opencode-auto-improve

Auto-capture skills and durable memory from your opencode sessions, present them for your approval, and prune what is no longer useful. opencode learns like Hermes.

Zero-config standalone plugin. MIT. No opencode core changes.

Install

Add it to your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-auto-improve"]
}

That's it. Opencode auto-installs it from npm (via Bun) at startup.

What it does

1. Capture — at every session end, the plugin classifies what happened:

   • wrote files / committed / installed → skill candidate
   • durable fact about your environment or preferences → memory candidate
   • neither → nothing Candidates are staged in ~/.agents/learning-staging/. The plugin never writes to your live skill library.

2. Approve — no command needed. At the end of a session that produced candidates, the plugin injects a chat message presenting each one ("[skill] — " / "[memory] ") and asks you to approve or reject each (yes/no per item, or "all"/"none"). You never browse folders: the agent curates, you approve. On approval:

   • skill → ~/.agents/skills/<name>/SKILL.md (description ≤ 200 chars to keep the discovery index lean)
   • memory → ~/.agents/memory/facts.md (injected at session start on the next launch)
   • The staged JSON is renamed .promoted (or .rejected for rejections) so it is not presented again. Duplicate candidates (same fact already in facts.md, or a skill with the same name already staged/live) are never staged twice.

3. Prune — the weekly checkup is /audit-skills (the companion skill): classify keep/remove/merge, propose removals, you approve. "Less is more" — a skill that is not useful must be removed.

Layout

opencode-auto-improve/
  index.js                        plugin entry (exports FmPrimaryLearning)
  plugins/fm-primary-learning.js  capture + memory injection (session.idle, session.created)
  learning/fm-learning-core.js    classify, stage, promote, dedup (shared logic)
  command/learn-review.md         optional /learn-review command (present + approve)

Runtime state (independent of any other tool's data dirs):

~/.agents/
  learning-staging/skills/<name>.json   staged skill candidates
  learning-staging/memory/<id>.json     staged memory candidates
  skills/<name>/SKILL.md                live skills (promoted only)
  memory/facts.md                       live memory (promoted only)

Why

Your agent improves with every session: procedures become skills, facts become memory, and the library stays lean because the audit removes what the models no longer need.

License

MIT