Omniscio

CLI Pending Actions (approve external AI changes)

Almost everything an AI asks Omniscio to do over the CLI control server now just happens — no card. Three risky kinds (deleting what cannot come back, sending or publishing as you, and security switches) make the agent read a warning and confirm before they run, and then tell you. A short list still waits for your click. This page covers that queue, how to turn a family's "ask me" switch back on, and how to read and answer a card.

On this page

What it is

When an external AI (Claude Code in another project, ChatGPT with the omniscio-control skill, a curl script) asks Omniscio to take an action that still needs your click (a family whose switch you turned on, or one of the person-only actions below), the request does not apply immediately. Instead it lands in a shared approval queue called cli_pending_actions, and a row appears in your inbox as a yellow item titled with the request's preview text (no "Approve:" prefix) alongside your messages and notifications. You click the row, a modal opens showing exactly what's about to happen — the action kind, a humanized breakdown of the request (named fields, resolved targets like the real project / session / tag / recipe name, plain-English values), and a collapsible Show technical details section with the raw JSON for power users — and you approve or reject. Only after you approve does the change apply.

This is the user-in-the-loop guarantee for the approval-gated CLI capabilities. Each action family has its own on/off switch, and since 2026-11-19 every one ships off — app settings, tags, project edits, away-mode and the rest apply at once until you switch a family on. Session pause / unpause / snooze / archive work the same way: that family defaults OFF, so those calls apply immediately until you switch its gate on (see Which action families need your click below). Existing installs were switched over once, the first time they started on this version.

Guarded actions — the agent double-checks, then you are told. Three kinds never need your click but never run on the first ask either: deleting something that cannot be brought back, sending or publishing something under your name, and security switches (sign-ins, keys, guards, standing permissions). The first request is answered with a plain warning and a one-time code; it runs only when the same agent sends the exact same request back with that code within 10 minutes, and you then get one inbox notice saying what ran (repeats of the same kind in an hour fold into that one notice with a count). A session that has been reading web pages or emails still gets a normal card for these, so a booby-trapped page cannot talk it into one.

Still your click. Restoring from a backup, adding or upgrading an outside software package, letting an agent act as you in a terminal, and the switches that limit what agents themselves may do. The CliActionKind union is large and keeps growing (the complete list is CLI_ACTION_KINDS in cli-pending-types.ts); the most commonly hit ones are:

  1. Settings PATCH (PATCH /settings/:key) — change a single Omniscio setting (theme, notification volume, MemPalace toggle, etc.) — settings.patch
  2. Session pause (POST /session/:id/pause) — session.pause
  3. Session unpause (POST /session/:id/unpause) — session.unpause
  4. Session snooze (POST /session/:id/snooze) — session.snooze
  5. Session archive (POST /session/:id/archive) — session.archive
  6. Agent-driven session spawn (POST /agent/sessions) — session.spawn.agent_driven (initial approval gates the spawn; follow-ups inside the user-approved turn + dollar caps are pre-approved per spawn request)
  7. Project delete (DELETE /project/:id) — project.delete
  8. Project bug-intake update (PATCH /project/:id/bug-intake) — project.bug_intake_update (enable/disable a project's bug-report email intake slug — flipping it on opens an external email side channel that auto-spawns Claude sessions, so it's approval-gated; slug uniqueness is enforced case-insensitively at queue AND apply time)
  9. Project docs upload (POST /project/:id/docs) — project.docs_upload
  10. Project docs delete (DELETE /project/:id/docs/:filename) — project.docs_delete
  11. Recipe run (POST /recipe/configs/:id/run) — recipe.run (this route applies at once by default and queues for approval only when you switch on the recipe-run family's requireApprovalForCliRecipeRun gate; the agentTriggerable pre-approval bypass and its 30-second per-(recipeId, projectId) cooldown belong to the separate direct route POST /recipes/run, not to this one)
  12. Away-mode rule create (POST /away-mode/rules) — away_mode.create
  13. Away-mode rule update (PATCH /away-mode/rules/:id) — away_mode.update
  14. Away-mode rule delete (DELETE /away-mode/rules/:id) — away_mode.delete
  15. Cron failure alert (auto-emitted by Omniscio when a cron job fails) — cron.failure_alert (status semantics differ: pending = unread, approved = acknowledged, rejected = dismissed-as-noise with 24h cooldown; capped at 5 outstanding per-kind so cron-failure cards don't starve the global queue cap)
  16. Cron failure heal (auto-emitted escalation card — "needs human attention" — after 3 self-heals fail; the only inbox-visible heal row, since regular heals auto-fire without queueing) — cron.failure_heal (acknowledge-only, same status semantics as the failure alert)
  17. Email summarizer rule create (POST /email-summarizer/rules) — email_summarizer.rule_create
  18. Email summarizer rule update (PATCH /email-summarizer/rules/:id) — email_summarizer.rule_update
  19. Email summarizer rule delete (DELETE /email-summarizer/rules/:id) — email_summarizer.rule_delete
  20. Email summarizer rule toggle (POST /email-summarizer/rules/:id/toggle) — email_summarizer.rule_toggle
  21. Email summarizer setup (POST /email-summarizer/setup) — email_summarizer.setup_wizard
  22. Tags library create (POST /tags) — tags.create
  23. Tags library update (PATCH /tags/:id) — tags.update
  24. Tags library delete (DELETE /tags/:id) — tags.delete (cascade preview shows session count before approval)

Cron-job creates and automation-rule creates do not land in cli_pending_actions — they ride on a separate per-row approval_status='pending' flag on the cron_jobs / automations tables (migrations v88/v89) and surface through their own dedicated UI inside the Cron Jobs / Automations panes. This page covers only the unified-inbox flow above.

The queue has a hard cap of 500 outstanding pending rows across ALL action kinds (raised from 20 on 2026-09-14). When the cap is full, every new queue request returns HTTP 409 until you approve or reject some.

Where to find it

The queue shows up as the Approvals section of your inbox, beside your messages and notifications, with a yellow badge so it stands apart from both. The switches that decide which requests need your click live in Settings → CLI Control → Approval requirements, and everything you have decided never to be asked about again is listed under Settings → CLI Control → Always-allowed actions.

How it behaves

Configuring which actions require approval (2026-06-05)

By default every capability above behaves exactly as described — but you can now decide per action family whether an external cli-token caller's request waits in this queue or applies immediately. Open Settings → CLI Control → Approval requirements and flip a family's switch:

  • On — requests queue here for your approval. Every family ships off since 2026-11-19.
  • Off — requests from an outside CLI caller apply immediately, running the same code path an approved row would, so an un-gated action is identical to one you'd have approved by hand. Rate limits, idempotency, and the queue cap are unchanged. Paid / destructive families (recipe run, session spawn, Nighty-Tidy run-now, AI-coaching interview, project delete, bug-intake enable) carry a ⚠️ badge and a confirm step before you can turn approval off.

Every switch ships off, so a family asks you only once you flip it on. Sessions Omniscio spawns itself are unaffected — the switches govern the external cli-token boundary only; an in-app session keeps applying an ordinary setting immediately. One exception: a setting that switches a safety guard off, changes one of these approval switches, widens what an agent may do, gives your consent, or arms spending or landing is treated exactly like an outside AI's request — the same card, following your App-settings switch — even from a session Omniscio spawned. Those sessions are AI agents, not you; your own Settings screen never goes through this route, so your own clicks still apply at once. The deep-link session spawn reuses the existing "Require approval for AI session spawn" switch. Mechanism + invariants: cli-server-gating.md and the contract cli-approval-policy-contract.md.

Always allow one specific action (narrower than a family). Right on an approval card, the green Approve button has a small arrow attached to its right. Approve itself (or Enter / Ctrl+Enter) answers only the item in front of you. Click the arrow to open the always-allow panel, headed Approve and stop asking about, for that one exact action — e.g. "Upload docs file" only, not the whole "project edits" family. Pick where: This session only and This overseer (the overseer post, whichever session holds it) appear only when Omniscio has verified which session sent the card — never for a request that merely names a session — then Projects (offered when the action belongs to a project; tick one or several) or Everywhere. Pick how long: 1 hour, Today (the default, until midnight), 1 week or Forever. Approve & stop asking approves the item in front of you and immediately clears any other items already waiting in your inbox that the new allowance covers (a session or overseer allowance clears only waiting cards from that verified session or post). The Undo toast removes what you just allowed and puts back anything it replaced. A panel left open closes when the card moves on to the next item, and the keyboard never grants a standing approval on its own. Money / delete actions confirm first. Almost every card has the arrow (since 2026-11-04). Only a short list never does, and those get a plain Approve button: cards whose action needs your own click to run at all (remote control of this computer, rotating your access key, raising a cloud box limit, a setting held for you, a request Ask Omniscio brokered, and a few more), and cards where a standing yes would let an agent give itself power (pausing a session, owner-power grants, acting as you in a terminal, removing your grants, raising its own crew limits, releasing a flagged email, letting spam through, switching guards off). The arrow works on the phone too (owner decision 2026-11-04): a standing approval can be saved, undone and removed from the paired phone as well as the desktop app, so the Always-allowed actions list and the agent-recording auto-approve switch appear there too. A "Remove the workspace" card for a folder outside every worktree home has its own arrow: once you choose Approve & stop asking there, the background cleanup approves each new card of that kind itself and the folder is removed, with its branch kept, untracked files copied to the rescue folder first, and never while a session is working in it. Allowing branch retires does not cover it, and a card raised before 2026-11-04 still needs its own click. A plugin action keeps its single Always for this <plugin> action choice. Everything you allow this way is listed at Settings → CLI Control → Always-allowed actions, with where it applies, how long it lasts and a Remove; a timed allowance switches itself off at its deadline. If an always-allowed action fails when it runs on its own (it times out, or hits an error), it is closed with its reason and you are not asked about it again; the agent that asked is told it failed and can try again. It layers over the family switches above and applies through the same approve path (rate limits, idempotency, and the queue cap unchanged). Note for AI agents: this is human-only — there is no CLI route to grant it, so an agent can never turn off its own approvals.

Cloud helpers can be always-allowed too. The card A session running in the cloud is asking to start a helper session carries the same arrow (it confirms first, because each helper rents a cloud machine). Pick Everywhere for Forever in the arrow's always-allow panel and every future helper request from a cloud session starts with no card, and the ones already waiting are approved on the spot. Your helper limits still apply first and refuse a request before any card would exist: 3 helpers per session (requests still starting count), helpers can't start their own helpers, $5 per helper group, $20 a day, and the cloud-started on/off switch. Remove it at Settings → CLI Control → Always-allowed actions to go back to being asked.

So can the "run its checks on this machine" card. The cloud-enforce window card has the same arrow; always-allowing it opens each later ask at once for that agent only and the length it asked for.

Plugin actions get a narrower scope than either of those. A plugin action (plugin.cli_dispatch) is never covered by a project or an everywhere scope — those two don't apply to it at all. Its own approval card's arrow menu instead offers "Always for this <plugin> action", which always-allows exactly that one method + path of that one plugin, nothing wider (a confirm step, with a stronger warning when the plugin itself flagged the action as needing confirmation every time). It's listed alongside every other always-allowed action at Settings → CLI Control → Always-allowed actions, and shows "Not active" if the plugin is later removed or stops declaring that action. Same human-only rule: there is no CLI route to grant it.

Keyboard-shortcut settings are always instant. One settings carve-out sits below the App-settings switch: a settings.patch that only binds or toggles a keyboard shortcut — any global hotkey (focus, Omni, Quick Launch + per-tab, KMS, voice), the screen-recorder hotkeys, the push-to-talk hold keys, plus each one's on/off toggle — applies immediately and never queues here, regardless of the App-settings switch. Rebinding a shortcut is a benign personal preference, the same trust level as Omniscio's in-app keybindings editor. Feature on/off switches (e.g. "Enable KMS", "Enable Voice") and contact data are NOT treated as hotkeys and still queue for approval. hotkey-settings-are-always-instant in the cli-approval-policy contract.

Cards only you can decide (2026-10-29, widened 2026-10-30). Two card kinds can never be always-allowed or switched off by a setting (they are on the short list above). "Run an action Ask Omniscio asked for" (ask.brokered_request) carries one request Ask was refused on — a session launch, an integration write — with the exact path, its contents and Ask's reason; approving it runs exactly that request once, with your authority, and the agent only learns whether it ran. A session-launch card is named from what it will start — "Start a new session in <project>: <session>" — because every launch shares one route, so a title built from the route gave two different launches the same line of API documentation and neither could be recognised as the session you were asked to approve. A request the action would refuse is turned away when it is ASKED — the agent gets the reason and no card is raised — so a card is only ever raised for something that will run when you approve it. The first comes only from the Ask Omniscio help agent. "Change a setting held for you" (settings.person_held_patch) carries a sign-in, second-factor, network allow-list or trusted-sheets setting that ANY agent session asked to change (a weaker land-safety standard only Ask may ask for); approving it applies that value, unless the setting changed after the card was raised, in which case it is refused and nothing changes. Consent records and the control server's own switch are never offered as cards. Only your own click in the app runs either one. A card like this belongs to the session that asked for it, so when that session's seat changes hands the card and the approval behind it move to the new holder together, and a card you can still see in your inbox is always one you can still act on. A session that ends with nobody taking its seat still has its waiting cards withdrawn, and its card then opens to say so.

What the confirm gate does on the Approval-requirements screen (F074)

Turning a paid or destructive family's approval off is the risky direction, so the switch does not move until you confirm: a dialog names the family ("Turn off approval for Project / group delete?") and offers Turn off approval / Keep approval. Turning approval back on for any family, or turning a non-danger family off, applies at once with no dialog. A destructive family's dialog is a destructive confirm and carries role="alertdialog" (destructive-confirm-is-an-alertdialog in the frontend-modal contract); a paid-but-not-destructive (cost) family gets an ordinary role="dialog".

Its spec, tests/unit/features/settings/cli-approval-settings.test.tsx, pins exactly that decision and takes its family fixtures from CLI_APPROVAL_FAMILIES so it tracks the catalog — the first generic family that is destructive and seeds approval ON, and the first non-danger generic family that seeds ON (the first danger row in the catalog is "Project / group delete", so the danger case's real role is alertdialog, never dialog). It mocks only ipc and framer-motion: the full ipc transport returning an EMPTY override list, because the overrides store's load() reads CLI_APPROVAL_OVERRIDE_LIST at mount and calls set({ overrides: result.data }), so an undefined payload crashes on .length; and framer-motion, so the portal dialog mounts synchronously in jsdom. ConfirmDialog, DialogShell and i18n are the real ones, which is what makes a green run prove the gate rather than a stub.

One switch for agent recordings

"Start a screen recording" renders no generic approval row — screen-recording-start is left out of the generic list so this screen carries ONE honest Auto-approve agent recording requests switch. The trap it closes: a second generic switch could un-gate the kind through the family key (approval reads that key directly) while the dedicated row kept reading only the grant store and still showed OFF — the person read "auto-approve: off" while an agent's recordings started with no card. That one switch tells the truth about both arming roads, and switching it back off restores the family key AND revokes the grant on the same click.

Restores from a backup — only you can approve

A restore (data.import, backup.restore, backup.mirror_restore, portable-backup.import, portable-backup.cloud_restore) replaces your local data and brings the backup's settings with it. Only you can approve one, from any surface you have authorized — the desktop window, or a paired phone or computer. A chat is never offered it and it has no place in a bulk approve. A restore that restarts the app (backup.restore, backup.mirror_restore) needs the settings switch on for the device approving it. Restoring from Settings yourself is unchanged.

Erasing data, rolling back and deleting a project — never from a chat

A chat is never offered privacy.erase-subject, privacy.erase-all, app.rollback or project.delete, and it cannot answer one, not even to deny it. The chat says the request is waiting, and you answer it in the app. Everywhere else these work as they did before.

Learning Loop cards — always asked, one at a time

The Learning Loop raises four kinds of card. They come from Learning Loop rather than from an outside AI, and none of them has a family switch, an always-allow arrow, or a place in a bulk approve: each one is a single change only you can decide.

  • A memory held for review (learning.memory-held) — a memory that may contain a password, a key, login details or planted instructions. The card shows why it was held, the title, the scope, and the text that will be saved, with the sensitive parts already hidden. Approving saves exactly that hidden-parts copy; rejecting saves nothing. Its target is the session the memory came from.
  • A skill change (learning.skill-apply) — a skill for all projects, a change to a skill you wrote, or a change the checks flagged. The card shows the skill, where it applies, why it waits, and the change as a before-and-after view (or the whole new text when the current version is too long to compare). For a new skill meant for all projects, Use in this project only narrows it before you approve. If the skill changed after the card was raised, approving changes nothing and says so. Its target is the session the lesson came from.
  • Undo and restore (learning.undo, learning.restore) — an agent asked over the control server to undo, or bring back, one entry on the learning timeline. The card names the entry; it has no session target.

Approving one of them a second time — a retry after a slow approve — never applies it twice: a held memory with the same content is saved once, and a skill change whose text is already in place counts as done.

How to use it

  1. An external AI tells you it queued something. A typical chat reply: "I asked Omniscio to set theme to dark — open the inbox and approve to apply." The AI got HTTP 202 from the CLI server, which means "queued, not applied". The response carries disposition: "awaiting-user-approval", requiresUserAction: true, the approval link, and an instruction for the AI to end its current turn without retrying. Your approval can reach the AI only after that turn ends. Retrying while the card is pending returns the same stop disposition and link; once the row is approved or rejected, completed replays omit those fields. A plugin action is the one kind whose approved replay carries a result: the answer the plugin itself gave for that call, which is handed back only to the session that queued it (a call still being dispatched says so, and a declined call answers its refusal).

  2. Open Omniscio's inbox. Click the Inbox button in the sidebar (or press your inbox shortcut). The pending row appears with the title "<preview>" — for example, "Pause session: My Claude Project". Settings changes get a humanized, change-aware headline rather than the raw key + JSON, naming the setting by its real Settings-screen label (not a title-cased key): "Set Theme to dark", "Set Quick Launch shortcuts (Calendar) to Ctrl+Alt+J", "Turn on Plain Speak" (a toggle reads "Turn on <name>" / "Turn off <name>"), or "Change AI Manager (3 settings)" — never "Turn Keep Db In Private Cache on" (the mangled key) or "Set quickLaunchTabHotkeys: {…}". The row carries a yellow accent badge so it stands apart from chat messages and notifications.

  3. Click the row to open the approval modal. For consent requests the modal's title is just the row's preview text plus a "?" (no "Approve:" prefix): "<preview>?" (e.g. "Archive session "My Claude Project"?", "Run recipe?"); the five informational kinds keep "Acknowledge: <preview>" (no question mark) — cron failure alert, cron failure heal (escalation), session budget warning, session budget exceeded, and Nighty Tidy summary. The title names the specific target, not just the action kind. The body shows:

    • Note (only when the agent supplied one) — if the requesting AI attached an X-AMC-Approval-Note header (see below), a labeled Note closes the approval's details as their last labeled row, in the same bordered card style as the pane's other facts (never a separate callout box), carrying the agent's stated reason — the why behind the request — so the concrete change reads first and the agent's reasoning last. Absent when no note was sent. It is the agent's claimed reason, not Omniscio's verification; the concrete change and the requesting session are still shown below.
    • Details block — the card body is one bordered block of identical rows (icon + label + value). It leads with Type (the kind — "Settings change", "Run recipe", "Archive session", etc.), then the target rows, and then — for the generic kinds — the request's own payload fields (each an icon + label + value row), so the entire card reads as one unified list, closing with the "Generated by <session>" provenance row and, as its own separate labeled field just below it, a "Submitted <date>" row (carrying the full queued date + time — "July 11, 1:48 PM"; the year only when it differs from now) at the very bottom. (The non-settings override kinds — recipe.run, drip.update, Nighty-Tidy run-now — instead keep their bespoke payload layout just below the block.) A Settings change card takes this further — the whole card is field rows (Type · About · Change · "Requested by <session>" · "Submitted <date>", then Note last when the agent supplied one). The created date is its OWN "Submitted" row on every approval card — a separate labeled field, never crammed onto the "Requested by" byline (approval-standard-ui-contract I7).
    • Target context — for session and project actions, that same block continues with rows naming exactly what is acted on: the session (or project) name, its project, and the repo folder on disk — plus the worktree path + branch when the session is isolated — resolved live from the row's target. This is what tells you which session "Archive session?" means, even when the action carries no payload of its own. Spawn approvals likewise show the folder the new session will open in. When the title already names the target — the consent session/project kinds whose preview is a naming sentence (Archive / Pause / Unpause / Snooze a session, Delete / Docs / Bug-intake a project) — the redundant name row is dropped and the block leads with Project + Folder (+ Worktree): the name is shown once (in the title), not twice. Budget alerts keep their Session row, because their title is the generic "Session budget …" and doesn't name the session.
    • Friendly payload view — the heart of the modal. Instead of raw JSON, every field of the request is rendered as a labeled row with a human-readable value: keys are title-cased ("New value", "Project", "Items per release"), ID fields are resolved against the live store to show the real project / session / tag / cron job / recipe / automation / coaching artifact / email rule name (with the truncated id beside it), timestamps are formatted ("May 20, 2026 9:43 AM"), dollar amounts and *Cents fields are formatted as currency, booleans render as "Yes" / "No", string lists join as "a, b, and c", and long strings (>240 chars) truncate with a pointer to the technical details. A cron expression never appears as a raw labeled field — it renders as a single plain-English Schedule row ("9:00 AM daily", or the submitter's own natural-language phrasing), and the raw 0 9 * * * survives only in Show technical details (the redundant cron-natural-language key folds into that one Schedule row). Three kinds get a tailored layout. Settings change computes a leaf-level diff against the prior value Omniscio stashed at queue time and renders only the fields that actually change — a top-level scalar swap reads as "Change: <old> → <new>", a nested-object change (e.g. a tweak inside aiManager) renders a "Will change N field(s):" list with one row per changed leaf path (router.enabled: No → Yes) plus an "(M other fields unchanged)" tail, a patch that exactly matches the current value shows "No effective changes — this patch matches the current value." instead of an empty list, and the rare case where Omniscio can't recover the prior value falls back to showing the raw new value with a hint to open Show technical details. Run recipe lists the recipe, project, optional custom message, and each variable on its own line. Drip edit (a drip.update approval) unwraps the nested { fields } change-set and renders only the changed fields through the same friendly view, so an edited schedule reads as a Schedule row rather than a raw-JSON blob. When an action carries no payload of its own (e.g. archive / pause / unpause), the Target context block above stands in for the field list; the "(no details to display)" placeholder now appears only for the rare kind that has neither a target nor a payload.
    • Friendly payload view — the heart of the modal. Instead of raw JSON, every field of the request is rendered as a labeled row with a human-readable value: keys are title-cased ("New value", "Project", "Cron expression"), ID fields are resolved against the live store to show the real project / session / tag / cron job / recipe / automation / coaching artifact / email rule name (with the truncated id beside it), timestamps are formatted ("May 20, 2026 9:43 AM"), dollar amounts and *Cents fields are formatted as currency, booleans render as "Yes" / "No", string lists join as "a, b, and c", and long strings (>240 chars) truncate with a pointer to the technical details. ID fields resolve to names across eleven entity types now (projects, sessions, tags, cron jobs, recipes, automations, coaching artifacts, email rules, plugins, drips, saved prompts), so a plugin / drip / away-mode card shows a real name instead of a raw id. A few kinds get a tailored layout. Settings change computes a leaf-level diff against the prior value Omniscio stashed at queue time and renders only the fields that actually change — a top-level scalar swap reads as "Change: <old> → <new>", a nested-object change (e.g. a tweak inside aiManager) renders a "Will change N field(s):" list with one row per changed leaf path (router.enabled: No → Yes) plus an "(M other fields unchanged)" tail, a patch that exactly matches the current value shows "No effective changes — this patch matches the current value." instead of an empty list, and the rare case where Omniscio can't recover the prior value falls back to showing the raw new value with a hint to open Show technical details. Run recipe lists the recipe, project, optional custom message, and each variable on its own line. Run Nighty Tidy audit (nighty_tidy.run_now / nighty_tidy_2.run_now) leads with the Project + Folder target block, then shows the human Audit name and a plain-English Mode line (read-only reassurance vs read-write "can edit files to apply fixes") — no "Audit Slug" jargon, no doubled slug, no raw project id, and no raw-JSON drawer. When an action carries no payload of its own (e.g. archive / pause / unpause), the Target context block above stands in for the field list; the "(no details to display)" placeholder now appears only for the rare kind that has neither a target nor a payload.
    • Show technical details — a collapsed <details> section at the bottom. Expand it to see the raw pretty-printed JSON request body exactly as the server received it. It is hidden entirely when there is nothing to show — a no-arg action (archive / pause / unpause) carries an empty {}, and a delete carries only the project id the card already names, so the drawer would reveal nothing but noise. (Payloads over 50,000 characters render a stub message instead of inlining, to protect the renderer — sized for the realistic worst case where a settings.patch carries both the new value and the prior value side by side, e.g. an aiManager twin at ~45kB.)
    • Approve / Reject buttons — green and red, with the small arrow described above attached to the right of Approve (a kind that can't be always-allowed shows a plain Approve with no arrow). Approve always approves just this item; on a phone the footer wraps onto a second row when it needs to. For the informational (acknowledge) kinds the labels read Acknowledge and Dismiss instead.
    • Keyboard: with the pane open, Enter approves (just this item — never a standing approval) and X rejects-immediately — the keyboard equivalent of the green / red buttons (a muted "Enter … · X …" hint sits in the footer on desktop, and hovering a button shows its key as a key-cap right on the button). Typing in a text field never triggers them, and holding a key can't mass-approve the queue. These two keys are fixed (not rebindable).
    • Esc closes the modal without changing the row's state (you can come back later)
  4. Approve or reject. Approve applies the change immediately — for settings, you'll see the UI react (theme flips, volume adjusts); for session lifecycle, the session's status badge changes. When approving a settings change empties your inbox, Omniscio then takes you straight to that setting in Settings — scrolled into view and briefly highlighted — plus a one-tap "How do I use it?" link that opens the in-app help agent for that feature (shown only when the help agent is enabled; it opens nothing until you click it). When other items still wait in your inbox it keeps advancing through them instead, so a triage burst is never interrupted. Single-source-of-truth: invariant 15 in settings-patch-approval-contract.md. Reject dismisses the request in a single click (the informational "Acknowledge" kinds show a Dismiss button instead). Either way the external AI's request is permanently resolved.

  5. The row disappears from the inbox. Both approve and reject remove the row from the pending list. You can view your full history in the queue by listing rows via the CLI server (see cli-control.md) or by checking the Omniscio main process logs.

    If the action can't reach the backend (e.g. a transport stall under heavy load), it does NOT silently do nothing. The optimistic removal is rolled back, the row returns to the inbox, and a "Couldn't approve — it's back in your inbox to retry" toast appears with a Retry button. A stuck call self-heals (a retry is safe — the backend applies each approval at most once) rather than leaving a button that does nothing until you reload. The reject request is bounded by a 60-second timeout; the approve request uses a LONGER bound that exceeds the backend's per-dispatch deadline, because approving can kick off a slow PAID action (a session / recipe / master-debt spawn) that legitimately runs for minutes — a shorter bound would falsely report "back in your inbox to retry" while the backend is still working and the row is not actually back. Implementation invariants: inbox-approval-load-resilience-contract.md I7 (the bound self-heals a stuck call) and I9 (the approve bound exceeds the backend dispatch deadline).

    A card that could never succeed is never put in front of you. Two kinds used to come back with that "retry" toast every time, because no retry could ever work, and neither is raised any more. An agent asking to pause a cloud session the server controls is refused straight away and told to use Stop instead, and a pause card already waiting closes with that reason when you approve it. A "Remove the workspace" card is not raised while a session is still working in that folder. (If a session starts in the folder after the card was raised, approving it still puts the card back for later, so the removal can happen once that session ends.) Invariants: cloud-session-lifecycle-authority-contract.md a-controller-session-is-stopped-never-paused and worktree-clear-unreachable-workspace-contract.md a-live-session-holds-the-folder.

  • cli-control.md — the CLI server itself, token handling, read-only endpoints
  • create-cron-job-with-ai.md — sister AI flow with the same user-in-the-loop guarantee but a different storage path: cron job creates ride on cron_jobs.approval_status='pending' (migration v88) and surface through the Cron Jobs pane, not this queue
  • omniscio-control skill — the external-AI side of the contract: how an AI queues these actions
  • learning-loop.md — the background reviewer behind the four Learning Loop cards

This page is split across two parts: part 2 covers what happens after you decide — how the session that asked is told, what the queue does when a card is never answered, which settings can never be patched from outside, and the rules that govern the note attached to a request.

Last verified 2026-10-02