty seconds
npx @youdede/toolkit add commit-guard --target claude-code --scope project
Then stage a fake secret and try to commit. The hook blocks it, names the rule, and never prints the secret back to you.
Prefer to look first:
npx @youdede/toolkit info commit-guard # what it does, what it refuses
npx @youdede/toolkit add commit-guard --target claude-code --dry-run
--dry-run prints every file it would write and exits. Nothing on disk changes.
Six runtimes, honestly
| Claude Code | Codex | Gemini CLI | GitHub Copilot | Cursor | OpenCode |
|---|
| Skills (170) | 170 | 169 | 169 | 169 | 169 | 169 |
| Hooks (64) | 64 | 22 | 22 | 0 | 24 | 22 |
| Subagents (62) | 62 | 0 | 62 | 62 | 0 | 50 |
| Commands (10) | 10 | 0 | 0 | 0 | 0 | 0 |
| Contexts (8) | 8 | 0 | 0 | 8 | 8 | 0 |
| MCP servers (5) | 5 | 0 | 0 | 0 | 5 | 0 |
A count below the total is a refusal, not a gap. An artifact is withheld from a runtime whose event, permission, or failure contract is not equivalent, because a gate that silently fails open is worse than no gate. docs/targets.md explains every withheld adapter.
Browse the catalog
Every artifact is listed, with what it does and which runtimes carry it, so you can read before you install:
| | |
|---|
| Skills | 170 | Procedures an agent loads when a task matches |
| Hooks | 64 | Deterministic gates that block a command before it runs |
| Subagents | 62 | Specialists with a narrow brief and no authority to act |
| Packs | 13 | The above composed into versioned, evidence-backed outcomes |
Those pages are generated from the catalog and checked by the build, so they cannot drift from what the repository contains.
Find what you need
npx @youdede/toolkit search secret # name, slug, summary and refusals
npx @youdede/toolkit list --problem commit-safety
npx @youdede/toolkit list --domain security --risk high
npx @youdede/toolkit list --kind hook --json # every read command takes --json
npx @youdede/toolkit packs
npx @youdede/toolkit pack-info daily-driver
npx @youdede/toolkit verify-pack-evidence daily-driver --report .youdede/evidence/daily-driver.json
search matches what an artifact refuses to do as well as what it does, so you can find a gate by the thing you want prevented. A filter value the taxonomy does not carry is refused rather than answered with an empty list, and every read command can answer in JSON. docs/install.md describes the full command surface, the removal dry run, shell completion, and the exit codes.
The artifact catalog, pack catalog, and taxonomy are generated, stable, and suitable for a search index or an API of your own. The catalog format describes their fields, including the links between artifacts. The pack evidence contract explains what a verified report proves and what it does not.
Each artifact also carries risk_reason and operational.latency_reason. Risk describes what using the artifact can do, not the severity of a problem it reviews. operational.latencyClass describes the work involved, not a measured response time. The classification audit records the evidence and unresolved cases.
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
