The session detail page is the full record of a single agent run. Click any row in the Sessions view to open it.
What you see
Two views cover the run:
- Chat — the conversation spine you saw: your prompts and the agent’s responses.
- All activity — everything the harness did on your behalf, with a kind filter to focus in: permission checks, title generation, subagent calls, context compaction, and more.
Messages render as markdown with syntax-highlighted code, and long messages collapse.
Alongside the conversation, the detail page surfaces:
- Tool calls — every tool the agent invoked, shown next to its result, with subagents nested inline.
- Prompt and reference thread — the original prompt plus any reference thread the session forked from.
- Session Info — model, tokens in and out, cost, turn count, and a main vs shadow split showing how much spend went to the conversation versus the machinery around it.
- Model chip — the header carries the same model chip as the sessions table: the cost-dominant model with its logo, a
+nbadge when more than one model served the run, and a hover breakdown of what each model cost. - Turn histogram and task checklist — a per-turn token histogram gives you the shape of the run, and the session’s task checklist tracks what got done.
- Status badge — tells a completed session from a failed or in-progress one at a glance.
- Author — the team member who ran the session.
- Session name — click the title in the header to rename the session. The name you set is durable — it survives a live session and shows everywhere the session is listed — and clearing it returns to the automatically derived title.
Explore the Session Info railLive demo — telemetry, model split, turn histogram, and task checklist
Use the detail page to understand how a result was produced, not just what was produced.
See what a session produced
Under Session Info, the Outcomes section lists what the session created, linked to the real thing: pull requests, repositories, GitHub issues, and Linear issues. If the session produced nothing, it says no outcome plainly instead of leaving you to scroll the transcript and guess.
An outcome is something the session created, not something it looked at — a session that opened a pull request gets credit for it; a session that merely viewed one doesn’t. Live sessions count too: a pull request opened mid-session shows up while the session is still running, and the section waits until detection has something to say, so a running session is never labeled “no outcome” prematurely.
Export a session
Session export is available on the Scale plan. On it, every session detail page has an Export button — click it and the session downloads as a dated JSONL file (session-<id>-<date>.jsonl): one JSON record per line, nested with trace and span granularity, the same shape tapes stores. On other plans the button’s tooltip says where export lives.
To export more than one session at once, use the Export last 30 days control on the Sessions list — it bundles every session from the past 30 days into a single sessions-last-30-days-<date>.jsonl download.
A few details worth knowing:
- Exports are full fidelity. What you download is the record, not a summary of it.
- If an export fails, nothing downloads. You get the error instead of a truncated file.
- Exports are scoped to your org, like the rest of the sessions surface.
Ways teams use this data
Captured sessions give your team something AI-generated code usually hides: how the work happened. Use that visibility to:
Delete a session
A trash icon sits in the header of every session detail page, next to Export. Click it and a confirm dialog names the session before you commit — deleting is permanent and can’t be undone.
Delete removes more than the one session: the session, its traces, and any subagent sessions it spawned are all removed together. On success you land back on the sessions list.
Deletion is scoped to your org: any member can remove any of the org’s sessions.