bug: Anthropic 类型供应商输出 token 计数错误

[ding113]

问题描述：
当前 v0.3.27 版本中，引入了 5m/1h 缓存区别计费功能。因此修改了提取 Anthropic 类型供应商的计费提取逻辑。
当前版本使用请求体的消息开始字段提取 token 用量数据，如：

event:message_start
data:{"type":"message_start","message":{"model":"claude-opus-4-5-20251101","id":"msg_01K63dMRtugrT9KH6hURm3Vv","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":8,"cache_creation_input_tokens":800,"cache_read_input_tokens":77249,"cache_creation":{"ephemeral_5m_input_tokens":800,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}}          }

这导致了输出 token 计数错误。因为这一字段位于流式输出响应体开头，此时主要的输出内容还未完成输出。

正确的方式是，提取流式输出响应体末尾字段，如：

event:message_delta
data:{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":8,"cache_creation_input_tokens":800,"cache_read_input_tokens":77249,"output_tokens":356}       }

event:message_stop
data:{"type":"message_stop"    }

中的output_tokens字段作为输出 token 计数依据。
同时，保留使用message_start提取输入 token 的逻辑，因为只有message_start字段包含了 5m/1h 区分计费所需的分时缓存创建计数字段。

[github-actions]

Thank you for this detailed bug report.

Analysis:

I've confirmed the issue in the codebase. The root cause is in the parseUsageFromResponseText function in src/app/v1/_lib/proxy/response-handler.ts (lines 1272-1391).

Problem identified:

The current SSE parsing logic (lines 1356-1388) iterates through all SSE events and uses an applyUsageValue function with early-return semantics:

const applyUsageValue = (value: unknown, source: string) => {
  if (usageMetrics) {
    return;  // ⚠️ Stops at first found usage
  }
  // ... extraction logic
};

When processing Anthropic streaming responses, it processes the message_start event first (which appears at the beginning of the stream), extracts usage from data.message.usage (line 1369-1372), and immediately returns - never reaching the message_delta event at the end of the stream that contains the correct final output_tokens count.

Affected code path:

• src/app/v1/_lib/proxy/response-handler.ts:1369-1372 - Claude message_start format handler (extracts incomplete output_tokens)

Issue reporter's correct analysis:
As you noted, the correct approach is:

1. Keep using message_start for input token fields (including cache_creation.ephemeral_5m_input_tokens and cache_creation.ephemeral_1h_input_tokens for differentiated cache billing)
2. Use message_delta for the final output_tokens count

Required fix strategy:
The parsing logic needs to be modified to:

• Track output_tokens separately from message_delta events (typically with stop_reason present)
• Merge the final output_tokens from message_delta with the input/cache metrics from message_start

Related helper functions:

• extractUsageMetrics (lines 1181-1270) - Extracts individual usage fields
• src/lib/utils/sse.ts:parseSSEData - Parses SSE text into structured events

---

Automated response from Claude AI

[ding113]

@claude fix it

[ding113]

I've created a fix for this issue.

Pull Request: #313

Changes made:

• Modified parseUsageFromResponseText in src/app/v1/_lib/proxy/response-handler.ts to properly handle Claude SSE streaming responses
• Now extracts input tokens and cache fields (including ephemeral_5m_input_tokens and ephemeral_1h_input_tokens) from message_start event
• Extracts final output_tokens from message_delta event (which contains the accurate count at stream end)
• Merges both to provide complete and accurate usage metrics

How the fix works:

1. When processing Claude SSE streams, the code now tracks both message_start (for input/cache metrics) and message_delta (for output_tokens) separately
2. After iterating through all events, it merges the final output_tokens from message_delta into the usage metrics from message_start
3. This preserves the 5m/1h differentiated cache billing feature while fixing the output token counting

Please review and let me know if you need any adjustments.

---

Response from Claude AI