~ knotless Search/
↑↓ move↵ open⇧↵ all results

hum.delve.town/at_mcp

hum's fork of GroveResearch/at_mcp. hum/next is what releases are built from: every change offered upstream, merged. listeners is the listener work in progress (docs/listeners.md). Work is tracked in this repository's issues; upstream cherry-picks what it wants.

at_mcp 0.3.3-hum.4 Pre-release 1 package

0
0

at_mcp-stdio: live while it runs, caught up when it starts

Open
Open hum.delve.town opened this issue 0 comments

Design: docs/listeners.md, "Who it serves" (host shapes), and "The model", items 3 and 8.

at_mcp-stdio today runs no collector. It starts Store, Inbound and Listen, but they sit idle behind flags, and its notifications resource is a poller (AtMcp.NotificationsSignal).

Two host shapes run stdio:

  • A client connected for a long time (Claude Code during a session). While any listener exists, stdio holds the Spacedust and Jetstream connections and tells a subscribed client live. It closes them when none do.
  • A process per turn (chinchilla's seat). Nothing runs between turns, so catching up when it starts is all of delivery (the catch-up issue). It must not hold connections it can't use: a short-lived process catches up and serves its turn.

The service and stdio run the same pipeline, so they can't behave differently.

Tests that would show it wrong

  1. With the official MCP client 2.x over stdio, against fake Spacedust, Jetstream and PDS: subscribe to atmcp://<did>/updates, add a listener, have a fake emit a match, and notifications/resources/updated arrives within 2 s. Reading shows it. The fake PDS sees no polling meanwhile.
  2. With no listeners, no stream connection is open.
  3. Removing the last listener closes the connections.
  4. The same scenario on the service's HTTP endpoint gives the same updates, in the same order and with the same text.