Continually improve your Agent Skills — and keep what was learned.
Get started · 中文 · Public example · Paper
Give your coding Agent this prompt:
Install SkillHone from
https://github.com/Tencent/SkillHone, followingdocs/install/skillhone.md. Make the CLI and Skills available to this Agent, use the documented--install-links=trueGit installation command, and verify the installation.
Prefer the terminal? Start with the installation guide for installation, provider configuration, and verification.
Accepted at EMNLP 2026 Industry Track. The paper reports +15.8 points on GAIA and +3.2 points on WebWalkerQA-EN against its commercially backed deep-research baseline, without a pre-integrated search stack. Read the paper and evaluation context.
An Agent hits a missing script, a broken reference, or an outdated instruction. SkillHone captures the evidence, repairs the complete Skill repository, and returns a tested local PR. The Issue, test, diff, run and decision stay together, so the next Agent can build on what was learned.
Works with CLI-capable Agents, including Codex, Claude Code, Cursor, Pi and ZCode. DeepSeek Harness runs the repairs; your Agent reports the problem.
skillhone-launch.mp4
Inspect the current workbench
See which Skills need attention, expand their repair records, and inspect each local PR before merging.
Reproduce a real repair. The Phoenix tracing walkthrough starts with four referenced documents that were never shipped, then follows the failure, regression test and repair evidence. Public source: Issue #2567 · merged fix #2568.
| You need | SkillHone keeps |
|---|---|
| A problem the next Agent can reproduce | One Issue with sanitized evidence and a linked regression test |
| A complete repair | Changes to SKILL.md, scripts, references, assets and tests in one Git diff |
| A result you can review | Test outcomes, repair run, commits and a local PR |
| A reason the next Agent can reuse | Per-Skill Wiki records and persistent decision history |
Repairs use isolated branches. Review is the default; automatic local merge is opt-in and requires linked tests to pass. Pushing remains a separate user action. The local workbench brings Issues, runs, PRs, approval state and Wiki records into one view.
| Runtime maintenance | Evaluation-driven optimization | |
|---|---|---|
| Start with | A reproducible failure encountered during real work | A frozen dataset and verifier in a separate Eval repository |
| Use it for | Missing files, broken scripts, stale instructions and API drift | Broader capability and quality improvements |
| Review | Issue, regression, repair and local PR | Baseline, candidate iterations, validation and local PR |
| Learn more | Runtime workflow | Benchmark workflow |
Zhiwei Li and Yong Hu. SkillHone: A Harness for Continual Agent Skill Evolution Through Persistent Decision History. EMNLP 2026 Industry Track. ACL Anthology · arXiv:2606.08671
SkillHone uses DeepSeek Harness as its repair executor. It works on an isolated branch and runs the linked tests; SkillHone keeps the resulting local PR, evidence and decision history together.