Adding a New AI Integration

Conventions First

Span ops and attributes are specified outside this repo. Never invent or hardcode either:

• gen_ai attributes and gen_ai ops — normative; import from @sentry/conventions/attributes and @sentry/conventions/op
• RFC 0153 — why Sentry's gen-AI conventions diverge from the OTel gen-ai semconv

Derive the op with getGenAiSpanOp() from ai/core/utils.ts rather than picking one by hand. ai/core/gen-ai-attributes.ts is for gap-fillers only — keys with no @sentry/conventions equivalent — so check it last, not first.

Which Pattern

Does the SDK publish its own `diagnostics_channel` telemetry?
|- YES (ai >= 7) -> Pattern 1: Native tracing channel
+- NO -> Does the SDK expose callback/exporter hooks?
    |- YES (LangChain, Mastra) -> Pattern 3: Callback/Exporter
    +- NO (OpenAI, Anthropic, Google GenAI, ai < 7) -> Pattern 2: Orchestrion-injected channels

What the reference files won't tell you:

• A provider can need two patterns at once: vercelAIIntegration subscribes to native ai:telemetry for ai >= 7 and runs orchestrion injection for v4-v6.
• Pattern 1 subscribers are safe to register unconditionally — subscribing is a no-op on SDK versions that never publish.

Where The Code Goes

• Instrumentation -> packages/server-utils/src/ai/{provider}/
• Integration -> packages/server-utils/src/integrations/{provider}.ts
• Runtime packages (node, cloudflare, bun, ...) re-export from @sentry/server-utils — they never define their own
• Exception: Workers AI is client-wrapped in packages/cloudflare/src/instrumentations/worker/instrumentEnv.ts

Gotchas

1. Detect streaming from the result shape — an async-iterable or the SDK's stream object — not from params.stream. Only the manual instrument{Provider}Client() API keys off params.stream === true.
2. Never set streamed response attributes by hand. Accumulate into a StreamResponseState and call endStreamSpan() (ai/openai/streaming.ts for an async generator, ai/anthropic-ai/streaming.ts for a listener-based stream).
3. Never truncate message payloads. Truncation was removed in v11 (#23045) and nothing downstream caps them; size limiting is server-side.
4. Never roll child token usage up onto parent spans. Tree totals are computed product-side from the full span tree.
5. Never read dataCollection.genAI directly. Gate input/output recording on resolveAIRecordingOptions().
6. LangChain must be registered first in getTracingIntegrations(), so it can disable the provider integrations before they instrument.
7. Set SEMANTIC_ATTRIBUTE_SENTRY_ORIGIN = 'auto.ai.{provider}' (alphanumerics, _, . only).

Checklist

• Instrumentation in src/ai/, integration in src/integrations/, registered in getTracingIntegrations() (LangChain first)
• Exported from packages/server-utils/src/index.ts, re-exported from the supported runtime packages
• E2E tests in dev-packages/node-integration-tests/suites/tracing/{provider}/ (and cloudflare-integration-tests/ if supported)
• Ops and attributes from @sentry/conventions, op derived via getGenAiSpanOp()
• Recording gated on resolveAIRecordingOptions(); no truncation, no token rollup
• JSDoc names the channels subscribed to, the supported SDK versions, and — for Pattern 2 — that it requires the Sentry runtime hook or bundler plugin
• Patching happens only once the target package is imported (zero cost if unused)

When in doubt, follow the pattern of the most similar existing integration.