continuous-learning

Pattern extraction happens automatically at session end when your session meets the configured minimum length. The Skill detects error_resolution, user_corrections, workarounds, debugging_techniques, and project_specific patterns, then saves them to ~/.opencode/skills/learned/ for reuse in future sessions.

The Skill identifies five pattern types: error_resolution (how you fix errors), user_corrections (manual fixes you apply), workarounds (alternative solutions), debugging_techniques (approaches to troubleshooting), and project_specific (context-dependent methods). Which patterns to detect is configurable via patterns_to_detect in config.json.

Configure behavior in config.json with fields like min_session_length (minimum session size to trigger extraction), extraction_threshold (confidence level for pattern detection), auto_approve (whether to save patterns without review), and ignore_patterns (patterns to exclude). The Stop hook runs evaluation automatically at session end.

Learned patterns are saved to ~/.opencode/skills/learned/ by default, though you can customize the path with learned_skills_path in config.json. Once stored, these patterns become reusable skills available for future opencode sessions.

Yes—configure a Stop hook in ~/.opencode/settings.json to trigger pattern evaluation at session end. The Skill also requires config.json with settings like min_session_length and patterns_to_detect. Run a sample session with at least 10 messages to see learned skills created.

With auto_approve enabled, detected patterns save automatically to ~/.opencode/skills/learned/ without review. Disabled requires manual approval before patterns persist, giving you control over which learned skills become active for future sessions.