ag-ui-protocol/ag-ui release/2026-10-30

Release 2026-10-30

on GitHub

Packages Published

.NET (NuGet) - published at 20:53:01 UTC

AGUI.ClaudeManagedAgents@0.0.4

• Child result delivery now persists across initial runs and reloads.
• Runtime context is forwarded across adapters.
• Canonical child results are recovered after approval.
• Text attachments are sent as UTF-8 text sources.
• Empty frontend results are now supported.

Breaking changes

None.

TypeScript (npm) - published at 21:04:55 UTC

@ag-ui/mastra@2.0.2

• RUN_FINISHED now carries Mastra's finish reason unchanged on metadata.mastra.finishReason when reported; cancelled runs carry none.
• Added subagentVisibility: "hidden" | "attributed" option (default "hidden"); attributed mode maps each Mastra delegation to an AG-UI 1.0 subagent invocation.
• Run usage now counts delegated sub-agents' token usage by adding each forwarded finish once, in both visibility modes.
• Fixed loss of sibling tool calls across suspension.
• Fixed tool ownership not being preserved across text segments.
• Fixed use() middlewares not carrying onto cloned agents, so they now run on clones CopilotKit serves.

Breaking changes

• AbstractAgent.middlewares is now protected; subclasses building their own clones must account for this access change.

@ag-ui/watsonx@0.0.4

• Added WatsonxRunsAgent for the /v1/orchestrate/runs/stream endpoint, surfacing run steps and raw tool responses as TOOL_CALL_* events.
• Added optional onToolResult hook that emits tool responses as STATE_SNAPSHOT events, plus a pluggable thread store mapping AG-UI threads to watsonx threads.
• Fixed loss of tool arguments when a tool response arrives before its call record; responses for unknown calls now wait for the call record.
• RAW-only events no longer flush queued tool responses early; pending responses flush only at text or run end.
• Streams that end without run.completed or done now finish with RUN_ERROR instead of RUN_FINISHED.
• The 1MB limit now applies per line instead of per read, so large multi-response reads no longer abort the run; reader is cancelled on error.
• Documented WatsonxRunsAgent.

Breaking changes

• Interrupted streams that previously ended as RUN_FINISHED now emit RUN_ERROR; verify handling of partial runs.
• Tool response/argument queueing and flush timing changed, which may alter the ordering and content of emitted TOOL_CALL_ARGS and onToolResult events.

@ag-ui/a2ui-middleware@0.0.14

• Completed SDK middleware tools.
• Retain recovery input across middleware processing.

Breaking changes

None.

@ag-ui/mcp-apps-middleware@0.1.3

• Completed SDK middleware tools.
• Recovery input is now retained during processing.

Breaking changes

None.

Python (PyPI) - published at 21:04:57 UTC

ag_ui_adk@0.11.1

• Fixed streaming of function-call arguments from non-Gemini ADK adapters (OpenAI, Anthropic, LiteLLM); first-chunk partial args are now processed and the open stream is closed, avoiding duplicate TOOL_CALL_START and invalid concatenated JSON.
• Fixed dropping of parallel same-name client tool calls streamed in separate Gemini SSE partials (#2856).
• Added option to disable session cleanup: session_timeout_seconds=None means sessions never expire and the cleanup task does not start.
• Added track_sessions option (default True) to skip the in-memory session index and cleanup task.
• Fixed runner teardown closing plugins owned by the caller's long-lived App (#2219).

Breaking changes

• track_sessions=False cannot be combined with max_sessions_per_user, which depends on the session index.
• session_timeout_seconds=None now disables session expiry and the cleanup task; previously the task still ran.