SkillHone CLI
A Tencent CLI and two skills that repair broken agent skills: they capture a failure, make a tested fix and open a local PR.
- 192 ★ on the repository
- Last commit
Broad access
Why this level
- Installation runs a build from GitHub outside the agent sandbox and edits the agent's global instruction file.
- The repair agent edits skill files and creates branches and commits in their repositories.
Install
- Works in
Manual install
npm install -g --install-links=true git+https://github.com/Tencent/SkillHone.git#mainThe command from the install guide. There is no npm registry package. Then `skillhone setup --agent codex` or another agent.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Installation runs a build from GitHub outside the agent sandbox and edits the agent's global instruction file.
- The repair agent edits skill files and creates branches and commits in their repositories.
This is third-party code. Review the repository files before installing.
What it does
SkillHone watches the skills an agent works with. When an agent hits a missing script, a broken reference or an outdated instruction, SkillHone records the evidence, repairs the whole skill repository (`SKILL.md`, scripts, references, assets, tests) and returns a local pull request. The issue, test, diff, run and decision stay together so the next agent builds on what is known. The `skillhone` tool works with Codex, Claude Code, Cursor, Pi and ZCode and uses the current agent's model and account. A web workbench shows which skills need attention, repair history and local PRs. A person reviews the fix by default, automatic merge is a separate opt-in requiring linked tests to pass, and pushing stays with the user.
Who it is for. People who maintain their own agent skills and want to close their failures with verifiable fixes instead of manual patches.
Good fit when
- A skill references documents that never made it into the repository.
- After an API update the skill instruction is stale and the agent repeats the error.
- Failures from several agents must be gathered into one issue with a regression test.
- A skill fix is reviewed in a local PR and decision history before merging.
- The agent is asked to install SkillHone and attach its skills to the runtime itself.
Not a fit when
- You need to fix application code; SkillHone repairs the files of skills themselves.
- Node.js 22.19 and Git are missing on the machine.
- You want fixes pushed automatically with no review; by default a person reviews each repair.
Example request
My skill references a reference file that does not exist. Record the failure in an issue, add a regression test and prepare a local PR with the fix.Limitations
The package is not published on npm: it installs from the GitHub main branch with `npm install -g --install-links=true git+https://github.com/Tencent/SkillHone.git#main`, and installation runs a TypeScript build outside the agent sandbox. It needs Node.js 22.19 or newer and Git. The version is labeled 2.0.0-rc.1, a release candidate. Installation adds a marked routing rule to the agent's global instruction file and links two skills. Repository metadata lists the license as NOASSERTION, so read the LICENSE file.
How to disable. Uninstall the package with `npm uninstall -g` using the name from `package.json`, and remove the skill links and the marked rule from the agent's global instruction file.
FAQ
How do I install SkillHone?
Ask the agent to follow `docs/install/skillhone.md` or run `npm install -g --install-links=true git+https://github.com/Tencent/SkillHone.git#main`. Then `skillhone setup --agent codex` connects it to the chosen agent.
Is an API key needed?
Not for installation, importing skills, recording issues or opening the workbench. By default the current agent makes repairs with its own model.
Which agents are supported?
Codex, Claude Code, Cursor, Pi and ZCode. The `--agent all` option connects all five at once.
Does it edit files itself?
Fixes go to separate branches and into a local PR. A person merges by default, automatic merge is a separate opt-in and pushing stays with you.
What is evaluation-driven optimization?
A second path for broader improvement: a frozen dataset and verifier in a separate evaluation repository, then candidate iterations and a local PR.
Discussion
No questions or reviews yet.
Related
CLI toolsSkill Zoo finds, installs, updates and audits skills for Claude Code, Codex, Cursor and other agents in a window and a CLI.
A set of independent skills against drift in large codebases: hard-cut, root-cause tracing, lean testing and CI right-sizing.
The author's personal set: shared agent rules and dozens of skills for GitHub, the browser, macOS, messaging and releases, plus a sync script
Tech Leads Club Agent Skills
Agent Skills
A vetted skill registry with a CLI that installs into twenty-plus agents and an MCP server for on-demand skill lookup
Brooklyn skills: 25 portable agent skills for pull requests, triage, cleanup, UI checks and research
Skill and agent manager with a lock file: installs one set into Claude Code, Codex, Cursor and others and keeps your edits.
README in short
The docs split the work into two paths: runtime maintenance from a reproducible failure and evaluation-driven optimization on a frozen dataset. There is a public Phoenix tracing example and a link to an EMNLP 2026 paper.
README badge
Original SKILL.md
--- name: skillhone description: Local Issue, pull-request, and Wiki workbench for agent skills. Use when a user asks what skill problems, Issues, PRs, repair trajectories, work records, completed optimizations, pending fixes, or pending synchronization exist; when an Agent needs to report a reproducible skill defect; or when the user wants to initialize, inspect, optimize, review, merge, or synchronize a local skill repair. If a managed Skill has a missing referenced file, failing test, or broken command, invoke skillhone-auto-optimization before editing implementation files or installing dependencies. Uses a local SQLite/Git backend, the current Agent for new homes, and optional DeepSeek Harness, with no hosted Git service. --- # SkillHone Use SkillHone as the local maintenance layer for Agent-skill repositories. One Skill always maps to one Git repository and therefore owns an isolated set of Issue numbers, PR numbers, Wiki records, branches, tests, and repair Runs. The Catalog indexes those repositories but never combines their state. Any Agent runtime can call the CLI. The default `auto` executor prepares a repair assignment for the current Agent, preserving its model, account, and provider. `native` selects this path explicitly; `dsh` selects the optional managed DeepSeek Harness. No executor switch or fallback is automatic. Repair authorization is independent from the Skill ownership and merge boundaries. ## Route the user's intent first Treat inspection and synchronization as different operations. Never turn a status question into a write: - “修好了吗 / did the repair succeed / 测试通过了吗 / 优化到哪了” means inspect only. Run `skillhone --skill <name> --json status`, then inspect the linked Issue, Run, and PR when needed. Report whether repair ran, tests passed, and the PR is open or merged. Do not merge and do not run `sync apply`. - “已经同步了吗 / 我现在用的是新版本吗 / 需要同步吗” means inspect copy-back state only. Run `skillhone --json sync status <name>` and interpret the state table below. Do not run `sync apply`. - “帮我同步 / 把修复应用回去 / update my installed Skill” explicitly authorizes copy-back. First run `sync status`; run `skillhone sync apply <name>` only when it returns `ready_to_sync` with `can_apply: true`. `sync` never merges a PR. A repair can have passed tests while still waiting in an open PR, and a merged repair can still be waiting for copy-back. ## Import existing Skills Import one Skill directory or discover the Skills already installed for an Agent runtime. Before importing, ask the user which ownership mode they want: