Learn From Session

Orchestrates the complete learning pipeline to extract lessons from model sessions and optimize CLAUDE.md rules.

Usage

/rules-learning-pipeline:learn-from-session <session-file> [--dry-run] [--verbose]

Pipeline Phases

1. Phase 1: Evaluate (model-evaluator)

   • Parse action sequences
   • Calculate metrics (F1, Precision, Recall)
   • Detect wasteful patterns (BDN, CIP, Verification Theater)
   • Extract lessons with scope classification

2. Phase 2: Review (critics-reviewer)

   • Validate scope classifications
   • Check lesson specificity (score ≥ 2)
   • Flag missed opportunities
   • Return verdict: APPROVED | NEEDS_REVISION | REJECTED

3. Phase 3: Optimize (prompt-optimizer)

   • Route lessons by scope
   • Apply peripheral bias
   • Create domain files if threshold met
   • Write optimized rules

4. Phase 4: Report

   • Generate summary report
   • Update changelog

Execution

When invoked, perform these steps:

Step 1: Validate Input

1. Verify input file exists
2. Check file has content
3. Detect input type (session_log | trajectory | failure)
4. Create checkpoint directory

Step 2: Run Phase 1

Task(
  subagent_type: "model-evaluator",
  prompt: "Evaluate session at {input_file}:
    1. Parse gold vs predicted actions
    2. Calculate metrics
    3. Identify error types including wasteful patterns
    4. Extract lessons with specificity validation
    5. Detect dismissive language patterns
    Output to: docs/evaluations/{session}-evaluation.md"
)

Step 3: Run Phase 2

Task(
  subagent_type: "critics-reviewer",
  prompt: "Review lessons at {lessons-raw.md}:
    1. Validate scope classifications
    2. Check lesson specificity
    3. Detect missed wasteful patterns
    4. Flag dismissive language not caught
    Output to: docs/evaluations/{session}-critics.md"
)

IF verdict == "REJECTED" AND retry_count < 2:
  Re-run Phase 1 with feedback

Step 4: Run Phase 3

Task(
  subagent_type: "prompt-optimizer",
  prompt: "Optimize lessons at {lessons-validated.md}:
    1. Route by scope (user-general, project, domain)
    2. Apply peripheral bias
    3. Enforce line limits
    Output to: docs/evaluations/{session}-optimized.md"
)

Step 5: Generate Report

1. Write updated rule files (unless --dry-run)
2. Update router index
3. Add changelog entries
4. Generate summary report

Output Files

Phase	Output	Consumed By
1	{session}-evaluation.md	Phase 2
1	{session}-lessons-raw.md	Phase 2
2	{session}-critics.md	Phase 3
2	{session}-lessons-validated.md	Phase 3
3	{session}-optimized.md	Phase 4
4	{session}-summary.md	User

Wasteful Patterns Detected

• BDN (Broad-Dismiss-Narrow): Full check → dismiss → narrow check
• CIP (Check-Ignore-Proceed): Errors found → no action → next task
• Verification Theater: "let me verify" → errors → "looks good"
• Redundant Tool Chains: Multiple tools when one suffices