learn

/learn Command

Architecture Note

Commands handle ALL user interaction. The learning-extractor agent runs in two phases.

Phase 1: Agent analyzes conversation, returns candidate learnings as structured text. Phase 2: Agent creates only user-approved components with user-specified types and scopes.

Execution

Step 1: Check Toolkit

Use Bash to check: cat $HOME/.claude-evolve/active 2>/dev/null

If empty: "Run /evolve init first." and STOP.

Step 2: Check for Explicit Intent (Shortcut)

Before running discovery, check if the user's request contains explicit type and scope:

• Explicit type from: "rule", "skill", "agent"
• Explicit scope from: "universal", "project", "this project only"

Examples: "save this as a universal rule", "remember this as a project skill"

If both type and scope are explicit: Run Step 3 (discover) to identify what to capture, then skip Steps 4-6 (presentation and selection) and go directly to Step 7 with the user's explicit type and scope applied to all discovered candidates.

If only scope is explicit (e.g., "remember this universally"): Pre-set scope, continue to Step 3 for normal discovery and selection.

If no explicit intent detected: Continue to Step 3.

Step 3: Discover Learnings (Phase 1 -- Agent)

Spawn claude-evolve:evolve-learning-extractor via Task tool with:

• action: "discover"
• topic: The user's trigger text if they provided a hint (e.g., "save what we learned about Redis")

The agent analyzes the full conversation context available through its Task invocation. It also checks for session signals at $HOME/.claude-evolve/signals/${CLAUDE_SESSION_ID}.json.

The agent returns at most 4 candidates, ranked by learning value. Each candidate includes: name, summary, detail, suggested_type (skill|agent|rule), suggested_scope (universal|project), reasoning, and consolidation target if applicable.

If agent returns no candidates: Output "No significant learnings found in this session." and STOP.