Assistant¶

The assistant is a person's own conversation partner in the console: a panel that opens from any screen, sees what the person is looking at, and answers from the core's data. This page is for those who work with it and for the administrator who enables it. Rationale: TAI-ADR-0058 (rev. 2).

What it is¶

The assistant panel is one more surface of the person's conversation, not a separate console chat. A person has a single conversation, run by the assistant engine; the console only shows it and passes messages along. The console does not store the conversation and does not write it to its log. The engine is the person's own container (see Personal workspace). The same conversation continues in Telegram: what you write to the bot is visible in the panel, and vice versa. The assistant has no separate web window — the /harness/ address leads to the console with the panel open.

The assistant reads tasks, approvals, runs, processes, and company memory with Control Plane tools and answers from them, not from the conversation's memory. Its permissions are the person's own: the assistant cannot do more than the person can.

Opening the panel¶

How	What happens
The "Assistant" button in the header	The panel opens to the right of the screen; the screen stays in place
⌘J / Ctrl+J	The same from the keyboard; pressing again closes the panel
An address with `?assistant=open`	The console opens with the panel already open; the parameter is removed from the address, so reloading or going "Back" does not open the panel again
"Discuss" on an item in Waiting for you on the Today screen	The panel opens with the context of that item; the screen does not change

Escape closes the panel. The message draft is not lost when the panel closes. The live conversation stream stays open while the panel is open; after reopening, it continues from where it stopped.

Screen context¶

With every message the assistant receives the screen context — so you do not have to name the object in your question: "why is this stuck?" on a work screen refers to that work.

Above the input field there is a "Context" chip with the screen's label. You can remove it — then the context is not sent until the person presses "Send again" or moves to another screen. When the screen changes, the context changes by itself.

The console server collects the context on behalf of the signed-in person, from the same data the screen shows. The browser sends only the screen address; anything else it might send as context is discarded.

Screen	What goes to the assistant
Work	a condensed provenance chain: where the work came from, its runs, its verification
Process instance	open steps, time on the step, the latest decision log entries
Run	the outcome, the latest actions, and the latest checkpoint
Rule, agent, artifact	label and status
Overview screens	only the screen kind and address

The context contains only identifiers, statuses, short labels, and timestamps. It has no run input or output, no process data, no artifact content or address; values that look like tokens and passwords are scrubbed. Of the address parameters only the tab, the status and kind filters, the workspace, and the revision remain — search text is not sent. The context is no larger than 8 KB: lists are shortened first, and in the worst case only the screen kind and address remain. If the data is not collected within 3 seconds, the message goes out with the screen kind and address and nothing else.

Conversation¶

History. The panel shows the latest messages of the conversation; "Show earlier" loads earlier ones. Your messages sent from outside the console are marked with where they came from.
The assistant's turn. While the assistant is answering, the panel shows "The assistant is answering…" and a summary "Read and done: N" — which tools it called and with what outcome.
Queue. A message sent while the assistant is answering is queued and goes out when it finishes the current turn.
Retry. A message that failed to send can be retried: the retry goes out with the same request number and does not duplicate the message.

Confirmations¶

The assistant does not perform an action that changes state in the core — creating a task, accepting a result, deciding an approval, invoking a skill — without confirmation. The panel shows a card "The assistant asks for confirmation" with the action, the waiting deadline, and the Allow and Decline buttons. The same request arrives in Telegram: the first answer counts, a repeated decision is rejected, and the card shows where the decision was made. A request that gets no answer expires; an interrupted turn withdraws its requests.

Waiting for you on the Today screen¶

The personal part of the Today screen is the Waiting for you list: decisions addressed to the person or their role, results under review, approaching and missed deadlines, blocked work, and assignments that keep failing for their executor. The core computes the list (GET /api/v1/me/attention), and every item has a reason.

An item has buttons: "Discuss" (opens the assistant with the item's context) and "No need to show" — feedback to the core; an approval can be decided right in the list. Hidden items can be shown and brought back.

Privacy¶

Only its owner sees the conversation. The console stores neither messages nor answers: the console server's access log has only the request number, without text. The organization sees core objects and the action log (tasks, approvals, artifacts, events), but not the correspondence. Each person talks only to their own assistant: the console server takes the principal from the session, not from the request.

How it works¶

flowchart LR
    B["Browser"] -- "/console/api/assistant/*" --> C["Console server"]
    C -- "Bearer IAM, harness:use" --> L["harness-launcher"]
    L --> H["Person's container"]
    T["Telegram"] --> N["notification-service"]
    N -- "harness:inbound" --> L
    H -- "cp_* on behalf of the person" --> CP["Control Plane"]

The console server obtains for the person a token for the human-harness audience with the harness:use scope through the same federation:exchange as the other tokens, and calls the launcher over the compose network: …/_launcher/internal/principals/{principal}/surface/{message,stream,history,approval}. The launcher checks that the token was issued to the same principal as in the address.
If the container was asleep, the launcher wakes it up, and the stream sends a wake-up event — the panel shows "The assistant is waking up". An open stream does not keep the container from falling asleep: only messages and decisions reset the idle timer.
From the outside, only the launcher's service routes (/harness/_launcher/internal/* and health) remain behind it. Any other address under /harness leads to /console/?assistant=open.
The "Open window" link in long Telegram answers leads to the same place. The launcher builds the address from LAUNCHER_PUBLIC_URL (<host>/console/?assistant=open) and passes it to the container in the HARNESS_WINDOW_URL variable; LAUNCHER_WINDOW_URL sets your own address. People's containers are created once, so they get a new value only after being recreated (docker rm -f harness-<principal-id>; the volume with the conversation remains).

Enabling¶

The harness profile and the people registry — see Personal workspace. The model in the container has a shell: the Docker proxy is visible only to the launcher, and people's containers are in a separate network, see Isolation and networks.
The console — see Installing the console. Compose sets the launcher address for it by itself: CONSOLE_LAUNCHER_URL=http://harness-launcher:8080/harness.
IAM: the human-harness audience with the harness:use scope in federation — it is registered by deploy/bootstrap.py.

If something is wrong¶

What the panel shows	Cause and what to do
"The assistant is waking up…"	The assistant's container is starting; the console keeps working. A cold start takes seconds
"The assistant did not wake up in time"	The engine did not start within the allotted time — "Retry"; if it repeats, check the engine's log
"Could not load the conversation: …"	The engine is unavailable or refused; the error code is in the message
"The assistant stream is not open: …"	The engine refused the stream (for example, missing permissions); the stream does not reconnect by itself — after fixing, press "Retry"
"The assistant is open in too many tabs."	The limit of concurrent streams per person is exceeded — close the panel in the extra tabs
"The screen context is too large"	Remove the context chip and retry
"The session has ended — going to sign-in"	The console session has expired; after signing in, the console returns you to the same screen
The assistant answers "Claude Code is not authorized"	The subscription token in the person's secrets is missing or expired — see Personal workspace