feat: add support for vercel/mastra sdks (#1548)

[dinmukhamedm]

Note

Medium Risk
Modifies LLM span classification and attribute extraction by normalizing new aisdk.* and operation-prefixed keys into existing gen_ai.*/ai.* fields; mistakes could impact token/cost calculation and stored trace payloads. Added tests reduce but don’t eliminate risk across diverse telemetry inputs.

Overview
Adds support for newer Vercel AI SDK / Mastra telemetry by normalizing aisdk.* and operation-prefixed attributes (e.g. stream.*, generateText.*) into the existing gen_ai.*/ai.* keys so the current input/output and usage extraction pipeline works.

Updates LLM detection and storage behavior: spans can be reclassified to LLM after normalization, llm.usage.total_tokens is now written when setting usage, and the original operation-prefixed attributes are dropped via should_keep_attribute to reduce stored payload size. Cost calculation now warns when gen_ai.system exists and tokens are present but no model name is available, and provider detection for AI SDK spans also recognizes aisdk.model.provider.

^{Reviewed by Cursor Bugbot for commit 46df0cf. Bugbot is set up for automated code reviews on this repo. Configure here.}

[greptile-apps]

Greptile Summary

This PR adds first-class support for Vercel AI SDK v5+ and Mastra SDK spans by normalizing their aisdk.*, stream.*, and generateText.* attribute namespaces into the existing gen_ai.* standard that the rest of Laminar's pipeline already understands. The implementation is clean, non-destructive (all normalization uses insert-if-absent semantics), and backed by thorough unit tests covering provider lowercasing, no-op behavior for non-aisdk spans, and multiple operation prefixes.

• normalize_aisdk_attributes() maps aisdk.model.id → gen_ai.request.model, normalizes aisdk.model.provider → gen_ai.system (lowercased first segment, e.g. "openai.chat" → "openai"), and remaps operation-prefixed usage/prompt/response attributes to their ai.*/gen_ai.* equivalents
• detect_aisdk_operation_prefix() auto-detects which prefix (stream/generateText/streamText/generateObject/streamObject) is in use — but currently omits {prefix}.response.toolCalls and {prefix}.response.object from its detection criteria even though it normalizes them
• is_ai_sdk_llm_span() extended with aisdk.model.provider check to guard Mastra spans from the Langchain-specific conversion path
• gateway → vercel_ai_gateway provider transform added for LiteLLM gateway spans; comment wording in this case is slightly misleading

Confidence Score: 4/5

Safe to merge; all issues are non-blocking style/clarity concerns with no runtime impact

Well-implemented feature with comprehensive test coverage (6+ new test cases), non-destructive insert-if-absent normalization semantics, and correct handling of all key edge cases. The only remaining items are: a minor inconsistency between detection and normalization scope in detect_aisdk_operation_prefix() (unlikely to matter in practice since real LLM calls with tool calls always include token usage data), and a backwards comment in utils.rs.

app-server/src/traces/spans.rs — review the detect_aisdk_operation_prefix() detection criteria gap for {prefix}.response.toolCalls and {prefix}.response.object

Important Files Changed

Filename	Overview
app-server/src/traces/spans.rs	Core aisdk normalization: adds normalize_aisdk_attributes(), detect_aisdk_operation_prefix(), and insert/normalize helpers; updates span_type() and should_keep_attribute() — well-tested, with a minor detection-criteria gap noted
app-server/src/traces/span_attributes.rs	Adds AISDK_MODEL_ID and AISDK_MODEL_PROVIDER constants for the new aisdk.* attribute namespace — minimal and correct
app-server/src/traces/provider/mod.rs	Extends is_ai_sdk_llm_span() to also check aisdk.model.provider, correctly preventing Langchain conversion for Mastra/newer Vercel SDK spans
app-server/src/traces/utils.rs	Adds gateway→vercel_ai_gateway provider transform for LiteLLM compatibility; existing comment wording is slightly backwards

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["Raw OTel span\n(aisdk.* attributes)"] --> B
    B["from_otel_span()\nspan_type() checks aisdk.* prefix\n→ SpanType::LLM"] --> C
    C["parse_and_enrich_attributes()"] --> D
    D["normalize_aisdk_attributes()"] --> E & F
    E["aisdk.model.id\n→ gen_ai.request.model\n→ ai.model.id"] --> G
    F["aisdk.model.provider\n→ ai.model.provider\n→ gen_ai.system\n   (lowercased 1st segment)"] --> G
    G["detect_aisdk_operation_prefix()\nstream / generateText / streamText\ngenerateObject / streamObject"] --> H
    H{Prefix found?} -->|Yes| I
    H -->|No| J
    I["Remap usage/prompt/response\n{prefix}.usage.inputTokens → gen_ai.usage.input_tokens\n{prefix}.prompt.messages → ai.prompt.messages\n{prefix}.response.toolCalls → ai.response.toolCalls"] --> J
    J["Re-evaluate span_type()\nafter gen_ai.system is populated"] --> K
    K["convert_span_to_provider_format()\nis_ai_sdk_llm_span():\nai.operationId OR ai.model.provider\nOR aisdk.model.provider\n→ skip Langchain conversion"] --> L
    L["tranform_model_and_provider()\ngateway → vercel_ai_gateway"]

Loading

Comments Outside Diff (1)

1. app-server/src/traces/utils.rs, line 351-352 (link)

   Comment describes the mapping backwards

   The comment reads "LiteLLM stores "gateway" as "vercel_ai_gateway"", which implies LiteLLM's internal label is "vercel_ai_gateway". The reality is the opposite: LiteLLM sends "gateway" as the provider name, and this code remaps it to "vercel_ai_gateway" for model-pricing lookup.

_{Reviews (1): Last reviewed commit: "remove redundant warnings" | Re-trigger Greptile}

[dinmukhamedm]

@laminar-coding-agent the normalized spans now have gen_ai.usage.* attributes for input_cost, output_cost, input_tokens, and output_tokens. Turns out there is frontend functionality that depends on the totals for these values. Let's make sure that after (during) normalization, total_cost is reflected as a gen_ai.usage.cost and total_tokens are reflected as llm.usage.total_cost.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit f9b4cfe. Configure here.}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
2 out of 3 committers have signed the CLA.

✅ PranshuSrivastava
✅ dinmukhamedm
❌ greptile-apps[bot]
_{You have signed the CLA already but the status is still pending? Let us recheck it.}