Executions

Executions is the journal of everything that ran in the workspace: chat turns, webhooks, Telegram, schedule fires and runs started from the editor.

The list

Five tiles sit on top: Runs · total, Completed, Partial, Failed and Avg duration. They are aggregated on the server over the whole filtered set rather than the page below them, so paging does not change their numbers.

The search box matches three things: the workflow name, the run id and the error text. It deliberately does not look inside inputs and outputs — those payloads are redacted on the way out, and searching them would let you confirm a value the API refuses to show.

The workflow filter lives in the sidebar: Filter runs lists the workflows, picking one puts ?workflow=<id> in the address, and All workflows clears it. The page subtitle then changes to “Runs for …”. Filter and search apply to the table and the tiles at once.

The table is served in pages of 50 rows; the counter below shows the range and the total. The Trigger column names the source of the run: chat, public chat, Telegram, webhook, cron, API. Failed runs print the error line under the workflow name.

Statuses

• completed — no node ended in an unhandled error;
• partial — there was an error, but the run reached an exit node;
• failed — there was an error and the run never reached an exit;
• queued and running — not finished yet;
• stopped and cancelled — interrupted by hand.

Skipped nodes do not affect the run status: a branch of an if that was not taken is ordinary control flow, not a failure. The exact status contract and the skip reasons are covered in run statuses.

A single run

Clicking a row opens the run. The header carries the status, start time, duration, total tokens split into input and output, credits, and an Open in editor link. Below it come the error text and the run output, when there are any. To the right of the tabs sit chips counting nodes per status.

Node Executions. A flat list of the nodes that ran, in start order. A row expands into its input, output, tool calls with arguments and results, error, and start/finish times. Nodes that execute other nodes inside themselves — ai_agent, for one — become a group, with nested calls and iterations shown inside the parent.

Workflow. The same run drawn as a graph: loop iterations unfolded, colour by status. Clicking a node opens an inspector with Inputs, Outputs and Info tabs (plus State for nodes that keep any). For a skipped node it also prints Why skipped — for example “An upstream branch went the other way” or “Every path here died on an upstream failure”.

Timeline. A Gantt view over nodes: the time axis starts at the first node start, one bar per node, nested calls indented. Clicking a bar expands that node’s card under the chart.

Logs. Server-filtered entries: level chips (debug, info, warning, error), a node dropdown, a message search and a time-window slider.

Where tokens and credits are shown

Tokens appear in two places: in the run header as a total (split ↑input / ↓output) and on every node — a “tok” pill on the collapsed row and a Tokens line in the expanded one. Only nodes that called a model have them.

Credits come from the usage ledger: some rows are attached to a node, others to the run as a whole. The number in the header is therefore per-node usage plus run-level charges, and it is larger than the sum of the per-node values. The list table shows the same figure. How the price is computed is described in plans and limits.

“Removed by retention”

Run details are kept for a limited time. When the window passes, a background pruner deletes the per-step logs, the node inputs and outputs and the graph snapshot, but keeps the run row: status, timestamps, duration and token totals stay visible in the list.

On such a run the tabs are replaced by the notice Detailed data removed by retention policy, with the date it happened. It cannot be undone — this is deletion, not archiving.

The window is set by the plan of the workflow’s owner and differs per owner; the current values are in plan limits.

Where to go next