Runs & Observability

A run is one agent session in a workspace: a goal, the turns that followed, every tool the agent called, and the changes it left behind. Runs are recorded as they happen, so nothing about a session is reconstructed after the fact. Each run opens as a tab in the workspace. Multiple runs can be open at once — see multiple sessions.

​

What A Run Records

The system prompt, config, and tool policy in force are snapshotted too, so an old run stays reproducible.

​

Reading The Transcript

​

Turn rail

Once a run has more than a few exchanges, a rail of ticks appears down the left edge — one per user message. It’s a map of the whole conversation, not a scrollbar: hovering a tick previews what was asked and how the answer opened, and clicking scrolls the transcript there.

​

Context usage ring

A circular indicator shows how much of the model’s context window the session has consumed. Hovering breaks it down. It turns amber as the auto-compact threshold approaches and red once the window is nearly full.

​

Usage tooltip

Each response is stamped with how long it took. Hovering that duration shows the turn’s input, output, cache-read, and cache-write tokens, plus its cost.

​

Run detail level

Settings → General → Run Detail controls how much of this the transcript shows — collapse tool calls and their outputs down to a summary, or expand everything.

​

Resuming And Forking

Fork is the ⑂ button next to a response. It starts a fresh run seeded from that point, leaving the original untouched — useful for trying a second approach without losing the first.

​

Beyond A Single Run

Runs can be archived when you’re done with them — they leave the tab strip but stay queryable.