Records Reference
The events recorded under a project's “Records (audit trail)”, its CSV, and what the “Export” of a run record contains. Item names follow what the screen shows. How to trace the records is covered in the Record Guide.
The Records (audit trail) list
Under the Execution history tab ＞ “Records (audit trail)”, events are listed newest first.
| Column | Content |
|---|---|
| Time | When the event happened |
| Event | The kind of event (see the table below). A tool operation is shown by its operation kind, the same kinds as in the execution rules (Text write, External integration and so on) |
| Who | A user's name (their email address or ID when there is no name). Records that do not identify the user show “User”. An external sender who started it by email shows “External sender ([EXT])”. Otherwise “Agent” or “System”. An automatic action at a deadline is “System” |
| Channel | Set on review decisions and on the records of review-request, review-overdue and similar emails. app (decided on screen), mail (decided through the personal link in an email, or a record of sending the email), system (an automatic action at a deadline) |
| Content | The main points for each kind of event: the review request's title, the decision, the reason, the step name and so on |
- You can filter by “Event”. Only the kinds present in this project's records can be chosen.
- “Search records” finds entries by a word. It searches the stored record data (the event name, the reason, the attached reference information and so on); the names of users, step names, review request titles and event labels shown on screen are not searched.
- “Export CSV” exports them as CSV. Only the event filter carries into the CSV; the search term does not.
The events that are recorded
| Event | event_type in the CSV | When it is recorded |
|---|---|---|
| Confirmation requested | gate.requested | A review request was created (on reaching a review step, when an agent's result was “Not allowed by rules” or “Failed (reported by agent)” and the step is set to ask for review, or when the agents on a step set to “Run at the same time” point to different destinations) |
| Review result recorded | gate.resolved | A review request was decided (Content shows “Approved” for a person's decision to proceed, “Automatic deadline action” for one made at the deadline, or “Sent back” / “Rejected”). Automatic actions at a deadline are included |
| Confirmation overdue | gate.overdue | A review request passed its deadline |
| Notification queued for sending | delivery.accepted | A notification, such as a review-request or review-overdue email, was handed over to be sent |
| Delivery Acked | delivery.acked | Sending the notification was confirmed |
| Delivery Failed | delivery.failed | The notification could not be sent |
| Delivery Unconfirmed | delivery.unconfirmed | The result of sending the notification could not be confirmed |
| Execution rules changed | policy.changed | The execution rules (basic rules and forbidden operations), the reviewers, the review method or the email start settings were changed |
| Member / role change | policy.changed | A member was added or removed, or a role was changed |
| Run started | run.started | A workflow run started |
| Lanes joined | run.lane_joined | In a workflow run, every lane finished and they joined |
| Stop requested | stop_signal.received | The running agent received a request to stop |
| Body Viewed | forensic.viewed | The body of a run record was opened. The reason is kept |
| Execution records exported | forensic.exported | A run record was exported. The reason is kept |
| Improvement ideas requested | consult.failure_requested | Improvement ideas were consulted on from a run that did not go well |
| Improvement applied | improvement.applied | An improvement idea was applied (the changes saved) |
| Stale Completion | completion.stale | A run's result arrived after its step or its workflow run had already ended (for example, after it was cancelled) |
| (the operation kind) | tool.access | An agent's operation was judged by the execution rules. When it was stopped, decision is deny |
| (the operation kind) | tool.invoked | An agent was about to call a tool |
| (the operation kind) | tool.effect | The result of a tool call |
| (the operation kind) | a2a.invoked | An agent was about to send a request to another agent (a record that it tried; it does not show that the request succeeded) |
- “Execution rules changed” covers only the four settings above. Changes to steps, agents, time triggers, connectors and other settings are not recorded here. However, when an improvement idea is applied from a run record, “Improvement applied” is recorded for each saved change.
- The email addresses that notifications were sent to are not recorded.
The Records (audit trail) CSV
“Export CSV” exports the records of the event you have chosen, newest first, up to 5000 (the count is in the Settings reference). When it is cut off at the limit, the screen says “Exported the newest ◯ entries (◯ in total).”
| Column | Content |
|---|---|
occurred_at | The time (UTC) |
event_type | The kind of event (the event_type in the table above) |
actor | A user's name (their email address or ID when there is no name). Otherwise, the name of the part of the platform that recorded it, or the kind of actor |
decision | The value of the decision or judgement (for a review decision: proceed = Approved, send_back = Sent back, reject = Rejected) |
reason | The reason. For a body viewed and an export of run records, the reason that was entered |
step | The step name |
workitem | The review request's title |
run_id | The ID of the workflow run |
root_execution_id | The ID of the run |
metadata | The reference information attached to the record (JSON) |
- The CSV can include kinds of rows that the on-screen list does not show. The screen lists only the kinds that have a label. The CSV also carries events inside a run, such as
guardrail.blocked(stopped by the guardrail) andkill_switch.triggered(cut off at the time limit). - Tool operations appear under their
event_typenames. They are not turned into operation kinds as on screen.
Exporting a run record
“Export” on a run record writes the execution of the open run record, and the records related to it, to a JSON file (for a workflow, the flow and step rows of the same run are included). A reason is required, and it is kept in Records (audit trail) as “Execution records exported”.
| Included | Not included |
|---|---|
| The run's state, result, reason, times and retry count | The body of the exchange with the agent |
| The execution rules in effect at the time (basic rules and version), and the name of the model used | The comment in the closing report |
| The counts of tool calls and external-service calls, and whether the closing report said it finished, with its category | The values passed to tools, and the body written |
| The Records (audit trail) rows (kind, who, decision, reason code, reference IDs) | Review request titles, and the text of decision reasons |
| A verification hash for each record, and a hash of all the records exported | Information about files created or received (names, locations, sizes and so on) |
| The workflow run's flow (the order of steps, Run at the same time, lanes, splits and joins, references to destinations), the rows for each step (placement, lane, round, state, reason, times), and the state of each lane | Step names, descriptions, instructions, lane names, display names of destinations |
| The review requests created from that run (state, decision, who decided, references to the chosen destination and the step sent back to) | (as “Review request titles, and the text of decision reasons” above) |
This is not the same as “Export CSV” on Records (audit trail). The CSV needs no reason, and exports the project's records as a list.