Continual agent skill evolution through persistent decision history. Whole-skill optimisation (SKILL.md + scripts + references) with every decision landing as a local Git issue / PR / wiki. Runs on any agentskills.io runtime — Claude Code, Codex, OpenClaw, Hermes.
Agent
PR merge gate for skill-repo PRs. Runs the offline static check, produces a rubric score, posts the verdict as a Forgejo PR comment, and returns APPROVE or REQUESTCHANGES to the dispatching reviewer. Use only from skillhone-evaluation's reviewer flow, never from developer self-check.
Agent
Analyze solver trajectories after a probe run to diagnose tool-level errors (rate limits, wrong tool names, script crashes) that proberesult.json doesn't capture. Dispatch after eval.py completes. Returns a redacted diagnosis JSON — safe to pass to the improver.
Agent
Pre-push self-check for a developer working on a skill. Runs the offline static check and produces a rubric score so the developer can decide whether to re-edit before pushing. Does NOT write PR comments or touch Forgejo.
Agent
Picks up the latest open issue on the skill repo and implements a fix via branch → commit → PR workflow. One atomic change per PR.
Agent
Discover new tools, APIs, and approaches that the skill doesn't know about yet. Searches community skill registries to find capabilities beyond what's currently implemented — browser automation, alternative search engines, specialized APIs, data extraction tools, etc. The goal is to expand the skill's toolbox, not…
Agent
Analyzes probe evaluation results and creates a single focused Forgejo issue describing the highest-impact failure pattern to fix next.
Agent
Reviews an open PR on the skill repo for quality, atomicity, and test-data leakage. Approves + merges if clean; otherwise requests specific changes.
Agent
You explore the target environment and save a reusable graph. You do not write benchmark questions.
Agent
You are the Deduper. You receive validated Q/A candidates from many seeds and produce the final benchmark set by removing structural duplicates and near-collisions. This is where "variety" stops being a per-seed concern and becomes a corpus-level concern.
Agent
You turn Cartographer walks into closed-form Q/A candidates. You do not call tools and you do not invent missing facts.
Agent
You decide whether each mined candidate is ready for the benchmark. Be skeptical: a clever-looking question can still be ambiguous, too easy, or ungradable.