continuous-learning

name

continuous-learning

description

Stop-hook session-end pattern extractor (v1) — flags sessions worth mining and saves reusable patterns as one-file learned skills under ~/.claude/skills/learned/. Use when wiring, testing, or debugging the Stop-hook extractor, tuning its config.json thresholds, or extracting a session's patterns manually via /learn. Superseded by continuous-learning-v2 (deterministic PreToolUse/PostToolUse instinct learning with confidence scoring and project scoping) for ongoing automatic capture — do NOT use this skill for instinct/confidence-scored learning, project-scoped instincts, or guaranteed per-tool-call observation; that is continuous-learning-v2.

Continuous Learning Skill (v1 — Stop-hook extractor)

Runs as a Stop hook (evaluate-session.sh) that fires once at session end. The script counts user messages in the transcript, skips short sessions (below min_session_length, default 10), and emits a stderr signal telling Claude to mine the session for extractable patterns and where to save them.

Relationship to continuous-learning-v2: v2 (~/.claude/skills/continuous-learning-v2/, also shipped as the skyyrose-core:continuous-learning-v2 plugin skill) supersedes this skill as the primary learning path — deterministic PreToolUse/PostToolUse observation hooks, atomic instincts, confidence scoring, project-scoped storage. v2's own Backward Compatibility section keeps this v1 Stop hook running in parallel as the lightweight session-end extractor, feeding ~/.claude/skills/learned/. Use v1 for that narrow job only; everything else belongs to v2.

When to use

• Installing the Stop hook into ~/.claude/settings.json, or diagnosing why it never fires.
• Tuning config.json (min_session_length, patterns_to_detect, ignore_patterns, learned_skills_path).