Session failed
What a failed PM session means (nothing saved, nothing sent), what causes one, where it shows, and when it is worth reporting.
- Who it is for
- Owners & operators
- Reading time
- 5 min read
- Updated
A session of the PM (shown as on ) "fails" when it ends without a result the system can trust. The rule is fail closed: no are saved and nothing reaches the . There is no such thing as a half-finished batch; either the structured output passed every check and was stored, or zero proposals were stored. Positions, orders and settings are untouched.
Where it shows
- Agent: the session's chip on the timeline reads "error · $cost" and "no output". Open it and the header carries an error badge and one line with the technical subtype. Tool calls made before the failure are still there to read.
- Overview: only failures that concern your decisions appear under Alerts, in plain words: "The analyst did not a " and "A scheduled analyst session did not run".
- Activity › Alerts: everything else (the diagnostics) is here, titled after its kind: "Analyst session error", "Analyst session failed", "Analyst output invalid", "Analyst timeout", "Analyst budget exceeded", "Daily token budget exceeded".
The causes
| Cause | How it reads | What to do |
|---|---|---|
| Invalid structured output: the model's answer did not fit the proposal schema | "Analyst output invalid", zero proposals | Nothing; the next session starts fresh |
| Timeout: 15 minutes for premarket, postmarket, briefs and ad-hoc runs, 13 for intraday, 10.5 for an event-driven session; the process is killed as a group | "Analyst timeout"; the session is closed as timeout killed | Nothing; if it repeats, report it |
| Per-session cost cap reached ($5 by default, or less when little of the day's budget is left) | "Analyst budget exceeded"; the session stopped where it was | Nothing; the cap is a deployment setting, shown on Desk › Portfolio Manager › Configure. Report it if it repeats |
| Daily token budget spent ($25 per book by default) | The session refuses to start and costs $0.00; "Daily token budget exceeded" | Wait for midnight CT; the budget is a deployment setting, and the Model spend tile on Overview shows where you stand |
| A steering note was not acknowledged, even after the PM was asked once more inside the session | "The analyst did not acknowledge a steering note" (Overview) | Reread the note on Strategy › Guidance; reword or retire it |
| Model API error, rate limit or an authentication problem | "Analyst session failed" with the API status in the message | Nothing for a one-off; report if it persists |
| The session crashed or the stream closed before a result | "Analyst session error" with the reason | Same |
| Scheduler could not run it: the process was not awake in the start window (5 minutes for intraday), or the previous run was still going for the whole window | "A scheduled analyst session did not run" (Overview) for a missed slot; an overlap skip is an info line on Activity | Nothing; the next slot runs |
| The platform restarted, or the process was stopped, while a run was in flight | The run is closed as failed with the reason "orphaned" the moment the scheduler is back (it does not wait out the timeout); a run you asked for from posts "Run failed (…): orphaned" into the thread; an event-driven session gets one more try and, if that fails too, its events are left for the next scheduled session to read, with no further alert to you | Nothing; ask again from Consult if you still want the ad-hoc run. Three event-driven failures in a row pause event-driven sessions for that book until the next day; scheduled sessions carry on |
Three things that look like failures but are not. A trimmed output warning on Activity › Alerts ("analyst output trimmed", naming the field and its original length) means the session's answer was valid except that a prose field (a summary, a , a journal entry, a handoff bullet) ran past its length cap; the runner cut that text to fit, ending in "[…]", checked the result once more and saved it. Proposals, symbols, quantities and prices are never touched by the trim, and any other defect still fails the session closed as "Analyst output invalid". It is worth a report only if it repeats every day (the prompt can be tightened). A denied tool call (the model asked for a tool outside its allowlist) is refused, logged as a warning after the session, and the session continues without it. A that errors or times out never fails the PM: the Critic is fail-open, and the proposals proceed with a note saying the Critic did not answer (a proposal that opens a new position or is $250 or more waits at most 90 seconds for the Critic's pre-trade pass, then moves on regardless).
What to do
Nothing is lost. The PM keeps no memory inside a session; everything it needs is in the database, so the next scheduled session (see the timeline on Agent) starts with the same book, the same notes and the same brief as if the failed one had never begun. A failed premarket session means the first intraday session does the day's first look instead.
If you do not want to wait, ask Consult. It gives its own read, and when the question amounts to "should we buy or sell" it offers an Ask for analysis card; confirming it queues an ad-hoc PM run now (or at the next open), capped at 6 Consult-triggered runs a day.
is not necessary. A failed session cannot have sent anything, and the Executor is unaffected.
When to report
Report when the same failure repeats across sessions, when a session fails every day at the same slot, or when the detail line names something you changed (a note, a model, a budget). Press the flag in the top bar; the screen you were on is attached, and what you write is read the next morning. One failed session on an otherwise normal day is not worth a report; the already lists it under Alerts & errors and flags it as an .