Agent Skills: Session Learnings Extraction

Session Learnings Extraction

Extract valuable lessons from completed work sessions and codify them into the reflection system.

When This Skill Activates

• After completing a work session with planning-with-files
• When findings.md or progress.md files exist with "Key Learnings" sections
• Before merging PRs that fixed bugs or discovered patterns
• When error logs show repeated issues that were resolved

Protocol

Phase 1: Scan Session Artifacts

Search for learning sources in order of priority:

1. Planning-with-files outputs:

   docs/plans/*/findings.md - "Key Learnings" sections
   docs/plans/*/progress.md - "Error Log" entries
   .taskmaster/docs/*/findings.md - Taskmaster findings
   

2. Conversation patterns (inspired by ConversationAnalyzer):

   • User corrections: "that was wrong", "I fixed it"
   • Repeated test failures on same error
   • Multiple attempts at same fix
   • Tool result patterns showing errors

3. Memory files:

   • CHANGELOG.md - Recent "fix:" entries
   • DECISIONS.md - Recent architectural decisions

Phase 2: Identify Learning Candidates

For each source, extract:

| Field | Source | | ------------ | ------------------------------------------ | | Title | "Key Learning" header or error description | | Anti-Pattern | "Problem" or "What went wrong" section | | Root Cause | "Why" or "Analysis" section | | Fix | "Solution" or "Resolution" section | | Impact | Financial/system consequences |

Learning Candidate Scoring:

• +3: Financial calculation bug
• +3: Repeated pattern (seen 2+ times)
• +2: Production impact documented
• +2: Security implication
• +1: Performance impact
• +1: Developer experience friction

Threshold: Score >= 3 suggests creating a reflection

Phase 3: Output Learning Report

# Session Learnings Report

**Session:** [DATE] **Sources Analyzed:** [count]

## Learning Candidates

### Candidate 1: [Title]

- **Score:** [X/10]
- **Source:** [file:line]
- **Category:** [Fund Logic | State | Math | API | Infrastructure]
- **Anti-Pattern:** [brief description]
- **Fix:** [brief description]
- **Recommendation:** CREATE REFLECTION / UPDATE EXISTING / SKIP

### Candidate 2: ...

## Actions

1. [ ] Create REFL-XXX for "[title]"
2. [ ] Update REFL-YYY with additional case
3. [ ] Skip "[title]" - already documented in REFL-ZZZ

Phase 4: Create Reflections

For each candidate with recommendation "CREATE REFLECTION":

1. Run duplicate check: python scripts/manage_skills.py check --title "Title"
2. Create reflection: python scripts/manage_skills.py new --title "Title"
3. Auto-populate from source:
   • Copy anti-pattern description
   • Copy fix implementation
   • Extract error codes
   • Set severity based on score

Integration with ConversationAnalyzer Patterns

Session Boundary Detection

Use 5-hour activity windows to group related learnings:

• Learnings within same session likely related
• Cross-session patterns indicate deeper issues

Tool Usage Correlation

Track tool invocations that led to discoveries:

• Failed Bash commands revealing environment issues
• Read tool revealing unexpected code patterns
• Grep tool finding similar issues elsewhere

Status Classification

| Recency | Classification | Action | | ---------- | -------------- | -------------------------- | | < 1 hour | Active session | Wait for completion | | < 24 hours | Recent session | Extract learnings now | | > 24 hours | Historical | Review for missed patterns |

Example Extraction

Input (from findings.md):

### Key Learning: Dynamic Imports Prevent Side Effects

**Pattern Success**: Dynamic import pattern works for preventing import-time
side effects.

**Key Learnings**:

1. Some files were already fixed
2. Not all skipped tests need dynamic imports
3. Client-side vs Server-side import differences

Output (learning candidate):

### Candidate: Dynamic Imports Prevent Side Effects

- **Score:** 4 (repeated pattern +3, DX friction +1)
- **Source:** docs/plans/integration-test-phase0/findings.md:53
- **Category:** Infrastructure
- **Anti-Pattern:** Static imports of server modules in test files cause
  initialization side effects
- **Fix:** Use dynamic imports in beforeAll for server modules
- **Recommendation:** CREATE REFLECTION

Usage

# Manual invocation
/session-learnings

# With specific session
/session-learnings --session "2026-01-15"

# Scan all recent sessions
/session-learnings --all-recent

File Outputs

• docs/skills/REFL-XXX-*.md - New reflections
• docs/session-learning-reports/YYYY-MM-DD.md - Learning report archive