Sessions & agents settings
Settings reference for how the app starts, runs, recovers and manages agent sessions, from automatic restarts and input focus through refill governors and deep forensics analysis.
On this page
- Auto-focus chat input
- Session engine V2
- Sessions can self-archive without approval
- Keep sessions with open work on the board
- Self-archiving master switch
- Clear finished agent and app sessions
- Clear finished app-generated runs too
- Clear finished sessions you started too
- Let another company's agent message archive itself
- Weekly forensics scan
- Forensics daily spend cap
- Forensics scan interval
- Session refill batch size
- Refill cooldown between revives
- Refill daily spend cap
- Refill target session count
- Allow refill to reopen closed sessions
- Refill daily revive limit
- Refill candidate age window
- Refill nudge message
- Name patterns excluded from refill
- Auto-continue when agent pauses
- Resume sessions after crash
- Auto-restart unresponsive sessions
- Retry transient launch failures
- Break agent read loops
- Pause endless code reviews
- Review rounds before pausing
- Auto-recover stalled context
- Fix late-thinking display bug
- Clear old sub-agent output
- Allow script-driven session spawning
- Agent Board
- Auto-switch to inbox when idle
- Bring app forward when a session needs you
- Require spawn source tracing
- Auto-detect project icons
- Keep sessions working while this computer sleeps
- Cloud-first helpers
These settings control how Omniscio starts, runs, recovers, and manages agent sessions — from automatic restarts and input focus to refill governors and deep forensics analysis.
Auto-focus chat input
When you switch to a session, Omniscio leaves your cursor OUT of the message box — you read first, then press R to start typing, Gmail-style. That keeps the single-key shortcuts (E archive, S star, C continue, H snooze) live while you read, instead of typing letters into a box you did not mean to be in. Turn this on if you would rather have the cursor placed in the message box automatically on every session switch. A brand-new session, and any session you launch yourself with Ctrl+T or the "+" button, still focus the box immediately either way — there is nothing to read yet. Accepted values: on / off. Off by default (it shipped on until 2026-08-25; if you had it on, your choice is kept).
Session engine V2
Runs the rebuilt session engine (V2) behind your session list, switching, and history. It behaves identically to the previous engine but is built on cleaner internals, which makes future improvements easier to ship. Leave this on unless you notice a problem after an update — if something looks off, turn it off and restart the app (the engine is chosen at startup). Accepted values: on / off. On by default.
Sessions can self-archive without approval
When a session finishes its work, it can archive itself straight away — even if your "Require approval for session archive" setting would normally put up an inbox card first. Only a session archiving itself gets this exemption; archiving a different session through the command-line API still goes through the normal approval path. Useful if you run many short-lived agent tasks and don't want archive confirmations cluttering your inbox. Off by default. Accepted values: on / off.
Keep sessions with open work on the board
When a session tries to archive itself while its branch still has work that is not on master, it stays on your board instead. Work counts as unfinished when its pull request is still open, when its pull request was closed without merging, or when it has commits and no pull request at all. A pull request that merged at the branch's latest commit, a branch whose content master already holds, or a branch tagged ready to merge all let the session go. The session is told why — naming the open pull request by number — and can sign off once the work lands. The same switch also keeps a session on the board when a package or release publish it started (npm publish, gh release create, and similar) failed or is still running; it can sign off once the publish succeeds or after you reply. Only a session archiving itself is checked; archiving a session yourself is unchanged. On by default. Accepted values: on / off.
Self-archiving master switch
The master switch for session self-archiving. When on (the default), a finished session may archive itself off your board once it decides there is nothing left for you to see, decide, or receive. When off, that self-archive signal is ignored and every session stays on your board until you archive it yourself. Turning this off also overrides Sessions can self-archive without approval above. Accepted values: on / off. On by default.
Clear finished agent and app sessions
When a session that an agent or Omniscio itself started — a scheduled job, an automation, a background round like the Resource auditor, or another agent's helper — finishes with a closing card that asks you nothing, names nothing you must do and hands you nothing to open, Omniscio takes it off your board for you and lists it in your daily digest. A session you started — from the app, Quick Launch, a "Start session" tap, or an Ask card you approved — is never cleared by this switch, and neither is any session whose starter Omniscio could not tell. A run the app itself generates — a nightly audit or fix wave, a friction triage or fix, a Sentry intake or other bug-report session, a Land … run, an alert follow-up ([Auto-fix] …, Investigate + fix …) or newsletter research — is never cleared by this switch either. An [Auto-fix] … session clears itself anyway when it finishes without a question and without telling you to do something (changed 2026-11-18), whatever this switch says. Off by default (changed 2026-11-10: finished scheduled jobs and helpers stay in your inbox until you clear them). Accepted values: on / off. Changed by name (PATCH /settings/finishedNoAskMachineAutoArchiveEnabled); it has no row in the Settings window.
Clear finished app-generated runs too
Lets the app-generated runs listed above — a Land … run, a nightly audit, a friction triage and the rest — leave your board when they finish with a closing card that asks you nothing, names nothing you must do and hands you nothing to open, exactly like the machine-started sessions above. It works together with Clear finished agent and app sessions, which must be on too. The daily maintenance run and a crew overseer always stay. Off by default, so these runs stay in your inbox unless you turn this on for your own computer (added 2026-11-14 for developers whose inbox fills with reports that say there is nothing for them to do). Accepted values: on / off. Changed by name (PATCH /settings/finishedNoAskAppRunAutoArchiveEnabled); it has no row in the Settings window.
Clear finished sessions you started too
The same clear as above, applied to every session, including the ones you started yourself — except the app-generated runs listed above, which always stay on your board even with this switch on. Off by default since 2026-10-08, when clearing every session took answers to questions people had asked off their board. Turn it on only if you want your own finished sessions cleared as well. Accepted values: on / off. Changed by name; it has no row in the Settings window.
Let another company's agent message archive itself
When an agent working for someone OUTSIDE your workspace sends a message into one of your sessions, the session that answers it normally stays on your board even when the reply leaves you nothing to receive or decide — so an outsider reaching you is always visible to you. Turn this on and an arrival like that ends its turn exactly as a teammate's message does: if the reply needs nothing from you, the session may archive itself, and the arrival never reaches your inbox. Off by default. Switching it on is an approval-gated change — an agent may ask for it over the command line, but it takes effect only once you approve the card, and only you can switch it on. It affects messages that arrive while Agent-to-agent messages is on, and nothing else: a teammate's agent, another of your own computers, and a person in Team Chat are unaffected either way. Accepted values: on / off.
Weekly forensics scan
When enabled, the deep AI-powered analysis of your past sessions runs automatically once a week (Monday morning) — no manual trigger needed. The deep run spawns a paid agent session and will respect the daily spend cap you set below. The free deterministic friction scan (which looks for command timeouts, tool errors, and similar patterns) always runs on its own schedule regardless of this toggle. Accepted values: on / off.
Forensics daily spend cap
The maximum amount (in US dollars) that the opt-in deep forensics analysis session may spend in a single calendar day before it is refused. Because the deep run spawns a real paid agent session, this cap prevents unexpected charges if the analysis runs more often than expected. Accepted values: 0.05 – 20. Set alongside the weekly auto-run toggle above.
Forensics scan interval
How often the free deterministic friction scan runs — the lightweight pass that looks for patterns like command timeouts, hook blocks, tool errors, and redundant directory changes in your recent session history. Lower values give you fresher friction data; higher values reduce background activity. Accepted values: 1 – 168 hours (1 hour to one week).
Session refill batch size
When the Session Refill Governor revives interrupted sessions on its five-minute check, this limits how many it brings back in a single pass. Keeping this low prevents a burst of simultaneous restarts that could strain your machine or blow through budget quickly. Raise it if you want faster recovery after a crash or restart. Accepted values: 1 – 20.
Refill cooldown between revives
After the governor revives a session, it will not poke that same session again for at least this many minutes. This prevents a session that keeps stopping immediately from being repeatedly relaunched in a tight loop. Accepted values: 1 – 1,440 minutes (up to 24 hours).
Refill daily spend cap
The most the Session Refill Governor may spend reviving sessions in a single local day. Only pay-per-use sessions count toward it — an API key or a per-token vendor; a session on one of your Claude Pro/Max subscription accounts costs nothing extra per turn, so it does not count (one whose account you have since removed does, because its billing can no longer be told). Once today's revives reach this dollar amount, the governor pauses until tomorrow, and the Queue page of the Bulk session manager says the spending limit is what stopped it. Set to 0 to disable the dollar cap entirely — the daily revive-count limit still applies even when the dollar cap is off. Accepted values: 0 – 100,000.
Refill target session count
When fewer than this many sessions are actively running, the governor looks for recently interrupted work sessions and revives them until your running count reaches this target. Think of it as your "always keep N sessions working" setting. The governor only revives interrupted work sessions by default — not sessions you explicitly closed. Accepted values: 1 – 500.
Allow refill to reopen closed sessions
By default the Session Refill Governor only revives interrupted sessions — it never reopens a session you explicitly closed yourself. Turn this on to grant it that extra permission, so it may also bring back sessions you closed if your running count drops below the target. Off by default. Accepted values: on / off.
Refill daily revive limit
A hard backstop on the total number of sessions the governor may revive per local day, regardless of cost. This bounds the absolute worst-case spend even if the dollar cap is off. Default is 200 revives per day. Accepted values: 1 – 5,000.
Refill candidate age window
The governor only considers sessions that were active within this many hours as candidates for revival — older or abandoned sessions are skipped. Lower this to revive only very recent work; raise it to cast a wider net on older interrupted sessions. Default is 24 hours. Accepted values: 1 – 720 hours.
Refill nudge message
The "continue" message the governor sends to a session it revives, for example: "Please continue where you left off and finish the task." You can customise this text if your sessions respond better to a different prompt — for instance, one that reminds them to check the branch status or summarise before continuing. Accepted values: any text up to 2,000 characters.
Name patterns excluded from refill
A list of case-insensitive patterns (regular expressions) matched against session names. Any session whose name matches is never revived by the governor — useful for keeping generic shells, intake sessions, or outward-facing bots off the revival list. Default patterns exclude names like ^Session 1, [BUG], [FEEDBACK], Email:, and Tweet:. Accepted values: up to 50 patterns, each up to 200 characters.
Auto-continue when agent pauses
When the agent pauses mid-task without asking a question — for example, it says "I'll continue now" but then stops — Omniscio automatically nudges it to keep going. This prevents unnecessary inbox interruptions during long jobs where the agent briefly idles between steps. Turn off if you prefer to see every natural pause land in your inbox. Accepted values: on / off.
Resume sessions after crash
When the app is closed unexpectedly or crashes, sessions that were actively running receive a "continue" nudge when Omniscio restarts; sessions that were waiting for your input are restored to their waiting state. This means you rarely lose work to an unexpected shutdown. Turn off if you prefer to manually decide which sessions to restart. Accepted values: on / off.
Auto-restart unresponsive sessions
When a session shows "thinking" but its underlying process has wedged and stopped responding, Omniscio detects the stall and restarts that turn automatically — instead of leaving it spinning forever. The restart picks up from where the session stalled. Turn off to leave a stuck session in place so you can investigate it yourself. Accepted values: on / off.
Retry transient launch failures
When a session fails to launch because the machine is momentarily overloaded — a brief OS hiccup, not a genuine error — Omniscio keeps retrying with an increasing gap between attempts instead of giving up after a few tries. Genuine errors (wrong credentials, bad config) still stop immediately and turn the session red. Accepted values: on / off.
Break agent read loops
When an agent gets stuck re-reading the same unchanged file over and over, Omniscio detects the pattern and re-launches that turn so it can make forward progress — instead of letting it spin and stall. This covers a known failure mode where an agent loses its place and loops on tool reads indefinitely. Accepted values: on / off.
Pause endless code reviews
Every code-review round starts fresh reviewer agents, and those use your Claude limit. A session left to review and fix its own work can run round after round without ever checking with you. When this is on, a session that reaches the round limit below with no message from you stops and lands in Needs You, with a note saying how many rounds it ran. Two or more reviewers sent out together count as one round, and a review the session asks Omniscio for counts too. Your reply lets it continue and starts the count again; messages from other agents and automatic wake-ups do not. Accepted values: on / off. On by default.
Review rounds before pausing
How many code-review rounds a session may run without a message from you before it pauses. Shown when Pause endless code reviews is on. Accepted values: a whole number from 2 to 20. Default 5.
Auto-recover stalled context
When a session fills its context window and stops mid-task without finishing, Omniscio automatically frees space and continues it — so you see a finished reply instead of a blank or truncated one. This is most useful for long-running coding or research sessions. If you prefer to see the context-full warning and decide manually, turn this off. Accepted values: on / off.
Fix late-thinking display bug
Addresses a rare display issue where a background tool returning late caused an "Extended thinking" line to appear after the agent's real answer — pushing that answer into the collapsed activity section and leaving only a short acknowledgement visible. Turn on to recover those hidden answers. Off by default because the problem is rare; the clean folded view is the normal case. Accepted values: on / off.
Clear old sub-agent output
When on, Omniscio removes the raw hidden sub-agent (Task helper) output from sessions you archived more than 14 days ago. Your conversation messages are untouched; only the collapsed "subagent result" text cards in those old archived sessions are cleared. A full backup is taken before anything is removed. Useful for reducing database size on installs with a large archive. Accepted values: on / off.
Allow script-driven session spawning
Lets external scripts spawn Omniscio sessions and send them follow-up messages programmatically. Off by default — leave it off unless you have a trusted script or automation that needs this access. When on, use the limits below (agentDrivenMaxConcurrent, agentDrivenMaxDollarsPerSession, agentDrivenMaxTurnsPerSession) to bound what those scripts can do. Accepted values: on / off.
Agent Board
Shows a live board of every agent Omniscio has launched — its workspace, current status, dev-pipeline phase, and to-do checklist — each with one click to jump straight to that session. Enabling this adds an Agent Board icon to the toolbar. Useful when you're running many concurrent agents and want a single-glance overview of what's happening. Accepted values: on / off.
Auto-switch to inbox when idle
When another session needs your attention, Omniscio automatically navigates you to the inbox — but only if you've been idle for at least five seconds and are not currently scrolling, clicking, or typing. This keeps you from being interrupted mid-action while still surfacing urgent items quickly during natural pauses. Accepted values: on / off.
Bring app forward when a session needs you
Brings the Omniscio window to the foreground whenever a session transitions to "Needs you" — for example, when an agent asks a question or hits a required approval. Useful if you work in other apps while agents run in the background. Turn off if you prefer to check in on your own schedule. Accepted values: on / off.
Require spawn source tracing
When on, a session spawned via the command-line API is rejected unless the caller declares which session requested it (the X-AMC-Source-Session-Id header — Omniscio-spawned agents have it automatically as $AMC_SESSION_ID). This keeps every spawned session traceable back to its origin, which is important for auditing and debugging complex multi-agent workflows. Besides a session, the app also accepts a scheduled job (its job id) and, on POST /project/:name/new only, an app-run ops job that its ops manifest marks as allowed to start sessions (today the PR problem router). Turn off to allow anonymous script or manual spawns. On by default. Accepted values: on / off.
Auto-detect project icons
Scans hub folders for favicons and logos and uses them as the project icon in the sidebar. When on, projects tied to a web app or known codebase often get a recognisable icon without you having to set one manually. Turn off if you prefer to assign icons yourself or if the auto-detected results are inaccurate. Accepted values: on / off.
Keep sessions working while this computer sleeps
Lives in Settings → SSH Remotes, on the Remotes tab. When it is on and this computer sleeps or closes, a session running on one of your own SSH-connected Linux machines keeps working there instead of stopping — and keeps spending while it does. That machine approves only what your normal approval level already allows and waits for you on the rest; when you come back, the session catches up and shows what it did. It applies only to your own Linux machines (a cloud machine has its own switch, and a Windows machine is never included), and the machine needs Node.js installed — if the helper cannot be installed there, the session refuses to start with a clear message rather than quietly running without it. Off by default. Accepted values: on / off.
Cloud-first helpers
When an agent asks Omniscio to start another agent and does not say where it should run, the new one runs in the cloud instead of on your computer — provided it is a helper: any spawn by a member of a mission crew, including the ones the crew's own lead starts, on a worker engine (that is nearly every helper a crew starts). This covers a plain spawn, a crew swarm's workers, and a crew reviewer. Your machine is kept for the sessions that read the fleet and hand out work, so helpers no longer fill your computer up. Coordinators (a Claude genius or smart engine), and anything you start yourself or that a script or scheduled job starts, are never moved — and asking for a specific place always gets exactly that place: send "location":"local" when a helper must stay on this computer (it needs local master or the live database, say) together with a "localReason" — box-bound-test, live-checkout-read, landing-mint, box-credentials or owner-approved — because a crew worker's local run now has to name a real reason or the request is refused (409) with the allowed ones listed. If the cloud cannot take the helper — no funded account, the day's machine allowance spent, or Omniscio cannot confirm it has room (when the Boat CLI is not installed on this computer the refusal says so and gives the install line) — the helper runs on this computer instead and tells you it did and why (in the new session's log and in the answer to the request), so a local run is never reported as a cloud one. Turn this off and every agent spawn behaves exactly as it did before this setting existed. Accepted values: on / off. On by default.