ag-ui-protocol/ag-ui release/2026-10-21 on GitHub

Packages Published

Package	Version	Install
@ag-ui/claude-agent-sdk	0.0.8	`npm install @ag-ui/claude-agent-sdk@0.0.8`
@ag-ui/claude-managed-agents	0.0.4	`npm install @ag-ui/claude-managed-agents@0.0.4`

Fixed queued SDK recovery turns to wait for each turn to finish before proceeding.
Completed batches are now cleared before later failures to avoid stale state.
Deferred batches are persisted before interrupted streams.
Unsupported pending tool batches are retained and rejected rather than silently dropped.
Unresolved child tool identities are retained during recovery.
SDK middleware tools now complete correctly and recovery input is retained.
Historical errors and results are preserved in metadata during approval recovery.

None.

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 completions are sent as content: [] without text blocks; zero-length text parts are removed while nonempty text, errors, and supported media are retained.
Example updated to use a system-injection-supported model.

Empty completions now serialize as content: []; consumers relying on prior empty text-block output must re-verify.
Zero-length text parts are stripped from outbound completions.

Package	Version	Install
ag_ui_adk	0.11.0	`pip install ag_ui_adk==0.11.0`
ag-ui-claude-sdk	0.2.3	`pip install ag-ui-claude-sdk==0.2.3`
ag-ui-claude-managed-agents	0.2.2	`pip install ag-ui-claude-managed-agents==0.2.2`

Fixed first-run history import so client history reliably reaches the model on new threads without an ADK session.
Seeded history events now spaced 1 ms apart to prevent reordering under DatabaseSessionService timestamp ties.
Tool results now named from the session's persisted FunctionCall when the assistant message is absent from client history.
Client-supplied state in RunAgentInput.state is now scoped to the session; keys beginning with app: or user: are discarded by the HTTP endpoint before merging.
Migrated A2UI helper to published SDK 0.9 APIs using client Catalog instances and the Direct JSON processor.
Updated test compatibility for google-adk 2.12 and google-genai 2.x.

Fixed an access-control flaw that allowed client-supplied state to modify shared ADK app-wide or user-wide state (#2927). Upgrading is recommended.

HTTP requests now discard client-supplied state keys beginning with app: or user:. The keys and their values are removed rather than renamed or moved into session scope.
To persist trusted app-wide or user-wide state, supply it through the server-side extract_state_from_request hook and set allow_scoped_state_keys=True on ADKAgent or ADKAgent.from_app(). Direct ADKAgent.run() callers also require this opt-in. The provided HTTP endpoint still discards client-supplied scoped keys when the flag is enabled.
Dropped legacy A2UI SDK compatibility; requires a2ui-agent-sdk>=0.9.0,<0.10.0 and a2ui-core>=0.4.0,<0.5.0. Google ADK remains supported at >=1.28.1,<3.0.0.

Clear completed tool batches before later failures to avoid stale state.
Persist deferred batches before interrupted streams so recovery works correctly.
Retain and reject unsupported pending tool batches instead of leaving them unresolved.
Complete SDK middleware tools and retain recovery input during recovery.
Retain unresolved child tool identities across parallel tool calls.
Preserve historical errors and results in metadata during approval recovery.

None.

Persist child result delivery across initial runs and reloads.
Forward runtime context across adapters.
Recover canonical child results after approval.
Send text attachments as UTF-8 text sources.
Send empty frontend completions as content: [], removing only zero-length text parts while retaining nonempty text, errors, and supported media.
Updated example to a model that supports system injection.

Empty frontend completions now serialize as content: [] with zero-length text parts stripped; verify handling of empty tool results and continuation/replay.