All were captured from the isolated demo instance described in
Self-development. It runs with OPENSESSION_DEV=1,
OPENSESSION_DEMO=1 and an isolated OPENSESSION_STATE_DIR. Everything in
frame is synthetic: fictional teammates, a fictional acme-todo repo and a
fictional PR #128.
The transcript streams as the agent works: plans, tool calls, commands and recovery. The composer can queue your next prompt behind the running turn.
A question pauses the run and surfaces as a card. Pick an option or write your own answer, and the turn continues.
Repository-backed sessions show their branch and working-tree status. When a PR exists, the workspace also shows its merge state, review and changed files.
Open a file chip to see exactly what a turn changed, inline with the transcript that produced it.
The review surface combines a split diff with the PR status, checks, reviewer approval and merge action.
Pull requests across your repos, with search and workspace filters. The sidebar keeps active workspaces close by.
Choose a mode and model, then describe the job. Start from a current workspace or create a new one.
Scheduled and webhook-triggered agent runs, each with its own history.
The same UI, installed as a PWA. Read a running session and steer it from anywhere.
