You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
- Added end-to-end multimodal user message support across TypeScript and Python SDKs, LangGraph integrations, and docs, including new `InputContent` schemas and example agents.
7
+
- Introduced runtime improvements such as `connectAgent`, safer cloning, and run lifecycle fixes, plus expanded event metadata (`parentRunId`, embedded inputs).
8
+
- Delivered a new `compactEvents` utility for consolidating streaming deltas and hardened backwards compatibility layers for legacy clients.
9
+
- Updated integrations (LangGraph, Vercel AI SDK, Mastra, Google ADK) to translate multimodal content and align with the latest protocol expectations.
10
+
- Bumped package versions to the `0.0.40` pre-release series (TypeScript) and `0.2.0a0` (Python) to ship these capabilities.
11
+
12
+
### TypeScript SDK
13
+
- Extended `UserMessage` to accept multimodal `InputContent[]`, added `TextInputContent`/`BinaryInputContent` schemas, and exported the associated types for consumers.
14
+
- Added optional `parentRunId` and embedded `input` payloads to `RunAgentInput` and `RunStartedEvent` schemas, plus surfaced a dedicated `AGUIConnectNotImplementedError`.
15
+
- Refined `AbstractAgent` by making `run` protected, introducing `connectAgent`, tracking `isRunning`, and ensuring `clone` copies agent state; adjusted `HttpAgent`, Mastra, and Vercel agents to clone safely.
16
+
- Updated `defaultApplyEvents` to handle non-string message content, merge `runStarted.input.messages` into local state, and improved event processing tests (`run-started-input`, cloning, multimodal, backwards compatibility).
17
+
- Added and exported a `compactEvents` helper (with comprehensive tests) for consolidating streaming text/tool call deltas before replaying them to subscribers.
18
+
- Improved legacy converters and integrations to flatten multimodal content when talking to text-only surfaces, guarding the behavior with new backwards compatibility tests.
19
+
- Updated the React SDK chat surface to render multimodal user messages, including inline attachment previews and safer clipboard handling when no text is present.
20
+
- Incremented package versions for `@ag-ui/core`, `@ag-ui/client`, CLI, encoder, and proto packages to `0.0.40-alpha.6`.
21
+
22
+
### Python SDK
23
+
- Introduced `TextInputContent` and `BinaryInputContent` models, allowing `UserMessage` instances to carry ordered multimodal content alongside traditional strings.
24
+
- Added optional `parent_run_id` and `input` fields to `RunStartedEvent` plus `parent_run_id` to `RunAgentInput`, mirroring TypeScript schema changes.
25
+
- Relaxed the base model configuration to allow extra fields for backwards compatibility and added validation to ensure binary payloads provide an ID, URL, or data source.
26
+
- Expanded the test suite to cover multimodal serialization, binary payload validation, run input parsing, and extra-field tolerance.
27
+
- Documented multimodal usage in the Python README and bumped the `ag-ui-protocol` package to `0.2.0a0`.
28
+
29
+
### Integrations
30
+
-**LangGraph (Python & TypeScript):** Added bidirectional converters for multimodal content, new vision-friendly example agents, updated tests, and bumped the package to `0.0.18a0` with matching dependency pins.
31
+
-**Google ADK Middleware:** Flattened multimodal message content into text parts when translating to ADK events and updated helper utilities accordingly.
32
+
-**Mastra:** Preserved constructor config when cloning agents and flattened AG-UI messages into the formats expected by Mastra clients.
33
+
-**Vercel AI SDK:** Added safe cloning, converted user content into SDK-compatible parts, and ensured multimodal inputs degrade gracefully to text.
34
+
-**General:** Updated converters and utilities across integrations to handle `InputContent` arrays without breaking existing text-only flows.
35
+
36
+
### Documentation
37
+
- Updated core concepts and SDK reference docs to describe multimodal user messages, the new input content schemas, and extended `RunStartedEvent` properties.
38
+
- Documented `connectAgent`, `connect`, and the replayable `events$` stream on `AbstractAgent`, clarifying how persistent connections are implemented.
39
+
- Marked the multimodal messages specification as implemented (October 16, 2025) and added README snippets showing multimodal message creation in both SDKs.
40
+
41
+
### Miscellaneous
42
+
- Added an explicit `@ts-expect-error` in the A2A middleware noting the intentional call to a protected method until a public API exists.
43
+
- Updated Poetry lockfiles and dependency pins to align with the new Python package versions.
0 commit comments