@langchain/langgraph-cli@1.8.12026-11-20
Patch Changes
- #2744 7257a24 Thanks @wario2k! - fix(langgraph-api): preserve additional_kwargs in protocol state normalizer
Forward non-empty additional_kwargs on values snapshot messages so
server-authored message metadata (e.g. committed attachment paths) reaches
live subscribers.
- Updated dependencies [63ad546, 7257a24]:
- @langchain/langgraph-api@1.8.1
@langchain/langgraph-api@1.8.12026-11-20
Patch Changes
- #3420 63ad546 Thanks @hari-dhanushkodi! - Expose each run's LangSmith tracing project and the session-name tracing capability so Studio can look up traces from the JavaScript dev server.
- #2744 7257a24 Thanks @wario2k! - fix(langgraph-api): preserve additional_kwargs in protocol state normalizer
Forward non-empty additional_kwargs on values snapshot messages so
server-authored message metadata (e.g. committed attachment paths) reaches
live subscribers.
- Updated dependencies [7257a24]:
- @langchain/langgraph-ui@1.8.1
@langchain/langgraph-ui@1.8.12026-11-20
Patch Changes
- #2744 7257a24 Thanks @wario2k! - fix(langgraph-api): preserve additional_kwargs in protocol state normalizer
Forward non-empty additional_kwargs on values snapshot messages so
server-authored message metadata (e.g. committed attachment paths) reaches
live subscribers.
@langchain/react@2.0.0-rc.3预发布2026-11-19
Patch Changes
- Updated dependencies [21b60ad]:
- @langchain/langgraph-sdk@1.17.1-rc.3
@langchain/langgraph-sdk@1.17.1-rc.3预发布2026-11-19
Patch Changes
- #3416 21b60ad Thanks @eliornl! - StreamOrchestrator now selects only run_id and kwargs when looking up a thread's running run on open, and avoids opening and immediately aborting the /threads/:id/stream watcher while joining a running run.
@langchain/react@2.0.0-rc.2预发布2026-11-18
Patch Changes
- #3410 adcf1d1 Thanks @eliornl! - Add ThreadCacheProvider. Every useStream and useSuspenseStream inside it shows a thread opened again from its last read at once, while it is read again. Wrap the part of the app whose streams read the same server as the same user, and key the provider by that user. See "Showing a reopened thread at once" in the useStream guide.
- Updated dependencies [84df9ca, adcf1d1]:
- @langchain/langgraph-sdk@1.17.1-rc.2
@langchain/langgraph-sdk@1.17.1-rc.2预发布2026-11-18
Patch Changes
- #3413 84df9ca Thanks @eliornl! - StreamOrchestrator now joins the running run a thread load's first lookup finds even when that lookup and the state read both answer before activate(). Bindings start the load during a render and call activate() from an effect, so this happens whenever the answers are already at hand, for example when an app sent those reads before the engine existed. Such a join was dropped and left to the lookup after the read, a whole runs list later. It is now held for activate(), as a join the lookup after the read finds already is, and dropped if that lookup finds no run going.
- #3410 adcf1d1 Thanks @eliornl! - StreamOrchestrator takes a threadCache, a ThreadCache exported from @langchain/langgraph-sdk/ui, that keeps what thread loads read. A thread opened again shows its last read at once while it is read again, and what the read finds replaces it as on a first open. Until then the cached head counts as one no read has confirmed: its interrupts and task errors stay hidden, its headless tool calls don't run, and respond(), and a submit() made before the read lands, wait for the load. A read that fails takes the cached thread off the screen and out of the cache. A read in flight when clear() is called, or one a disposed stream makes, stays out of it. A run that starts on a thread takes it out of the cache, and a read begun during a run, or before one began, stays out, except the read after the run, so a thread left while a run streamed opens as on a first open. The cache holds threads up to maxBytes characters of strings (32 M by default) and drops the ones opened least recently first. Without a threadCache, loads work as before.
@langchain/react@2.0.0-rc.1预发布2026-11-17
Patch Changes
- 98e0aaa Thanks @eliornl! - Picks up the stream engine changes in the matching @langchain/langgraph-sdk release candidate: with stateKeys: [], runs no longer stream the whole state after every step, a thread opened during a run joins it a round trip sooner, and /info is read once per server instead of once per client.
- Updated dependencies [4befdb6, e139165, ed8f2cd]:
- @langchain/langgraph-sdk@1.17.1-rc.1
@langchain/langgraph-sdk@1.17.1-rc.1预发布2026-11-17
Patch Changes
- #3404 4befdb6 Thanks @eliornl! - StreamOrchestrator now shares a server's /info read for 5 minutes among the clients built with the same config, instead of reading it once per Client. An app that builds a Client for each thread it shows read /info again on every thread it opened. A client with another URL, API key or headers still reads its own. A 404 still means the server has no /info, and a read that fails or times out is still read again by the next run that needs it.
- #3405 e139165 Thanks @eliornl! - StreamOrchestrator now looks for a thread's running run while it reads the thread's state, not only after. A run it finds is joined as soon as both answer, so a thread opened while a run is going shows the run's progress, isLoading and the stop button a round trip sooner. The lookup after the read still runs, and only it can tell that no run is going: the thread watcher opens once the read lands and doesn't replay a run that started before it, so a lookup sent with the read can miss a run that started during the read. A thread load sends one more runs list than before.
A join queued behind another stream, such as a joinStream() called during a run, no longer starts once the engine has switched to another thread, and the server's queued runs stay in the queue until such a join has streamed.
- #3402 ed8f2cd Thanks @eliornl! - StreamOrchestrator runs no longer ask for values when the messages are the only state key shown (stateKeys: []) and messages-tuple streams them, as with @langchain/react 2.0. A values event carries the whole state after every step, so on a long thread each step sent the whole thread again: one 22-message turn on a 1,350-message thread took 70 MB in the browser, and 3.6 MB without values, most of it one read of the thread as the run ends. The run's messages come from messages-tuple and its interrupts from its root updates, and a run that streamed no values reads the thread's head once as it ends, joined runs included, so the order its parallel steps saved and anything the stream missed come from the server. A stream's first output drops the interrupts an earlier run left, as its first root values did. A panel on a subagent or subgraph opened during such a run shows nothing until it has read what that subgraph saved, then that with what streamed meanwhile on top, and reads it again when the run ends, also when the engine can't tell which modes a joined run streams, and subgraphs are found from their nodes' updates. An optimistic input shows as sent once the run's first step streams. Runs the engine joins keep the modes they were started with. Apps that don't set stateKeys, list other keys, or pass values in submit()'s streamMode get values as before.
A message whose messages event names no namespace now belongs to the graph its metadata names without the segment of the task that wrote it, as in LangGraph. The tools node's own ToolMessage (tools:<task id>) was taken for a subagent's: the main conversation left it out until the next values event, and a subagent panel on a server that doesn't filter by namespace showed its task result.
@langchain/angular@0.4.102026-11-17
Patch Changes
- Updated dependencies \[4d11ea7]:
- @langchain/langgraph-sdk@1.8.13