github ag-ui-protocol/ag-ui release/2026-11-21
Release 2026-11-21

38 minutes ago

Packages Published

TypeScript (npm) - published at 18:04:23 UTC

Package Version Install
@ag-ui/langgraph 0.0.47 npm install @ag-ui/langgraph@0.0.47

@ag-ui/langgraph@0.0.47

  • Added a durability agent option; the adapter sends it with runs and uses it to read boundary checkpoints instead of a client-controlled forwardedProps.durability.
  • Client forwardedProps now reach runs.stream only via an allowlist (streamSubgraphs, durability); context, checkpoint, checkpointId and others are dropped with a warning.
  • forwardedProps.config is no longer merged over assistantConfig; forwardedProps.command accepts only resume (goto/update dropped with a warning).
  • Continue mode now resumes only from the node that thread history shows ran in the latest step, using it as asNode, and filters client state by the graph input schema.
  • A continue that would change an input channel now starts a new run so prior gates re-evaluate the changed input.
  • A continue that proceeds unchanged no longer writes client state, avoiding reducer channels (e.g. operator.add) doubling stored values.
  • Fixed a pending interrupt being dropped when a continue was requested on an interrupted thread.
Breaking changes
  • Client-sent forwardedProps.durability, context, checkpoint, checkpointId and other non-allowlisted run options are now ignored; use the new durability agent option instead.
  • forwardedProps.config is no longer merged over assistantConfig.
  • forwardedProps.command now accepts only resume; goto and update are dropped.
  • Continue behavior changed: continues with altered input now start a new run, and resume node/state are constrained by thread history and input schema; re-verify continue flows.

Python (PyPI) - published at 18:04:24 UTC

Package Version Install
ag_ui_adk 0.14.0 pip install ag_ui_adk==0.14.0
ag-ui-langgraph 0.0.48 pip install ag-ui-langgraph==0.0.48
ag-ui-openai-agents 0.2.0 pip install ag-ui-openai-agents==0.2.0

ag_ui_adk@0.14.0

  • add_adk_fastapi_endpoint() now forwards dependencies to /agents/state and /capabilities, closing an auth bypass (GHSA-jhvp-j8pf-5vh5).
  • Request headers now take precedence over client state.headers in extract_headers, preventing identity spoofing.
  • Run state that is not a JSON object or null is rejected with HTTP 422.
  • ADKAgent.run() accepts any collections.abc.Mapping as run state.
  • HTTPException from the state extractor is now re-raised on /capabilities and /agents/state, preserving status, detail and headers.
  • Removed dead _NonObjectStateError 422 handlers on /capabilities and /agents/state.
  • Docs updated to read identity from the authenticated request and fail closed (HTTP 401) when absent.
Breaking changes
  • /agents/state and /capabilities now enforce the dependencies passed to add_adk_fastapi_endpoint(); previously-open routes may now require authentication.
  • Non-object run state (e.g. arrays) is now rejected with HTTP 422.
  • Client-supplied state.headers values no longer override request headers; clients relying on that behavior must change.
  • /agents/state uses the first registration's dependencies per app; all calls on one app must pass the same dependencies.
  • Re-raised HTTPException changes responses on these routes from 500 to the raised status.

ag-ui-langgraph@0.0.48

  • Continue-mode now starts a new run when forwarded input differs from what a prior node (e.g. permission gate) checked, preventing bypass of gating nodes.
  • Continue no longer writes client state when it goes on, avoiding reducer channels (e.g. operator.add) double-combining values.
  • Node name for continue is honored only when the thread actually ran that node last; client-supplied node_name can no longer skip graph nodes.
  • Fixed resume logic under durability="exit" where one checkpoint per run made paused/gate nodes look already run.
  • Continue-mode state writes now apply the input-schema filter, so clients cannot set channels outside the graph's input schema.
  • Failed tool calls now stream TOOL_CALL_RESULT (and START/ARGS/END for non-streaming models) instead of only appearing in MESSAGES_SNAPSHOT.
  • Pending tool failures are keyed by run plus tool_call_id, fixing lost/crossed results across concurrent graphs sharing an id.
  • Interrupted tool calls (GraphInterrupt/GraphBubbleUp) are no longer recorded as failed.
  • Error handlers returning a Command carrying an error ToolMessage now stream the result correctly.
  • Tool messages are presented under their streamed id in MESSAGES_SNAPSHOT regardless of consumer timing; stopped mutating tool message objects in place.
  • Tool messages without an id get a stable id so TOOL_CALL_RESULT matches the snapshot.
  • Cancelled runs now synthesize a cancelled ToolMessage for dangling tool calls so the next turn is not rejected by the provider.
  • lg_interrupt_to_agui reads Interrupt.response_schema first, falling back to value-level keys; non-mapping values ignored.
  • Middleware-internal model calls (e.g. SummarizationMiddleware from langchain 1.4.0) no longer leak into the client stream as assistant text.
  • RUN_STARTED declares protocol_version when the SDK provides PROTOCOL_VERSION, guarded for the 0.1.22 floor.
  • A2UI recovery workers isolated and subagents kept on the caller event loop.
  • Locked langgraph moves 1.2.8 -> 1.2.12.
Breaking changes
  • Continue with a changed input-schema value now starts a new run instead of resuming from the named node; thread flows relying on the old resume behavior must be re-verified.
  • Client forwardedProps.node_name is honored only when the thread last ran that node; previously-accepted skips will now be rejected.
  • Continue no longer writes client state when it goes on, and applies input-schema filtering; code depending on those writes or on setting out-of-schema channels must change.
  • Tool message ids and TOOL_CALL_RESULT timing changed; id-based snapshot merges should be re-verified.
  • Requires langgraph 1.2.12 for Interrupt.response_schema.

ag-ui-openai-agents@0.2.0

  • Paused tool calls are now signed with an HMAC over thread, call id, tool name, and canonical arguments; resumes run a tool only if this server paused it.
  • Approval tokens are bound to the pausing agent and rejected when used by a different agent.
  • Approvals are now checked for expiry at consume time, before and after the store responds, preventing reuse of expired tokens.
  • OpenAIAgentsAgent and add_openai_agents_fastapi_endpoint accept consumed_tokens=, a shared ConsumedTokens store, so decided approvals are shared across workers and survive restarts.
  • Pausing or resuming approvals now refuses an unscoped (default) signer to prevent cross-agent approval reuse.
Breaking changes
  • Pausing or resuming approvals with the default unscoped signer is now rejected; provide a properly scoped signer.
  • Resumed tool calls are validated against a server-signed paused call; client-modified arguments, injected calls, or replayed approvals are refused.
  • Expired approval tokens are now refused at consume time, not only at resume start.
  • Multi-worker or post-restart deployments must pass a shared consumed_tokens= store to avoid duplicate tool execution.

Don't miss a new ag-ui release

NewReleases is sending notifications on new releases.