fix(core): prevent premature tool_call parsing on empty SSE args

[Giulio Leone (giulio-leone)]

Description

Fixes #35514

When streaming tool calls via SSE, some LLM providers (OpenRouter, DeepSeek) split tool_call arguments across multiple chunks. The first chunk carries name + id but empty args (""). Previously, the empty string fell through to else {} in init_tool_calls, producing a valid tool_calls entry with args={}. Downstream code (ToolNode, user streaming handlers) treated this as a complete tool call and executed the tool prematurely with empty arguments.

Root Cause

In AIMessageChunk.init_tool_calls() (line 543):

args_ = parse_partial_json(chunk["args"]) if chunk["args"] else {}

chunk["args"] is "" (empty string) → falsy → args_ = {} → creates a valid tool_call with empty args.

Fix

Skip tool_call_chunks whose args is explicitly the empty string "":

• args="" (empty string): streaming not started yet → skip (no premature tool_call)
• args=None: no args info (backward compat) → treat as {} (unchanged)
• args="{}" (explicit empty dict): valid no-arg tool call → parse normally (unchanged)

The final accumulated chunk (after merge_lists concatenation across all SSE fragments) carries the full args string and parses correctly.

Verification

# Before fix: chunk1.tool_calls = [{name: "my_tool", args: {}}]  ← WRONG
# After fix:  chunk1.tool_calls = []                              ← CORRECT

# Accumulated (chunk1 + chunk2 + chunk3 + chunk4):
# tool_calls = [{name: "my_tool", args: {"url": "http://example.com"}}]  ← CORRECT (unchanged)

Tests

• Updated existing test expectation for empty-string args behavior
• Added regression test test_sse_fragmented_tool_calls_no_premature_parse — verifies individual chunks don't produce premature tool_calls but accumulated result is correct
• Added test_no_arg_tool_call_still_works — verifies tools with args="{}" still work
• All 1703 core unit tests pass
• All 308 OpenAI partner unit tests pass

Issue link

Fixes #35514

[codspeed-hq]

Merging this PR will degrade performance by 10.56%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

❌ 1 regressed benchmark
✅ 12 untouched benchmarks
⏩ 23 skipped benchmarks¹

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	test_import_time[InMemoryVectorStore]	637.1 ms	712.3 ms	-10.56%

---

_{Comparing giulio-leone:fix/streaming-tool-call-empty-args-35514 (19355f7) with master (527fc02)}

Footnotes

1. 23 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[Giulio Leone (giulio-leone)]

Friendly ping — CI is green, tests pass, rebased on latest. Ready for review whenever convenient. Happy to address any feedback. 🙏

[Giulio Leone (giulio-leone)]

Friendly ping — rebased on latest and ready for review. Happy to address any feedback!