A run has a kind: agent, a model turn with a transcript session (sessionId), or workflow, a script executing in the engine with a journal and no session. Its origin records the mechanism that started it, never the person: user (a chat message), trigger (a firing), agent (a model child spawned with delegate), workflow (spawned by a script), or system. Every run is created with a title: the message excerpt, the child's brief title, or the script name.

Runs form trees. The run that a message or a firing started is a root run; every run it spawns, and every run those spawn, carries the same rootRunId, its own parentRunId, and a depth (0 for the root). The tree is what ListRuns {rootRunId} returns and what a runs/<rootRunId> WebSocket subscription replays, in creation order. A child spawned by a model turn also records parentToolCallId, the tool call in the parent's transcript that spawned it, which is how a delegate row finds its child while it is still working. A run started by ctx.continueAsNew is linked to its predecessor by continuedFromRunId and shares its root. Depth and fan-out are bounded by the delegation budget of the root (maxDepth, maxChildren), set by the session's thoroughness.

The pages of a tree read like one piece of work: the root run page shows its children under the plan steps they were attached to, each child row opens that child's own page, and cancelling a run cancels its whole subtree.

wait.reason is one of children (parked on spawned children, with pendingChildren), timer (a sleep or an event wait's timeout, with wakeAt), event (ctx.waitForEvent, which an activity event or a SignalRun answers; answerWith names the signal a person can send), or budget-pause (the run stopped rather than spend more and is waiting for a person). See park and wait and wake source.

A run that ends with work it promised and did not deliver, an undelivered typed result or plan steps neither finished nor written off, carries unmetObligations on the row rather than being quietly written off.

The same table is the queue. An executor claims the next queued run whose not_before has passed, fair-share across accounts (the account holding the fewest claimed or running runs goes first), interactive runs before background ones, oldest first. Claiming sets status = 'claimed', records the executor as lease_owner with a lease_expires_at sixty seconds out, and bumps attempt. When the server starts it requeues every run it left claimed or running, so a crash never wedges a run and the attempt already counted is kept; a run that fails with a retryable error is requeued with backoff until max_attempts (1 by default) is spent. A parked timer writes its wake time to not_before, and one sweep requeues every waiting run whose time has come. Operations has the queue settings and how to watch it.

A run may own a plan, the visible checklist of steps with their status; the server stamps ownerRunId and records settledAt when every step has stopped moving. usage is the run's token usage with its children's rolled up. sourceText on a workflow run is the exact module it executes, shown under Code on the run page. childRunCount is how many children it spawned.