v0.8.9

v0.8.9

The room bridge moves to massrelay v0.3.1, whose wire format changed, and Bridge.Notify now sends
one message per room instead of one per tab. Two 2026-07-28 wire fixes come along with it. This is
also the first release where @panyam/mcpkit-rooms reaches npm since 0.8.7, because v0.8.8's
publish failed (below).

Changed (room bridge)

• massrelay v0.3.1 and servicekit v0.1.12. massrelay v0.3.0 renamed AppMessage.to_client_id
  to a repeated to_client_ids. It kept the field number, so binary protobuf still matches, but
  the page's CollabClient speaks protojson, where the name changes from toClientId to
  toClientIds. A page on @panyam/mcpkit-rooms 0.8.7 or earlier, talking to a relay on massrelay
  v0.3, sends replies the relay can't address, and the agent's call times out with no error saying
  why. The page client now depends on @panyam/massrelay ^0.3.1 and
  @panyam/servicekit-client ^0.1.12. (PR 1629)
• Room.SendEach(ctx, to, kind, payload) delivers one message to several peers. A partial
  delivery returns *rooms.NotMembersError naming the peers that weren't in the room, and the rest
  still get the message. memroom and both massrelay transports implement it. Anything else that
  implements Room has to add the method. (PR 1629)
• Bridge.Notify sends one message per room. A big notify used to queue one message per tab
  behind the dialled transport's 25-a-second pacer. A tab the relay reports gone lands in
  rep.Failed on its own. Fan-out tool calls still go one per tab, since each tab answers its own
  call. (PR 1629)

Fixed

• @panyam/mcpkit-rooms publishes from a release tag again. v0.8.8's publish-npm-rooms run
  failed with a missing go.sum entry, because the tagged tree has no in-repo replace lines and
  the workflow never put them back. It restores them first now, and
  check-release-workflows-selftest fails any tag workflow that sets up Go without doing that.
  0.8.8 was never published to npm, so go from 0.8.7 straight to 0.8.9. (PR 1626)
• 2026-07-28 always gets ttlMs and cacheScope. That schema makes both required on every
  cacheable result. WithoutListCacheControl, WithoutReadResourceCacheControl and a negative TTL
  dropped them, and so did agents/list and agents/get without Config.TTLMs. On 2026-07-28 an
  opted-out hint now goes out as ttlMs: 0 with cacheScope public for lists and private for
  reads. 2025-11-25 and earlier still omit them. (issue 1546)
• The stateless tools/list strips x-mcp-file for clients without fileInputs.
  core.HasFileInputs read session capabilities only, so it answered false for every stateless
  client, and the stateless list didn't strip anyway. It reads the request's _meta too now, and
  both wires build tools/list through one helper, which also gives the stateless list the session
  wire's pagination. (issue 1515)

Upgrading downstream

• agni:
  • Go: experimental/ext/rooms, …/rooms/massrelay and …/integrations/roomsevents at
    v0.8.9, plus github.com/panyam/massrelay v0.3.1 if you require it directly.
  • Browser: @panyam/mcpkit-rooms@0.8.9, in the same change as the Go side. npm users get the
    matching massrelay through ^0.3.1, but an embedder that bumps only the Go module, or keeps
    serving a page bundle built before the bump, gets the silent timeout above.
  • Your own Room implementations, if any, need SendEach.
• Anyone relying on WithoutListCacheControl to keep ttlMs off the wire: on 2026-07-28 it
  now sends ttlMs: 0, which a client reads the same as an absent field.