Episode Log Steps

Episode Log Steps

Log execution steps during episode execution for detailed tracking and learning.

Purpose

Record individual execution steps to build a complete picture of task execution for pattern learning.

When to Log Steps

1. Tool usage: Each time a significant tool is used
2. Decision points: When making architectural or implementation choices
3. Error handling: When encountering and resolving errors
4. Milestones: Key progress points (test passing, module complete)

Step Structure

pub struct ExecutionStep {
    pub tool: String,           // Tool/​action used
    pub action: String,         // Description of action
    pub latency_ms: u64,        // Time taken
    pub tokens: Option<u32>,    // Tokens used (if applicable)
    pub success: bool,          // Whether step succeeded
    pub observation: String,    // Outcome/​observations
}

Logging Guidelines

1. Batch When Appropriate

• Don't log every tiny operation
• Batch related steps when many occur quickly
• Log significant operations individually

2. Include Context

• Tool: cargo, rustfmt, clippy, git, etc.
• Action: Specific command or operation
• Observation: Result, output summary, or error

3. Track Performance

• Record latency for slow operations
• Note token usage for LLM calls
• Mark success/​failure clearly

Examples

Build Step

let step = ExecutionStep {
    tool: "cargo".to_string(),
    action: "build --all".to_string(),
    latency_ms: 12500,
    tokens: None,
    success: true,
    observation: "Build successful, 0 warnings".to_string(),
};
memory.log_step(episode_id, step).await?;

Test Step

let step = ExecutionStep {
    tool: "cargo".to_string(),
    action: "test --test integration_test".to_string(),
    latency_ms: 3200,
    tokens: None,
    success: false,
    observation: "2 tests failed: test_batch_insert, test_concurrent_writes".to_string(),
};
memory.log_step(episode_id, step).await?;

Code Generation Step

let step = ExecutionStep {
    tool: "claude".to_string(),
    action: "generate async batch implementation".to_string(),
    latency_ms: 4500,
    tokens: Some(2800),
    success: true,
    observation: "Generated batch.rs with async Tokio patterns".to_string(),
};
memory.log_step(episode_id, step).await?;

Error Resolution Step

let step = ExecutionStep {
    tool: "clippy".to_string(),
    action: "check --all".to_string(),
    latency_ms: 1500,
    tokens: None,
    success: true,
    observation: "Fixed 3 warnings: unused imports, unnecessary clone".to_string(),
};
memory.log_step(episode_id, step).await?;

Batching Strategy

When multiple related steps occur in a burst (e.g., fixing multiple test failures):

let step = ExecutionStep {
    tool: "cargo".to_string(),
    action: "fix 5 test failures in batch_test.rs".to_string(),
    latency_ms: 18000,
    tokens: Some(3500),
    success: true,
    observation: "Fixed: test_batch_insert (await missing), test_concurrent (lock order), test_timeout (semaphore added)".to_string(),
};
memory.log_step(episode_id, step).await?;

Notes

• Steps are stored as JSON array in episode record
• Use clear, actionable observations
• Include error messages when relevant
• Keep observation strings concise but informative