e2e-chatbot-app-next: duplicate/unresolved tool parts can reach the provider and trigger 400s (tool_use / tool_result mismatch)

[AndreasJaeck]

Summary

We run a fork of e2e-chatbot-app-next against Databricks model serving endpoints and periodically hit provider 400s caused by a tool_use / tool_result mismatch in the converted model messages.

Caveat up front: I have not reproduced this against the unmodified template. Please treat this as a field report plus an inferred mechanism rather than a confirmed template bug. I am filing it because the relevant code path looks unguarded upstream and the mitigation is cheap.

Symptom

Anthropic-family endpoints reject the request with a 400 complaining that tool_use blocks have no corresponding tool_result blocks. There is an OpenAI / AI Gateway flavour of the same failure. It appears mid-conversation, after tool calls have happened, and tends to recur for the affected thread because the bad state has been persisted.

Inferred mechanism

server/src/routes/chat.ts L251 calls:

const modelMessages = await convertToModelMessages(uiMessages);

with no options and no de-duplication. uiMessages is assembled from persisted DB messages and/or requestBody.previousMessages (L169-198). Two states appear able to survive that round-trip:

1. Duplicate tool-<name> parts sharing one toolCallId, at different states.
2. Parts stuck at input-available — a call with no result, e.g. the stream ended early, the tool errored, or an approval was denied.

Either one produces a tool_use with no matching tool_result once converted.

Mitigation we use

Two changes, both cheap and drop-only:

1. De-duplicate UI tool parts by toolCallId before conversion, keeping the most-resolved state (output-available > output-error > input-available > input-streaming). Never synthesise a part, only drop.
2. Pass ignoreIncompleteToolCalls: true to convertToModelMessages.

Since adding those we have not seen the 400 recur.

Ask

Would you consider guarding this path upstream? Happy to send a PR if that would be useful. Equally, if you already know this is handled elsewhere or is specific to how our fork persists messages, feel free to close.

Version

Template e2e-chatbot-app-next at ab30424be796009d34a0b760c25395211fe08e91