> ## Documentation Index
> Fetch the complete documentation index at: https://docs.statebase.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Debugging with Replay

> Reproducing production failures in 1:1 fidelity.

"It worked on my machine" is the death of AI agent development. LLMs behave differently at 3 AM in production than they do on your local terminal.

StateBase provides **Decision Replay** to solve this.

## The Workflow

### 1. Identify the Failure

You see a failed session in the **StateBase Dashboard**. The agent gave a weird answer at Turn #14.

### 2. Export the Session

Fetch the session data using the SDK or CLI.

```bash theme={null}
sb export sess_abc123 --turn 14 > failure_turn.json
```

### 3. Replay Locally

Using the exported state, you can re-run Turn #14 locally against your latest code or a different model (e.g., GPT-4o instead of Gemini) to see if it fixes the issue.

```python theme={null}
# Initialize session with the EXACT state from the failure
session = sb.sessions.create(
    initial_state=imported_state_from_turn_13
)

# Run the turn
result = my_agent.process(user_input_from_turn_14)
```

## Why this matters

* **Exact Reproducibility**: You are not guessing what the state was. You have the exact snapshot of the agent's "brain" at the millisecond of failure.
* **Regression Testing**: After you fix the prompt, you can replay the last 100 failed sessions against the new code to ensure they now pass.
* **Confidence**: Ship updates knowing you've actually solved the edge cases.