Fix ChatReqLLM streaming tool-call merge for OpenAI-style chunks

[brainlid]

Problem

Streaming tool calls through ChatReqLLM against OpenAI-compatible backends (direct OpenAI, LiteLLM proxy, and similar) fail at finalize with:

delta_conversion_failed: "Error applying delta message: \"tool_calls: call_id: can't be blank; name: can't be blank\""

The error surfaces only when the model invokes a tool while streaming. Non-streaming calls and Anthropic streaming were unaffected.

The root cause is an index mismatch during MessageDelta merging. ReqLLM's default OpenAI decoder emits two kinds of chunks for a single tool call:

1. An initial %StreamChunk{type: :tool_call, name: ..., arguments: %{}, metadata: %{id: ..., index: ...}} chunk.
2. A stream of %StreamChunk{type: :meta, metadata: %{tool_call_args: %{index: ..., fragment: "..."}}} argument-fragment chunks.

ChatReqLLM.translate_stream_chunk/1 was reading id from the initial chunk's metadata but dropping index, and was emitting the call as :complete with arguments: %{}. Fragment chunks then arrived with index: 0 and no call_id/name. Because MessageDelta.merge_tool_calls/2 keys lookups by index, the fragments did not merge into the initial tool call — they landed in a separate slot. At finalize, Message.new ran ToolCall.complete/1 on every entry, and the orphan fragment ToolCall failed validate_required([:call_id, :name]).

Solution

Added a new process_stream_chunk/2 clause that handles OpenAI-shaped initial :tool_call chunks (those whose metadata carries :index). The clause:

• Carries :index through onto the ToolCall, so subsequent fragment chunks merge into the same slot.
• When arguments is an empty map (placeholder; fragments will follow), emits the ToolCall as :incomplete with arguments: nil. ToolCall.append_arguments/2 then string-concatenates the incoming JSON fragments, and ToolCall.complete/1 JSON-decodes the accumulated string at finalize.
• When arguments is already a non-empty map (single-shot delivery — no fragments will follow), emits the ToolCall as :complete with the args inline.

The existing Anthropic %{start: true} clause still matches first, so Anthropic streaming is unchanged. Chunks without :index in metadata still fall through to the original translate_stream_chunk/1, preserving the previous single-shot behaviour for providers that deliver the whole call in one chunk.

Changes

• lib/chat_models/chat_req_llm.ex — Added OpenAI-style :tool_call chunk clause to process_stream_chunk/2; preserves :index from metadata and chooses :incomplete/:complete based on whether args fragments are expected to follow.
• test/chat_models/chat_req_llm_test.exs — Two regression tests:
  • Fragmented case: initial chunk with empty args + index + id, followed by two tool_call_args fragments, then a terminal :meta. Asserts the assembled Message has a single ToolCall with the correct call_id, name, and JSON-decoded arguments.
  • Single-shot case: initial chunk already carrying non-empty args. Asserts the assembled ToolCall is :complete with the inline args.

Testing

• mix test test/chat_models/chat_req_llm_test.exs — 85 tests, 0 failures (7 live-tagged tests excluded).
• mix test — full suite passes: 31 doctests, 1767 tests, 0 failures.
• No live API tests were added for this path because the failure mode requires an OpenAI-compatible streaming backend (and was originally observed through a LiteLLM proxy that the test harness is not wired up to talk to). The regression tests exercise the exact StreamChunk sequence ReqLLM emits, which is what ChatReqLLM actually consumes.