WikiSkill Evolve
Run Google's WikiSkill evolution loop (arXiv:2608.27454) — a faithful
implementation with Hermes as the reference backend. The agent evolves its own
skills: raw sessions → a maintainer distills failure patterns into a
persistent wiki → a proposer writes a candidate skill → a validation gate
accepts it only if R_val > R_best (git rollback otherwise).
When to use
- You want an agent's own experience (traces) turned into reusable skills
- You want to test whether a candidate skill actually helps, with statistics instead of vibes
- You're running the paper's protocol on your own tasks
Install
pip install wikiskill # Python ≥ 3.10; works with Hermes, Claude Code
Run the loop
wikiskill init myws # workspace + auto-graded bench
wikiskill evolve myws --iters 3 # train → maintain → propose → gate
wikiskill status myws # baseline, r_best, skill state
wikiskill compare wsA wsB --iters 5 # paired exact-binomial comparison
wikiskill transfer src dst # copy accepted skills to another ws
Reading the output
runs/state.json—baseline(S₀ on val),r_best,next_iterwiki/log.md— every maintenance/proposal/gate decision with evidencewiki/patterns/— distilled failure patterns (the raw material)wiki/skill-impact.md— rejected proposals stay visible (paper requirement)- Gate verdicts:
ACCEPTED(R_val > R_best, git commit),REJECTED(rolled back),no_action(proposer declined — a valid outcome)
Backends
wikiskill init myws --backend claude # Claude Code as the worker
Hermes is the reference backend; Claude Code ships in the box; codex/opencode are on the roadmap. All speak open SKILL.md, so evolved skills transfer.
Honest-expectation notes
- Each iteration costs ~$0.09 on free-tier models (gemini-lite class) —
turn budgets and
--max-turnsbound the spend - A weak model may produce
no_actioniterations — that's the gate working, not a failure; skill accumulation needs a reasonably strong proposer - The gate has rejected harmful skills in live runs — a rejection is a win