Episode Complete

Complete and score a learning episode to extract patterns and update heuristics.

Purpose

Finalize an episode with outcome scoring, reflection generation, and pattern extraction for future retrieval.

Steps

1. Gather outcome data:

   • Final verdict (success, partial_success, failure)
   • Total time spent
   • Total tokens used (if applicable)
   • Key artifacts produced
   • Errors encountered

2. Create TaskOutcome:

   let outcome = TaskOutcome {
       verdict: Verdict::Success,
       time_ms: total_time,
       tokens: total_tokens,
       artifacts: vec![/* paths to created/modified files */],
       errors: vec![/* any errors encountered */],
   };
   

3. Call complete_episode:

   memory.complete_episode(episode_id, outcome).await?;
   

4. System processes:

   • Computes RewardScore based on:
     • Success/failure
     • Time efficiency
     • Code quality
   • Generates Reflection:
     • What worked well
     • What could be improved
     • Key learnings
   • Extracts Patterns:
     • Tool sequences
     • Decision points
     • Common pitfalls

5. Update storage:

   • Store in Turso (permanent record)
   • Update redb cache
   • Index by task_type and timestamp
   • Update related patterns and heuristics

6. Validation:

   • Verify episode was scored
   • Check patterns were extracted
   • Ensure heuristics were updated

Pattern Types Extracted

• ToolSequence: Common tool usage patterns
• DecisionPoint: Key decision moments and outcomes
• ErrorPattern: Common errors and resolutions
• PerformancePattern: Optimization opportunities

Scoring Rubric

• Success: Task completed, tests pass, meets requirements
• Partial Success: Task mostly complete, minor issues
• Failure: Task incomplete, major issues, tests failing

Example

let outcome = TaskOutcome {
    verdict: Verdict::Success,
    time_ms: 45000,
    tokens: 12000,
    artifacts: vec![
        "src/storage/batch.rs".to_string(),
        "tests/integration/batch_test.rs".to_string(),
    ],
    errors: vec![],
};

memory.complete_episode(episode_id, outcome).await?;

Post-Completion

• Patterns are now available for future retrieval
• Heuristics updated for similar tasks
• Episode stored for long-term learning
• Embeddings computed (if service configured)