[dependabot]

Bumps the minor-and-patch group with 24 updates:

Package	From	To
@aws-sdk/client-bedrock-runtime	3.1120.0	3.1158.0
@langchain/anthropic	1.5.11	1.6.0
@langchain/aws	1.4.4	1.5.1
@langchain/core	1.2.9	1.3.1
@langchain/google	0.2.3	0.2.12
@langchain/langgraph-checkpoint-sqlite	1.0.4	1.1.0
@langchain/openai	1.6.0	1.8.0
@langchain/openrouter	0.4.10	0.4.21
@langchain/protocol	0.0.19	0.0.20
@modelcontextprotocol/sdk	1.32.1	1.32.2
cron-parser	5.10.1	5.10.2
cronstrue	3.27.0	3.38.0
deepagents	1.13.2	1.16.0
google-auth-library	11.1.0	11.3.1
langchain	1.5.10	1.6.0
langsmith	0.10.5	0.11.3
marked	18.0.14	18.1.2
posthog-node	5.51.4	5.57.2
react	19.3.0	19.3.1
@types/node	26.4.0	26.6.5
eslint	10.11.0	10.13.0
jsdom	30.1.1	30.1.3
prettier	3.9.9	3.9.10
typescript-eslint	8.70.1	8.72.1

Updates @aws-sdk/client-bedrock-runtime from 3.1120.0 to 3.1158.0

Release notes

Sourced from @​aws-sdk/client-bedrock-runtime's releases.

v3.1158.0

3.1158.0(2026-10-27)

Chores

• codegen: smithy-aws-typescript-codegen 0.55.0 (#8385) (4a149acb)

New Features

• client-license-manager: AWS License Manager now supports delegate grants for read-only visibility into license distributions. CreateGrant accepts a new GrantType (BuyerGrant or DelegateGrant) and a desired status at distribution time. (c2b033fe)
• client-ecs: ECS now supports GPU sharing via NVIDIA Multi-Process Service (MPS) on ECS Managed Instances and ECS EC2, allowing multiple ECS tasks to share a single Nvidia GPU for improved utilization. (82c0bea6)
• client-bedrock-agentcore-control: Adds the ResetTokenVaultKeyMaterial API, which restores a token vault whose customer managed key was deleted by moving it to an AWS owned key. The reset permanently deletes the stored OAuth 2.0 tokens, users must consent again. Adds on-behalf-of token exchange client authentication override option. (caa886f6)
• client-inspector2: Adding support for Machine Image scan (44b1f0cb)
• client-lightsail: Added the attributeUpdates parameter to the Lightsail StartInstance API, which allows customers to update the attributes of a stopped instance when starting it. (f01e4a58)
• client-ec2: EC2 AMIs can now show vulnerability data and be targeted directly for scanning. (4f53b1cf)
• lib-sqs: add SQSPoller for auto queue polling (#8339) (cb0d4f01)

Bug Fixes

• lib-sqs: sync version with current monorepo release (#8384) (82c877fa)
• client-s3: isolate special character e2e test keys (#8382) (d9c091fa)

---

For list of updated packages, view updated-packages.md in assets-3.1158.0.zip

v3.1157.0

3.1157.0(2026-10-26)

Chores

• codegen: sync for service default socket timeout and RPC v2 JSON (#8378) (f9d9119f)

Documentation Changes

• client-deadline: fix the sdk documenation issue only (e3bf3c6c)

New Features

• clients: update client endpoints as of 2026-10-26 (a95658fd)
• client-connect: Amazon Connect now automatically saves agents' in-progress email drafts preserving the message, recipients, and subject during browser or network interruptions so agents can return to the mail and continue instead of starting over. (d03e5700)
• client-appstream: Add Windows 11 to PlatformType enum (52666c31)
• client-lightsail: This release adds Web Application Firewall (WAF) support for Lightsail distribution. (1fee104e)
• client-marketplace-catalog: This release adds support for the SubsidiaryAuthorization entity type in the ListEntities operation, including entity-type filters with partial name search, sort options, and a summary returned for each subsidiary authorization. (a3f6ea1b)
• client-eks: This release adds support for EKS management of Gateway API CRDs (Custom Resource Definitions) (67c963fe)
• client-sso-admin: Added support for the RegionScope parameter on ListInstances, allowing you to list IAM Identity Center instances in the current Region or across all Regions. (6cc5db1b)
• client-medialive: AWS Elemental MediaLive now lets you control whether explicit weighted prediction is applied to H.264 outputs. It also introduces Perceptual Filter, a new noise reduction filter designed to improve video quality by increasing VMAF scores, now with Auto, None, and strength options. (b515e6cd)
• client-billing: Adds support for managing private pricing contracts and contractual investments including listing and associating contracts, retrieving contractual investments and their summaries, and creating, listing, and canceling contractual investment attestations (b4e7d786)
• client-taxsettings: Add Mexico Additional Info changes (88e365c5)

... (truncated)

Changelog

Sourced from @​aws-sdk/client-bedrock-runtime's changelog.

3.1158.0 (2026-10-27)

Note: Version bump only for package @​aws-sdk/client-bedrock-runtime

3.1157.0 (2026-10-26)

Note: Version bump only for package @​aws-sdk/client-bedrock-runtime

3.1156.0 (2026-10-22)

Note: Version bump only for package @​aws-sdk/client-bedrock-runtime

3.1155.0 (2026-10-21)

Note: Version bump only for package @​aws-sdk/client-bedrock-runtime

3.1154.0 (2026-10-20)

Note: Version bump only for package @​aws-sdk/client-bedrock-runtime

3.1153.0 (2026-10-19)

Note: Version bump only for package @​aws-sdk/client-bedrock-runtime

3.1152.0 (2026-10-14)

... (truncated)

Commits

• 2d08e1c Publish v3.1158.0
• e5988ec Publish v3.1157.0
• f9d9119 chore(codegen): sync for service default socket timeout and RPC v2 JSON (#8378)
• 163a566 Publish v3.1156.0
• 7c2f089 Publish v3.1155.0
• 5f53137 chore(codegen): sync for JSON string serializer buffer fix and retryMode docs...
• f6ff312 Publish v3.1154.0
• 3508584 Publish v3.1153.0
• 3e350b5 chore(codegen): sync for JSON codec and schema server plugin (#8353)
• 3ce8aa7 Publish v3.1152.0
• Additional commits viewable in compare view

Updates @langchain/anthropic from 1.5.11 to 1.6.0

Release notes

Sourced from @​langchain/anthropic's releases.

@​langchain/anthropic@​1.6.0

Minor Changes

• #11986 2caf64d Thanks @​richardscarrott! - feat: stream tool-call args as tool-call-delta slices

  Tool-call and server-tool-call args now stream as tool-call-delta events, which carry only the new slice of the args, instead of block-delta snapshots of the args so far. A large tool call now costs about the same bytes on the wire as the same length of text.

  • This covers the compat bridge (any model that only implements _streamResponseChunks), the OpenAI-compatible converter (OpenAI Completions, Groq, Mistral, IBM, OpenRouter, Perplexity, DeepSeek, Fireworks, Together AI and xAI Completions), OpenAI and xAI Responses, Anthropic (client and server tool calls) and Bedrock Converse. Google and Ollama still send one complete snapshot per call.
  • A tool call's id, name and index go on content-block-start, which carries no args. An id or name that arrives later reaches consumers on content-block-finish.
  • If you read raw events and their block-delta fields.args mid-stream, or write agent stream transformers that do, read tool-call-delta instead. await stream, .toolCalls and the assembled message are unchanged.
  • Upgrade your clients first. Older clients mishandle tool-call-delta. You need @langchain/langgraph-sdk 1.12.6 or later; @langchain/react, @langchain/vue, @langchain/svelte or @langchain/angular 1.2.6 or later; and langchain-core 1.6.13 or later for Python clients.
  • To hold back, pin @langchain/core below 1.3.0. That also keeps you on @langchain/openai, @langchain/anthropic, @langchain/aws and @langchain/xai versions that send snapshots.

@​langchain/anthropic@​1.5.13

Patch Changes

• #11922 d71c542 Thanks @​richardscarrott! - feat(anthropic): declare mid-conversation support in model profiles. Claude Opus 4.8, Opus 5, Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5 and 5.1, and Mythos 5 and 5.1 set midConversationSystemMessages and midConversationTools. Sonnet 5 sets system messages only.

  Regenerating the profiles from models.dev also adds Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, removes Claude Opus 4.1, and lowers Claude Sonnet 4.5's maxInputTokens to 200,000.

@​langchain/anthropic@​1.5.12

Patch Changes

• #11707 4109812 Thanks @​richardscarrott! - feat(anthropic): support mid-conversation tool changes. A SystemMessage can now carry Anthropic's tool_addition and tool_removal blocks, either bare in content or wrapped in a non_standard block, and they reach the provider verbatim. The required beta is added automatically: mid-conversation-tool-changes-2026-07-01 for tools named by reference, or inline-tools-2026-09-15 when a tool_addition defines a tool by value (tool: { type: "tool_definition", definition }).

  System content is now narrowed to what Anthropic accepts there. text, tool_addition and tool_removal are sent; any other block is dropped instead of being forwarded for the provider to reject. A system message left with no content is omitted.

Commits

• 38eb4d8 chore: version packages (#12018)
• dae3b1f chore(core): fill in client versions in tool-call-delta changeset (#12019)
• 2caf64d feat(core,openai,anthropic,aws,xai): stream tool-call args as `tool-call-delt...
• 8a008ea test(core): accept either order of a chunk's stream events in chat model chai...
• f65acc3 chore(deps): restore missing @modelcontextprotocol/core@2.3.0 lockfile entr...
• 6a0ab38 chore: version packages (#12008)
• 654c7c6 chore(deps-dev): bump @​modelcontextprotocol/server from 2.3.0 to 2.3.1 (#12005)
• f096a32 fix(langchain): deliver final answers after todo updates (#11989)
• c690992 fix: update Node.js to October 2026 security release (#12007)
• 0c5e7bc fix(langchain): avoid full-state copies in parallel tool tasks (#11990)
• Additional commits viewable in compare view

Updates @langchain/aws from 1.4.4 to 1.5.1

Release notes

Sourced from @​langchain/aws's releases.

@​langchain/aws@​1.5.1

Patch Changes

• #11463 52d2ba8 Thanks @​syedsohailhussain1! - fix(aws): use this.caller in ChatBedrockConverse to support maxRetries, timeout, and onFailedAttempt

@​langchain/aws@​1.5.0

Minor Changes

• #11986 2caf64d Thanks @​richardscarrott! - feat: stream tool-call args as tool-call-delta slices

  Tool-call and server-tool-call args now stream as tool-call-delta events, which carry only the new slice of the args, instead of block-delta snapshots of the args so far. A large tool call now costs about the same bytes on the wire as the same length of text.

  • This covers the compat bridge (any model that only implements _streamResponseChunks), the OpenAI-compatible converter (OpenAI Completions, Groq, Mistral, IBM, OpenRouter, Perplexity, DeepSeek, Fireworks, Together AI and xAI Completions), OpenAI and xAI Responses, Anthropic (client and server tool calls) and Bedrock Converse. Google and Ollama still send one complete snapshot per call.
  • A tool call's id, name and index go on content-block-start, which carries no args. An id or name that arrives later reaches consumers on content-block-finish.
  • If you read raw events and their block-delta fields.args mid-stream, or write agent stream transformers that do, read tool-call-delta instead. await stream, .toolCalls and the assembled message are unchanged.
  • Upgrade your clients first. Older clients mishandle tool-call-delta. You need @langchain/langgraph-sdk 1.12.6 or later; @langchain/react, @langchain/vue, @langchain/svelte or @langchain/angular 1.2.6 or later; and langchain-core 1.6.13 or later for Python clients.
  • To hold back, pin @langchain/core below 1.3.0. That also keeps you on @langchain/openai, @langchain/anthropic, @langchain/aws and @langchain/xai versions that send snapshots.

@​langchain/aws@​1.4.7

Patch Changes

• #11916 73a739c Thanks @​zduric-langchain! - Report Bedrock thinking tokens as output_token_details.reasoning

  ChatBedrockConverse now accepts additionalModelResponseFieldPaths (constructor and call option, matching additional_model_response_field_paths in Python) and forwards it to Converse and ConverseStream. When the response includes Anthropic's native usage (additionalModelResponseFields.usage.output_tokens_details.thinking_tokens, returned when requesting ["/usage"]), the count is mapped to usage_metadata.output_token_details.reasoning for invoke(), .stream() and streamEvents(). output_tokens is unchanged, and nothing changes when the field is absent.

Commits

• e8c5dc8 chore: version packages (#12047)
• b5135cc feat(openai,langchain): support Decisions API and agent middleware (#11930)
• 52d2ba8 fix(aws): use this.caller in ChatBedrockConverse to support retries (#11463)
• 178af31 fix(core): deliver callbacks once to a handler registered twice by merged con...
• 304d190 fix(core): make drawMermaid node metadata labels safe (#12048)
• 50343ba fix(core): include node metadata in Graph.toJSON() (#12042)
• d1d0692 chore: version packages (#12044)
• 95b191b feat(openai): refresh model profiles from models.dev (#12046)
• 81577b3 fix(langchain): re-throw a drained tool from the agent's ToolNode (#12043)
• 4cbf0f5 chore(deps): bump the npm_and_yarn group across 2 directories with 1 update (...
• Additional commits viewable in compare view

Updates @langchain/core from 1.2.9 to 1.3.1

Release notes

Sourced from @​langchain/core's releases.

@​langchain/core@​1.3.1

Patch Changes

• #12026 178af31 Thanks @​eliornl! - fix(core): deliver callbacks once to a handler registered twice by merged configs

  Merging two runnable configs that both carry the same callback handler, as happens when an agent's streamEvents() runs inside a LangGraph node with tracing on, registered that handler twice, so it received every callback twice: every streamed chunk arrived twice. The configured callback manager now keeps each handler once, in order, including a handler passed as a plain object of methods that merged configs wrapped twice.

• #12042 50343ba Thanks @​sera-wi! - fix(core): include node metadata in Graph.toJSON() output when set, matching Python

• #12048 304d190 Thanks @​sera-wi! - fix(core): escape drawMermaid node metadata labels and leave out object values, so metadata with brackets or objects no longer breaks Mermaid parsing

@​langchain/core@​1.3.0

Minor Changes

• #11986 2caf64d Thanks @​richardscarrott! - feat: stream tool-call args as tool-call-delta slices

  Tool-call and server-tool-call args now stream as tool-call-delta events, which carry only the new slice of the args, instead of block-delta snapshots of the args so far. A large tool call now costs about the same bytes on the wire as the same length of text.

  • This covers the compat bridge (any model that only implements _streamResponseChunks), the OpenAI-compatible converter (OpenAI Completions, Groq, Mistral, IBM, OpenRouter, Perplexity, DeepSeek, Fireworks, Together AI and xAI Completions), OpenAI and xAI Responses, Anthropic (client and server tool calls) and Bedrock Converse. Google and Ollama still send one complete snapshot per call.
  • A tool call's id, name and index go on content-block-start, which carries no args. An id or name that arrives later reaches consumers on content-block-finish.
  • If you read raw events and their block-delta fields.args mid-stream, or write agent stream transformers that do, read tool-call-delta instead. await stream, .toolCalls and the assembled message are unchanged.
  • Upgrade your clients first. Older clients mishandle tool-call-delta. You need @langchain/langgraph-sdk 1.12.6 or later; @langchain/react, @langchain/vue, @langchain/svelte or @langchain/angular 1.2.6 or later; and langchain-core 1.6.13 or later for Python clients.
  • To hold back, pin @langchain/core below 1.3.0. That also keeps you on @langchain/openai, @langchain/anthropic, @langchain/aws and @langchain/xai versions that send snapshots.

@​langchain/core@​1.2.20

Patch Changes

• #11922 d71c542 Thanks @​richardscarrott! - feat(core): add midConversationSystemMessages and midConversationTools to ModelProfile. They say whether a chat model sends a system message after the first turn in place, and whether it can add a tool through a message rather than the request's tool list. Either way the prompt prefix before it stays unchanged.

@​langchain/core@​1.2.19

Patch Changes

• #11982 5af4684 Thanks @​richardscarrott! - feat(core): read tool-call-delta in ChatModelStream

  • ChatModelStream appends tool-call-delta slices to a tool call's args, and still replaces them on block-delta
  • It ignores delta types it doesn't recognize instead of throwing Unknown delta type

@​langchain/core@​1.2.18

Patch Changes

• #11956 93d62b0 Thanks @​hntrl! - Require langsmith >=0.11.0 to include the restored NullableHeaders type declarations.

• #11822 d3e6fe6 Thanks @​dqbd! - Add explicit agent address support to LangChainTracer, inherit project and address destinations together, and delegate default destination resolution to the LangSmith SDK.

• #11150 cea15cf Thanks @​mdnanocom! - fix(core): defer tool call chunk collapsing in AIMessageChunk to fix O(n²) streaming aggregation

  AIMessageChunk eagerly re-ran collapseToolCallChunks in its constructor on every streamed delta concat, re-parsing all accumulated tool args each time — O(n²) over a stream, which can block the event loop for seconds on long tool-call streams. Collapsing now runs lazily, once, on first read of tool_calls / invalid_tool_calls.

@​langchain/core@​1.2.17

Patch Changes

... (truncated)

Commits

• e8c5dc8 chore: version packages (#12047)
• b5135cc feat(openai,langchain): support Decisions API and agent middleware (#11930)
• 52d2ba8 fix(aws): use this.caller in ChatBedrockConverse to support retries (#11463)
• 178af31 fix(core): deliver callbacks once to a handler registered twice by merged con...
• 304d190 fix(core): make drawMermaid node metadata labels safe (#12048)
• 50343ba fix(core): include node metadata in Graph.toJSON() (#12042)
• d1d0692 chore: version packages (#12044)
• 95b191b feat(openai): refresh model profiles from models.dev (#12046)
• 81577b3 fix(langchain): re-throw a drained tool from the agent's ToolNode (#12043)
• 4cbf0f5 chore(deps): bump the npm_and_yarn group across 2 directories with 1 update (...
• Additional commits viewable in compare view

Updates @langchain/google from 0.2.3 to 0.2.12

Release notes

Sourced from @​langchain/google's releases.

@​langchain/google@​0.2.12

Patch Changes

• #11968 84eca2b Thanks @​afirstenberg! - refactoring in preparation for Interactions API

  • Protocol-specific operations are now done through GoogleChatApiHandler subclass
  • Converters for generate_content and shared converters
  • Types moved into protocol-specific generated file
  • Includes support Gemini 4 Argon parity (thinking config, model profile, usage-only streaming)

@​langchain/google@​0.2.10

Patch Changes

• #11696 824d07b Thanks @​dvd233! - Support Gemini's VALIDATED function-calling mode through tool_choice.

@​langchain/google@​0.2.9

Patch Changes

• #11806 241a046 Thanks @​thushanth-bengre-langchain! - fix(google): send thinkingLevel LOW instead of MINIMAL to Gemini 3 models that reject MINIMAL (e.g. gemini-3.7-flash, gemini-3.8-flash), so thinkingBudget: 0 and reasoningEffort: "minimal" no longer 400

Changelog

Sourced from @​langchain/google's changelog.

0.2.12

Patch Changes

• #11968 84eca2b Thanks @​afirstenberg! - refactoring in preparation for Interactions API

  • Protocol-specific operations are now done through GoogleChatApiHandler subclass
  • Converters for generate_content and shared converters
  • Types moved into protocol-specific generated file
  • Includes support Gemini 4 Argon parity (thinking config, model profile, usage-only streaming)

0.2.11

Patch Changes

• #11940 57d0e02 Thanks @​hntrl! - Support Gemini 4 Argon model capabilities and Gemini 3+ compatibility; preserve usage-only streaming chunks.

0.2.10

Patch Changes

• #11696 824d07b Thanks @​dvd233! - Support Gemini's VALIDATED function-calling mode through tool_choice.

0.2.9

Patch Changes

• #11806 241a046 Thanks @​thushanth-bengre-langchain! - fix(google): send thinkingLevel LOW instead of MINIMAL to Gemini 3 models that reject MINIMAL (e.g. gemini-3.7-flash, gemini-3.8-flash), so thinkingBudget: 0 and reasoningEffort: "minimal" no longer 400

0.2.8

Patch Changes

• #11641 12626c2 Thanks @​ccurme! - Add LANGSMITH_GATEWAY direct routing for Bedrock Converse and Vertex AI chat models.

• #11680 29caa17 Thanks @​jackjin1997! - Fix service account authentication on the Node client. NodeApiClient.fetch interpolated the unresolved access token promise, sending Bearer [object Promise], and the JWT it exchanges carried no scope claim, which Google requires on the assertion for an access token request. Credentials supplied through GOOGLE_CLOUD_CREDENTIALS now reach Vertex AI instead of failing with invalid credentials and then an invalid scope error.

0.2.7

Patch Changes

• #11664 16c5a5e Thanks @​jackjin1997! - Configure the required OAuth scopes when the Node.js client falls back to Application Default Credentials.

0.2.6

Patch Changes

• #11433 bc88b75 Thanks @​jackjin1997! - Fix Vertex AI endpoint routing for multi-region locations such as eu and us.

... (truncated)

Commits

• 8a3e1e5 chore: version packages (#11969)
• 2c40a63 refactor(google): Step 5 - Cleanup (#11936)
• 6711142 refactor(google): Step 4 - Decoupling (#11935)
• b4bb1e6 refactor(google): Step 3 - API Handler (#11934)
• 1d46259 refactor(google): Step 1 - Typegen Namespace Setup (#11932)
• f39d79b revert(google): support Argon thinking and preserve streaming usage (#11940) ...
• ba363b1 chore: version packages (#11937)
• 57d0e02 fix(google): support Argon thinking and preserve streaming usage (#11940)
• b2c1134 chore: version packages (#11881)
• 824d07b fix(google): support Gemini VALIDATED tool choice (#11696)
• Additional commits viewable in compare view

Updates @langchain/langgraph-checkpoint-sqlite from 1.0.4 to 1.1.0

Release notes

Sourced from @​langchain/langgraph-checkpoint-sqlite's releases.

@​langchain/langgraph-checkpoint-sqlite@​1.1.0

Minor Changes

• #2961 1631fd8 Thanks @​soarez! - order parallel writes by task path, like langgraph (Python) — one sort key for live execution and every saver, instead of the task-id order #2544 added for DeltaChannel writes

  When two or more tasks in one superstep write the same DeltaChannel, their writes are now applied and replayed in (task_path, task_id, idx) order — the order Python's langgraph#8544 established — instead of task-id order. Task ids are uuid5 hashes that include the checkpoint id, so the old order permuted parallel writers from run to run and disagreed with Python.

  • BaseCheckpointSaver.putWrites takes an optional taskPath (the serialized task path, taskPathStr), and getTuple/list must return pendingWrites in the new writesSortKey order. writesSortKey and compareWritesSortKeys are new exports of @langchain/langgraph-checkpoint; every saver (memory, SQLite, Postgres, Redis incl. ShallowRedisSaver, MongoDB) stores the path and sorts in JS — no database collation is involved.
  • SQLite and Postgres gain a task_path column (migrated automatically on setup(); read-only pre-column SQLite databases keep reading, with their rows keeping the old order). Redis and MongoDB documents gain a task_path field. Rows written before the upgrade keep their old order.
  • Exit-durability resumes replay delta writes in live order (port of langgraph#9114): the resumed superstep's writes interleave with the loaded ones by task path, later supersteps sort after every real write, and writes loaded with the checkpoint are no longer stored twice.
  • bulkUpdateState seals the first update superstep: DeltaChannels touched by the base checkpoint's pending writes (and whose version moved) snapshot on the update checkpoint, so a finished sibling's writes and the update's can't replay in the wrong order (partial port of langgraph#8548).
  • bulkUpdateState/updateState update tasks carry (__interrupt__, i) task paths, as in langgraph#9128, so several updates in one superstep writing the same DeltaChannel replay in the order given.
  • Send task paths become [PUSH, i, false] and node-error-handler paths become [*failedTask.path[:3], "node_error_handler", false] (the failed task's first three path elements), matching Python's shapes — langgraph_path task metadata changes accordingly, and task ids of functional task() calls under a Send or handler task change with them. The SDK's Send-task detection accepts both the old and new shapes.

  Upgrade notes:

  • PostgresSaver: run setup() once after upgrading, from one process, before the new version serves traffic. The task_path column comes from the first new migration since the saver shipped, and until it runs, reads and writes fail with column "task_path" does not exist. Several processes running setup() at once on a database that still needs the migration can fail with a duplicate key error on checkpoint_migrations.
  • Third-party checkpoint savers remain source-compatible but must now persist the taskPath they are given and return pendingWrites in writesSortKey order; the conformance suite's new base tests enforce this. A saver that ignores the new argument replays in task-id order and can rebuild a different DeltaChannel value than the run produced.
  • An incomplete superstep containing stored delta writes that must interleave with resumed or retried task writes is affected by the upgrade, whether the incompleteness came from an interrupt, a task error, a crash, or a drain: finish such threads with the old runtime before upgrading writers. A single interrupted node with no finished delta-writing siblings is safe to resume. The same applies to resuming work that includes functional task() calls under a Send or error-handler task, whose child task ids change.
  • @langchain/langgraph-checkpoint 1.2.0 must be released before any package requiring its new exports; the consumers' minimum ranges are raised accordingly.

Changelog

Sourced from @​langchain/langgraph-checkpoint-sqlite's changelog.

1.1.0

Minor Changes

• #2961 1631fd8 Thanks @​soarez! - order parallel writes by task path, like langgraph (Python) — one sort key for live execution and every saver, instead of the task-id order #2544 added for DeltaChannel writes

  When two or more tasks in one superstep write the same DeltaChannel, their writes are now applied and replayed in (task_path, task_id, idx) order — the order Python's langgraph#8544 established — instead of task-id order. Task ids are uuid5 hashes that include the checkpoint id, so the old order permuted parallel writers from run to run and disagreed with Python.

  • BaseCheckpointSaver.putWrites takes an optional taskPath (the serialized task path, taskPathStr), and getTuple/list must return pendingWrites in the new writesSortKey order. writesSortKey and compareWritesSortKeys are new exports of @langchain/langgraph-checkpoint; every saver (memory, SQLite, Postgres, Redis incl. ShallowRedisSaver, MongoDB) stores the path and sorts in JS — no database collation is involved.
  • SQLite and Postgres gain a task_path column (migrated automatically on setup(); read-only pre-column SQLite databases keep reading, with their rows keeping the old order). Redis and MongoDB documents gain a task_path field. Rows written before the upgrade keep their old order.
  • Exit-durability resumes replay delta writes in live order (port of langgraph#9114): the resumed superstep's writes interleave with the loaded ones by task path, later supersteps sort after every real write, and writes loaded with the checkpoint are no longer stored twice.
  • bulkUpdateState seals the first update superstep: DeltaChannels touched by the base checkpoint's pending writes (and whose version moved) snapshot on the update checkpoint, so a finished sibling's writes and the update's can't replay in the wrong order (partial port of langgraph#8548).
  • bulkUpdateState/updateState update tasks carry (__interrupt__, i) task paths, as in langgraph#9128, so several updates in one superstep writing the same DeltaChannel replay in the order given.
  • Send task paths become [PUSH, i, false] and node-error-handler paths become [*failedTask.path[:3], "node_error_handler", false] (the failed task's first three path elements), matching Python's shapes — langgraph_path task metadata changes accordingly, and task ids of functional task() calls under a Send or handler task change with them. The SDK's Send-task detection accepts both the old and new shapes.

  Upgrade notes:

  • PostgresSaver: run setup() once after upgrading, from one process, before the new version serves traffic. The task_path column comes from the first new migration since the saver shipped, and until it runs, reads and writes fail with column "task_path" does not exist. Several processes running setup() at once on a database that still needs the migration can fail with a duplicate key error on checkpoint_migrations.
  • Third-party checkpoint savers remain source-compatible but must now persist the taskPath they are given and return pendingWrites in writesSortKey order; the conformance suite's new base tests enforce this. A saver that ignores the new argument replays in task-id order and can rebuild a different DeltaChannel value than the run produced.
  • An incomplete superstep containing stored delta writes that must interleave with resumed or retried task writes is affected by the upgrade, whether the incompleteness came from an interrupt, a task error, a crash, or a drain: finish such threads with the old runtime before upgrading writers. A single interrupted node with no finished delta-writing siblings is safe to resume. The same applies to resuming work that includes functional task() calls under a Send or error-handler task, whose child task ids change.
  • @langchain/langgraph-checkpoint 1.2.0 must be released before any package requiring its new exports; the consumers' minimum ranges are raised accordingly.

Commits

• 4510774 chore: version packages (#2977)
• 1631fd8 fix(langgraph): order parallel writes by task path, like langgraph (Python) (...
• 6fb9a3f chore(deps): bump vitest from 4.1.10 to 4.1.11 (#2818)
• See full diff in compare view

Updates @langchain/openai from 1.6.0 to 1.8.0

Release notes

Sourced from @​langchain/openai's releases.

@​langchain/openai@​1.8.0

Minor Changes

• #11930 b5135cc Thanks @​hntrl! - Add a beta OpenAIDecisions runnable with typed predicate, choice and score questions, plus model-routing and tool-risk agent middleware.

@​langchain/openai@​1.7.1

Patch Changes

• #12046 95b191b Thanks @​thushanth-bengre-langchain! - feat(openai): refresh model profiles from models.dev

  • Add GPT-6.1 Terra
  • Remove gpt-5.2-chat-latest and gpt-5.3-chat-latest, which OpenAI shut down on Aug 10, 2026

@​langchain/openai@​1.7.0

Minor Changes

• #11986 2caf64d Thanks @​richardscarrott! - feat: stream tool-call args as tool-call-delta slices

  Tool-call and server-tool-call args now stream as tool-call-delta events, which carry only the new slice of the args, instead of block-delta snapshots of the args so far. A large tool call now costs about the same bytes on the wire as the same length of text.

  • This covers the compat bridge (any model that only implements _streamResponseChunks), the OpenAI-compatible converter (OpenAI Completions, Groq, Mistral, IBM, OpenRouter, Perplexity, DeepSeek, Fireworks, Together AI and xAI Completions), OpenAI and xAI Responses, Anthropic (client and server tool calls) and Bedrock Converse. Google and Ollama still send one complete snapshot per call.
  • A tool call's id, name and index go on content-block-start, which carries no args. An id or name that arrives later reaches consumers on content-block-finish.
  • If you read raw events and their block-delta fields.args mid-stream, or write agent stream transformers that do, read tool-call-delta instead. await stream, .toolCalls and the assembled message are unchanged.
  • Upgrade your clients first. Older clients mishandle tool-call-delta. You need @langchain/langgraph-sdk 1.12.6 or later; @langchain/react, @langchain/vue, @langchain/svelte or @langchain/angular 1.2.6 or later; and langchain-core 1.6.13 or later for Python clients.
  • To hold back, pin @langchain/core below 1.3.0. That also keeps you on @langchain/openai, @langchain/anthropic, @langchain/aws and @langchain/xai versions that send snapshots.

@​langchain/openai@​1.6.3

Patch Changes

• #11922 d71c542 Thanks @​richardscarrott! - feat(openai): declare mid-conversation support in model profiles. GPT-4.1, GPT-5.x, GPT-6.x and o3 set Description has been truncated