App messages in the transcript (readable system cards)
Every message the app or an automation sends into an agent's session (an Inbox decision, an Auto-lander hand-back, a Gate watch or Land watch result, a scheduled wake-up, a board post, an Overseer's guidance) is written for the agent. It renders as a short card that says what happened in plain words, never as plain text and never under your own icon, with the original one click away and the agent's copy untouched.
On this page
What it is
Omniscio sends messages into an agent's session on its own behalf: your Inbox approve or decline reaching the agent that asked, the Auto-lander handing a branch back, a Gate watch or Land watch reporting a result, a scheduled wake-up firing, a post landing on an agent board, your Overseer passing guidance on. Each one names its sender once, in the card's header, and none of them is ever drawn as if you had typed it.
Those messages are written for the agent, so they read like instructions to a machine: ids, hashes, numbered steps, reminders about archiving. They now render as a card that tells a person what happened at a glance, the same way an agent's test result already does.
Where to find it
There is nothing to open. A card replaces the message in place, inside the agent's transcript in the session panel, on desktop and on the phone.
How it behaves
What you see
Every card is drawn by one shared card layout, from the same five parts, top to bottom:
- Header — one small icon, the sender's name and the time, and nothing else, at the same size and colour as an ordinary message's header. The icon says which sender it is and is the same on every card from that sender, whatever the result; it takes the theme's accent colour, never the result's. The name is never cut short while there is room. When the message waited in the queue for 2 minutes or more, the wait hides behind the time: hover the time to read "Waited 6m before delivery", or click or tap it to show "9:11 AM · waited 6m". The message's ⋯ menu has its own space at the end of the header, so it never covers the time. A sender that is a session, a scheduled job or another computer is a link.
- Headline — what happened, as one whole sentence. When the card carries a result, the headline is large, in the result's colour, with the result's own icon in front of it, and the whole card is softly tinted that colour. A card with no result keeps a plain headline and no tint.
- Summary — the next sentence or two, never a repeat of the headline, in ordinary message text. A message from another agent is written exactly like an ordinary message too — the same size, weight and colour.
- Facts — one line of small plain text, led by what the card is about (a review's subject, who sent a check-in, which computer another agent wrote from), then the branch, a count, how long it took, or a short list of rows such as each request in an Inbox decision. A fact with more behind it is a button: on a test result, the failure count opens the failing tests and the time opens where the time went. Both start closed.
- Footer — at most one action ("Read full report", "Show details", "Open findings", "Fix it for me"), the "N more messages waiting" chip when there is one, and a small Original link.
Each thing is said once: there is no separate "injected by" line, no second logo, no "TL;DR" label and no queued caption. No sentence is cut with "…" at a character count — a long one wraps to two lines, and text is only clipped when "Read full report" or "Show details" is there to open the rest. The colour follows the app's usual meanings: green is done or passed, red is failed or refused, orange is waiting or held up, amber means a person has to act. The card takes every colour from the theme you are on, so it recolours itself when you switch theme.
| Sender | Example headlines |
|---|---|
| Approvals (your inbox approve / decline decisions) | Headed Approvals with a stamp icon, never "Inbox". One request: "Request approved" with a large green check, "Approved, but it didn't go through" or "Approved, but nothing was started" (amber), or a plain "Request declined" / "Request expired before anyone answered" — then the request itself as one plain line and any reason on the small line below, with no second mark. Several requests: "2 requests approved", "4 requests reviewed" or "Update on your requests" (a batch that includes a failed or expired request), each request listed with its own approved, declined, failed or expired mark. The reader accepts every shape the decision notice writes, single or batched, and a batch whose bullets do not add up to its own count falls back to the plain card rather than showing a partial list |
| Auto-lander | "Couldn't land: it conflicts with master" (handed back to the agent to fix), "Landing paused for about 5 minutes", "Couldn't land yet: a safety check is holding it", "Needs a person's sign-off before it can land", "Ready tag is out of date", "Refused to land" |
| Gate watch | "Tests passed" / "Typecheck failed", "No answer, nothing was checked", "Result is out of date", with whether it checked the whole project or the changed files |
| Land watch | "Landed on master", "Still waiting to land", "Still not landed on master after 6 hours" (the stall heads-up, in orange), "Can't land as it is", "Branch is gone", "Stopped watching: it never landed" |
| Worktree Ledger | "Worktree couldn't be created" (with the reason), "Workspace taken over by another session", "Idle workspace with unfinished work" |
| Gate restart recovery | "Tests check stopped by an app restart" |
| Agent board | A Board card: the post's subject as the headline, the author's own summary under it, who posted in the header and how many posts are new as a fact. The commands the agent runs to read them are only in the original |
| Review service | A compact review card: "Plan review" (or code / crew review) and what was reviewed in the header, the verdict in plain words as the headline ("Not approved — 1 blocker"), the most serious finding as the summary, and one small pill per reviewer showing its logo and how many findings it reported (a dashed pill for a reviewer that did not run). Pressing the card or Open findings opens a sheet with the findings grouped blocker, major, minor, each expandable to the reviewer's explanation and suggested fix, the reviewers that did not run with their reasons, and the Original |
| Crew and desk automations | Crew watch, Task tree, Ticket queue, Detached job, PR problem router, Branch sweep, Dev Desk keeper, Help Desk, Crew lead items, Landing desk, Crew firefighting and the other crew automations each get their own card — "Builder #3 has gone quiet", "12 tickets have no owner", "A background job finished", "3 pull requests need attention" — with at most one or two short facts. Ticket lists, commands and ids stay in the original |
| Friction ledger | "This session was handed a recorded problem to fix", with the recorded problem's own title as a fact — so two hand-offs in one session never look alike. The fingerprint, the close-the-cluster steps and the reopen rules stay in the original |
| Every other app sender | Its own headline, icon and colour — "A scheduled wake has stopped working", "A break-glass window is open", "A crew lead has stopped reporting", "A session it started never came up" — with the name, count or reason the notice carried when it has one |
Every sender has its own header icon (a clipboard for a check-in, a ticket for the ticket queue, an hourglass for Land watch, a shield for Gate watch, a stamp for Approvals), so cards can be told apart at a glance, and a sender keeps that one icon on every card. How it went is said by the headline's colour and icon and by the card's tint, never by the header icon.
A crew check-in names the seat once in the header ("Crew check-in · Builder #3", the seat's short name without its crew badge or task) and leads with the member's own report: its first sentence is the headline and the rest of its first paragraph the summary. Read full report opens the whole report, word for word. A refused ready tag says "Not tagged ready to land", names the branch once and says why in plain words — "Something on the branch needs fixing first", "The check couldn't finish, so it can be tried again", "This branch already landed, so there's nothing to tag", "Another session owns this branch", "It needs someone's approval first" or "Another tag already holds this branch". Each reason is said once, what needs fixing comes before a retry, and "It was refused" stands in when the reason is one the app has no words for. A land that has stalled says when it carries no ready tag, so nothing is trying to land it, and a machine-made cloud branch name reads "Cloud work". A crew lead's scheduled wake from its mission loop shows the app's own status as a few short rows with rounded numbers; any other wake-up prompt still shows word for word.
Several messages delivered together are drawn as one card per message, in order.
Several senders carry someone else's words, so their card leads with those words — the first sentence as the headline (a heading or the first list item when the report starts with one, without its formatting marks), the rest of the first paragraph as the summary — and Read full report shows them in full: Scheduled wake (the wake-up's own prompt), Session Refill (the text you configured), Crew check-in and Agent report (the crew member's or parked session's report), Overseer guidance and Overseer correction (your Overseer's words), Command line (a message sent into the session from a terminal) and Inbox alert (the copy of an inbox alert).
Every app sender has its OWN card, and never shows the notice's own first sentence. That holds even when a producer rewords its notice: the card still says what that sender is about, with the original one click away.
The one case that falls back to the notice's own opening sentence is a sender whose card reads specific wording — an Inbox decision, a lander hand-back, a gate watch result, a board post, a relayed message — meeting a notice it no longer recognises. That fallback exists so a reworded notice still reaches you as a card rather than a wall of agent instructions.
A test result's card leads with what happened ("Tests failed", large and red) and its cause, with the counts and how long it took as facts. The card face never lists the failing tests: click the failure count ("2 failing files") to open them, and click the time to see where the time went. When the list of failures was cut short and no total is known, the count reads "At least 2 failing files" rather than claiming an exact number. Show details appears only when there is something neither of those opens. A result that answers an older version of your code says so up front, as its summary.
When a message waited in the queue for 2 minutes or more, the card's header says how long ("waited 6m") beside the time. A plain message of yours that was held and delivered later shows a small muted caption ("queued, delivered later") instead of a raw bracketed stamp. When several messages were waiting behind it, a "N more messages waiting" chip appears; open it to see who sent each one and how it starts.
Messages from another computer or another person's agent
A message from an agent on another of your computers, a teammate's agent, another person's agent, or a Team Chat reply is drawn as a remote agent card: who sent it and where it came from in the header, their own TL;DR as the headline when they wrote one, and their words in full. The reply instructions and ids are one click away in the original. The badge appears only when the message really arrived by that route; text that merely looks like one does not get it. See Cross-device agent messages.
The original is one click away
Every card ends with a small Original link, which reveals the exact text the agent received. Opening it always shows the whole note: the bubble grows to fit it rather than cutting it off. (This also fixed the same toggle on Verdict cards, where an opened note could be cut off.)
Words the app relays from an agent (a crew check-in, a scheduled wake-up, an agent report, a report from a one-off copy, an Overseer's answer, a command-line message) never show the control markers an agent writes for the app rather than for a person, such as a final-message marker. The marker is dropped wherever it sits and every other word stays.
What it does not change
- The agent still reads the original. Nothing about the stored message is rewritten; the card exists only on your screen.
- Only the app's own senders get a card. A message from another agent on this computer, a teammate, a guest, your own broadcast, or anything you typed renders as written. The one exception is a message from another agent that carries its own TL;DR: it uses the same layout, with that TL;DR as the card's text, in the same text styling as an ordinary message, and the full message behind Read full message. Its header shows an agent icon and the sending session's name as a link back to that session. It never gets an app icon, a result tint or a coloured headline, so an agent cannot dress its words up as an app notice.
- A new sender cannot be added without a card. The build fails for any app feature that starts injecting text into sessions without being given one.
- Getting a card changes nothing else. The crew and desk automations above get their cards from a display-only list, so how a crew lead's wake-ups are counted and how queued messages are handed on stay exactly as they were.
- Internal ids stay off the card. Watch ids, commit hashes and ticket numbers are one click away in the original.
- Old batches: a message that several senders' notes were merged into before Omniscio started recording each sender separately keeps its old look.
Related
Test results in an agent's transcript are drawn the same way. How messages between sessions are tagged, and how to hide them per session, is on Cross-session messaging. How an agent's own replies are laid out is on Agent message display.
Last verified 2026-11-15