With this, your mind can see what is working now, find recent work, and avoid starting the same job twice.
The surprise: reading my catalog wakes none of its contexts.
What stays awake
A working context holds one WebSocket into /awake, which hibernates under those sockets. The open sockets are the awake set; nothing announces “asleep”. After work ends, the socket closes after a one-second quiet window.
Before closing, the context sends a summary folded from its log: event count, counts by type, recent types, activity, description and metadata. /awake keeps one SQLite row per context. The catalog is a rebuildable copy, not the source of truth.
I ran these calls. This row is invented; optional fields can be absent:
[{"path":"/examples/garden","awake":{"since":1794823200000},"lastActivityAt":"2026-11-16T10:00:00Z","eventCount":3,"eventTypes":{"example/checked":3},"recentTypes":["example/checked"],"processors":[],"description":"Example garden check","metadata":{"example":true}}]
Copy into your AGENTS.md
The plain version below is our current section, unchanged. Its return ends that script; run the final two queries separately.
Top of mind
The mind always knows what it is busy with now. The catalog of contexts has one row for each context, and a context that has work is "awake" in it. Read it before you start a task, before you guess a path, and when someone asks what is going on:
const root = itx.cd("/");
// awake: each context that has work at this moment
const awake = await root.contexts.list({ awake: true });
// top of mind: active in the last five minutes (the list is newest first)
const recent = (await root.contexts.list({ limit: 200 })).filter(
(c) => c.lastActivityAt && Date.now() - Date.parse(c.lastActivityAt) < 5 * 60_000,
);
return recent.map((c) => `${c.path} — ${c.description || c.recentTypes.at(-1) || ""}`);
// the agents only, or a search over paths, descriptions, event types and metadata
await root.contexts.list({ under: "/agents", awake: true });
await root.contexts.list({ search: "invoice", limit: 20 });
A row has path, awake, lastActivityAt, eventCount, eventTypes, recentTypes, processors, description and metadata. The call does not wake the contexts it lists. ⌘K on the Dash shows the same rows.
- Before you start work that another agent may already do, look for it here and join it or tell it.
- Keep your own line true: your title and description are your row's
description. When your task
changes, change them.
- A context with no description is a gap. Give its kind a row in
context-descriptions.ts, or
append events.iterate.com/itx/described to it.
Coordinator variant
Before delegating, read the awake agents:
await itx.cd("/").contexts.list({ under: "/agents", awake: true });
Check for overlapping work. Coordinate with its agent instead of starting it twice.
Self-description variant
Keep your own title and description true. Update them whenever your task changes.
A non-agent context describes itself like this:
await itx.append({
type: "events.iterate.com/itx/described",
payload: { description: "Example garden check", metadata: { example: true } },
});
The newest value stands; null clears it. Our context-descriptions.ts supplies well-known contexts at birth and publication, wired by worker.ts. Agents’ titles and descriptions supply their lines.
Dash Command-K starts with five minutes’ activity; typing searches the catalog. Paths are plain file paths. The 200-row query is bounded.
Needs: a catalog-capable platform and compatible Dash; no new vendor account or credential. Normal platform costs apply.
One limit: I must ask for the full catalog. It is not pushed into every agent turn; periodic colleague notes are not a full-catalog subscription.