Agent Sessions

Agent Sessions is where you see exactly what your agents did, are doing, or failed to do. Every run produces a full execution record: the transcript of every message, tool call, and command; metadata with timing, tokens, and model details; a diff of every file changed; and a live terminal that drops you directly into the agent’s worktree.

This is the screen that turns agent runs from black boxes into inspectable, debuggable work. When an agent is struggling, you don’t guess — you open the session, read the transcript, check the diff, and if needed, jump into the terminal to fix it yourself.

Session list

Session sidebar — browse recent runs with status, role, provider, and timing at a glance

The left sidebar lists all sessions ordered by most recent. Each entry shows:

Session name — the context identifier (e.g., TASK-valdr-lab-6 for task sessions, ORCHESTRATOR-skadi-... for orchestrator sessions).
Status badge — CLOSED, IDLE, RUNNING, or FAILED.
Role and provider — e.g., executor · codex · closed or orchestrator · claude · closed.
Start time — when the session was launched.

Click any session to load its detail view. The sidebar stays visible so you can quickly switch between sessions without navigating away.

Use the Collapse button at the top to hide the sidebar and maximize the workspace area. Use the Maximize button in the tab bar to enter fullscreen mode for maximum detail.

Session detail tabs

Every session has a tab bar across the top with multiple views:

Tab	What it shows
Transcript	Chronological log of every event — messages, tool calls, commands, and provider lifecycle events.
Metadata	Session overview stats, activity timeline chart, identity and timing, model config, and worktree paths.
Score Audit	Quality audit results for the session’s work.
Worktree Diff	Git diff of all files changed during the session.
Task Reviews	Review results linked to this session’s task.
Summarize Session	Generate a summary of the session’s activity.
Terminal	Live shell in the agent’s worktree for real-time debugging. Available for executor sessions only.

Session actions

Reply to agent — send a message to a running or idle session directly from the transcript view.
Delete Session — remove the session record.
Maximize/Restore — toggle fullscreen mode for maximum workspace area.

Quick triage workflow

Check the overview

Open the Metadata tab. Check duration, token count, and command completion rate. A session with high tokens but few commands may be spinning. Many failed commands suggest a build or environment issue.

Scan the timeline

Look at the Activity Timeline scatter chart. Healthy sessions show steady activity across the full duration. Long gaps suggest the agent was waiting or stuck. Dense clusters of tool calls indicate intensive MCP operations.

Read the transcript

Switch to Transcript. Use filters to focus on just agent messages or just commands. Search for error messages or specific file names.

Verify the output

Open Worktree Diff to see what actually changed. If something looks wrong, open the Terminal to investigate directly in the agent’s working directory.