Metadata

The Metadata tab gives you the complete execution profile for a session — from headline stats and timing down to the exact model, temperature, and worktree path. This is where you go to understand the shape of a session before diving into the transcript.

Session metadata — overview stats, activity timeline, identity, and runtime configuration

Session overview

The top section surfaces the most important numbers at a glance:

Metric	What it tells you
Duration	Wall-clock time from session start to end. Long durations may indicate the agent was stuck or the task was complex.
Events	Total event count across all types. A rough measure of session complexity.
Turns	Conversation turns completed vs total (e.g., `4/4 completed`). Incomplete turns suggest the session was interrupted or timed out.
Reasoning	Reasoning token count and time spent. Higher reasoning means the agent spent more time thinking before acting.
Commands	Shell commands executed with completion count (e.g., `98/98 completed`). Failed commands appear here — if the count shows `95/98`, three commands failed.
Tool Calls	MCP tool invocations and cumulative time. High tool call counts with zero time indicate fast MCP responses.
Tokens	Total tokens consumed, split into input and output. This is your cost signal — high input tokens often mean large context windows or repeated prompts.

Reading the overview signals

High tokens, low commands — the agent spent most of its budget reasoning and making tool calls rather than executing. Could indicate it was planning or stuck in a loop.
High commands, some failures — the agent was active but hit build or test failures. Check the transcript for error messages.
Duration far exceeding reasoning time — the gap is wall-clock time spent waiting for model responses, tool calls, or command execution.
Incomplete turns — the session ended before finishing all turns. Check if it was aborted, timed out, or hit a rate limit.

Activity timeline

The scatter chart plots every event by type over the session’s duration. Each dot represents one event, color-coded by category:

Color	Event type	What to look for
Message	User and agent messages	Conversation rhythm — steady exchanges vs long gaps
Command	Shell commands	Dense clusters = heavy execution phases
Tool Call	MCP tool invocations	Bursts of tool calls = planning or data gathering
Provider	Model lifecycle	Turn boundaries — when the agent started/stopped thinking
Reasoning	Internal reasoning	Visible only for models that expose reasoning tokens
System	System events	Session start, prompt delivery

Reading the timeline

Steady dots across the full width = healthy session with consistent activity.

Dense cluster followed by silence = the agent did a burst of work then stopped. Check if it completed or got stuck.

Long horizontal gap = the agent wasn’t doing anything. Could be waiting for a slow command, a rate-limited API, or it could be stuck.

All dots concentrated in the first minute = fast session. The agent finished quickly.

Toggle event types on/off using the legend buttons below the chart to isolate specific patterns. The event count legend below shows totals per type.

Identity and timing

Detailed identification for cross-referencing and debugging:

Field	Purpose
Session ID	Unique session identifier (ULID). Copyable.
Spec ID	The launch specification that created this session.
Thread ID	Provider conversation thread. Use this to correlate with provider logs.
Started / Ended	Exact timestamps for the session window.
Launch Reason	Why the session started (e.g., `queued-input`, `task-launch`, `orchestrator-dispatch`).
Max Runtime	Configured maximum duration before automatic timeout.

Runtime and model

The model configuration that was active during this session:

Field	What it shows
Provider	Which AI provider ran this session (e.g., `claude`, `codex`).
Provider Version	Provider integration version.
Model	The specific model used (e.g., `opus`, `gpt-5.4`).
Temperature	Sampling temperature. Lower = more deterministic.
Reasoning Effort	For models that support it, the reasoning budget.
Role	Session role (e.g., `executor`, `orchestrator`, `reviewer`).
Context	The context identifier (task key or orchestrator name).

Paths and artifacts

For task executor sessions, this section shows the worktree details:

Field	What it shows
Worktree	Full path to the agent’s git worktree. Open this in your editor or terminal to see the agent’s working directory.
Branch	The git branch the agent worked on.
Base Ref	The branch the worktree was created from.
Head at Start	The commit SHA the agent started from. Compare with the current HEAD to see what changed.
Transcript	Path to the raw NDJSON transcript file on disk.

Every field has a Copy button for quick use in terminal commands or git operations.

Advanced configuration

Expand the Show raw spec JSON disclosure to see the full session launch specification — the complete configuration that was used to start this session, including prompt assembly, tool policy, provider settings, and local filesystem paths.

Treat raw spec JSON as sensitive local diagnostics. Review and redact before sharing it outside the trusted local environment.

Next steps

Read the Transcript to see the full event-by-event detail.
Check Worktree Diff to see what code changed.
Use Terminal to investigate the worktree directly.

Transcript Worktree Diff