Get Help (Helpdesk)
Omniscio's in-app help assistant: ask anything and get an answer drawn from its own documentation with clickable source chips, escalate to a human when that is not enough, and follow the conversation from question to resolution.
On this page
What it is
The Get Help panel is a conversation-style assistant inside Agent Mission Control. Ask anything — questions about the app are answered using its own help documentation with clickable source chips linking to the exact help page, while general questions (writing a follow-up email, explaining a concept) get a helpful answer without any source citations. The model self-identifies which kind of answer it's giving.
Get Help opens as a small floating chat widget in the corner of the screen — it's non-modal, so the rest of the app stays visible and interactive behind it, and it never takes over the whole window unless you ask it to. On desktop it's anchored to the bottom-left corner by default (escalating here is how anyone reaches a person), a move button in the widget's header flips it to the other side of the screen (left ↔ right — the choice persists across restarts), and a full-screen button expands it into a console that fills the app below the title bar (so Omniscio's own window controls stay put): your conversation list on the left (admins get a Conversations / Needs-your-reply switch) and the open chat beside it. For an admin with a thread open, a details rail fills a third column on the right — the sender's contact card (avatar — their Google profile photo when they signed in with Google, initials otherwise — name, email, plan), then the conversation's private tags (add, colour and remove them right there — they sit at the top so a conversation can be tagged without scrolling), then the same user's other conversations (click one to switch to it), who it is assigned to, the conversation's status and activity times, the sender's local time / time zone / language and their app version + OS (when the escalating app stamped them; older threads fall back to the "System info" note in the thread), which admins have replied, and — at the foot — the admin's private notes. The two outer columns are yours to size: drag the divider beside the conversation list or the one beside the details rail and each keeps its own width across restarts (double-click a divider to put that column back to its default), or focus a divider and use the arrow keys. The chat in the middle always takes whatever space is left. Who the conversation is with and its status also show at the top of the chat itself. While the console is full-screen the "Help Desk Console" row in the app's far-left navigation is highlighted to show that's where you are — and clicking anything else there (a project, the Inbox, Settings) simply closes the Help Desk and takes you to that page, so you're never stuck on the Help page when you try to navigate away. The console never shrinks into the small corner widget: it has no exit-full-screen button, and ✕ or Esc closes it. Inside, a small header has: Chat (the default view — a composer ready to type into), Help Center (browse the help articles), Your reports (the reports you have sent and support's replies to them — see Report conversations), Insights (a small analytics dashboard — see below), and — for developers only — Needs your reply (the escalation queue, with a count badge; each waiting row shows the question, the reporter's name and email when they were signed in, and the date and time it arrived). Your Get Help is a single ongoing conversation — there is no History list and no "New conversation" button; whenever you open Get Help you land back in that one continuous conversation and anything you type is a follow-up in it. A developer thread in Needs your reply opens right in the chat body — while it loads (and while a first question is being answered) the chat shows just your message with a small typing indicator. Messages render through the same chat rendering a regular Omniscio conversation uses — the same bubble shapes, the same markdown engine — with your own on the right and the assistant or developer on the left. They look like Team Chat: the first message from each sender carries a small picture, the sender's name and the time inside the bubble (your profile photo when you're signed in, a friendly bot mark for AI answers, initials otherwise), your own messages show just the time, and follow-on messages from the same sender are plain bubbles. AI answers wear Team Chat's green agent outline, so they never look like a person's reply. In the developer console a customer who signed in with Google shows their Google profile photo — on their row in the queue, on their contact card and on their messages; a customer who wrote in by email or was not signed in keeps the initials circle. That includes tickets opened from the bug-report form: the photo shows when the sender's own app proves who they are, and tickets that existed before this was switched on were linked to their sender's account once. A report another computer opened on the sender's behalf keeps the initials rather than risk showing the wrong person. Only an admin's console can look the photo up, and it only ever loads a Google-hosted picture.
You can send follow-up messages within a conversation — the AI answers with the full thread context. If the AI answer didn't help, a "Send to developer" button escalates the conversation. For a developer (with the dev-console gate enabled), escalated questions from users show up in the "Needs your reply" view, switched to via its own header icon. Opening a thread there shows it in the same chat body (the developer's own replies sit on the right, like any normal chat). The reply arrives as a toolbar badge + OS notification and appears in the chat view within a few seconds — while a conversation is waiting on the developer the app checks for replies on a fast cadence, and opening the Get Help panel checks immediately.
Where to find it
On a phone
On the phone web client the Help Desk is an ordinary screen, just like every other integration — it never covers the whole display. Tap Help Desk in the projects list (or the toolbar's Help Desk button) and it opens under the app's usual navigation and header:
- Two tabs along the top switch between your Conversation (or the Queue, for a support operator, with the number of people waiting) and the Help Center. Your reports (the reports you have sent and support's replies to them — see Report conversations), Insights (and Routing, for an operator) sit beside them as buttons. Once a support operator opens a ticket those tabs step aside, so the ticket's own row is the only bar above the conversation.
- Your phone's Back steps out one level at a time — from an open conversation back to the queue, from the Help Center back to the conversation, then out of the Help Desk.
- With a ticket open there is one bar, not two. It carries Back, the person's name, how they reached you, the subject and the assign button, plus a ⋯ menu with Start session, Open session (once a session is working the ticket) and Mark resolved — so the conversation keeps the great majority of the screen.
- The reply box keeps just Reply and Note; Shortcuts, Knowledge Base, AI reply and the Formatting (Aa) toggle move behind the ⋯ button in the box's own toolbar.
- Tapping a Help Desk notification opens that conversation directly.
How to open it
- Press Ctrl+Shift+H — the Get Help widget pops in at the bottom-left of your screen (bottom-right if you've moved it via the header's move button).
- Click the LifeBuoy icon in the toolbar (appears when the feature is enabled), or find it in the overflow menu.
- Click Help Desk in the projects sidebar — an ordinary hub row alongside the other integrations, not a fixed slot of its own. The row is always there while Get Help is enabled, highlights while the panel is open, and carries the same unread count as the toolbar button. (A separate Help Desk Console row sits right beside it for a seat that holds the admin console — see part 2.)
Note: Get Help is on by default for everyone — new installs and existing ones alike (an existing install that had it off is switched on once, when it updates). To turn it off, flip its toggle in Settings → Lab; once you turn it off it stays off. The separate Help Desk Console stays off by default and is for admins only.
How it behaves
Full lifecycle
User asks a question
→ AI answer (Omniscio-scope with source chips, or general-scope without)
→ The answer shows in your single ongoing conversation
→ Rate 👍 or 👎 — feedback only; rating never changes the conversation's status
User sends follow-up
→ AI answers with thread context (up to 20 prior messages)
→ Scope may widen general → Omniscio (never narrows)
If the AI answer didn't help (Omniscio-scope only):
→ User clicks "Send to developer"
→ Status: waiting_dev
→ Escalation email sent via resend-service, resolved server-side; its subject
carries a snippet of the question so each alert is its own email conversation
(temporarily routed to a single developer inbox;
normally the account owner plus each admin account)
→ Subsequent messages in the chat go to the developer, not AI
Developer (on their machine, gated "Needs your reply" view):
→ New question (or a user follow-up re-queuing a thread) raises a desktop
notification with a chime, even when Omniscio is in the background —
clicking it brings the app forward and opens the waiting thread
(plus an in-app inbox alert and a phone push if a phone is paired)
→ Opening that in-app inbox row lands on the console showing **that** thread.
An explicit "open this question" outranks the console's load-on-open habit of
resuming the developer's own most-recent conversation — needed because those
two share one selection slot, so the resume would otherwise discard the thread
the row pointed at (and on a single-user install the developer IS a user, so
they have a most-recent conversation to resume)
→ Opens Get Help widget → clicks the "Needs your reply" header icon → lists waiting questions, load-on-open + Refresh
→ Opens thread → reads question + AI attempt → types reply → sends
→ Status: answered (in cloud); thread refreshes in list
→ When the exchange is done, clicks "Mark resolved" — or presses **E** — to
close it, and the console moves straight on to the next conversation in the
queue (the one below, else the one above; none left → an empty console) — and
"Mark unresolved" reopens a resolved thread, landing it where the last
speaker left it (customer spoke last → back in the Pending queue;
developer spoke last → Awaiting customer). Typing a reply
into a resolved thread reopens it automatically, so a follow-up never
lands on a closed conversation.
User receives reply:
→ Push listener refreshes the chat view + conversation list in real time
→ Desktop notification with a chime: "Developer replied to your question" —
clicking it brings the app forward and opens the replied conversation
(a phone push is sent too if a phone is paired)
→ Inbox notification: "Developer replied to your question"
→ Developer reply shown as accent-tinted bubble in chat
→ Reading the reply (here, on your phone, or by opening the ticket
anywhere) clears that inbox row everywhere, so no surface keeps
showing "an agent replied" for a reply you have already read
User sends follow-up on escalated thread:
→ Status flips back to: waiting_dev (in cloud AND locally)
→ Thread reappears in the developer's "Needs your reply" view on next open/refresh
→ Developer follow-up alert email sent
Developer resolves (or reopens):
→ "Mark resolved" (or **E**) closes the thread and opens the next conversation
in the queue; "Mark unresolved" reopens it to wherever the last speaker left
it (waiting_dev or answered)
→ **E** is the console's resolve key: it works on the conversation that is open,
never while you are typing in a box (or in the moment after), never with
Ctrl/Alt/Shift/Cmd held, and never inside a menu or dialog. Hovering Mark
resolved shows the key. While the console hub is open, single-letter
keys never reach the app behind it — E there can never archive a session
→ A hairline divider marks the resolve/reopen in the developer console only;
the user's chat shows a single muted "Marked Resolved" note while resolved
→ In the console, a resolved conversation's row (the queue, the Resolved view,
the person's other conversations, and search) previews its question, not
the resolve note or a closing line; an open one previews its latest message
→ The user's local copy flips to resolved too; their next message on it goes
back to the AI (and they can re-escalate afresh)
→ Resolution is a developer action now — the user no longer has a resolve button
No AI is involved after escalation — developer messages come from the developer console only (or the in-memory fake seam in tests).
Who sees which developer replied. Every developer reply is stamped server-side with the replying admin's identity (from their verified sign-in — a client can't forge it). The user sees the developer's first name only — a reply reads "Alex" instead of the generic "Agent" label, which is friendlier without identifying the teammate: the relay that delivers replies to users strips the full identity (email, account id) server-side and passes through just the first name, on both the push and the poll delivery paths. The full attribution is admin-console-only: in the "Needs your reply" thread view, your own replies are labeled "You" and another admin's replies are labeled with their email, so a team of developers can see who answered what. A reply an agent session or the desk itself sent through your computer is labeled from this computer’s own record of what it sent instead: Agent · <session name> (click it to open that session) or Automatic reply (see Starting a session from a Help Desk ticket). Replies sent before this feature carry no name and render as "Agent".
Conversations are private to your account. Your Get Help history belongs to the account that asked: each conversation is tagged with the signed-in account that created it, and every list and lookup returns only your own. A second person signing into the same computer sees their own (empty) history, not yours — and switching accounts also clears the developer-side incoming-alert cache on that machine. Conversations started while signed out belong to the signed-out state only: they show while you are signed out and are not carried into any account when someone signs in — signing back out shows them again. If a send to the agent fails (for example, offline), the message is not kept locally — retrying sends it once, never as duplicates.
Resolve/reopen markers. When the developer marks your conversation resolved (or reopens it), your copy updates quietly: the conversation's status flips, with no desktop notification or chime. While a conversation is resolved, a single muted "Marked Resolved" note shows at the end of your chat; reopening removes it.
You are also told in words, once: a short note on the ticket that greets you by name, names the ticket number and what it was about, and says whether it is closed or back open. When whoever closes the ticket says what was actually done, that sentence is in the note too, right under the opening. Reopening and closing the same ticket again sends a fresh note, because it is a new closure; a second close of an already-closed ticket, a retry, or a second door firing does not, so you never get the same closure twice. The per-event resolve/reopen hairline dividers are developer-console-only — the back-and-forth of resolving and reopening is support-side bookkeeping and is deliberately not replayed in your chat. Once a conversation is resolved, typing a new message hands it back to the AI — you get an instant AI answer again, and you can send it to the developer afresh if you still need a human.
Who gets told when a reporter is waiting
A ticket that needs a person does not sit silently in a queue. Once it is waiting, the crew that owns the Help Desk is told, and the message goes to whoever can act on that ticket right then — not a name picked when the ticket arrived. It names the ticket, the reporter and how long they have waited.
Who hears first. The session that owns the ticket (the one working its task) hears first, as long as it can take a message now. If it has finished, been paused or archived, crashed, or there is no owner, the crew's Foreman hears instead, and the crew's lead only when no Foreman can be reached. Each message says why that session got it. If a delivery is refused, the next one in line gets it at once. The crew having no lead no longer silences everyone: the "no live lead" card appears only when nobody at all can be reached.
Work whose owner has gone goes to the Foreman. When fix work belongs to a session that has ended, the Foreman gets one message listing those tickets and asking for each to be given a live owner or closed, instead of a card sitting in your inbox. It is sent again only when the list changes.
The promise list is the ledger. Every ticket row in the Help Desk ledger shows the date we promised the reporter and when we last told them something, so who is waiting and since when comes from the app rather than from an agent's notes.
One message per reporter message. A new message from the reporter gets one message to whoever is in line for that ticket. Replaying the same ticket, reading it again, or any other activity on it gets nothing, because there is no new message to tell anyone about. If nobody answers, one short "still unanswered" reminder follows every fifteen minutes, and it names the same wait.
Answering stops it, for good. The moment a person's reply is the newest thing on the ticket, nobody hears about it any more — the reminder included. It does not depend on the ticket's status changing, on the console being open, or on anyone marking it resolved. A ticket answered by one person is never announced to another as unanswered, however many times the list behind it is refreshed.
A receipt is not an answer. The automatic "thanks, we've logged this" note goes out on every ticket, and it does not count: no person has spoken yet, so a ticket whose only desk message is that receipt keeps being announced until a person actually replies.
A chase shows up straight away. If you write again on a ticket that was already answered, the ticket goes back to owing you a reply the moment your message arrives — not whenever the next background check happens to get to it. The console shows it as waiting on the desk again, and the desk's own to-do list carries it, even when a hundred other tickets are also in flight.
If the crew has no live lead, the desk does not page anybody — the waiting tickets stay in the console queue and in the Inbox, and one card says the crew has no lead. On a computer with no Help Desk crew at all, nothing is sent.
How answers work
When you submit a question, the app:
- Searches its bundled help documentation (
docs/llm-library/) for the most relevant pages. - Sends your question — along with those pages as context — to an AI model
(
source: 'helpdesk-ask'; a singlellmProviderService.chat()call, no session spawn). - Returns a plain-language answer together with source chips: small tags naming the exact help pages the answer was drawn from. Click any chip to open that help page in the Help & Docs panel (chips are hidden if the page cannot be resolved).
If your question is about Omniscio but not covered by the docs, the answer will say so
honestly rather than guessing (covered: false, empty source list — a
"couldn't find it" answer never shows source chips). If your question is general
(not about the app), the assistant answers it as a helpful AI — no source chips,
scope: 'general'.
What the AI answer path needs
The model call pins provider: 'openrouter', whose baked company key is
retired — the call is developer-paid through the cloud gateway, which needs
the Omniscio cloud (global-auth) sign-in. Without a gateway session the provider
layer degrades to its Anthropic backstop, which needs an Anthropic API-key
account; a subscription-only (OAuth) Claude sign-in cannot use the Messages
API directly. When BOTH of those are missing but a healthy Claude
subscription login is signed into Omniscio, the assistant still answers: it runs
the question through a one-shot Claude Code print call on the user's own plan
(the same subscription bridge the Council and Writer features use) — no API
key, no extra dollars, slightly slower than the direct API. Only when there is
no gateway session, no API-key account, AND no subscription login does the
assistant reply with a friendly "can't reach its AI service" message (never
the raw AnthropicFallbackUnavailableError). That error area still helps two
ways, because neither needs an AI credential: a "Send to developer" button
(one click creates the conversation from your question and sends it through
the helpdesk relay) and, when the local doc search found candidates, a row of
clickable help-page links ("these help pages might have the answer") that
open the Help & Docs panel at the matching section. In a dev checkout the missing-gateway case most often means
default-credentials.local.json was never materialized
(npm run secrets:pull), so the global-auth config that enables the gateway
sign-in is missing.
Things you can do with a conversation
- Copy an answer. Every AI answer and every developer reply has a small copy button — click it to put the full message text on your clipboard; it briefly shows a "Copied" confirmation.
- Rate — and change your mind. The 👍 / 👎 rating is reversible: your chosen rating stays visibly pressed, and you can switch it or clear it at any time. Rating is feedback only — it no longer resolves the conversation (whether a support conversation is resolved is the developer's call).
- Click a source chip. Source chips always open the Help & Docs panel at the exact page the answer drew from (opening the panel if it's closed). A chip is hidden whenever its page can't be resolved — including when the cited doc has no page on the public help site — so you never get a dead click.
- One ongoing conversation. Get Help keeps a single continuous conversation for you — there's no list of past conversations to browse, no "New conversation" button, and no per-conversation delete or history search on the user side. Opening Get Help always returns you to that one conversation, and anything you type continues it. (Developers still get the full queue browse, search, and filters below.)
- Search & filter the queue (developers). The developer console's queue has its own search box beside the queue-view menu. The menu offers All, Unread, Pending, Unresolved, Resolved, Most Recent, and Longest Waiting views. Every view shows one row per person (their newest conversation) EXCEPT Resolved, which is the history of past conversations: every resolved conversation gets its own row, newest first, even when the same person has an open one. Search matches the question, the person's name or email, the row's preview, and your private tags. One clickable chip appears per tag in use; pick one or more to narrow the current view (chips combine with search). Tags and their filters are private to the console and never leave the developer's machine.
- Drafts survive a failed send. If a message fails to send (yours or, for developers, a reply), your typed text stays in the composer so you can retry without re-typing — the retry re-sends that exact message. Pressing Esc while the composer still has text just moves focus out of it rather than closing the whole panel, so you can't lose a draft by accident.
What gets attached when you escalate
When you send a conversation to the developer, Omniscio automatically attaches a short
system-info line — your Omniscio version and operating system, e.g.
Omniscio v1.2.3 · win32 10.0.19045 — to both the alert email and the conversation
the developer sees, so they can triage without asking you what you're running.
Escalation is only offered for questions about Omniscio. On a general-knowledge conversation the panel shows a short caption explaining that "Send to developer" is for app questions only.
What a person on our side can and can't do for your account
"Can't you just fix it on my account?" is the most common question after an escalation, and the honest answer has two halves.
What a support person can see and do — all of it cloud-side and all of it audited:
- Read the conversation you escalated, including its messages and any file you attached to it.
- Reply to it, resolve or reopen the ticket, and see when a reply has reached you.
- Look up your account: its role, plan tier, add-ons, and whether it is disabled or still on an invite — that is what lets them answer "why can't I get in" without guessing.
- Change those account-level settings (a plan tier, an add-on, a disabled account), and issue credit grants. These are deliberate, confirmed actions, never silent adjustments.
- Suspend, restore or take back a public agent address you own, if it is being abused.
What nobody on our side can do:
- Open your sessions, projects, messages, files or vault. Those live on your computer. They reach us only when you attach them — an escalated conversation, a diagnostic bundle, or a file you send — and that is the whole picture we can see.
- Drive your app. There is no remote control. Any change to your machine happens because you clicked it.
- Read your data without leaving a record. Every operator read of a customer's support or helpdesk data writes an audit row naming who read what and when, in both the desktop and the cloud consoles, and the record is reviewable rather than an in-app courtesy.
So the rule of thumb: if the fix is inside your machine (a setting, a stuck session, a local file), an agent in a session or the steps on this page do it with you. If the fix is inside your account (sign-in, plan, credits, access), support can do it directly. If you are not sure which, say what you are seeing in the escalation — whoever picks it up can tell you, and will ask for the diagnostic bundle when the answer has to come from your machine.
Rate limit
The panel is limited to 10 questions per minute. If you reach the limit, an inline message shows a live countdown of the seconds remaining, ticking down until you can ask again — so you're never left guessing when the panel will accept another question. The countdown recovers correctly even if your computer's clock jumps backward.
Insights
The Insights view (the bar-chart icon in the header) is a small, read-only dashboard summarising how Get Help is being used on this computer. It shows:
- Questions — how many questions were asked in the selected window.
- Handled by AI — the share the AI answered without ever escalating to a human (the self-service deflection rate), with how many still needed a person.
- Satisfaction — of the answers you rated 👍/👎, the share that were 👍 (with how many of the total were rated).
- Questions over time — a simple trend line of daily volume.
A 30 days / 90 days / All time switch at the top scopes every figure. Until a few questions have been asked, it shows a friendly empty state.
Insights is entirely local and free: it only reads the questions already stored on your machine — it makes no AI call, sends nothing to the cloud, and costs nothing. It shows only your own questions, so (unlike "Needs your reply") it needs no developer/admin access — anyone who can see Get Help can open it.
Phase 1 note. Insights currently summarises just this install's own questions. A future phase may roll the numbers up across all users of a deployment; that needs additional cloud plumbing and is tracked separately.
The "How did we do?" survey after a ticket is closed
When someone on the team marks your ticket resolved, the "marked resolved" note you get may carry one more line asking how we did, on a scale of 1 (poor) to 5 (great).
- In the app the note shows five small number buttons under it. Tap one and you are done: the ticket stays closed and nobody is alerted. The buttons stay for 14 days.
- By email the note asks you to reply with just a number. A reply that is only a
number (
4,4/5,4 out of 5or5 stars) is recorded as your rating and the ticket stays closed. Anything with words in it reopens the ticket exactly as before, so "4, but the login still fails" reaches a person. - A number only counts as a rating while the survey is the last thing we sent. Once the conversation has moved on (say we asked "which step failed, 1, 2 or 3?"), a bare number is an ordinary reply that goes to the team.
- We try not to pester. You are never asked if you opened three or more tickets in the last 30 days, you are asked at most once every 30 days, and a ticket is never asked twice. Tickets closed automatically for going quiet, declined tickets and spam are never surveyed.
- Only the first answer counts. A later rating for the same ticket is ignored.
For the team, the results show in the Help Desk console's Dashboard: the period's customer score
in This period at a glance, and the all-time Resolution surveys (surveys sent, answered, the
response rate and the average score) under Your operator activity at its foot.
The switch is Help Desk resolution survey in Settings → Lab
(helpdeskResolutionSurveyEnabled, on by default). Turning it off stops new questions; a
rating sent back for a question already asked is still recorded, so it never reopens a
ticket.
Feature gates
There are two separate gates, both using the standard unreleased-feature registry:
| Gate | env var | settings key | Purpose |
|---|---|---|---|
helpdesk |
AMC_SHOW_HELPDESK=1 |
helpdeskEnabled |
Entire Get Help feature (user-facing) — defaultOn, ON unless switched off |
helpdesk-dev-console |
AMC_SHOW_HELPDESK_DEV_CONSOLE=1 |
helpdeskDevConsoleEnabled |
Adds the "Needs your reply" header icon + view to the widget (developer only) |
The helpdesk-dev-console gate stays 'in-development' even after helpdesk
ships — ordinary installs never see the developer console. Both gates are checked
in Main (not just in the renderer), so hiding them client-side is not the only
barrier. Access is now server-enforced, not just client-gated: every
developer-console read/write routes through the admin-authed operatorConsole
Cloud Function (verified Firebase ID token + requireAdmin from signed claims),
so the shipped Firebase admin service account key is retired from the helpdesk
operator surface.
Related
- helpdesk-ticket-ledger.md — the desk's own record of every ticket: its priority, the reporter's wait, what was actually delivered, and the bulk doors that triage hundreds of tickets in one call
- keyboard-shortcuts.md — full list of Omniscio keyboard shortcuts, including Ctrl+Shift+H
- ask-amc.md — a separate feature that lets you have a full back-and-forth chat with Claude about the app, with the ability to configure settings for you
- Omniscio’s unreleased-feature (“Lab”) gate — how Omniscio hides in-development features until they ship
- report-conversations.md — every report sent from the Feedback button opens a conversation the support team answers here; the reporter reads the replies under "Your reports", even with Get Help off
Last verified 2026-11-17