Replay
Replay the step that failed.
An observability tool can show the failure. It never ran the graph, so it has no checkpoint to replay from. Linea does.
How it works
The fix loop starts at the checkpoint
When a node misses, keep the steps that already succeeded. Change the prompt or model on that node and replay it against the payload it originally received.
01.
Capture the execution
As the graph runs, Linea records the input and output for each node, including the memory and tool results in scope at that moment.
02.
Find the step to change
Inspect the execution, identify the node that needs work, and update its prompt, model, configuration, or code.
03.
Replay from the checkpoint
Linea loads the payload captured for that node and runs the updated step against the same context. Upstream nodes remain unchanged.
04.
Compare the new result
Review the new output against the original execution with the same thread, facts, and tool results in scope.
Example
Replay starts with captured input
This is the payload the response node received during the original execution. Linea restores this checkpoint when you replay the step.
{
"runId": "run_4f2a",
"step": "draft_reply",
"input": {
"thread": "cust_8891",
"facts": ["plan is annual", "last invoice failed"],
"draft": null
}
}Compare
Iterate without repeating completed work
A full rerun creates a new execution. Step-level replay uses the recorded context from the execution you are already investigating.
| Full workflow run | Step-level replay | |
|---|---|---|
| Work executed | Every node runs again | Only the selected node |
| Input to the broken node | Newly generated upstream context | Captured input from the original run |
| Upstream nodes | Repeated | Preserved |
| Best for | Testing the workflow end to end | Debugging or refining one step |
Designed for precision
A controlled loop for debugging and evaluation
Replay is grounded in a checkpoint from a completed or interrupted execution, so every iteration starts from known input.
01.
Target a single node
Choose the exact step you want to run again instead of queuing a new execution of the whole graph.
02.
Keep the same input
A new trigger creates new context. Replay restores the input from the execution you selected.
03.
Work from production context
The checkpoint comes from a real hosted execution, including the data available to that node at the time.