[wujiaming-ai]

Model traces currently lose the first generated delta's receipt time, preventing A360 from separating request preprocessing from model first-output latency. This change adds request-start, model SDK request-start, and first-token timestamps to model spans and existing trace exports/dumps.

• Capture request start at VeADK HTTP ASGI ingress or direct Runner execution, with isolated context propagation and an explicit source.
• Observe LiteLLM Chat and Ark Responses raw SDK streams before ADK aggregation. Record the first nonempty text, reasoning, or tool-call delta once, using UTC Unix nanoseconds and a child span for each SDK request.
• Distinguish the actual SDK streaming mode: non-streaming responses have no first-token timestamp; no-output and unsupported paths expose explicit status. Preserve attempt evidence through retries, errors, cancellation, and stream closure.
• Document timing definitions/scopes in bilingual component specifications and tracing documentation. Client-render timing and TTFT/TPOT histogram recording remain outside this requirement.

Validation

• Changed executable production-line coverage: 100% (208/208) against 23014771; no added coverage exclusions.
• 244 affected tests passed, including 55 focused timing/adapter cases. Re-ran the affected suite and all applicable pre-commit hooks after fetching/rebasing onto the unchanged origin/main.
• 6/6 real cloud E2E cases passed in cn-beijing using doubao-seed-2-1-pro-260628: Chat and Responses streaming/non-streaming, plus concurrent Chat/Responses streaming. All nine changed production files matched the deployed branch wheel's SHA256. Verified exported timestamps, SDK mode/status, parent relationships and request isolation. The isolated Runtime was deleted.
• Documentation type checking and production build passed; bilingual identifiers/links and diff whitespace checks passed.
• Default parallel regression after test stabilization: 7,942 passed, 0 failed, 42 skipped, 4 xfailed, plus 6 passing subtests. The original Sidecar baseline failure is resolved by isolating the fake runtime from parent upstream environment variables.
• CI repair: exercise both Sandbox inspection orders, assert exact call multiplicities, and wait for both cached model states to appear in /web/system-info. This fixes an order-dependent test while preserving production concurrency and every existing state/endpoint assertion. Three targeted cases and all 189 affected cases passed; production code and the cloud-validated source hashes are unchanged. See the bilingual CI repair design and evidence.

Design and verification (English) · 需求与验证结论（中文）

[feng-95]

@wujiaming-ai @yaozheng-fang 请按 APMPlus 规范补齐下面的属性和指标：

位置	名称	类型 / 单位
LLM Span 流式标记	gen_ai.request.stream	bool，实际模型 SDK 请求是否流式
LLM Span TTFT 属性	gen_ai.chat_completions.streaming_time_to_first_token	double，秒
TTFT 指标	gen_ai.chat_completions.streaming_time_to_first_token	Histogram，单位 s

TTFT 是单次 LLM 流式请求开始到首个输出 token 到达的耗时。可以基于本次采集的模型请求开始时间和首 Token 时间计算：(first_token_time_unix_ns - model_start_time_unix_ns) / 1e9。Span 属性应填写这个秒级耗时，不能直接填绝对时间戳，也不能用整个 Agent 请求或 LLM Loop 的耗时代替。

目前只记录了 veadk.* 时间戳，portal_metrics.py 中 TTFT 的 .record() 仍是 TODO。请同时写入上述 Span 属性并实际记录同名 Histogram；每次 LLM 请求取得观测值时记录一次。非流式或没有首 Token 观测值时省略，不用 0 占位。现有自定义时间戳可以作为附加信息保留，并补充属性值、单位和指标记录的测试。