Failed server tool call is saved without `error`, so it reloads as complete

[aayushbtw]

TanStack AI version

@tanstack/ai 0.58.0. Also 0.55.0. The code is the same on main.

Framework/Library version

Framework-agnostic (Node 24)

Describe the bug and the steps to reproduce it

A server tool call fails when the handler throws, when the input fails its schema, or when the tool does not exist. chat() streams the TOOL_CALL_RESULT with metadata.tanstack.state: 'output-error'. The client then shows a tool-result part with state: 'error' and error, as the client tools docs describe.

But the tool message that chat() adds to its message history has no error:

{ role: 'tool', content: '{"error":"lookup is down"}', toolCallId: 'call_1' }

Middleware reads that history in onFinish, and @tanstack/ai-persistence saves it. On a reload, modelMessagesToUIMessages gives the tool-result part state: 'error' only when msg.error is set (messages.ts#L1128-L1141). As a result, the part comes back with state: 'complete' and no error. The UI shows the call as failed while the run streams, and as successful after a reload.

buildToolResultChunks uses result.state to tag the chunk, but not to build the message (index.ts#L3157-L3185).

Steps:

1. In the linked repo, run npm install && npm run repro. No API key is necessary. A scripted adapter makes the tool call.
2. For each of the three failures, the script prints the streamed metadata, the saved tool message, and the rebuilt tool-result part. This is the output for the handler that throws:

## handler throws
1. streamed TOOL_CALL_RESULT metadata: { tanstack: { state: 'output-error' } }
2. saved tool message: {
  role: 'tool',
  content: '{"error":"lookup is down"}',
  toolCallId: 'call_1'
}
3. rebuilt tool-result part: { state: 'complete', error: undefined }

Expected: the saved tool message has error, and the rebuilt part is { state: 'error', error: 'lookup is down' }. That is the same state that the stream processor sets live (processor.ts#L1655-L1682).

When result.state === 'output-error', set error on that message. This fixes all three cases in the repro. A PR follows.

Related: #1386 and #1401 are about denied and cancelled calls. #1401 adds toolResultOutcome to their metadata. It does not mark a plain failure. I ran a test of the three cases against the head of #1401, and all three still rebuild as complete. Both change the same block of code.

Your Minimal, Reproducible Example - (Sandbox Highly Recommended)

https://github.com/aayushbtw/tanstack-ai-tool-error-repro

Do you intend to try to help solve this bug with your own PR?

Yes, I am also opening a PR that solves the problem along side this issue

Terms & Code of Conduct

• I agree to follow this project's Code of Conduct
• I understand that if my bug cannot be reliable reproduced in a debuggable environment, it will probably not be fixed and this issue may even be closed.