Replay a run

Replay takes a finished run and executes it again. You can repeat it exactly, swap in a different prompt, or run it against a different model. If the workflow changed since the original run, you map any renamed nodes to their replacements before the replay starts.

When to use this

Replay a run when you want to reproduce or vary a past execution without waiting for the trigger to fire again:

• Re-run a failed run after you fixed the underlying cause, using the same trigger payload.
• Test a prompt change against a real, historical input instead of inventing test data.
• Compare how a different model handles the same run.

Replay only applies to runs that have finished. A run must be in completed, failed, or cancelled status before you can replay it. You cannot replay a run that is still pending or running.

Where replay lives

The replay controls are on the run trace viewer at /[workspaceSlug]/ai/runs/[runId], not on the monitoring run detail page under /[workspaceSlug]/monitoring/workflow-runs. The monitoring detail page handles pause, resume, and cancel, but it has no Replay button.

Replay a run as-is

Replaying as-is re-executes the run against its original trigger payload and the workflow version it ran on. Nothing changes except that a new run is created.

1. Open the run

   Go to /[workspaceSlug]/ai/runs/[runId] for the run you want to repeat. The run must be completed, failed, or cancelled.

2. Open the replay dialog

   Select Replay, or press r. The dialog opens with the title "Replay run" and the description "Re-execute this run. Pick a replay mode below."

3. Keep the default mode

   Leave Replay as-is selected. It is the default mode and needs no extra input.

4. Check the mutation banner, then submit

   Read the banner at the top of the dialog (see below), then select Submit replay. On success you get a toast reading "Replay submitted" with the new run id.

Mutating tool calls default to a dry run

Every replay dialog shows a banner about mutating tool calls. By default, replays dry-run any mutating calls instead of executing them, so a replay does not duplicate side effects like sending a message or creating a record. The banner reads:

Mutating tool calls will dry-run. Set policies.replay.allowMutations: true per-node to opt in.

If a workspace policy has opted in, the banner turns red instead:

policies.replay.allowMutations: true: this replay WILL invoke mutating tool calls.

When you see the red banner, the replay performs real mutations. Confirm that is what you want before you submit.

Replay with a new prompt

This mode keeps the original trigger payload but runs against a different prompt revision. It expects two internal identifiers, not a friendly picker, so you need the prompt artifact id and the exact content hash you want to replay against.

1. Select the prompt mode

   In the replay dialog, select Replay with new prompt. Two fields appear.

2. Enter the prompt artifact id

   In Prompt artifact id, paste the artifact identifier. It looks like prompt_a1b2c3d4.

3. Enter the content hash

   In Content hash, paste the exact prompt revision hash, formatted like sha256:9f86d081884c7d65....

4. Submit the replay

   Select Submit replay. The button stays disabled until both fields are filled.

Replay with a new model

This mode re-runs against a different model. Like the prompt mode, it takes two internal identifiers: the model id and the provider-qualified model ref.

1. Select the model mode

   In the replay dialog, select Replay with new model.
2. Enter the model id

   In Model id, enter the internal model id, for example claude-opus-4-7.

3. Enter the model ref

   In Model ref, enter the provider-qualified reference, for example anthropic:claude-opus-4-7.

4. Submit the replay

   Select Submit replay. Both fields are required.

The model you pick is re-checked against your workspace and account model allowlist on the server. If the model or its provider is not allowed, the replay is rejected (see Troubleshooting).

Resolve renamed nodes for a cross-version replay

When you replay against a workflow version that renamed or restructured nodes since the original run, the platform cannot match the old node ids to the new ones on its own. The replay dialog shows a red alert like:

2 renamed node(s) need mapping before this cross-version replay can run.

Submit stays disabled until every renamed node is mapped.

1. Open the mapping dialog

   Select Resolve mapping in the alert. The dialog opens with the title "Map renamed nodes."

2. Pick a target for each renamed node

   Each renamed node is listed by its original id. For each one, open the select (it starts on "Pick target") and choose the node in the new version that should receive that input.

3. Choose who can reuse the mapping

   Under Mapping visibility, choose Operator (private) to keep the mapping to yourself, or Workspace (shared) so others replaying the same version pair reuse it. The default is Operator (private).

4. Save the mapping

   Select Save mapping. The button is disabled until every renamed node has a target. You return to the replay dialog, where the alert clears and Submit replay becomes available.

The platform first checks for a saved mapping for this version pair (you may briefly see "Checking saved node mappings..."). If a saved mapping already covers every renamed node, it is applied for you and you go straight to submitting. A saved mapping holds up to 1000 node pairs.

Verify the replay started

A successful submit shows a toast titled "Replay submitted" with a description reading "New run: {newRunId}". That new run id is the replay. Open it at /[workspaceSlug]/ai/runs/{newRunId} to watch it execute and read its steps. The replay is a separate run with its own timeline, so the original run stays unchanged for comparison.

Troubleshooting

If a submit fails, the dialog shows "Replay submission failed" with one of the messages below, and a matching error toast.

A few server-side limits apply even when the dialog looks ready:

• A workflow that is paused for safety reasons rejects replays until it is unpaused.
• You can only replay runs that belong to your own workspace and account. Cross-boundary replays are blocked with a 403.

Related

• Read a run in detail
• Pause, resume, or cancel a run
• The run model