cross-task-learner

What this skill does

# Cross-Task Learner Skill

Enable agent loops to learn from similar past tasks and share discovered patterns across multiple concurrent or sequential loops.

**Research Foundation**: REF-013 MetaGPT - 159% improvement with shared state

**Version 2.0**: Multi-loop awareness with loop_id tracking

---

## Overview

This skill provides two core capabilities:

1. **Pattern Extraction** - On loop completion, extract reusable patterns from execution history
2. **Pattern Injection** - On loop start, inject relevant patterns from previous loops

### Benefits

| Benefit | Impact |
|---------|--------|
| Faster resolution | Patterns eliminate redundant debugging |
| Higher success rates | Proven approaches applied automatically |
| Accumulated wisdom | System gets smarter over time |
| Anti-pattern detection | Failed approaches flagged and avoided |

### Research Basis

From REF-013 MetaGPT:
- **159% improvement** with shared state across agents
- **Publish-subscribe pattern** enables knowledge sharing
- **Structured outputs** become inputs for other agents
- **Memory persistence** critical for cross-session learning

---

## Pattern Extraction (On Loop Completion)

### Trigger

- Agent loop completion (success, partial, or failure)
- Manual extraction request via `aiwg ralph-extract-patterns {loop_id}`

### Process

```yaml
extraction_steps:
  1_analyze_loop_history:
    - Load loop state from .aiwg/ralph/loops/{loop_id}/state.json
    - Load iteration analytics
    - Load debug memory
    - Load reflection history

  2_identify_error_fix_pairs:
    - Scan iterations for test failures
    - Identify fixes that resolved errors
    - Extract error signature + fix approach
    - Compute initial success rate (1.0 for first occurrence)

  3_identify_successful_approaches:
    - Analyze task category (testing, debugging, refactoring, etc.)
    - Extract step sequence that led to success
    - Note tools used and iteration count
    - Identify preconditions and benefits

  4_identify_failure_patterns:
    - Detect repeated same errors (anti-patterns)
    - Note approaches that led to scope creep
    - Flag patterns that caused quality degradation
    - Record better alternatives if discovered

  5_extract_code_templates:
    - Identify successful code changes
    - Generalize with placeholders
    - Document use case and placeholders
    - Tag by language and purpose

  6_check_for_duplicates:
    - Compare against existing patterns in registry
    - Merge if >80% similar
    - Update usage count and success rate if duplicate

  7_store_in_registry:
    - Add to .aiwg/ralph/shared/patterns/{type}-patterns.json
    - Update patterns index for semantic search
    - Link to source loop_id

  8_update_effectiveness_metrics:
    - Increment pattern counts
    - Update cross-loop benefit statistics
    - Log extraction event
```

### Example Extraction

**Input** (from loop `ralph-fix-auth-a1b2c3d4`):
```yaml
iteration_2:
  error:
    type: "TypeError"
    message: "Cannot read property 'email' of null"
    location: "src/auth/validate.ts:42"

iteration_3:
  fix_applied:
    description: "Added null check for user object"
    diff: |
      + if (user == null) {
      +   throw new ValidationError("User is required");
      + }
  test_results:
    passed: 12
    failed: 0
```

**Output** (extracted pattern):
```yaml
pattern_id: "pat-error-null-check-015"
type: "error_pattern"
error_signature:
  error_type: "TypeError"
  error_pattern: "Cannot read property '.*' of null"
  error_location_hints:
    - "*.ts:validate*"
fix_approach:
  description: "Add null check before property access"
  fix_category: "add_null_check"
  code_template: |
    if ({{variable}} == null) {
      throw new ValidationError("{{message}}");
    }
  code_template_language: "typescript"
source_loops:
  - loop_id: "ralph-fix-auth-a1b2c3d4"
    timestamp: "2026-02-02T15:00:00Z"
    contributed_by: "software-implementer"
success_rate: 1.0
usage_count: 1
first_discovered: "2026-02-02T15:00:00Z"
last_used: "2026-02-02T15:00:00Z"
tags:
  - "typescript"
  - "null-safety"
  - "validation"
```

### Configuration

```yaml
# In aiwg.yml or .aiwg/config.yml
ralph:
  cross_loop_learning:
    extraction:
      enabled: true
      auto_extract_on_completion: true
      min_success_rate_threshold: 0.6
      min_usage_count_for_evaluation: 3
      extract_code_templates: true
      merge_similar_patterns: true
      similarity_threshold: 0.80
```

---

## Pattern Injection (On Loop Start)

### Trigger

- Agent loop start
- Manual injection request via `aiwg ralph-inject-patterns {loop_id}`

### Process

```yaml
injection_steps:
  1_analyze_task_description:
    - Extract task text
    - Identify task category (testing, debugging, refactoring, etc.)
    - Generate task embedding for semantic matching

  2_search_error_patterns:
    - Query error patterns by task category
    - Match error signatures to likely error types
    - Filter by min success rate (default 0.6)
    - Sort by effectiveness

  3_search_success_patterns:
    - Query success patterns by task category match
    - Use semantic similarity on task description
    - Filter by min success rate
    - Sort by average iterations (lower is better)

  4_search_anti_patterns:
    - Query anti-patterns by task category
    - Identify failure modes to avoid
    - Include better alternatives

  5_search_code_templates:
    - Query templates by language and task type
    - Filter by success rate
    - Sort by usage count

  6_filter_and_rank:
    - Combine all pattern types
    - Remove duplicates
    - Rank by relevance × effectiveness
    - Take top-k (default k=5)

  7_inject_into_context:
    - Format patterns for display
    - Add to loop context as "Cross-Loop Learning Context"
    - Track which patterns were injected (for effectiveness measurement)

  8_track_usage:
    - Log pattern injection event
    - Record loop_id and injected pattern_ids
    - Enable later effectiveness analysis
```

### Example Injection

**Input** (loop `ralph-fix-validation-b2c3d4e5` starting):
```yaml
task: "Fix TypeScript type errors in user validation"
category: "debugging"
```

**Patterns Retrieved**:
```yaml
retrieved_patterns:
  error_patterns:
    - pattern_id: "pat-error-null-check-015"
      relevance: 0.95
      success_rate: 1.0
      usage_count: 1

    - pattern_id: "pat-error-type-mismatch-003"
      relevance: 0.88
      success_rate: 0.94
      usage_count: 18

  success_patterns:
    - pattern_id: "pat-success-type-fixing-002"
      relevance: 0.92
      success_rate: 0.87
      average_iterations: 2.8

  anti_patterns:
    - pattern_id: "pat-anti-premature-optimization-001"
      relevance: 0.65
      failure_rate: 0.85
```

**Injected Context** (top 5 patterns):
```markdown
## Cross-Loop Learning Context

Patterns from previous loops that may help with this task:

### Error Patterns

1. **TypeError: null property access** (100% success, 1 use)
   - Error: "Cannot read property '.*' of null"
   - Fix: Add null check before property access
   - Template:
     ```typescript
     if ({{variable}} == null) {
       throw new ValidationError("{{message}}");
     }
     ```
   - Source: ralph-fix-auth-a1b2c3d4

2. **TypeError: type mismatch** (94% success, 18 uses)
   - Error: "Type 'X' is not assignable to type 'Y'"
   - Fix: Update interface definition or add type assertion
   - Source: 18 previous loops

### Success Patterns

1. **Type error fixing approach** (87% success, avg 2.8 iterations)
   - Steps:
     1. Run tsc --noEmit to see all errors
     2. Group errors by file/type
     3. Fix interface definitions first
     4. Fix usages second
     5. Verify with tsc again
   - Source: 12 successful loops

### Anti-Patterns to Avoid

1. **Premature optimization** (85% failure rate)
   - Don't optimize before tests pass
   - Complete primary task first
   - Better alternative: pat-success-refactor-module-002
```

### Configuration

```yaml
ralph:
  cross_loop_lear