feat(core): Align GenAI telemetry with ARMS

[doudouOUC]

What this PR does

This PR aligns the first batch of Qwen Code LLM, tool, and agent span attributes with the fields whose names, types, and meanings agree between the version-pinned OpenTelemetry GenAI semantic conventions and Alibaba Cloud ARMS LLM Trace. It adds canonical operation, provider, conversation, response model, output type, finish reason, cache usage, tool-call, and agent-description attributes while preserving the existing span topology and private fields that have no exact standard equivalent.

Provider response models, finish reasons, cache usage, and token counts now retain their source provenance. OpenAI-compatible responses that report only a total token count no longer receive a synthetic input/output split, and Anthropic partial streams retain input/cache usage known at message_start. Provider and operation resolution is URL-boundary-safe, stays consistent with internal DashScope routing, recognizes Azure OpenAI, does not expose credentials or arbitrary proxy hosts, and only emits output type when the effective Gemini or Vertex request carries an unambiguous response format.

Exact-equivalent private aliases and invalid GenAI aliases are removed instead of dual-written. The change also documents the pinned standards baseline, migration mapping, omission rules, deferred scope, and the ARMS resource attribute users must configure for automatic GenAI application recognition.

Why it's needed

Qwen Code previously emitted a mixture of private aliases, incomplete GenAI attributes, and several keys whose names or semantics did not match both ARMS and the OpenTelemetry GenAI conventions. This made the same trace require different queries across backends and could present estimated token splits or request-model fallbacks as provider-reported response data. A single truthful field contract makes traces portable and queryable in ARMS without misrepresenting provenance.

Reviewer Test Plan

How to verify

• Export telemetry to a file while running an LLM to tool to LLM flow against the fake OpenAI server. Confirm the JSONL spans contain the shared GenAI fields, preserve the provider tool-call ID, retain only private fields without exact standard replacements, and omit the removed aliases.
• Exercise OpenAI, Azure OpenAI, Anthropic, Gemini, Vertex AI, and Qwen OAuth request paths. Confirm provider and operation resolution, actual response models, multi-candidate finish reasons, explicit zero tokens, omission of input/output usage for total-only responses, cache read/create usage, partial stream failures, cancellation, and idle timeout behavior follow the documented omission rules.
• Exercise Core and ACP tool scheduling and subagent tracing. Confirm provider call IDs take priority with internal-ID fallback, and agent descriptions are propagated and surrogate-safe when truncated.
• Local verification for the latest alignment and review updates completed: 766 targeted Core tests and the GenAI telemetry integration test passed, together with Core build/typecheck, targeted ESLint, and bundle generation.

Evidence (Before & After)

N/A — telemetry schema, tracing behavior, tests, and documentation only; there is no TUI change.

Tested on

OS	Status
🍏 macOS	✅ tested
🪟 Windows	⚠️ not tested
🐧 Linux	⚠️ not tested

Environment (optional)

macOS, Node.js v22.22.3, npm 10.9.8, local fake OpenAI server, and the OpenTelemetry file exporter.

Risk & Scope

• Main risk or tradeoff: OpenTelemetry GenAI semantic conventions are still in Development status, so this implementation is pinned to commit 2e994c6d59a93bb4fc1752c5378eedb9b8e14d6b and future upgrades require a new compatibility review.
• Not validated / out of scope: Online ARMS staging validation was not performed because credentials were unavailable. Full GenAI span naming, CLIENT span kind, retry topology, sensitive message/tool payloads, embeddings, and fields without a truthful wire-request source remain out of scope. The local full build reaches the CLI package but is blocked because the installed Ink types do not include the selection extensions used by current main; Core build and typecheck, targeted lint and tests, bundle generation, and the telemetry integration test pass.
• Breaking changes / migration notes: LLM spans replace qwen-code.model with gen_ai.request.model, response_id with gen_ai.response.id, and the bare input_tokens, output_tokens, and cached_input_tokens aliases with their gen_ai.usage.* equivalents. Root tool spans replace tool.name with gen_ai.tool.name. Invalid aliases gen_ai.usage.cached_tokens, gen_ai.server.time_to_first_token, and gen_ai.usage.reasoning_tokens are removed. Interaction spans, logs, hooks, blocked-on-user spans, and private fields without an exact shared standard equivalent retain their existing schemas.

Linked Issues

N/A

中文说明

本 PR 做了什么

本 PR 将 Qwen Code 第一批 LLM、工具和 Agent Span 属性与固定版本的 OpenTelemetry GenAI 语义规范及阿里云 ARMS LLM Trace 中名称、类型和含义一致的字段对齐。它新增规范的 operation、provider、conversation、response model、output type、finish reason、cache usage、tool call 和 agent description 属性，同时保留现有 Span 拓扑以及没有精确标准等价项的私有字段。

Provider 返回的响应模型、结束原因、缓存用量和 token 数现在会保留数据来源。对于只返回 token 总量的 OpenAI 兼容响应，不再合成 input/output 拆分；Anthropic 部分流也会保留 message_start 时已知的 input/cache usage。Provider 与 operation 解析采用安全的 URL hostname 边界匹配，与内部 DashScope 路由保持一致，支持识别 Azure OpenAI，不会暴露凭据或任意代理主机名；仅当实际 Gemini 或 Vertex 请求携带明确且唯一的响应格式时才写入 output type。

与标准字段完全等价的私有别名和无效的 GenAI 别名会直接移除，不进行双写。本次变更还记录了固定的标准版本、迁移映射、省略规则、延期范围，以及用户为 ARMS 自动识别 GenAI 应用所需配置的资源属性。

为什么需要这个 PR

Qwen Code 此前混合上报私有别名、不完整的 GenAI 属性，以及名称或语义无法同时匹配 ARMS 和 OpenTelemetry GenAI 规范的若干字段。这导致同一条 Trace 在不同后端需要不同查询方式，也可能把估算的 token 拆分或请求模型 fallback 表现成 provider 实际返回的数据。统一且真实的字段契约可让 Trace 跨后端使用，并能在 ARMS 中直接查询而不歪曲数据来源。

Reviewer 测试计划

如何验证

• 使用 fake OpenAI server 跑通 LLM 到工具再到 LLM 的流程并将 telemetry 导出到文件。确认 JSONL Span 包含三方共同的 GenAI 字段、保留 provider tool-call ID、只保留没有精确标准替代项的私有字段，并且不再出现已删除的别名。
• 覆盖 OpenAI、Azure OpenAI、Anthropic、Gemini、Vertex AI 和 Qwen OAuth 请求路径。确认 provider 与 operation 解析、真实 response model、多 candidate finish reasons、显式零 token、仅有 total 时省略 input/output usage、cache read/create usage、流式部分失败、取消和 idle timeout 均遵循文档中的省略规则。
• 覆盖 Core 与 ACP 的工具调度和 subagent tracing。确认 provider call ID 优先并在缺失时回退到内部 ID，同时确认 agent description 可正确传递并在截断时不会破坏代理对。
• 最新字段对齐与 review 修复的本地验证结果：766 个定向 Core 测试和 GenAI telemetry 集成测试通过，同时 Core build/typecheck、定向 ESLint 与 bundle 生成通过。

前后对比证据

N/A——本 PR 仅涉及 telemetry schema、tracing 行为、测试和文档，没有 TUI 变更。

测试平台

OS	状态
🍏 macOS	✅ 已测试
🪟 Windows	⚠️ 未测试
🐧 Linux	⚠️ 未测试

环境（可选）

macOS、Node.js v22.22.3、npm 10.9.8、本地 fake OpenAI server，以及 OpenTelemetry file exporter。

风险与范围

• 主要风险或取舍：OpenTelemetry GenAI 语义规范仍处于 Development 状态，因此实现固定到 commit 2e994c6d59a93bb4fc1752c5378eedb9b8e14d6b；未来升级需要重新进行兼容性审查。
• 未验证或不在范围内：由于没有凭据，未进行 ARMS 在线 staging 验证。完整 GenAI Span 命名、CLIENT SpanKind、retry topology、敏感的消息或工具 payload、Embedding，以及缺少真实 wire request 数据源的字段均不在本次范围内。本地全仓 build 可执行到 CLI package，但已安装的 Ink 类型不包含当前 main 使用的 selection 扩展，因此在此处被阻塞；Core build 与 typecheck、定向 lint 与测试、bundle 生成和 telemetry 集成测试均通过。
• 破坏性变更与迁移说明：LLM Span 使用 gen_ai.request.model 替代 qwen-code.model，使用 gen_ai.response.id 替代 response_id，并使用对应的 gen_ai.usage.* 字段替代裸 input_tokens、output_tokens 和 cached_input_tokens 别名；根 Tool Span 使用 gen_ai.tool.name 替代 tool.name。无效别名 gen_ai.usage.cached_tokens、gen_ai.server.time_to_first_token 和 gen_ai.usage.reasoning_tokens 已移除。Interaction Span、日志、hook、blocked-on-user Span，以及没有精确共同标准等价项的私有字段继续保留现有 schema。

关联 Issue

N/A

[qwen-code-ci-bot]

Thanks for the PR!

Template looks good ✓

Problem: This is a feature enhancement aligning telemetry span attributes with the OpenTelemetry GenAI semantic conventions and Alibaba Cloud ARMS LLM Trace fields. The motivation is concrete — existing traces use a mix of private aliases (qwen-code.model, response_id, bare input_tokens) and incomplete GenAI attributes that aren't portable across backends, and estimated token splits can misrepresent provenance. Not a bug fix, so no reproduction is expected.

Direction: Telemetry standardization is a natural evolution of the existing tracing infrastructure. The broader ecosystem is actively adding structured OTel attributes (Claude Code's CHANGELOG shows ongoing OTel attribute work — message.uuid, tool_source, workflow.run_id, etc.). Aligning with the GenAI semantic conventions and ARMS makes traces queryable in Alibaba Cloud monitoring without backend-specific queries. That said, this touches telemetry schema — escalating to the maintainer for direction confirmation since telemetry is a sensitive area.

Size: This PR touches core paths (packages/core/src/**). Breakdown: ~665 production logic lines, ~1244 test lines, ~154 docs lines across 30 files. Production lines exceed 500 — flagging for maintainer awareness per the core module policy. Total production lines also exceed the 1000-line large PR advisory threshold when including docs. This is a feat-type PR, so no hard block applies.

Approach: The scope feels focused on one concern (GenAI attribute alignment). The new gen-ai-provider.ts module (180 lines) centralizes provider/operation resolution with proper hostname boundary checking — worth a close look in code review. The WeakMap-based usage provenance tracking in gen-ai-usage.ts is an elegant solution that avoids changing the public response type. The design doc documenting the pinned standards baseline and migration mapping is a good addition. No unrelated changes or drive-by refactors visible. The two follow-up commits (removing estimated token splits, addressing review feedback) are well-scoped refinements. One question from the prior review remains relevant: the PR removes existing aliases (qwen-code.model, response_id, tool.name) rather than dual-writing — has the downstream impact of this breaking change been considered (e.g., existing dashboards or queries relying on the old field names)?

Moving on to code review. 🔍

中文说明

感谢贡献！

模板完整 ✓

问题：这是一个功能增强 PR，将 telemetry span 属性与 OpenTelemetry GenAI 语义规范及阿里云 ARMS LLM Trace 字段对齐。动机具体——现有 trace 混合使用私有别名（qwen-code.model、response_id、裸 input_tokens）和不完整的 GenAI 属性，无法跨后端移植，且估算的 token 拆分可能歪曲数据来源。不是 bug 修复，因此不需要复现。

方向：Telemetry 标准化是现有 tracing 基础设施的自然演进。更广泛的生态系统正在积极添加结构化 OTel 属性（Claude Code 的 CHANGELOG 显示了持续的 OTel 属性工作——message.uuid、tool_source、workflow.run_id 等）。与 GenAI 语义规范和 ARMS 对齐可使 trace 在阿里云监控中直接查询，无需后端特定查询。不过，这涉及 telemetry schema——由于 telemetry 是敏感领域，转交 maintainer 确认方向。

规模：本 PR 触及核心路径（packages/core/src/**）。明细：约 665 行生产逻辑、约 1244 行测试、约 154 行文档，共 30 个文件。生产行数超过 500——按核心模块政策标记供 maintainer 知悉。这是 feat 类型 PR，不适用硬性阻止。

方案：范围聚焦于一个关注点（GenAI 属性对齐）。新的 gen-ai-provider.ts 模块（180 行）集中了 provider/operation 解析，具有正确的主机名边界检查。gen-ai-usage.ts 中基于 WeakMap 的 usage 来源追踪是优雅的解决方案，避免了更改公共响应类型。记录固定标准版本和迁移映射的设计文档是好的补充。未见无关改动或顺手重构。两个后续提交（移除估算 token 拆分、处理 review 反馈）范围合理。

进入代码审查 🔍

— Qwen Code · qwen3.8-max-preview

_{Reviewed at c1123192c44649ab38516ee9871a981acf2f0a4f · re-run with @qwen-code /triage}

[qwen-code-ci-bot]

Code Review

Independent proposal: For aligning GenAI telemetry with ARMS/OTel GenAI conventions, I would: (1) define target attribute names from the pinned spec commit, (2) create a pure provider/operation resolution function with safe hostname matching, (3) update each content generator to emit standard attributes with provenance tracking (distinguishing provider-reported vs estimated values), (4) update session-tracing to use standard names and remove invalid aliases, (5) add comprehensive tests, (6) document the migration mapping and deferred scope.

Comparison: The PR matches this approach closely and executes it well. Two follow-up commits since the last review refined the implementation: one removed the synthetic input/output token split for total-only OpenAI responses, and the other addressed review feedback (Anthropic partial-stream usage retention, cache-creation token tracking, Azure OpenAI recognition).

The implementation is clean and well-structured:

• gen-ai-provider.ts (180 lines): Pure function for provider/operation resolution with proper hostname boundary checking (isHostOrSubdomain prevents suffix attacks like api.deepseek.com.attacker.example). Resolution priority is clear: Qwen OAuth → DashScope proxy match → hostname → env key hint → protocol fallback. Azure OpenAI is now recognized via openai.azure.com and services.ai.azure.com. No credentials or arbitrary proxy hosts are exposed. The normalizedEndpoint helper strips trailing slashes for reliable proxy comparison.

• gen-ai-usage.ts (30 lines): WeakMap-based provenance tracking attached to usage objects without changing their public shape. Records cachedInputTokensReported and cacheCreationInputTokens (Anthropic-specific). This lets the telemetry layer distinguish estimated vs. provider-reported values while allowing GC to follow the usage object. Elegant.

• session-tracing.ts: Removes old aliases (qwen-code.model, response_id, tool.name on LLM/tool spans) and invalid GenAI aliases. Adds standard attributes with proper guards — token counts only emitted as standard GenAI fields when they're non-negative safe integers. Cache read tokens are gated on cachedInputTokensReported so estimated cache values aren't presented as provider-reported. Surrogate-safe truncation for agent descriptions via the renamed truncateSpanText.

• OpenAI converter: modelVersion now comes from openaiResponse.model / chunk.model (not request-model fallback). The hasTokenBreakdown guard prevents synthesizing an input/output split when only total_tokens is reported — this was the key fix in the second commit. Usage provenance is attached via setGenAiUsageProvenance.

• Anthropic generator: message_start input and cache usage is now attached to the first subsequent yielded chunk, so partial stream failures retain provider-reported usage. Cache-creation tokens are tracked with a cacheCreationTokensReported flag. The buildAnthropicUsageMetadata discriminator logic (cache_creation reported → Anthropic semantics, sum all three; else OpenAI-style) is well-documented with a comment explaining the earlier false-positive bug.

• Agent tool: agentDescription is now propagated from subagentConfig.description through the invocation spec to the subagent span, where it's emitted as gen_ai.agent.description with surrogate-safe truncation.

No critical blockers. No AGENTS.md violations.

Unit tests: All changed test files pass — 192 telemetry + 1072 core (Anthropic, OpenAI, logging, scheduler, agent) + 391 CLI Session = 1655 tests total, all green.

Real-Scenario Testing

Ran the built bundle against a local fake OpenAI server with QWEN_TELEMETRY_OUTFILE file export. The LLM span output confirms the PR's claims:

=== qwen-code.llm_request ===
  finish_reason: STOP
  gen_ai.conversation.id: af42155e-db7a-4681-9c0f-42978771328d
  gen_ai.operation.name: chat
  gen_ai.provider.name: openai
  gen_ai.request.model: request-model-test
  gen_ai.response.finish_reasons: ['STOP']
  gen_ai.response.id: chatcmpl-test-7536
  gen_ai.response.model: provider-model-actual
  gen_ai.usage.cache_read.input_tokens: 5
  gen_ai.usage.input_tokens: 15
  gen_ai.usage.output_tokens: 8
  Removed aliases still present: NONE (correct)

Key observations:

• gen_ai.response.model is provider-model-actual (from the server response), NOT request-model-test (the request model) — provenance is correct
• gen_ai.usage.cache_read.input_tokens: 5 — cache read properly reported from prompt_tokens_details.cached_tokens
• All removed aliases (qwen-code.model, response_id, gen_ai.usage.cached_tokens, gen_ai.server.time_to_first_token, gen_ai.usage.reasoning_tokens, bare input_tokens/output_tokens/cached_input_tokens) are confirmed absent
• gen_ai.conversation.id matches the session ID
• gen_ai.response.finish_reasons is an array ordered by candidate index

中文说明

代码审查

独立方案： 要将 GenAI telemetry 与 ARMS/OTel GenAI 规范对齐，我会：(1) 从固定规范提交定义目标属性名，(2) 创建带安全主机名匹配的纯 provider/operation 解析函数，(3) 更新每个 content generator 以发出带来源追踪的标准属性，(4) 更新 session-tracing 使用标准名称并移除无效别名，(5) 添加全面测试，(6) 记录迁移映射和延期范围。

对比： PR 与此方案高度一致且执行良好。上次审查以来的两个后续提交改进了实现：一个移除了仅有 total 的 OpenAI 响应的合成 input/output 拆分，另一个处理了 review 反馈（Anthropic 部分流 usage 保留、cache-creation token 追踪、Azure OpenAI 识别）。

实现干净且结构合理：

• gen-ai-provider.ts（180 行）：纯函数实现 provider/operation 解析，具有正确的主机名边界检查。解析优先级清晰。Azure OpenAI 现在通过 openai.azure.com 和 services.ai.azure.com 识别。不暴露凭据或任意代理主机名。

• gen-ai-usage.ts（30 行）：基于 WeakMap 的来源追踪，记录 cachedInputTokensReported 和 cacheCreationInputTokens（Anthropic 特有）。设计优雅。

• session-tracing.ts：移除旧别名和无效 GenAI 别名。Cache read token 以 cachedInputTokensReported 为门控，确保估算值不会被表现为 provider 实际报告值。

• OpenAI 转换器：hasTokenBreakdown 守卫防止在仅有 total_tokens 时合成 input/output 拆分——这是第二个提交的关键修复。

• Anthropic 生成器：message_start 的 input 和 cache usage 现在附加到第一个后续产出的 chunk，部分流失败时保留 provider 报告的 usage。

无关键阻塞项。无 AGENTS.md 违规。

单元测试： 所有变更的测试文件通过——共 1655 个测试，全部绿色。

真实场景测试

使用构建后的 bundle 对本地 fake OpenAI server 运行，启用 QWEN_TELEMETRY_OUTFILE 文件导出。LLM span 输出确认了 PR 的声明：provider 响应模型正确使用、cache read 正确报告、所有已移除别名确认不存在、gen_ai.conversation.id 匹配 session ID。

— Qwen Code · qwen3.8-max-preview

_{Reviewed at c1123192c44649ab38516ee9871a981acf2f0a4f · re-run with @qwen-code /triage}

[qwen-code-ci-bot]

Confidence: 3/5 — clean review across every stage, but the core-module size escalation (665 production lines) and telemetry-scope sensitivity need a maintainer's sign-off.

This is a well-executed telemetry alignment PR that has improved meaningfully since the last review. The two follow-up commits addressed the right things: removing the synthetic token split for total-only OpenAI responses was a real provenance fix (not just hygiene), and the Anthropic partial-stream usage retention closes a genuine data-loss gap on aborted streams.

Going back to my independent proposal — the PR matches it closely and exceeds it in a few places. The WeakMap approach for usage provenance is better than what I would have proposed (adding fields to the type or threading provenance through parameters). The hasTokenBreakdown guard is a clean way to avoid the estimated-split problem without special-casing every provider. The buildAnthropicUsageMetadata discriminator logic — using cache_creation_input_tokens presence as the primary signal for Anthropic vs OpenAI-on-Anthropic semantics — is well-reasoned and the comment explaining the earlier false-positive bug is exactly the kind of "why" documentation that prevents regressions.

The real-scenario test confirms every claim: response model provenance is correct (provider-model-actual, not the request model), cache read is properly gated on provider-reported values, all removed aliases are confirmed absent, and the conversation ID ties spans to the session.

The breaking change (removing qwen-code.model, response_id, tool.name without dual-write) is documented and scoped — interaction spans, logs, hooks, and blocked-on-user spans retain their existing schemas. The design doc's migration table makes the mapping explicit. The PR is honest about what it doesn't cover (span naming, CLIENT span kind, retry topology, sensitive payloads, embeddings).

Why 3/5 and not higher: policy, not code quality. This touches core telemetry schema at 665 production lines — the core-module policy requires maintainer awareness at 500+, and telemetry is a sensitive area where direction confirmation is appropriate. The code itself would be a 4/5 (solid, well-tested, no blockers).

中文说明

置信度：3/5 — 每个阶段审查均干净，但核心模块规模升级（665 行生产代码）和 telemetry 领域敏感性需要 maintainer 签核。

这是一个执行良好的 telemetry 对齐 PR，自上次审查以来有实质性改进。两个后续提交解决了正确的问题：移除仅有 total 的 OpenAI 响应的合成 token 拆分是真正的来源修复（而非仅仅是卫生问题），Anthropic 部分流 usage 保留关闭了中止流上的真实数据丢失缺口。

回到我的独立方案——PR 与之高度一致且在若干方面超越。WeakMap 方法用于 usage 来源比我提议的更好。hasTokenBreakdown 守卫是避免估算拆分问题的干净方式。buildAnthropicUsageMetadata 判别逻辑——使用 cache_creation_input_tokens 存在作为 Anthropic 与 OpenAI-on-Anthropic 语义的主要信号——推理合理，解释早期误报 bug 的注释正是防止回归的那种"为什么"文档。

真实场景测试确认了每一项声明。

为什么是 3/5 而不是更高：政策原因，非代码质量。代码本身是 4/5。

— Qwen Code · qwen3.8-max-preview

_{Reviewed at c1123192c44649ab38516ee9871a981acf2f0a4f · re-run with @qwen-code /triage}

[qwen-code-ci-bot]

⏸️ Deferring to @yiliang114 — this PR touches core telemetry schema (665 production lines across packages/core/src/telemetry/, content generators, and tool scheduler) and introduces a breaking change to span attribute names. The code review is clean at the updated head (c1123192) and the real-scenario test confirms correct behavior, but two policy triggers require a maintainer's call:

1. Core module size: 665 production logic lines in core paths (up from 549 at the prior review — the two follow-up commits added Anthropic partial-stream usage retention and cache-creation tracking) → maintainer awareness per the two-tier gate.
2. Telemetry scope: Schema changes to tracing attributes affect downstream dashboards and queries — direction confirmation is appropriate.

The PR removes qwen-code.model, response_id, and tool.name from spans without a dual-write period. If existing ARMS dashboards or alerting rules depend on those field names, this is a breaking change that needs coordination. The code quality is high (would be 4/5 on merit alone) — this is purely a policy escalation. Needs a human call on this one.

[qwen-code-ci-bot]

⚠️ Downgraded from Approve to Comment: CI failing: Test (ubuntu-latest, Node 22.x). Reviewed.

— qwen3.7-max via Qwen Code /review

[qwen-code-ci-bot]

Reviewed. Suggestions are inline. 1 Suggestion-level finding(s) could not be anchored to a changed line and were dropped; nothing further to act on here. Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.

— qwen3.7-max via Qwen Code /review

[gwinthis]

Review & Local Verification Report

代码审查

设计评价：系统性的遥测对齐工程。 本 PR（30 文件，+1697/-119）将 GenAI 遥测字段与 ARMS（Application Real-Time Monitoring Service）对齐。

变更范围：

• 3 个设计文档（field alignment、LLM request timing、subagent spans）
• 遥测核心：gen-ai-provider.ts、gen-ai-usage.ts、session-tracing.ts
• 内容生成器：OpenAI pipeline、Anthropic generator、logging generator
• 转换器：OpenAI/Anthropic converter 的 usage 提取
• 工具调度：coreToolScheduler.ts、agent.ts 的 span 追踪
• 集成测试：gen-ai-telemetry.test.ts、fake-openai-server.ts
• CLI 层：ACP Session 的遥测集成

设计亮点：

• 设计文档先行，明确字段映射和 span 层级
• 跨 provider 统一 usage 提取（OpenAI/Anthropic/Gemini）
• subagent span 追踪支持多 agent 工作流的可观测性
• 集成测试用 fake OpenAI server 验证端到端遥测

结论

LGTM。 大规模但系统化的遥测对齐，设计文档完整，跨 provider 统一，测试覆盖充分。

[yiliang114]

Review Summary

High-quality telemetry alignment PR. The OTel GenAI semantic convention mapping is precise, the WeakMap usage-provenance design is elegant, and provider hostname parsing has proper boundary checks. Test coverage is comprehensive.

One concern needs author confirmation before merge:

(1) Breaking attribute removal without dual-write (session-tracing.ts): qwen-code.model, response_id, tool.name, gen_ai.usage.cached_tokens, gen_ai.server.time_to_first_token, gen_ai.usage.reasoning_tokens, gen_ai.agent.id, and subagent gen_ai.provider.name are all removed in one shot. The design doc says "removed without a dual-write period." If any existing ARMS/Grafana dashboards depend on these keys, they break immediately. Suggest at minimum a prominent breaking-change note in the release, or one release of dual-emit.

(2) cacheCreationInputTokens guard order (session-tracing.ts ~L680): Number.isSafeInteger(x) && x !== undefined — the !== undefined is redundant since isSafeInteger(undefined) is already false. Consider swapping order to match the cachedInputTokens guard style above.

(3) gen_ai.response.model omission when provider doesn't return it (converter.ts:354): correct per spec, but some OpenAI-compatible proxies never return a model field, leaving the ARMS column permanently empty. Worth a doc note.

(4) finishReasons Map uses last-wins semantics for same candidate index (loggingContentGenerator.ts ~L638): reasonable but worth a comment clarifying the override behavior for multi-chunk scenarios.

(5) gen_ai.output.type only maps Gemini/Vertex (gen-ai-provider.ts:145): OpenAI response_format: json_object is not mapped. Design doc covers this, noting for future extensibility.

Minor: truncateSpanError is now a single-line wrapper of truncateSpanText — consider exporting the latter directly. Hostname list in gen-ai-provider.ts is hardcoded (12+ domains) — could be registry-driven later.

— qwen3.8-max-preview via Qwen Code /review

[yiliang114]

The main alignment direction looks right, but the provider mismatch above affects supported configurations and should be fixed before merge. I also left one non-blocking partial-stream telemetry issue inline.

[wenshao]

Local Build & Real-Flow Verification (maintainer merge reference)

Independent local verification of head 726618f — real build, changed-file test suites, a full end-to-end telemetry run through the built CLI, and a base-vs-PR A/B on the actual exported telemetry.log. This complements the existing code reviews (which did not include local run evidence). Environment: macOS, Node 22.23.1, isolated git worktree at the PR head with a separate base bundle built at the merge-base 32c491fc6.

TL;DR — ✅ verified, merge-ready

Build clean, all exercised suites green, and the emitted spans match the documented field contract exactly — every added gen_ai.* key present, every removed alias gone, private compat fields retained. The one pre-merge item is the release-note breaking-change callout already raised by @yiliang114; the A/B below makes those removed keys concrete.

1. Build & tests

Stage	Result
npm run build:packages (tsc, all workspaces) + npm run bundle	✅ exit 0 (PR & base bundles)
Unit — changed files (vitest, packages/core)	✅ 10 files, 1257 tests
Unit — packages/cli ACP telemetry (Session.test.ts)	✅ 391 tests
Integration — cli/gen-ai-telemetry.test.ts (real E2E: built dist/cli.js + fake OpenAI server)	✅ 1 test
Security — URL-boundary-safe provider resolution	✅ spoof / look-alike rejected

Upstream CI Test (ubuntu-latest, Node 22.x) is also green at this head (the ci-bot's earlier "downgrade" reflected a since-resolved transient failure).

2. Real-flow A/B — the field contract on live telemetry

I ran an identical LLM → tool → LLM turn through the base bundle (PR not applied) and the PR bundle, then read the spans back out of the exported telemetry.log. The server returns request-model request-model, a distinct response-model provider-model-tool, a provider tool-call id provider-call-123, and cached_tokens: 3 — so every migrated field is observable.

Confirmed on the wire, exactly as the design doc specifies:

• Added (OTel GenAI / ARMS aligned): gen_ai.operation.name, gen_ai.provider.name, gen_ai.conversation.id, gen_ai.response.model, gen_ai.usage.cache_read.input_tokens on the LLM span; gen_ai.operation.name=execute_tool, gen_ai.tool.name, gen_ai.tool.type=function, gen_ai.tool.call.id on the tool span.
• Removed (private aliases / invalid GenAI keys), no dual-write: qwen-code.model, response_id, gen_ai.server.time_to_first_token, gen_ai.usage.cached_tokens, gen_ai.usage.reasoning_tokens, tool.name.
• Provenance honored: request/response model split is real (request-model vs provider-model-tool); tool-call id prefers the provider id; private compat keys (input_tokens, cached_input_tokens, ttft_ms, thoughts_token_count, tool.call_id) are retained.

3. Provider resolution is spoof-proof

Exercising the built resolveGenAiProviderName directly (bottom panel of image #1): a real subdomain x.dashscope.aliyuncs.com → dashscope, but the spoof dashscope.aliyuncs.com.attacker.com and the look-alike xdashscope.aliyuncs.com both correctly fall through to the default rather than being mis-attributed. Unknown hosts fall back to the protocol provider with no host/credential leak; qwen-oauth and apiKeyEnvKey paths resolve as documented.

Notes for author / reviewers

1. @yiliang114 point (2) is cosmetic-only, not a bug. In session-tracing.ts the cacheCreationInputTokens guard Number.isSafeInteger(x) && x !== undefined has a redundant !== undefined (Number.isSafeInteger(undefined) === false), so it is harmless — no functional impact. Fine to tidy or leave.
2. Breaking-change note (@yiliang114 point (1)): the removed keys above are the concrete blast radius. Recommend a prominent release-note callout so any ARMS/Grafana dashboard keyed on qwen-code.model / response_id / tool.name / the three invalid gen_ai.usage.* keys is migrated. Intentional per design doc — flagging for visibility, non-blocking.
3. Minor: the PR body's "9 GenAI telemetry integration tests" — the new integration file gen-ai-telemetry.test.ts contains 1 end-to-end test; the broader GenAI coverage lives in the unit suites (gen-ai-provider 57, session-tracing 131, etc.). Wording only.

Verdict: build + unit + integration + real-flow A/B all pass; emitted telemetry matches the field contract. LGTM from a verification standpoint.

中文版本（点击展开）

本地构建与真实流程验证（维护者合并参考）

对 head 726618f 的独立本地验证：真实构建、改动文件的测试套件、通过构建后 CLI 的完整端到端遥测运行，以及基线 vs PR 在真实导出 telemetry.log 上的 A/B 对比。这补齐了现有代码评审所缺少的本地运行证据。环境：macOS、Node 22.23.1，在 PR head 处使用隔离 git worktree，并在 merge-base 32c491fc6 单独构建基线 bundle。

结论 — ✅ 已验证，可合并

构建干净，所执行的套件全绿，导出的 span 完全符合文档定义的字段契约——新增的每个 gen_ai.* 字段都在、被移除的每个别名都已消失、私有兼容字段全部保留。唯一的合并前事项是 @yiliang114 已提出的 release note 破坏性变更说明；下方 A/B 将这些被移除的字段具体呈现。

1. 构建与测试

阶段	结果
npm run build:packages（tsc 全 workspace）+ npm run bundle	✅ exit 0（PR 与基线两套 bundle）
单元测试——改动文件（vitest, packages/core）	✅ 10 文件, 1257 个测试
单元测试——packages/cli ACP 遥测（Session.test.ts）	✅ 391 个测试
集成测试——cli/gen-ai-telemetry.test.ts（真实 E2E：构建后的 dist/cli.js + fake OpenAI server）	✅ 1 个测试
安全——URL 边界安全的 provider 解析	✅ 伪造/近似域名均被拒绝

该 head 上游 CI Test (ubuntu-latest, Node 22.x) 同样为绿（ci-bot 之前的"降级"源自一次已恢复的临时失败）。

2. 真实流程 A/B——活体遥测上的字段契约

我以完全相同的 LLM → tool → LLM 回合分别跑基线 bundle（未应用 PR）与 PR bundle，再从导出的 telemetry.log 读回 span。server 返回请求模型 request-model、不同的响应模型 provider-model-tool、provider tool-call id provider-call-123、以及 cached_tokens: 3——因此每个迁移字段都可观测。

线上确认，与设计文档完全一致：

• 新增（OTel GenAI / ARMS 对齐）： LLM span 上的 gen_ai.operation.name、gen_ai.provider.name、gen_ai.conversation.id、gen_ai.response.model、gen_ai.usage.cache_read.input_tokens；tool span 上的 gen_ai.operation.name=execute_tool、gen_ai.tool.name、gen_ai.tool.type=function、gen_ai.tool.call.id。
• 移除（私有别名 / 无效 GenAI 键），无双写： qwen-code.model、response_id、gen_ai.server.time_to_first_token、gen_ai.usage.cached_tokens、gen_ai.usage.reasoning_tokens、tool.name。
• 来源保真： 请求/响应模型拆分真实（request-model vs provider-model-tool）；tool-call id 优先 provider id；私有兼容键（input_tokens、cached_input_tokens、ttft_ms、thoughts_token_count、tool.call_id）均保留。

3. Provider 解析可防伪

直接调用构建后的 resolveGenAiProviderName（图 1 底部面板）：真实子域 x.dashscope.aliyuncs.com → dashscope，而伪造的 dashscope.aliyuncs.com.attacker.com 与近似的 xdashscope.aliyuncs.com 都正确地回落到默认值，不会被误判。未知 host 回落到协议 provider，且不泄漏 host/凭据；qwen-oauth 与 apiKeyEnvKey 路径按文档解析。

给作者 / 评审的备注

1. @yiliang114 第 (2) 点仅为书写冗余，非 bug。 session-tracing.ts 中 cacheCreationInputTokens 的判断 Number.isSafeInteger(x) && x !== undefined 里 !== undefined 冗余（Number.isSafeInteger(undefined) === false），无功能影响，可清理也可保留。
2. 破坏性变更说明（@yiliang114 第 (1) 点）： 上述被移除的键即具体影响面。建议在 release note 显著提示，以便迁移任何依赖 qwen-code.model / response_id / tool.name / 三个无效 gen_ai.usage.* 键的 ARMS/Grafana 看板。按设计文档为有意为之，此处仅作提示，不阻塞。
3. 小项： PR 描述中"9 个 GenAI telemetry 集成测试"——新的集成文件 gen-ai-telemetry.test.ts 实际含 1 个端到端测试；更广的 GenAI 覆盖在单元套件里（gen-ai-provider 57、session-tracing 131 等）。仅措辞问题。

结论： 构建 + 单元 + 集成 + 真实流程 A/B 全部通过；导出遥测符合字段契约。从验证角度 LGTM。

_{Verified locally on an isolated worktree at 726618f with a base bundle at merge-base 32c491fc6; screenshots are real terminal output rendered from the actual test runs and telemetry export.}

[doudouOUC]

Addressed all three currently unresolved review threads in c112319: standardized the safe-integer guard order, aligned DashScope internal and Azure OpenAI provider resolution with regression coverage, and preserved provider-reported Anthropic message_start usage on partial stream failure without synthesizing output. Also removed the 70/30 total-only token split and the remaining private LLM token aliases in 4752b91. Verification: 766 targeted Core tests, Core build/typecheck, targeted ESLint, bundle generation, and the GenAI telemetry JSONL integration test all pass. The local full-repository build remains blocked only by the installed Ink types lacking current main selection extensions, as documented in the PR body.

[wenshao]

@qwen-code /triage

[qwen-code-ci-bot]

Triage re-run completed without a new review.

The stage comments above were updated with the latest result. View workflow run.

[qwen-code-ci-bot]

Reviewed — no blockers. 1 Suggestion-level finding(s) could not be anchored to a changed line and were dropped; nothing further to act on here.

— qwen3.7-max via Qwen Code /review

[yiliang114]

LGTM — the provider-resolution and partial-stream follow-ups are addressed, and the current CI checks are green.