fix: accumulate ChatUsage across Anthropic streaming events

[waterWang]

Fixes #2094

Problem

When AnthropicChatModel is used with streaming enabled, every ChatResponse has usage == null. The Anthropic streaming protocol sends token usage in message_start (input_tokens) and message_delta (output_tokens) events, but these events have empty content blocks and are filtered out by the .filter(response -> !response.getContent().isEmpty()) in parseStreamEvents().

Fix

1. parseStreamEvent(): Parse inputTokens from the message_start event's message.usage.input_tokens and create a ChatUsage object.

2. parseStreamEvents(): Use a mutable accumulator array to merge usage across events:

   • message_start → sets inputTokens
   • message_delta → sets outputTokens
   • content_block_delta → attaches the accumulated usage when content is emitted

This ensures that every content-bearing event carries the complete token usage (input + output tokens), while the filter still correctly removes empty events.

[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]

Code Review — PR #2467

fix: accumulate ChatUsage across Anthropic streaming events

Fixes #2094

Summary

Fixes ChatUsage being null in streaming responses from AnthropicChatModel by accumulating token usage from message_start (input tokens) and message_delta (output tokens) events.

Analysis

The approach is correct:

1. parseStreamEvent() already extracts usage from individual events
2. parseStreamEvents() now uses a mutable ChatUsage[] accumulator to merge usage across events
3. Content-bearing events get the accumulated usage attached before emission

The merge logic handles the case where message_start provides inputTokens first, then message_delta adds outputTokens, and subsequent content events carry the full accumulated usage.

Observations

• ✅ Correctly identifies that message_start and message_delta have empty content blocks and are filtered out
• ✅ Accumulator pattern is appropriate for the stateful nature of streaming usage
• ✅ Fixes a real user-facing issue (no token usage tracking for Anthropic streaming)
• ❌ CI failed — Spotless formatting in AnthropicResponseParser.java. Run mvn spotless:apply to fix.
• ⚠️ Minor: the ChatUsage[] array as mutable accumulator works but could be replaced with AtomicReference<ChatUsage> for clarity in concurrent contexts (though flatMap here is sequential per-event, so the array is fine).

Required Action

@waterWang Please run mvn spotless:apply in the agentscope-extensions-model-anthropic module and push the fix.

Verdict

Request Changes — logic is correct, but Spotless formatting must be fixed before merge.