Workflows audit logs
Workflows audit logs record what people, API keys, and service accounts do with your workflows, and how each task execution ends. They answer questions like "who canceled this task?", "who changed this schedule?", and "how did this task end?".
Before you begin
- Your Organization is on an Enterprise plan. See Audit logs.
- You can open the audit logs in the Admin Panel›Administration›Audit Logs ↗, or you have an Admin API key to use the API.
What gets recorded
We record two kinds of entries. Action entries are written the moment a request succeeds. Outcome entries follow a minute or two after an task ends.
An action log entry's created_at value records the time the request was made.
Every task execution produces an outcome log entry when the task starts and when it completes. The created_at value for these log entries capture the start or the end of the task execution.
Identify the actor
actor_type | Meaning | actor_metadata |
|---|---|---|
HUMAN | A signed-in user. | user_uuid |
API_KEY | An API key. For a user's API key, the actor is the key, not the user who owns it. | api_key_id |
SERVICE_ACCOUNT | A service account. | service_account_id, and api_key_id when it used a key |
OTHER | Our platform recording server actions. | None |
Read a log entry
An log entry records who performed an action, what action was taken, and which resources were impacted. This example shows the start of an scheduled action. It has been shortened for readability.
{
"log_id": 2210,
"created_at": "2026-11-12T08:00:00.214Z",
"actor_type": "OTHER",
"actor_metadata": {},
"event_type": "workflow.execution.started",
"event_metadata": {
"trigger_type": "schedule",
"schedule_id": "nightly-report",
"deployment_name": "invoice-worker",
"deployment_id": "019e8c85-43cf-7508-bc95-7c1b46d4720c"
},
"target_type": "WORKFLOW",
"target_metadata": {
"workflow_id": "7d3f1c9a-52e8-4b06-9a41-0e6c8d2b5f13",
"workflow_name": "nightly_report",
"execution_id": "3c8e2b7a-1f40-4d95-b6a2-9d0e5c71a8f4",
"run_id": "a41b9d06-7e2c-4f83-8c15-2b6d0f9e3a77"
},
"organization_uuid": "c7a41d2e-6f08-4b95-91a3-5e0d8b2f7c14",
"workspace_uuid": "5f2a8c41-0b7d-4e63-a9d4-2c8e1f6b3a70"
}Every execution entry has workflow_name, execution_id, and deployment_name. deployment_id is added when we can identify it. Schedule entries other than workflow.schedule.created don't name a deployment.
Note that the Admin Panel can display a name in place of an ID.
Reading log entries for cancelled tasks
Canceling or terminating a task execution through the API records two entries with the same event type:
| When | Actor | What it tells you |
|---|---|---|
| When the request arrives | The person, API key, or service account that requested the cancellation. | Who asked. |
| A minute or two after the execution ends | OTHER, with no metadata. | How the execution ended, with reason and initiator_identity. |
The second entry describes the outcome. A cancellation is a request, so the execution can still end as completed or failed. Its initiator_identity names the client that sent the request. For an API request, the client is the Workflows service. To identify the person, use the first entry.
Understanding successful execution requests
The log entries listed below identify successful execution requests. The actor is the person that made the request.
| Event type | Recorded when | Extra event metadata |
|---|---|---|
workflow.execution.signal | Someone sends a signal to an execution. | signal_name |
workflow.execution.query | Someone queries an execution. | query_name |
workflow.execution.update | Someone sends an update to an execution, including when they submit input to a workflow that waits for it. | update_name, and task_id when submitting input |
workflow.execution.canceled | Someone asks an execution to cancel. | batch_id for a batch cancel |
workflow.execution.terminated | Someone terminates an execution. Archiving a workflow also terminates its running executions. | batch_id for a batch terminate. reason is workflow_archived for an archive. |
A batch cancel or terminate records one entry for each successful execution.
Understanding execution results
We record the log entries listed below for every task execution. Regardless of the trigger, whether a request, a schedule, a parent workflow, or a Temporal client, the log entry is captured.
| Event type | Recorded when | Extra event metadata |
|---|---|---|
workflow.execution.started | An execution starts. | trigger_type, deployment_name, deployment_id, input_size_bytes, is_obo, plus the metadata for its trigger, below. |
workflow.execution.completed | An execution completes. | None |
workflow.execution.failed | An execution fails or times out. | error_type. For a timeout, it's timeout. |
workflow.execution.canceled | An execution ends canceled. | reason, initiator_identity, initiator_execution_id, when known |
workflow.execution.terminated | An execution ends terminated. | reason, initiator_identity, when known |
trigger_type captures how an execution was triggered:
trigger_type | How it started | Actor | Extra event metadata |
|---|---|---|---|
api | A request to the Workflows API. | The person, API key, or service account that made it. | None |
webhook | A webhook. | The principal recorded for the webhook. | None |
schedule | A schedule fired. | OTHER, with no metadata. | schedule_id, scheduled_start_time |
child_workflow | Another workflow started it. | OTHER, with no metadata. | parent_execution_id, parent_run_id, root_execution_id |
temporal_client | A Temporal client started it directly, not through the Workflows API. | OTHER, with no metadata. | initiator_identity |
The closing entries always have the actor OTHER with no metadata. input_size_bytes is the size of the input after encoding. When we can't read the input back, the history metadata is unavailable instead.
Retries, resets, and continue-as-new runs don't record a new started entry. A run that continues as new doesn't record a closing entry.
Schedules
These entries have the target type SCHEDULE, with the schedule_id as target metadata. The actor is whoever made the change.
| Event type | Recorded when | Extra event metadata |
|---|---|---|
workflow.schedule.created | A schedule is created, from the API or when a worker registers. | workflow_id, workflow_name, source (api or worker), deployment_name, deployment_id |
workflow.schedule.updated | A schedule's timing, policy, or input changes. | workflow_id, workflow_name |
workflow.schedule.paused | A schedule is paused. | workflow_id, workflow_name |
workflow.schedule.resumed | A schedule is resumed. | workflow_id, workflow_name |
workflow.schedule.triggered | Someone fires a schedule by hand. | workflow_id, workflow_name |
workflow.schedule.deleted | A schedule is deleted. | workflow_id, workflow_name, and reason when there is one |
workflow.schedule.triggered records only a manual trigger. When a schedule fires on its own cadence, you get a workflow.execution.started entry with the trigger_type schedule.
A schedule can also be deleted by something other than a direct request. reason is workflow_archived when archiving its workflow removes it, and removed_from_worker when a worker registers without it.
Archive and unarchive
| Event type | Recorded when |
|---|---|
workflow.archived | Someone archives a workflow. |
workflow.unarchived | Someone unarchives a workflow. |
Archiving a workflow terminates its running executions and deletes its schedules, so one archive records several entries: the workflow, each terminated execution, and each deleted schedule.
Workers and deployments
These entries have the target type DEPLOYMENT.
| Event type | Recorded when | Metadata |
|---|---|---|
workflow.worker.registered | A worker registers with a deployment. | worker_name, and location_type, k8s_cluster, k8s_namespace when the worker reports them, as event metadata. The target metadata has deployment_name and deployment_id. |
workflow.deployment.authorized_credential.granted | A credential is authorized to run workers for a hardened deployment. | The target metadata has deployment_id, deployment_name, credential_id, and principal_type. |
workflow.deployment.authorized_credential.revoked | A credential is no longer authorized. | The same as granted. |
Rotating a credential records the revoke first, then the grant. Starting, stopping, or restarting a hardened managed deployment can record these entries too, because we rotate the worker's credential. The actor is whoever made the change.
Find what you need
In the Admin Panel, use the Event, Actor, and Target filters. See Search and filter.
Every execution records at least a start and an end, so busy workflows produce many entries. To see only what people did, and hide the entries we record on our own, filter on the actor types HUMAN, API_KEY, and SERVICE_ACCOUNT. With the Admin API, pass one actor_type parameter per type:
curl 'https://api.mistral.ai/v1/admin/audit-logs?actor_type=HUMAN&actor_type=API_KEY&actor_type=SERVICE_ACCOUNT&target_type=WORKFLOW' \
-H "Authorization: Bearer $MISTRAL_ADMIN_API_KEY"To see how executions ended, pass event_type=workflow.execution.completed, event_type=workflow.execution.failed, event_type=workflow.execution.canceled, and event_type=workflow.execution.terminated instead. The Admin API reference lists the other filters, such as before and after.
Troubleshooting
| Problem | Cause | Solution |
|---|---|---|
| No Workflows events appear. | Your Organization isn't on an Enterprise plan, or Workflows audit logs aren't enabled for it yet. | Check your plan, then contact your Mistral account team. |
| Entries for actions appear, but not for starts and ends of executions. | Those entries are written one to two minutes after the fact. | Wait a few minutes and refresh. If they're still missing, contact support with an execution_id. |
| A cancel recorded two entries. | One records the request, the other the outcome. | See One cancel can record two entries. |
An entry has the actor OTHER with no user. | We record execution ends, scheduled starts, and child workflow starts on our own. | Use trigger_type and schedule_id to find the cause. For a cancel, look at the earlier entry for the same execution. |
| The actor is an API key, not the user who owns it. | A request made with an API key is attributed to the key. | Look up the key's owner with the api_key_id. |
| Resetting an execution recorded no reset entry. | A reset has no entry of its own. | Look for a workflow.execution.terminated entry for the previous run, which has a reason. |
deployment_id is missing. | We couldn't find the deployment by its name at the time. | Use deployment_name. |
| Executions that ran before audit logs were enabled have no entries. | Entries aren't recorded retroactively. | None. |
What isn't recorded
- Reads, such as listing or opening workflows, executions, or schedules.
- What happens inside an execution: activities, retries, and the steps of your workflow code.
- The helper executions that the SDK creates to run activities in parallel. Their parent execution is recorded.
- Messages in a Vibe session. Its start and its end are recorded.
- Starting, stopping, and deleting a managed deployment. See Managed deployments audit logs.
Next steps
- Audit logs reference: every field, filter, and event type.
- Managed deployments audit logs: the entries for the deployment that runs your workers.
- Hardened deployments: restrict which credentials can register workers.