WikiSkill: skill pack for Hermes Agent

Self-evolving skills for Hermes Agent: traces become a persistent wiki and gated skill patches

What WikiSkill does

WikiSkill runs a closed evolution loop. The agent executes training tasks with its current skills and produces raw traces. A Wiki Maintainer agent distills the traces into pattern pages covering root causes and fixes. A Skill Proposer agent reads the wiki and traces and proposes one skill change, either creating or patching a skill. The change is validated on held-out tasks and kept only if it scores strictly better than the best so far; otherwise it is rolled back with git reset --hard. The wiki is never rolled back.

On Hermes, the inference agent runs as hermes chat --oneshot in an isolated HERMES_HOME profile with bundled skills opted out and empty memory, so gating sees exactly the candidate skill set and the user's real profile is untouched. Sessions are exported with hermes sessions export. Claude Code, Codex and GitHub Copilot CLI backends ship in the box, and OpenCode is on the roadmap. The wikiskill init command creates a workspace with a 22-task auto-graded bench of 13 training and 9 validation tasks.

Key features

Closed loop of training runs, wiki maintenance, skill proposal and gating
Strict gating that keeps a skill change only if validation beats the best score
Each stage runs as a real agent turn in an isolated HERMES_HOME profile
Git-tracked wiki that is never rolled back
22-task auto-graded bench created by wikiskill init
Backends for Hermes Agent, Claude Code, Codex and GitHub Copilot CLI

When to use it

Letting an agent improve its SKILL.md files from its own failures
Running the paper's Algorithm 1 against real Hermes Agent runs
Building a persistent wiki of failure patterns and fixes

Who it is for: Researchers and agent builders who want to test skill evolution with real Hermes Agent runs.

How it fits with Hermes Agent

Hermes Agent is the reference backend and is built natively, using hermes chat --oneshot, isolated HERMES_HOME profiles and hermes sessions export.

How to install WikiSkill

These commands are copied from the project's README. Check the repository for the latest steps before you run them.

pip install wikiskill
wikiskill init demo
wikiskill status
wikiskill evolve --iters 3

Requirements: Python 3.10 or later and an agent CLI to evolve, such as Hermes Agent

FAQ

What is WikiSkill?

WikiSkill is an implementation of the paper of the same name that compiles agent experience into a persistent wiki and uses it to evolve skills. Only skill changes that score strictly better on held-out tasks are kept.

Does WikiSkill work with Hermes Agent?

Yes, Hermes Agent is the reference backend and every component runs as a real Hermes agent turn. Claude Code, Codex and GitHub Copilot CLI are also supported.

How do I install WikiSkill?

Run pip install wikiskill, then wikiskill init demo to create a workspace and wikiskill evolve --iters 3 to run the loop. The README says to upgrade and re-init any workspace created with a PyPI version before 0.1.5.