Continuous Learning

This is a runtime policy document, not an ordinary task skill. It configures the agent's learning behavior at session end. The agent client (OpenCode) loads this to enforce pattern-extraction-on-stop discipline. It does not implement the extraction logic itself — it prescribes the behavior the client must follow.

Continuous learning runtime (OpenCode)

This opt-in OpenCode runtime supplies the stop-hook assets used by plugins/continuous-learning-stop-hook.js; it is not a shared agent skill.

Automatically evaluates OpenCode sessions at session end and extracts reusable patterns into new “learned skills” for future reuse.

When to Activate

Use this runtime when you want:

• Automatic pattern extraction from OpenCode sessions
• A Stop hook that evaluates the just-finished session transcript
• A workflow to review/curate learned skills in:
  • ~/.config/opencode/skills/<learned-skill>/SKILL.md
  • ~/.config/opencode/skills/learned/.continuous-learning-index.json
• Tuning for extraction thresholds, categories, and auto-approval
• Comparing skill-based v1 extraction vs “instinct-based” v2 approaches

How It Works

This skill runs as a Stop hook at the end of each OpenCode session:

1. Session Evaluation
   • Ensures the session is “worth learning from” (default: 10+ messages)
2. Pattern Detection
   • Scans transcript for repeated solutions, corrections, debugging flows, and conventions
3. Skill Extraction
   • Writes one or more skill directories into:
     • ~/.config/opencode/skills/<descriptive-slug>/SKILL.md
   • Stores learning metadata in:
     • ~/.config/opencode/skills/learned/.continuous-learning-index.json
   • Each extracted skill is:
     • Named deterministically
     • Tagged by pattern type(s)
     • Saved with a short “When to use / Steps / Examples / Caveats” template

Learned File Naming

Learned skills are written with meaningful, descriptive names derived from the session content rather than generic pattern categories:

• <descriptive-slug>/SKILL.md (e.g., angular-facade-debugging/SKILL.md)
• <descriptive-slug>-2/SKILL.md (counter suffix for collisions)

The slug is built from the most distinctive terms found in the session transcript. The content also includes a signature: in frontmatter for deduplication.

Files / Layout

Recommended structure:

• ~/.config/opencode/runtime/continuous-learning/
  • README.md (this file)
  • config.json
  • hooks/stop.sh
  • bin/evaluate-session.js (or evaluate-session.py)

Learned output:

• ~/.config/opencode/skills/<descriptive-slug>/SKILL.md
• ~/.config/opencode/skills/learned/.continuous-learning-index.json

Configuration

Edit config.json:

{
  "min_session_length": 10,
  "extraction_threshold": "medium",
  "auto_approve": false,
  "skills_root_path": "~/.config/opencode/skills/",
  "learned_metadata_path": "~/.config/opencode/skills/learned/",
  "patterns_to_detect": [
    "error_resolution",
    "user_corrections",
    "workarounds",
    "debugging_techniques",
    "project_specific"
  ],
  "ignore_patterns": ["simple_typos", "one_time_fixes", "external_api_issues"],
  "max_skills_per_session": 3,
  "dedupe_window_sessions": 20
}

Legacy note: learned_skills_path is still supported for backward compatibility. If it points to .../skills/learned/, the generator will still write skills to .../skills/<slug>/SKILL.md and keep metadata in the learned directory.

Shared setup note: if you want learned skills shared between OpenCode and ClaudeCode, point skills_root_path to ~/.claude/skills/ and learned_metadata_path to ~/.claude/skills/learned/.