Merge streaming tool call deltas by index in OpenAiChatModel

[jewoodev]

OpenAI-compatible streaming endpoints can send continuation tool-call deltas with id: "" instead of omitting id. OpenAiChatModel.ChunkMerger treats any present id as a new tool-call boundary, so argument fragments become separate ToolCalls with empty id/name and tool execution receives empty or null arguments.

This PR merges tool-call deltas by the required index field instead. It fixes the empty-id case from gh-6374, keeps OpenAI-style omitted ids working, and removes the one-tool-call-per-chunk assumption so parallel or multi-delta chunks merge correctly. #6369 covers the empty-id continuation case; this PR uses the index-keyed path so those related cases stay covered together.

Tests cover omitted ids, empty ids, parallel tool calls, and multiple deltas per chunk. Verified with ./mvnw -pl models/spring-ai-openai package.

Fixes gh-6374

[jewoodev]

Rebased onto main to resolve a conflict with #6365.

#6365 changed mergeDeltas to preserve tool-call additional properties. This PR replaces that merge logic with by-index merging, so I carried the preservation over: ChunkMerger.mergeToolCalls now copies each delta's additional properties via putAllAdditionalProperties.

The streaming test added in #6365 (preserveUnmappedToolCallAdditionalProperties) verifies this. It fails without that call and passes with it, along with the rest of the chat model tests.

[freshman-wang]

Thanks for this PR — the index-based merging approach is much cleaner than my earlier attempt.

While testing with DashScope/Qwen, I also hit two NoSuchElementException cases in the same ChunkMerger that this PR does not yet cover. It would be great if they could be addressed here (or in a follow-up) so the streaming path is fully robust:

1. mergeDeltas — lastFromTc1.function().get()

When a continuation chunk carries arguments only (no function object), lastFromTc1.function() is empty and .get() throws:

// current
Function lastFromTc1F = lastFromTc1.function().get();

Suggested:

Function lastFromTc1F = lastFromTc1.function()
    .orElseGet(() -> Function.builder().build());

2. chunkToChatCompletion — tc.id().get(), tc.function().get().name().get(), tc.function().get().arguments().get()

All four .get() calls can throw when the corresponding fields are absent in the delta:

// current
toolCallBuilder.id(tc.id().get());
toolCallBuilder.function(ChatCompletionMessageFunctionToolCall.Function.builder()
    .name(tc.function().get().name().get())
    .arguments(tc.function().get().arguments().get())
    .build());

Suggested:

toolCallBuilder.id(tc.id().orElse(""));
ChatCompletionMessageFunctionToolCall.Function.Builder funcBuilder = ChatCompletionMessageFunctionToolCall.Function.builder();
tc.function().ifPresent(f -> {
    funcBuilder.name(f.name().orElse(""));
    funcBuilder.arguments(f.arguments().orElse(""));
});
toolCallBuilder.function(funcBuilder.build());

Reproduced on Spring AI 2.0.0 GA with DashScope Qwen3 streaming + tool_calls. Happy to open a separate issue if that would help track it. Thanks!

(Originally submitted as #6522, closed in favor of this PR and #6373.)

[freshman-wang]

@jewoodev Thanks for the great work on this PR!

Would you be open to me adding the NoSuchElementException fixes directly to this branch as a follow-up commit? That way the reviewer can see all the ChunkMerger fixes together.

If yes, could you add me as a collaborator (Write permission) to your fork at https://github.com/jewoodev/spring-ai/settings/access ? My GitHub username is freshman-wang.

If you prefer to keep the PR scope tight, I'm happy to wait and open a follow-up PR after this one merges. Just let me know which you prefer. Thanks!

[jewoodev]

Thanks for the feedback.

Added a small follow-up commit that guards the remaining chunkToChatCompletion Optional.get() calls and covers both arguments-only tool call deltas and deltas without a function object.

The mergeDeltas case is already covered by the by-index merge in this branch, since mergeToolCalls now builds a function when the previous delta does not carry one.

I'll keep follow-up changes on this branch on my side to keep the PR history straightforward. Feel free to leave any additional feedback in this PR.

[freshman-wang]

@sdeleuze Could you take a look at this PR when you have a chance? It fixes streaming tool-call merging for empty id continuations from OpenAI-compatible endpoints. The index-keyed approach handles parallel tool calls and multi-delta chunks correctly. Thanks!