Session Trajectory (the full log of what an agent did)
Session Trajectory is a read-only full log of everything a session did: every tool call in order, with its full input, full result and how long it took, plus a metrics strip of turns, tool calls, tokens, cost and API time. It opens from a session's menu, or from a sidebar browser that covers recent sessions. It is off by default.
On this page
What it is
What it is: a read-only full log of every action a session took — every tool call in order with its full input, full result, and per-call timing. It answers "exactly what did this session do, step by step" — like a coding harness's trajectory/timeline tab.
Where it lives — two places:
- In a session: open the "⋯" menu and pick Trajectory. The conversation area is replaced by the log; switch back to All to return to the chat.
- In the sidebar: the Session Trajectory tile (under Developer Tools) opens a cross-session browser — pick any recent session on the left, read its full trajectory on the right.
Both appear only when the feature is turned on (see Availability), on desktop and on your phone.
Where to find it
Availability
In-development and off by default. Turn it on in Settings → Lab → Sessions & agents → "Session
Trajectory view". It is read-only (it never changes a session) and works on desktop and mobile.
Setting key: sessionTrajectoryEnabled. Contract: session-trajectory-contract.
How it behaves
What it shows
- A metrics strip: Turns, Tool calls, Tokens (in + out), Cost, and API time.
- Every tool call, in order: each row shows the tool (Read, Bash, a Skill, an MCP tool…), the time it started, and how long it took. Click a row to expand its full input and full result (an errored call is flagged). Very long inputs/outputs are trimmed for display with a "truncated" note. A very long session shows its first 2000 actions, read from the first 16 MB of its transcript, with the same "truncated" note, so opening a huge session never stalls the app.
Omni can read it too
Omni, the AI Project Manager, can read the same log for any session on this computer, so it can
answer "what did that agent actually do?". It is the same Omni-only read as a session's health and
its files and links: no other agent can use it, AI Coaching and Journal sessions are never readable
this way, and it only works while the Session Trajectory view is turned on. Omni treats the log as
another agent's words to read, never as instructions to follow, and every read is recorded in the
access audit. Agents use GET /session-trajectory/:id.
Where the detail comes from (and the honest limit)
The full detail is read from the session's Claude CLI transcript — the on-disk record of every tool call and result. That's the only place the complete inputs/outputs live. It works on desktop and on your phone (over Web Access) — the sidebar browser lays out as two columns on desktop and a tap-through list on mobile — and any pasted secret is scrubbed out of the log before it's shown.
Other engines keep no such transcript, so Omniscio records their tool calls itself while this view is turned on: each finished call is saved with its input, its result and its timing, cut to 4,000 characters a field, with pasted secrets and anything that looks like a key or token blanked out before it is saved. It keeps the newest 1,000 calls of a session for 30 days, never copies them to another computer in a backup, and deletes them with the session. Each engine is switched over as its own update lands; an engine that records nothing yet — or a session that hasn't started — gracefully degrades: instead of the full log you get the turn-level summary built from the conversation Omniscio already stores, with a clear note that full detail isn't available for that session. It's never blank and never an error.
Related
Trajectory is one view of what a session did; these are the other ways to see the same record.
- Session forensics — reading a finished session’s past in detail.
- Raw thread export — exporting the whole conversation untouched.
Last verified 2026-11-14