Skip to main content

Sessions

Every time an agent works on a task, it runs a session. The Sessions tab of Audit lists those sessions, says how each one ended, and lets you download what the agent did and said: its transcript. A session that ended without explaining itself can also have a pane log: the last screen of output its node kept.

Open it from the sidebar: select Audit, then the Sessions tab. It has two parts: Sessions, the list, and Efficiency.

Sessions is a plan feature, Session transcripts. When your organization's plan leaves it out, both parts, Sessions and Efficiency, show Not on your plan in place of the list and its figures, for every role, the Org Owner included. An org admin sees a See plans link; anyone else reads Ask an admin to change the plan. See Plans and billing.

Not on your plan: Your plan doesn't include session transcripts, with a See plans link

How a session is recorded​

A run that ends before it writes a session is listed as an attempt, not a session. A session that is dispatched but never starts stays stale until it is deleted. A session that ends hands its transcript to its node to upload; one that ended without a recorded cause can also send its pane log.

The list​

Each row is one session.

Filters​

FilterWhat it does
WindowLast 24 hours (the default), Last 7 days, Last 30 days, All time or Custom range, which adds From and To boxes. A range counts sessions by when they started; times are in your browser's time zone, named under the filters while Custom range is picked.
StatusRunning, Ended or Unknown. Any status shows all.
TranscriptShows only sessions whose transcript is in one state. See Transcript states. Any transcript shows all.
End reasonShows only sessions that ended one way, such as Model not available to the account. Each reason shows how many sessions ended that way, and only reasons that occurred are listed. Any end reason shows all.
NodeShows only sessions that ran on one node. All nodes shows all.
SearchNarrows the list to sessions whose name, session id, row id, slot, branch or holder contains what you type. Hover or select the box to see that list.

When you arrive from a task, the list shows only that task's sessions and a Task #… ✕ button. Select it to clear that filter.

The filters, the sort and the page are kept in the page's address, so a copied link opens the same view. A Custom range's From and To are not, so such a link opens on the default window.

Attempts with no session​

When a task's run ended before it wrote a session, it is listed above the sessions under Attempts with no session, newest first.

ColumnMeaning
TaskThe task, by its title. Select it to open the task. +… tasks shows other tasks that ended the same way; select it to list each with its own last attempt.
WhenHow long ago the attempt ended. Hover it for the exact time.
CauseWhat ended the run, such as stopped while its question waits for an answer. A cause seen more than once is counted, such as 2 ×.
NodeThe node it was dispatched to. Shown only when an attempt listed names one.

Each row starts with a red ✕ mark: it ended with no session. Hover the row, or tap the mark on a touch screen, to read the runner's own reason. Select the arrow at the start of a row with +… tasks to list the other tasks beneath it, each with when it ended; a task that ended more than once is counted, such as ×2. On a phone the cause and node follow the task's title. Five attempts are shown at first; select Show all (…) for the rest and Show fewer to go back.

Each session​

ColumnMeaning
MarkThe transcript's state, as a symbol. Hover it for the reason.
SessionThe session, such as task 7725 · f374ddea: its task and the start of its session id. Beside it, its status: ▶ running, ■ ended, ? unknown.
TokensHow many tokens the session used, then what it cost when that is known, such as 1.2M · $3.40. Hover it for the split between input and output.
TimeHow long the session ran.
EndHow long ago it ended. Hover it for the exact time.
NodeThe node it ran on.
TaskThe task it worked on, by its title, or by its number when no title is shown. Select it to open the task.
TranscriptThe transcript's state, in words.
StartHow long ago it started. Hover it for the exact time.

Select a row to download its transcript, as Download in its ⋯ menu does. When the transcript can't be downloaded yet and the node sent a pane log, selecting the row opens the pane log instead. Both need a role that can download transcripts; without one, selecting a row does nothing.

On a narrow screen, only Session and Start stay, with the tokens and cost beside the session's name. A session's task can still be opened from its ⋯ menu.

Hover the row's mark or its name, or tap the mark on a touch screen, to read its notes. A note about a failure turns the mark red. The notes can be:

  • Started … · Ended …: the exact times, with the time zone. A running session shows only Started.
  • Agent: …: the agent the session ran, with its version, for a session run by an agent you built.
  • Engine: …: the built-in engine that ran it, such as Claude Code, when the page lists sessions from more than one engine.
  • Runtime: ACP host: the session ran on the ACP host runtime.
  • Account: …: the model account the session ran on, when it was recorded and you can see it.
  • Slot: …: the worktree slot the session ran in, which Search also matches.
  • Why it ended: its end reason and what was recorded about it, red when that explains a failed run.
  • No events received, or Events refused: …: the session sent nothing, or what it sent was turned away.
  • Node log: …: a line from the node's own log about why the run ended.
  • No pane log on node: the node had no pane log to send.
  • Why the transcript isn't ready or failed.

When the window holds few sessions or none, a Wider windows: line below the list counts the sessions in each longer window, such as Last 7 days: 12. Select a count to switch to that window.

When there are more sessions than fit on one page, use Previous and Next below the list.

Sort​

Select a column heading to sort by it; select it again to reverse the order, and a third time to go back to the default. The sort button (↑↓) at the end of the filters adds the orders no heading offers.

Default order (running first, then newest) is used until you pick another. Once you do, a Default order button appears beside the filters to go back.

Transcript states​

StateMeaning
pendingWaiting for the node to upload it.
uploadingThe upload is in progress.
uploadedUploaded and ready to download.
failedThe upload failed. The reason is shown under the row.
gave upThe upload was tried and abandoned.
unavailableThe transcript couldn't be read from the node.
n/aThe session has no transcript.
noneThe session ended with nothing to upload: it sent no events, or its MicroVM failed. Hover the mark for which.

The Transcript filter names these states, all but none, with a capital letter, and calls n/a Not applicable. It also offers Stale (never started, over …): sessions that were dispatched long ago and never started. See Stale sessions.

A session's menu​

Select ⋯ on a row.

ItemWhat it does
Open task #…Opens the task the session worked on.
DownloadDownloads the transcript. Shown only if your role can download transcripts. Unavailable until the transcript is uploaded; the reason is shown under the item.
Pane logOpens the session's pane log. Shown only if your role can download transcripts and the node sent one.
DeleteDeletes the session's row. Shown only if your role can download transcripts. Available only for a stale session; otherwise the reason it can't be deleted is shown under the item. To confirm, select Delete session. There is no undo.

Transcript redaction is best effort​

Download first shows this warning, with the archive's redaction confidence. Redaction removes the secrets it recognizes, but it can miss some: treat the transcript as if it still holds credentials, tokens and customer data. Download in the warning starts the download; Cancel goes back without it. See Follow a task's sessions.

Pane log​

A session's pane log is the last output its node's terminal showed, sent once when the session ended without a recorded cause. Open it with Pane log in the session's ⋯ menu; a node's health view and a task's activity link to it too. It needs a role that can download transcripts: without one the page says Needs sessions edit access.

ItemMeaning
TitleThe session's name. Sessions above it goes back to that session's row in the list.
NodeThe node it ran on. Select it to open the node.
TaskThe task it worked on. Select it to open the task.
End reasonWhy the session ended, in words.
Ended · last writtenWhen the session ended, which is usually when the log was last written. When the two differ, Ended and Last written are shown apart.
WrapWraps long lines instead of letting them run past the edge. On by default on a phone.
CopyCopies the whole log.

The log opens at its end, where the newest lines are. Above it, Secrets redacted best effort (… caught); treat as sensitive. says how many secrets were removed; others may remain. Notes under it can explain [×N], where repeated retry lines were folded into one, and Long lines cut at the edge, with a Wrap link.

MessageMeaning
Written after this run ended; may show a later run.Runs of the same task share one log, so a later run's output may be shown.
No pane log for this session.The node sent none.
Could not load the pane logIt couldn't be read. Try again.

Stale sessions​

A session that was dispatched but never started never ends, so nothing will ever finish it. When there are such sessions, a yellow banner above the filters says how many.

  • Select Show them to list them.
  • Select Delete … sessions to delete them all. Shown only if your role can download transcripts. A dialog states the count; select Delete … sessions again to confirm. There is no undo. One pass deletes a limited number; when more remain, the result says how many, so run it again.

A blue banner can also appear: it counts sessions that had no record until their node reported them.

Empty states​

MessageMeaning
No sessions in the last 24 hours (or the window you picked)No session ran in that window. Set Window to a longer one.
No sessions in this rangeWith Window at Custom range: no session started between From and To.
No sessions recorded yetWith Window at All time: no session has run. One appears as soon as a node reports it.
No sessions match these filtersNo session matches your filters. The line under it says how many sessions there are in total. Clear the filters, or set Window to All time.
Could not load sessionsThe list couldn't be read. It doesn't mean no sessions ran. Try again.

Efficiency​

The Efficiency part, under the heading Where the time goes, splits the agents' time over a period into phases. Pick the period, 7d, 14d or 30d, from the list beside the heading; the dates are shown in UTC.

ItemMeaning
LegendEach phase with its share of the time and its total time.
Genuine explorationMinutes per session spent reading the codebase, with waiting, gate and test runs left out.
Cached answer / Fresh scanWhether the figures were just computed or reused, how much was scanned, and how long ago.

Per phase​

PhaseTime spent…
ExplorationReading the codebase to find out what is there.
AnalysisDeciding what to do: planning, design and critique.
CodingWriting the change.
VerificationProving the change works: reviews, tests, browser checks.
GateRunning the project's own gate command.
IntegrationGit, the pull request, and waiting for the merge.
ReportWriting the report you read on the task.
CompletionFinishing the task.
WaitingDoing nothing on purpose, such as waiting on a job. Select it to split it into Gate, Tests, Subagent and Other.
Awaiting humanWaiting for someone to answer a question. Shown as excluded: it is people's time, not the agent's, so it counts in no share.
OtherClassified, but into none of the above.

Hover a phase's name for what it covers. The columns:

ColumnMeaning
TotalTime spent in the phase.
ShareIts share of all the time measured.
Median / runThe typical time per session.
SessionsHow many sessions spent time in it.
CallsHow many actions the agents took in it.
Tokens in / outTokens read and written. Hover for the cache figures.
CostWhat the phase cost. When some spend has no price, a note says so and no cost is shown.

Per model​

Per model shows the same period's tokens and cost for each model: Tokens in, Tokens out, Cache read, Cache write and Cost, most expensive first, or by Tokens out when no cost is shown. When no spend in the period is tied to a model, it reads No model spend was attributed in this window.

Gate churn​

Gate churn shows how often the project's gate had to run more than once: Median runs, p90 runs, Needed > 1, Red runs, Red → green, Sessions gone red, and Files re-edited after the first red gate, and how much of the time went on rework. It reads not measured when there is nothing to count.

Duplicated work​

Duplicated work shows how often agents read the same thing twice. Sessions, Median targets and Repeat reads count reads repeated within a session; Re-reads, Epics and Median / epic count files read again by a later stage of the same epic. Hover a figure's name for what it counts.

Subagent fan-out​

Subagent fan-out shows the helper agents a session started: Spawns, Median / session, Waves (spawns started together), Median wave, Failed spawns, Subagent time, and Subagent wall-clock per hour of run time, where above 1× means they ran in parallel.

Orientation​

Orientation shows how long sessions took to start changing code: Median calls is the typical number of actions before a session's first edit, and Edited and No edit count the sessions that made an edit and those that made none. It reads not measured when there is nothing to count.

Sessions​

The Sessions list shows each session in the period: Session, Persona, Repo, Start, Task, Gates (gate runs), Red (red gates), Spawns (subagents), Sub time (subagent time), Rereads (repeat reads), To edit (calls to first edit), Total, Phases and Cost. A cost marked agent-reported was reported by the agent, not priced from its tokens. Hover an abbreviated heading for its full name; select a heading with an arrow to sort by it. A narrow window keeps Session, Total, Phases and Cost and folds the rest. Select a session to show its task's sessions in the list.

Project memory​

Project memory, at the bottom of Efficiency, appears only when your plan includes project memory. It compares one project's sessions run with project memory on against those run with it off, over the period picked at the top of Efficiency, named under the heading, such as Last 7 days. Pick the project with Project, and narrow it to one persona with Persona (All by default). The line above the tables gives the dates and which sessions each side counts: with no comparison group it reads No comparison group: split by each session's memory setting; when Jaah held some tasks back from memory on purpose, it reads OFF: comparison group and default tasks · ON: default tasks.

ItemMeaning
ON vs OFFThe project's sessions with memory off beside those with it on. When the sessions got their setting in more than one way, each side's name has a line saying how, such as 13 by default, 2 set on the task. The ways are set on the task, from the epic, by default, in the comparison group, by the kill switch, not on your plan, set on project and from an unknown source.
Sessions / TasksHow many sessions and tasks each side counts.
Units injected per task (median)How many memory units a task was given.
Units read per task (median)How many of them it read.
Units cited per task (median)How many it cited in its work.
Tool calls to first edit (median)How many actions a session took before its first edit.
Duplicate reads (median)How often a session read the same thing twice.
Not countedSessions from before project memory, and tasks that ran with memory both on and off.
ON, split by deliveryThe ON side split into tasks given at least one unit (ON (≥1 injected)) and tasks given none (ON (0 injected)).

Each median shows its spread (IQR) and total beneath it. A side with too few sessions shows why, such as Not enough ON sessions in this window (n = 2), in place of its figures.

Read rate per kind lists each kind of memory unit: Injected, Used, Read rate (the share of injected units the task used), Written, Read, and Written per read.

When there is nothing to show​

MessageMeaning
No sessions in this windowNo session ran in the period. Pick a longer one.
Off — not in your plan.Under Project memory: your organization's plan doesn't include project memory, so there is nothing to compare.
No project has memory onUnder Project memory: no project has project memory turned on. The line under it says where to turn it on: Billing for an org admin, otherwise An admin can turn it on.
Not set upBeside Where the time goes: time breakdown isn't set up for your organization's Jaah, so there is nothing to show, and your organization can't change it.
Session efficiency is not available, or the reason in its placeIt can't be shown right now, such as while a scan is busy. Try again, or pick another period.