Bug: Streaming timeout causes wrong model name and zero tokens in usage logs

[ChaoXu1997]

Summary

When a streaming response hits the 120-second idle timeout, the proxy_request_logs record stores the original request model (e.g. claude-opus-4-8) instead of the mapped model (e.g. glm-5.1). Token counts and cost are also zeroed out, making the UI show incorrect data.

What I see (broken)

Time	Provider	Model	In	Out	Cache	Cost	Latency	Status	Source
06/07 20:28	Zhipu GLM	claude-opus-4-8	0	0	—	$0.0000	129.1s	200	proxy

The UI shows only claude-opus-4-8 with no mapping arrow, zero tokens and zero cost.

What I expect (normal)

Time	Provider	Model	In	Out	Cache	Cost	Latency	Status	Source
06/07 20:28	Zhipu GLM	claude-opus-4-8 → glm-5.1	720	R140,736	226	$0.0386	8.9s	200	proxy

Steps to Reproduce

1. Configure Claude Code to proxy through cc-switch with model mapping claude-opus-4-8 → glm-5.1 (Zhipu GLM)
2. Start a long conversation request
3. If the upstream provider goes silent for 120+ seconds after the first byte, the streaming idle timeout fires
4. The proxy_request_logs record now has model = claude-opus-4-8 (same as request_model) and all tokens = 0

Root Cause Analysis

Call chain

1. Claude Code sends request with model: claude-opus-4-8
2. RequestContext stores request_model = "claude-opus-4-8" (original)
3. Forwarder correctly maps to glm-5.1 and sends to Zhipu API ✅ — log shows [Claude] >>> 请求 URL: ... (model=glm-5.1)
4. Zhipu API starts streaming, some SSE events are received (first_token_ms is set)
5. 120-second idle timeout fires: [Claude] 流式响应静默期超时 (120秒)
6. SseUsageCollector::finish() is called, but collected SSE events are incomplete (missing message_delta which carries usage)
7. claude_model_extractor cannot extract the model name from incomplete events, falls back to request_model
8. Database record: model = "claude-opus-4-8", tokens = 0, cost = $0.0000

Problematic code

In src-tauri/src/proxy/handler_config.rs, claude_model_extractor:

fn claude_model_extractor(events: &[Value], request_model: &str) -> String {
    if let Some(usage) = TokenUsage::from_claude_stream_events(events) {
        if let Some(model) = usage.model {
            return model;
        }
    }
    request_model.to_string()  // ← Bug: falls back to original model on timeout
}

When the stream is cut by timeout:

• TokenUsage::from_claude_stream_events(events) returns None (no message_delta with usage)
• Falls back to request_model = "claude-opus-4-8" instead of the actual mapped model glm-5.1

The same pattern exists in openai_model_extractor, codex_auto_model_extractor, and gemini_model_extractor.

Database evidence

SELECT datetime(created_at, "unixepoch", "localtime") as time,
       request_model, model, input_tokens, output_tokens, total_cost_usd, latency_ms, status_code
FROM proxy_request_logs
WHERE model = request_model AND input_tokens = 0 AND output_tokens = 0
ORDER BY created_at DESC LIMIT 5;

time	request_model	model	in	out	cost	latency	status
20:34:37	claude-opus-4-8	claude-opus-4-8	0	0	0	120769ms	200
20:31:42	claude-opus-4-8	claude-opus-4-8	0	0	0	126104ms	200
20:28:51	claude-opus-4-8	claude-opus-4-8	0	0	0	129145ms	200
20:26:06	claude-opus-4-8	claude-opus-4-8	0	0	0	128027ms	200
20:23:29	claude-opus-4-8	claude-opus-4-8	0	0	0	39081ms	200

All broken records share: model == request_model, tokens = 0, latency > 120s, status = 200.

Corresponding log entries:

[2026-06-07][20:28:51][ERROR][cc_switch_lib::proxy::response_processor] [Claude] 流式响应静默期超时 (120秒)

Suggested Fix

Option A: Add intermediate fallback in model_extractor (recommended, minimal change)

Before falling back to request_model, try to extract the model from the message_start event:

fn claude_model_extractor(events: &[Value], request_model: &str) -> String {
    if let Some(usage) = TokenUsage::from_claude_stream_events(events) {
        if let Some(model) = usage.model {
            return model;
        }
    }
    // NEW: on timeout, try to extract actual model from message_start event
    events
        .iter()
        .find_map(|e| {
            if e.get("type")?.as_str()? == "message_start" {
                e.get("message")?.get("model")?.as_str()
            } else {
                None
            }
        })
        .unwrap_or(request_model)
        .to_string()
}

Option B: Store mapped model in RequestContext

Add a mapped_model field to RequestContext. The forwarder writes back the mapped model name after remapping, so the timeout handler can use it instead of the original.

Environment

• cc-switch version: latest (as of 2026-06-07)
• OS: Linux
• Providers: Zhipu GLM (glm-5.1) / Xiaomi MiMo (mimo-v2.5-pro)
• App type: claude (Claude Code)