Runs in the app
In the sidebar, under AI operations, choose Relay. The runs list is at /ai-operations/runs; one run opens at /ai-operations/runs/<run id>. Connect, in the same sidebar section, holds your Model Context Protocol (MCP) key and setup snippets.
You need relay.read and read access to the sessions dataset. You see your own user's agents unless your role has relay.read_all. The area appears for workspaces with logs. Local mode uses the tools and local database; these instructions describe the hosted app.
Runs list
The list shows runs in the selected time range, newest started first. The Status filter selects live, done or all runs. A run is live while an agent has written in the last 60 seconds.
| Column | Meaning |
|---|---|
| Run | Shared run id, or a session id when that session has its own run. |
| Agents | Agents with a record in the run. |
| Started | Time of the first write. |
| Duration | Time from the first write through the last write. |
| Posts | Notes posted by agents. |
| Claims held / refused | Successful and refused claims. |
| Done | Finished task names; file paths are never done. |
| Open claims | Holds that have not been released, finished or expired. |
| Duplicate edits | Keys edited by more than one agent. |
| Silent agents | Agents with no recent write in a live run. |
| Cost | Recorded session cost when supplied. The column is hidden when no listed run carries a cost. |
Sort by cost or duplicate edits to inspect those runs. Load more fetches the next page. A range wider than 366 days is clamped and labelled. Choose a row to open the run. The empty state links to Connect.
Run detail and tiles
A run shows its id, live state, start time, duration, timeline and detail panels. Tiles include agents, tasks done, duplicate edits, refused claims, time lost waiting, finished with evidence, open questions and cost. Cost says Not reported when records carry none, as with Codex.
The evidence tile distinguishes finishes with recorded evidence and, when returned, checks matching the task's own test. The questions tile shows open questions and the age of the oldest. Detector badges link to the relevant panel or rows.
Agent map opens the AI overview at /ai/overview, where the Summary view maps callers, agents, tools and models from trace data.
Timeline
Each agent has a lane on a shared time axis. Claim bars run from hold or reopen to finish, release or expiry. Live holds remain open. Refused claims are ticks; posts are markers with text on hover. Tool calls appear as a thin activity strip beneath the lane.
Choose an agent's name or a lane event to open that agent's session transcript at the selected point.
Resources and coordination panels
Resources lists each claimed or edited key, its holder, who finished it, refused claims and editors. Keys edited by several agents carry a duplicate badge.
Tasks shows registered work, files, state, holder or finisher, round, test command and completion evidence. It also shows stalled task information. Questions shows the asker, recipient, question and answer state. Contracts shows interface descriptions, their defining files and changes since the contract. These panels read the shared record; agents use Relay tools to update it.
When available, forecast lines under each task show contention in earlier runs, files changed together and the median time from hold to finish. The fifth panel, Patterns, groups recorded activity by template and tool, showing counts, shares of turns and anomaly flags such as retry loops. It loads when first opened and shows an unavailable state when the workspace cannot supply patterns.
Detectors
| Detector | Condition |
|---|---|
| Duplicate edit | At least two agents edited one normalized path. Supported edit-tool records supply these paths. |
| Silent agent | A live run's agent has written nothing for silentAfterSeconds, default 120. |
| Stalled task | Held or lease-expired unfinished work has no holder progress on its task or files for stallAfterSeconds, default 300. |
| Unanswered question | An open question is older than questionAfterSeconds, default 300. |
Detectors are computed when the run is read. They do not send alerts. Shell writes have no keyed edit paths, so they are absent from duplicate-edit detection.
Live updates
A live run refreshes every two seconds, less often while the browser tab is hidden. Claims and posts appear after their writes. Tool calls appear after the hook uploads the turn, so that strip can lag by one turn. Detail keeps the newest 5,000 events and shows a truncation note for older events.
Session drawer
The drawer lists transcript rows in sequence order: text, collapsed thinking, tool calls with commands or paths, and tool results with error state. Search this session needs at least 3 characters and searches those rows. Open in logs opens the sessions dataset in the Log Explorer, filtered to that session and its time window when available.
For sessions outside a shared run, use Sessions in the app.
Connection and settings
Connect lets you generate, rotate and copy your MCP key. It shows the key's expiry, your CtrlB URL and sessions dataset. Its tabs are Claude Code, Codex, OTEL only for OpenTelemetry (OTEL), and HTTP. The page displays a placeholder; the copy button inserts your key into the copied snippet.
If the workspace's MCP server is off, ask your CtrlB administrator to turn it on before uploading sessions or calling tools.
The run reads workspace settings for its dataset, enforcement mode and detector thresholds. A signed-in administrator with relay.manage changes settings through the settings endpoint; an MCP key can only read them. Claims and edit checks lists the settings and defaults.