Agent activity
A long generation can hold the inference server for minutes. Without a shared
view, it is visible only on the screen that started it. The agent activity card
fixes that. It is fed by ActivityService.SubscribeAgentActivity.
What you see
Each run shows its agent (for example pulse:whitepaper), its status
(running, done, or failed), and its steps.
A step is not a sentence the backend wrote. PULSE emits a code from the reasoning-step vocabulary shared with the rest of the Runink platform, plus the operands that go with it, and the client composes the line in your language. The codes PULSE emits say that a prompt was dispatched (with the model and the token ceiling it was dispatched with), that a source is about to be read or was read or could not be read, that tokens are coming back (count and elapsed time, marked approximate because a stream does not promise one frame per token), that a draft was rewritten or a reply refused, what was found, and that the run finished or failed. A step the client has no line for renders as nothing rather than as an internal token.
Every operand is a value the backend already held. Nothing is estimated from elapsed time or read off the model’s own account of itself, and a measurement that is missing is left out rather than sent as a zero — an unknown token count is not a cheap generation.
A step can also carry provenance: whether the answer is fresh or reused from a cache, and whether it was produced in a degraded mode. A healthy, fresh answer shows nothing extra. The card calls out only reused, offline, or degraded answers.
PULSE does not run a propose-then-critique loop, so a plain generation is a short sequence of these steps. Longer pieces of work have more: the one-phrase audit runs tracked stages, and each stage also reports its own health to the operator console as its own agent, so a stage that wedges goes stale there instead of disappearing.
Whose runs you see
By default you see your own session’s runs. The server uses the session
field of the request, or, when that is empty, the caller’s x-agent-trace-id
header.
Work started without a request behind it (a schedule, a WhatsApp message, an A2A call, a background engine) has no session. A scoped subscriber does not see those runs. Only a subscriber that sends neither a session nor a trace ID gets the ambient view of everything. An operator console does that on purpose.
Opening the page mid-run
When you subscribe, the server first replays every run in progress, with the
steps recorded so far. Then it sends snapshot_complete, and after that live
events. So a page opened halfway through a run still shows that run.
The featured strip
ActivityService.ListFeaturedMessages returns operator-written sponsor and
partner cards for the rotating strip. The server reads them from a file named by
FEATURED_MESSAGES_PATH. When that is unset, it looks for a file named
featured_messages.json in /app, the working directory, data, and
/app/data, in that order. No file means an empty strip, not an error. You
can change the cards without rebuilding the app.