continuous-learning-v2

Add PreToolUse and PostToolUse entries to ~/.claude/settings.json pointing to the observe.sh script with 'pre' and 'post' arguments. Then create the ~/.claude/homunculus directory structure with instincts and evolved subdirectories.

An instinct is an atomic learned behavior with one trigger and one action, stored as a Markdown file with YAML frontmatter. It carries a confidence score from 0.3 to 0.9, a domain tag like code-style or testing, and evidence of the observations that created it.

Skills trigger probabilistically based on Claude's judgment, firing only about 50-80% of the time. Hooks execute deterministically on every tool call, so no patterns are missed and learning is comprehensive.

Yes, v2 is fully backward compatible. Existing skills in ~/.claude/skills/learned/ continue to work, the Stop hook still runs, and both versions can operate in parallel during a gradual migration.

No, observations stay local on your machine. Only distilled instincts (patterns, not content) can be exported for sharing, and you control exactly what gets exported.

Confidence increases when patterns are repeatedly observed or suggestions go uncorrected, and decreases when the user explicitly corrects the behavior or contradicting evidence appears. Instincts at 0.7 or above are auto-approved for application.