GenAI and ARMS field alignment

Scope and standards baseline

This design aligns the first set of Qwen Code span attributes whose names, types, and meanings agree between OpenTelemetry GenAI semantic conventions and Alibaba Cloud ARMS LLM Trace. It does not change span names, span kinds, parenting, or retry topology. It also documents the opt-in ARMS-only end-user identity extension.

The OpenTelemetry GenAI convention is still Development status. This change is pinned to commit 2e994c6d59a93bb4fc1752c5378eedb9b8e14d6b:

• Inference spans
• Agent spans
• GenAI registry

The streaming attributes are a narrow supplement pinned to OpenTelemetry Semantic Conventions v1.41.0. This supplement adopts only gen_ai.request.stream and gen_ai.response.time_to_first_chunk; it is not a wholesale upgrade of the baseline above.

The ARMS baseline is LLM Trace field definitions. An upgrade to either baseline requires regenerating and reviewing this matrix.

Field contract

Private attributes without an exact standard equivalent remain available for compatibility unless explicitly listed for removal below. Exact-equivalent private aliases and invalid GenAI aliases are removed without a dual-write period:

gen_ai.response.finish_reasons now preserves the provider's raw strings for all candidates instead of the previous Gemini-normalized values. Existing queries that filter values such as STOP or MAX_TOKENS must migrate to the provider values, such as stop, length, tool_calls, or end_turn.

gen_ai.response.time_to_first_chunk uses a monotonic timer from immediately before the wrapped provider call to the first normalized GenerateContentResponse observed by LoggingContentGenerator. Provider adapters may filter or merge raw protocol frames before they reach the logging wrapper, so frames an adapter drops (for example, the OpenAI pipeline's empty-response filter) are excluded from this measurement and the recorded value may be later than the true first network frame. Metadata-only and usage-only normalized responses that survive adapter filtering count as chunks. The attribute is retained if the stream later fails, is aborted, or times out, and is omitted when no chunk arrives.

The internal ttftMs timer remains first-user-visible-output latency and continues driving ApiResponseEvent.ttft_ms, sampling_ms, output_tokens_per_second, and the API request breakdown metric. Therefore, duration_ms - gen_ai.response.time_to_first_chunk * 1000 is not sampling_ms.

Existing streaming-Span queries should replace llm_request.stream=true with gen_ai.request.stream=true; non-streaming spans are identified by the absence of gen_ai.request.stream (the old llm_request.stream=false filter now matches zero rows). Span ttft_ms remains available for first-user-visible-output latency; gen_ai.response.time_to_first_chunk is an independent standard attribute measuring first normalized chunk latency in seconds.

Provider and operation resolution

Resolution is a pure function over the effective content-generator config. It never returns a URL, credential, arbitrary proxy hostname, or a value inferred from the model name.

1. Qwen OAuth and an exact DASHSCOPE_PROXY_BASE_URL match resolve to dashscope.
2. A boundary-safe hostname match recognizes Alibaba Model Studio endpoints and internal Alibaba gateways, Azure OpenAI, and the supported third-party endpoints (DeepSeek, xAI, Mistral, MiniMax, Z.AI, ModelScope, MiMo, OpenRouter, and Requesty).
3. If the host is unknown, a known apiKeyEnvKey identifies the configured provider. Host identity wins on conflict.
4. Unknown endpoints fall back to the protocol provider: openai, anthropic, gcp.gemini, or gcp.vertex_ai.

OpenAI-compatible, Anthropic, and Qwen OAuth requests use operation chat. Gemini and Vertex AI requests use generate_content.

Request parameters

Request attributes are collected after provider adapters have applied defaults, overrides, unsupported-field removal, and output-window clamps, immediately before calling the provider SDK. This is the final SDK request object visible to Qwen Code, not the original logical configuration or the serialized HTTP body. A logical LLM span records only its first such request snapshot.

Finite numbers and safe integers are preserved exactly, including zero and negative values on failed provider requests. Choice count is omitted when it is one. Stop sequences must be a complete string array; OpenAI's single-string form is normalized to a one-element array. Empty arrays are retained and mixed arrays are omitted rather than filtered. Explicit adapter defaults are recorded, while implicit SDK or server defaults are not inferred.

When multiple OpenAI-compatible output-budget aliases are present, the standard maximum is emitted only if all present values are valid safe integers and equal. Conflicting values are omitted because compatible endpoints do not have a common precedence rule.

Content and tool payloads

Sensitive GenAI content is collected only when telemetry.includeSensitiveSpanAttributes is enabled. Qwen Code does not read OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT, so there is a single content-capture switch. OpenAI-compatible, Anthropic, Gemini, and Vertex adapters convert their provider-final SDK request and raw response structures to the JSON schemas pinned with this design.

The first physical request attempt supplies input messages, system instructions, and tool definitions. Responses are generation-bound: a provider fallback or required-thinking retry starts a new response accumulator, and late chunks from an older attempt are ignored. Streaming accumulators retain canonical parts rather than raw chunks. Partial failures mark unfinished candidates with error; a successful response with a candidate that lacks an explicit finish reason omits the complete output-message attribute.

Each JSON attribute is compactly serialized and independently limited by telemetry.sensitiveSpanAttributeMaxLength. Invalid, cyclic, incomplete, or oversized attribute values are omitted as a whole; JSON is never truncated. Within gen_ai.tool.definitions, type and name are required identities, so an invalid identity omits the complete attribute. parameters is optional in the standard schema; when a provider-supplied parameter schema cannot be normalized to Draft-07, only that optional property is omitted while the ordered tool identity list is retained. Empty arrays and objects are retained when the provider explicitly sends or returns them. With the default 1 MiB limit, the application-side theoretical maximum is about 4 MiB of sensitive attributes per LLM span and 2 MiB per Tool span. Collectors and backends can impose lower limits.

Tool arguments are captured from the final invocation parameters immediately before execution, after permission and edit hooks. A tool result is captured only after a successful call and successful post-processing, from the final FunctionResponse.response object returned to the model. Both roots must be JSON objects. gen_ai.tool.description comes from the static registry description and is not sensitive; it is limited to 4096 UTF-16 code units, preserves surrogate pairs, and appends …[truncated] when shortened. Agent descriptions and span errors retain their 1024-unit limits.

Response and usage provenance

Provider converters attach internal provenance to normalized Gemini usage objects with a WeakMap. It records whether a cache-read field was actually present and Anthropic cache-creation tokens. This preserves the public response JSON shape and lets garbage collection follow the normalized usage object.

When an OpenAI-compatible provider reports only total_tokens, the normalized total remains available to existing internal consumers, but no input/output split is synthesized and neither standard usage attribute is emitted.

OpenAI response.model/chunk.model and Anthropic message model are preserved as modelVersion. A missing provider model remains missing for tracing; request-model fallback remains limited to existing API logs and UI behavior. Stream merging carries the last known provider model and usage provenance into the terminal response. Anthropic message_start input and cache usage is attached to the first subsequent yielded chunk so partial stream failures retain provider-reported usage without synthesizing an output count.

ARMS configuration

ARMS automatic GenAI application recognition requires this resource attribute:

{
  "telemetry": {
    "resourceAttributes": {
      "acs.arms.service.feature": "genai_app"
    }
  }
}

Qwen Code does not inject that vendor-specific resource attribute or gen_ai.span.kind. ARMS can infer LLM, Tool, and Agent roles from gen_ai.operation.name.

ARMS end-user identity extension

gen_ai.user.id is an ARMS Span common attribute, not part of the pinned OpenTelemetry GenAI baseline above. Qwen Code emits it only when the operator explicitly configures telemetry.userId or QWEN_TELEMETRY_USER_ID. The value is placed on the interaction Span at creation and propagated through the existing in-process context to LLM, Tool, and Agent spans, including linked-root fork/background agents. Tool-result continuations resolve the same logical interaction by prompt ID without changing Span parenting; that minimal identity entry expires with the existing 30-minute Span safety-net TTL.

The value is never inferred, generated, written to Resource/logs/metrics, or placed in outbound Baggage. Qwen Code does not dual-write enduser.id or user.id. A previous telemetry.resourceAttributes.user.id remains a generic Resource dimension and must be removed explicitly when migrating. Because the setting is process-wide, it is supported only when one process represents one end user; request-scoped identity for shared daemon and channel deployments is deferred until their trusted caller identity can be wired end to end.

Deferred work

• seed and top_k have incompatible ARMS and GenAI types in the baselines.
• Embedding needs a correct requested-model lifecycle before tracing.
• ARMS time-to-first-token and OpenTelemetry time-to-first-chunk differ in name, unit, and meaning. Qwen Code emits the standard gen_ai.response.time_to_first_chunk alongside the private ttft_ms and does not promise automatic population of an ARMS first-token dashboard.
• Full GenAI span naming, CLIENT span kind, and logical retry topology are a separate compliance project.