SkillHone

What SkillHone does

When a coding agent such as Codex, Claude Code, Cursor, Pi or ZCode, or any CLI-capable agent, hits a reproducible skill problem, it can report it to SkillHone while it works. SkillHone queues a repair that can touch SKILL.md, scripts, references, assets and repository tests in one change. Every repair is tied to an issue, a regression test, a Git diff, a run record and a local PR, and failed or untested changes do not enter the review queue.

Each skill has its own repository and decision history, so the next agent can see what failed and why a candidate was accepted or rejected. Review mode waits for your merge decision, automatic local merge is opt-in and still needs the linked tests to pass, and SkillHone never pushes a skill for you. Quick mode starts from a failure found during work, while Full mode uses a frozen dataset and verifier in a separate Eval repository. It installs straight from GitHub, with no registry account or hosted service.

Key features

• Captures reproducible skill failures while an agent is working
• Repairs SKILL.md, scripts, references, assets and tests in one change
• Each repair is tied to an issue, regression test, Git diff and local PR
• Per-skill repository with decision history
• Review mode by default, with opt-in automatic local merge
• Quick mode for field failures and Full mode with a frozen Eval repository

When to use it

• Fix a skill whose referenced script or document is missing
• Update skill instructions after an API changes
• Keep a reviewable history of why each skill change was accepted

Who it is for: Authors and maintainers of agent skills who want fixes driven by real failures and gated by tests.

How it fits with Hermes Agent

Hermes is named as one of the agentskills.io runtimes SkillHone works with, alongside Claude Code, Codex and OpenClaw; it is not specific to Hermes.

FAQ