Patch Changes

• #82 1a4a74c Thanks @saulmc! - Fix normalizeMessageContent so a reply that carries a non-text payload renders its display string instead of a raw JSON dump of the decoded content. A reply containing a remote attachment now renders reply to "...": [remote attachment: photo.jpg (... bytes) https://...] rather than serializing the attachment envelope — which leaked the AES key material (secret/salt/nonce) into the message text. Inline attachments, reactions, and other nested content types render through the same path.

Patch Changes

• #79 08a76dd Thanks @insipx! - update xmtp

Patch Changes

• #73 d3cc3f2 Thanks @insipx! - update to xmtp 06-10-2026 pending self remove worker fix

Patch Changes

• #70 ea300be Thanks @insipx! - update to libxmtp 06-09-2026, event-driven disappearing messages

Patch Changes

• #67 29f7593 Thanks @insipx! - update xmtp to include init_logging panic fix

Patch Changes

• #64 88d3794 Thanks @insipx! - Pin @xmtp/node-sdk and @xmtp/node-bindings to dev build *-dev.f7b334d (node-sdk 6.0.0-dev.f7b334d, node-bindings 1.11.0-dev.f7b334d). Widen the node-sdk/node-bindings peer ranges to >=...-dev so dev-channel prereleases satisfy the peer.

Patch Changes

• #60 39c349a Thanks @insipx! - Declare @xmtp/node-sdk and @xmtp/node-bindings as peer dependencies so consumers install a single copy. Prevents native-binding segfaults from duplicate node-sdk trees.

Patch Changes

• #55 286753e Thanks @insipx! - make a release with the automated workflow

Added: Thinking content type

New silent codec for ambient agent thinking-status indicators, anchored to
a specific message (like read receipts):

• convos.org/thinking:1.0 — JSON payload {state, targetMessageId, content, resultMessageId?}.
  state is "start" or "stop" — agents pair a start with a matching stop
  on the same targetMessageId. content is a 3–5 word human-readable label
  shown alongside the indicator (e.g. "Designing your cycling guide").
  resultMessageId is optional and only meaningful on stop — the agent's own
  reply message that closed the thought, so receivers can link "thought about X"
  to "replied with Y".
• Silent (shouldPush: false), no fallback. Filtered from the chat stream by
  isDisplayableMessage.
• agent serve event surface: receive path emits a thinking event; live and
  catchup both surface, deduped by message id. Stdin command {"type":"thinking","state":"start|stop","targetMessageId":"...","content":"...","resultMessageId":"..."} sends a Thinking message. resultMessageId only valid with state: "stop".

Exports: ThinkingCodec, ContentTypeThinking, isThinkingMessage,
getThinkingContent, Thinking, ThinkingState.

Added: Remote-attachment decode helpers

agent serve's message event now carries structured attachment metadata
when the content type is one of XMTP's native remote-attachment types:

• xmtp.org/remoteStaticAttachment:1.0 — emits a remoteAttachment
  object alongside content.
• xmtp.org/multiRemoteStaticAttachment:1.0 — emits a
  multiRemoteAttachment: { attachments: [...] } object.

Each entry has url, contentDigest, scheme, secret/salt/nonce
(base64-encoded), and optional contentLength/filename — enough for
an agent to fetch and decrypt the bytes with decryptAttachment. Bytes
were previously Uint8Array on message.content, which doesn't
serialize to JSON cleanly. Encoding matches the agent serve remote-attach
stdin command so agents send and receive attachments through one mental
model. Live and catchup message paths both populate the new fields.

New exports: extractRemoteAttachment, extractMultiRemoteAttachment,
RemoteAttachmentJson, RemoteAttachmentInfoJson, MultiRemoteAttachmentJson.

Full Changelog: v0.9.1...v0.10.1

What's Changed

• feat: catch up to convos-ios#812 (multi-agent capability resolution) in #41

Full Changelog: v0.9.0...v0.9.1