Skip to content
Source
Sign inSign up
TanStack/ai

ci: Version Packages #2185

Mergedgithub-actions[bot](GitHub profile, opens in a new tab)merged 1 commit intomain fromrefs/pull/2185/headmerged by AlemTuzlak View on GitHub

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@tanstack/ai@1.6.0

Minor Changes

  • #2110 95c1d78 - Undo strict-mode nulls on tool arguments before the input check. An approved call whose history still has those nulls now runs when repairToolInput is off.

  • #2156 df6f5a9 - The OpenAI-compatible adapters support chat({ reasoning }). Add the levels of a model to its createModel definition (reasoning: { levels, budget }), or to the reasoning config of openaiCompatibleText. The adapter sends the level as reasoning_effort on Chat Completions, or as reasoning: { effort, summary } on Responses. chat() accepts only the levels of the model.

    OpenAICompatibleChatAdapter and OpenAICompatibleResponsesAdapter take the same levels in a new reasoning constructor option, and in their last type parameter. TypeScript shows an error when the option is missing or its levels are different from the type parameter. The new OpenAICompatibleAdapterOptions type is exported from @tanstack/ai-openai/compatible.

Patch Changes

  • #2163 40d1c61 - A chat loaded from stored messages now looks the same as the chat the user saw live.
    • A tool result that is a string stays a string, also when the string is JSON text. Before, a tool that returned '42' showed the number 42, live or after a reload. This applies to server tools, client tools, middleware skip results, and MCP text results from tools without an outputSchema. A server tool that returns JSON.stringify(value) now gives that string as part.output. Return the value itself to get it as an object.
    • Live messages no longer get the snapshot wire fields tanstack.createdAt and toolCallMetadata in their metadata.
    • A reload drops an assistant answer that holds only whitespace, as the live chat does.
    • When a run stops for a client tool, the messages that chat() stores keep the ids of the snapshot, so a reload gives tool results the same ids.
    • ThinkingPart.stepId is deprecated and is no longer set. Parts are the same live, stored, and reloaded.

@tanstack/ai-openai@1.2.0

Minor Changes

  • #2156 df6f5a9 - The OpenAI-compatible adapters support chat({ reasoning }). Add the levels of a model to its createModel definition (reasoning: { levels, budget }), or to the reasoning config of openaiCompatibleText. The adapter sends the level as reasoning_effort on Chat Completions, or as reasoning: { effort, summary } on Responses. chat() accepts only the levels of the model.

    OpenAICompatibleChatAdapter and OpenAICompatibleResponsesAdapter take the same levels in a new reasoning constructor option, and in their last type parameter. TypeScript shows an error when the option is missing or its levels are different from the type parameter. The new OpenAICompatibleAdapterOptions type is exported from @tanstack/ai-openai/compatible.

Patch Changes

@tanstack/ai-sandbox@1.4.0

Minor Changes

  • #2180 e5d9e0f - Delete native sandbox snapshots once nothing references them.

    SandboxProvider has a new optional deleteSnapshot({ snapshotId, signal }). defineSandbox calls it, best effort, when a sandbox record's latestSnapshotId is replaced (a re-create or a new after-run snapshot) and when the sandbox is torn down through destroy or reclaimSandbox. A failed delete is logged and never fails the run or the teardown. Providers that don't implement it behave as before.

    railwaySandbox() implements deleteSnapshot, so its tsai-* checkpoints are deleted once nothing references them instead of staying in the environment until pruned by hand.

Patch Changes

  • #2109 2ed7e21 - A reconnect during a long tool call no longer writes the rest of a durable sandbox run twice. The host that started the run holds the driver claim until the run ends, so the reconnect tails the log.

  • #2108 8813243 - Keep text and tool calls the first host already delivered in the saved transcript after a sandbox takeover. The event log still stores that prefix once.

  • Updated dependencies [95c1d78, df6f5a9, 40d1c61]:

    • @tanstack/ai@1.6.0
    • @tanstack/ai-harness@0.3.0
    • @tanstack/ai-persistence@1.0.1

@tanstack/ai-sandbox-railway@1.1.0

Minor Changes

  • #2180 e5d9e0f - Delete native sandbox snapshots once nothing references them.

    SandboxProvider has a new optional deleteSnapshot({ snapshotId, signal }). defineSandbox calls it, best effort, when a sandbox record's latestSnapshotId is replaced (a re-create or a new after-run snapshot) and when the sandbox is torn down through destroy or reclaimSandbox. A failed delete is logged and never fails the run or the teardown. Providers that don't implement it behave as before.

    railwaySandbox() implements deleteSnapshot, so its tsai-* checkpoints are deleted once nothing references them instead of staying in the environment until pruned by hand.

Patch Changes

@tanstack/openai-base@1.1.0

Minor Changes

  • #2110 95c1d78 - Undo strict-mode nulls on tool arguments before the input check. An approved call whose history still has those nulls now runs when repairToolInput is off.

Patch Changes

@tanstack/ai-bedrock@1.4.1

Patch Changes

  • #2078 fc93edc - Report a Converse stream that stops with stopReason: 'model_context_window_exceeded' as a RUN_ERROR with the context_overflow code. It was reported as a clean RUN_FINISHED with finishReason: 'stop', although the answer was cut off. This matches the Anthropic adapter.

  • Updated dependencies [95c1d78, df6f5a9, 40d1c61]:

    • @tanstack/ai@1.6.0
    • @tanstack/openai-base@1.1.0

@tanstack/ai-claude-code@1.0.3

Patch Changes

  • #2108 8813243 - Keep text and tool calls the first host already delivered in the saved transcript after a sandbox takeover. The event log still stores that prefix once.

  • Updated dependencies [95c1d78, df6f5a9, 2ed7e21, 40d1c61, e5d9e0f, 8813243]:

    • @tanstack/ai@1.6.0
    • @tanstack/ai-sandbox@1.4.0

@tanstack/ai-codex@1.0.2

Patch Changes

  • #2108 8813243 - Keep text and tool calls the first host already delivered in the saved transcript after a sandbox takeover. The event log still stores that prefix once.

  • Updated dependencies [95c1d78, df6f5a9, 2ed7e21, 40d1c61, e5d9e0f, 8813243]:

    • @tanstack/ai@1.6.0
    • @tanstack/ai-sandbox@1.4.0

@tanstack/ai-grok-build@1.0.2

Patch Changes

  • #2108 8813243 - Keep text and tool calls the first host already delivered in the saved transcript after a sandbox takeover. The event log still stores that prefix once.

  • Updated dependencies [95c1d78, df6f5a9, 2ed7e21, 40d1c61, e5d9e0f, 8813243]:

    • @tanstack/ai@1.6.0
    • @tanstack/ai-sandbox@1.4.0
    • @tanstack/ai-acp@1.0.2

@tanstack/ai-isolate-daytona@1.1.1

Patch Changes

  • #2170 24ad184 - Pin @daytona/sdk to ^0.231.1 in the install command. The usage example calls daytona.create(), which @daytona/sdk 2.0 moves to daytona.v0, so an unpinned install breaks the example once 2.0 becomes the default.

  • Updated dependencies []:

    • @tanstack/ai-code-mode@1.1.1

@tanstack/ai-mistral@1.1.3

Patch Changes

  • #2183 4c614ad - Mistral streams that end with finish_reason: "error" now end the run with RUN_ERROR instead of a clean RUN_FINISHED.

  • Updated dependencies [95c1d78, df6f5a9, 40d1c61]:

    • @tanstack/ai@1.6.0
    1. Pushed3ba3a1c

    2. Pushedd207e2a

    3. Pushed07d3303

    4. Pushed531eff7

    5. AlemTuzlak approved these changes

    6. AlemTuzlak merged this48c9ff4

    7. AlemTuzlak closed this

Sign in to join this conversation.