fix(core): repair null streaming tool args from complete raw JSON

[Blue-Berrys]

Summary

• Repair ToolCallsAccumulator so complete raw tool-call JSON fills missing/null input values after streaming.
• Prevent later null input maps from overwriting previously accumulated non-null arguments.
• Add regression tests for HTML-like multi-line content and null-overwrite protection.

Closes #768

Description

When streaming tool-call arguments for complex multi-line content (e.g. HTML), early chunks can leave ToolUseBlock.input with keys present but null values. Because build() previously only parsed raw content when finalArgs was empty, those nulls were never repaired even though content held valid complete JSON.

Checklist

• Code has been formatted with mvn spotless:apply
• Targeted tests pass (ToolCallsAccumulatorTest, 15 tests)
• Related documentation has been updated (N/A)
• Code is ready for review

Testing

JAVA_HOME=/path/to/jdk21 mvn -pl agentscope-core -am \
  -Dspotless.check.skip=true \
  -Dtest=ToolCallsAccumulatorTest \
  -Dsurefire.failIfNoSpecifiedTests=false \
  test

[CLAassistant]

All committers have signed the CLA.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[oss-maintainer]

Review — Approved ✅

Good fix for the null-value streaming tool-call accumulation issue (#768).

Code analysis:

• merge() now correctly skips null values from partial input maps, preventing early stream chunks from wiping previously accumulated non-null arguments.
• build() is updated to always attempt raw JSON parsing (not just when finalArgs is empty), which is the right call — the final accumulated raw content is indeed the source of truth.
• The merge logic in build() correctly preserves existing non-null values while filling in missing/null keys from parsed JSON.

Tests:

• testRepairNullInputValuesFromCompleteRawContent — directly reproduces the HTML streaming scenario from #768.
• testNullInputDoesNotOverwriteNonNullArgs — validates the ordering invariant.

Both tests are well-structured and target the exact failure modes.

LGTM. Please sign the CLA.

[codecov]

Codecov Report

❌ Patch coverage is 78.57143% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...e/core/agent/accumulator/ToolCallsAccumulator.java	78.57%	1 Missing and 2 partials ⚠️

📢 Thoughts on this report? Let us know!