feat: add OpenTelemetry bridge for AI observability

[posthog]

💡 Motivation and Context

• Go teams building agent backends had no supported path to PostHog AI observability. Over the last 30 days, not a single $ai_* event arrived with $lib = posthog-go, which matched the code: grepping the SDK for $ai_generation, $ai_span, or anything llm-shaped returned nothing. Those users got a flat "no" on a product they were otherwise ready to pay for.
• The ingestion side already worked. posthog-python ships a generic OpenTelemetry bridge that keeps gen_ai./llm./ai./traceloop. spans and forwards them to /i/v0/ai/otel. Google's Agent Development Kit (ADK) for Go emits exactly those gen_ai.* spans — there was just nothing on the Go side pointing at them.

This PR ports that bridge to Go.

• New github.com/posthog/posthog-go/otel module (posthogotel):
  • SpanProcessor — recommended integration; register it on a TracerProvider. Filters, batches, and exports AI spans.
  • Exporter — for frameworks that accept only a span exporter.
  • IsAISpan — the shared prefix filter (gen_ai., llm., ai., traceloop.), matched against the span name and every attribute key.
  • Both target {host}/i/v0/ai/otel over OTLP/HTTP with Authorization: Bearer <project api key>; WithHost switches cloud region.
• example/ wires the processor to the tracer provider an ADK Go agent emits through.

Why a separate Go module

The core SDK today has a lean dependency graph. Folding the OTel SDK, otlptracehttp, and their protobuf/grpc transitive tree into the root go.mod would burden every posthog-go consumer for an optional feature. A nested module keeps those dependencies opt-in, so only users who import .../otel pull them.

💚 How did you test it?

• Unit tests over IsAISpan: name-prefix match, attribute-key match, and a non-AI negative, using real ReadOnlySpans from a SpanRecorder.
• End-to-end tests against an httptest OTLP server that decodes the protobuf payload and asserts:
  • only the AI span (gen_ai.chat) reaches the endpoint when an AI and a non-AI span are emitted together, for both SpanProcessor and Exporter;
  • the request path is /i/v0/ai/otel and the header is Authorization: Bearer <key>;
  • a batch of only non-AI spans produces no HTTP request.
• Empty-API-key rejection and WithHost normalisation (blank falls back to default, trailing slash trimmed).
• go build, go vet, and go test -race pass in the module. Added an otel-bridge CI job (Go 1.25 and 1.x), since the root ./... does not reach a nested module.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed.
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran pnpm changeset to generate a changeset file

The root SDK and nested otel module share one release version and cadence. The release workflow creates both vX.Y.Z and otel/vX.Y.Z tags at the same release commit, matching the synchronized nested-module approach used by Sentry Go.

🤖 Agent context

Autonomy: Fully autonomous

• Authored by Claude (via PostHog Desktop) from an inbox report about a customer asking for ADK Go support in AI observability.
• Read the posthog-python posthog/ai/otel/ bridge (spans.py, exporter.py, processor.py) as the reference and ported its behaviour.
• Chose a separate Go module over inline .go files in the root package: the report suggested following the error_tracking*.go pattern, but that product added no third-party dependencies, whereas the OTel bridge pulls in grpc/protobuf. Isolating it keeps the core SDK lean.
• Pinned OpenTelemetry to v1.43.0 and the bridge module to Go 1.25 to avoid reachable vulnerabilities in older OTel releases; the separate core module remains on Go 1.21.
• Kept the example credential-free and runnable: real ADK wiring needs model credentials and a heavy dependency, so the example registers the processor on the tracer provider ADK Go emits through and emits one representative gen_ai.* span, with the ADK hookup documented.

---

Created with PostHog Desktop from this inbox report.

[posthog]

🦔 PostHog Review reviewed this pull request

Found 3 must fix, 8 should fix, 0 consider.

Published 11 findings (view the review).

Resolved comments: 8 fixed, 3 left for you

[github-actions]

posthog-go Compliance Report

Date: 2026-09-01 15:24:01 UTC
Duration: 169005ms

✅ All Tests Passed!

111/111 tests passed

---

Capture_V1 Tests

✅ 94/94 tests passed

View Details

Test	Status	Duration
Endpoint And Method.Targets V1 Endpoint	✅	22ms
Endpoint And Method.Does Not Use Legacy Endpoints	✅	6ms
Required Headers.Has Authorization Bearer Header	✅	5ms
Required Headers.Has Content Type Json	✅	16ms
Required Headers.Has Posthog Sdk Info Format	✅	5ms
Required Headers.Has Posthog Attempt Header	✅	6ms
Required Headers.Has Posthog Request Id	✅	4ms
Required Headers.Has Posthog Request Timestamp	✅	16ms
Required Headers.Has User Agent	✅	6ms
Body Format.Body Has Created At And Batch	✅	6ms
Body Format.No Api Key In Body	✅	5ms
Body Format.No Sent At In Body	✅	5ms
Event Format.Event Has Required Root Fields	✅	5ms
Event Format.Event Uuid Is Valid	✅	5ms
Event Format.Event Timestamp Is Rfc3339	✅	6ms
Event Format.Distinct Id Is String	✅	5ms
Event Format.Distinct Id At Root Not Properties	✅	5ms
Event Format.Custom Properties Preserved	✅	5ms
Event Format.Set Properties Preserved	✅	5ms
Event Format.Set Once Properties Preserved	✅	5ms
Event Format.Groups Properties Preserved	✅	5ms
Event Format.Sdk Generates Uuid If Not Provided	✅	5ms
Event Format.Event Has Required Root Fields Batch	✅	6ms
Event Format.Event Uuid Is Valid Batch	✅	8ms
Event Format.Event Timestamp Is Rfc3339 Batch	✅	6ms
Event Format.Distinct Id Is String Batch	✅	7ms
Event Format.Distinct Id At Root Not Properties Batch	✅	7ms
Event Format.Custom Properties Preserved Batch	✅	7ms
Event Format.Set Properties Preserved Batch	✅	7ms
Event Format.Set Once Properties Preserved Batch	✅	7ms
Event Format.Groups Properties Preserved Batch	✅	7ms
Event Format.Sdk Generates Uuid If Not Provided Batch	✅	6ms
Batch Behavior.Multiple Events In Single Batch	✅	30ms
Batch Behavior.Batch Envelope Smoke	✅	9ms
Batch Behavior.Flush With No Events Sends Nothing	✅	2ms
Batch Behavior.Flush At Triggers Batch	✅	1007ms
Batch Behavior.Created At Reflects Batch Creation Time	✅	6ms
Deduplication.Generates Unique Uuids	✅	13ms
Deduplication.Different Events Same Content Different Uuids	✅	7ms
Deduplication.Preserves Uuid On Retry	✅	5133ms
Deduplication.Preserves Timestamp On Retry	✅	5132ms
Deduplication.Preserves Uuid And Timestamp On Batch Retry	✅	5136ms
Deduplication.No Duplicate Events In Batch	✅	29ms
Header Behavior On Retry.Attempt Header Starts At One	✅	5ms
Header Behavior On Retry.Attempt Header Increments On Retry	✅	10139ms
Header Behavior On Retry.Request Id Preserved On Retry	✅	5131ms
Header Behavior On Retry.Different Requests Have Different Request Ids	✅	2022ms
Header Behavior On Retry.Request Timestamp Changes On Retry	✅	5135ms
Response Format Validation.Success Response Has Uuid Keyed Results	✅	7ms
Response Format Validation.Success Response Has Ok For Each Event	✅	26ms
Response Format Validation.Success No Retry After When All Ok	✅	27ms
Response Format Validation.Success Retry After Present When Retry Events	✅	128ms
Response Format Validation.Success No Retry After When Drop Only	✅	27ms
Response Format Validation.Response Echoes Request Id	✅	5ms
Retry Behavior.Retries On 408	✅	5133ms
Retry Behavior.Retries On 500	✅	5133ms
Retry Behavior.Retries On 503	✅	5135ms
Retry Behavior.Retries On 504	✅	5135ms
Retry Behavior.Retryable Errors Have Retry After	✅	2132ms
Retry Behavior.Respects Retry After On Retryable Error	✅	8133ms
Retry Behavior.Does Not Retry On 400	✅	2132ms
Retry Behavior.Does Not Retry On 401	✅	2131ms
Retry Behavior.Does Not Retry On 402	✅	2132ms
Retry Behavior.Does Not Retry On 413	✅	2131ms
Retry Behavior.Does Not Retry On 415	✅	2131ms
Retry Behavior.Non Retryable Errors Have No Retry After	✅	2133ms
Retry Behavior.Implements Backoff	✅	15131ms
Retry Behavior.Max Retries Respected	✅	15145ms
Partial Batch Handling.Handles 200 Full Success	✅	2009ms
Partial Batch Handling.Handles 200 With All Ok	✅	3030ms
Partial Batch Handling.Does Not Retry Dropped Events	✅	3028ms
Partial Batch Handling.Does Not Retry Limited Events	✅	3031ms
Partial Batch Handling.Prunes Ok Events On Partial Retry	✅	5134ms
Partial Batch Handling.Prunes Dropped Events On Partial Retry	✅	5135ms
Partial Batch Handling.Retries Only Retry Events From Partial	✅	5133ms
Partial Batch Handling.Partial Retry Preserves Uuids	✅	5133ms
Partial Batch Handling.Partial Retry Attempt Header Increments	✅	5136ms
Partial Batch Handling.Partial Retry Request Id Preserved	✅	5135ms
Partial Batch Handling.Respects Retry After On Partial	✅	5133ms
Partial Batch Handling.Unknown Result Treated As Terminal	✅	3029ms
Partial Batch Handling.Mixed Ok Drop Limited No Retry	✅	3032ms
Compression.Sends Gzip Content Encoding	✅	15ms
Compression.No Content Encoding When Disabled	✅	5ms
Compression.Compressed Body Is Decompressible	✅	5ms
Error Handling.Does Not Retry On Unknown 4Xx	✅	2131ms
Event Options.Cookieless Mode Override	✅	6ms
Event Options.Disable Skew Correction Override	✅	4ms
Event Options.Process Person Profile Override	✅	5ms
Event Options.Product Tour Id Override	✅	4ms
Event Options.Unset Options Omitted	✅	5ms
Event Options.Options Override In Batch	✅	7ms
Geoip And Historical Migration.Geoip Disable Injected Into Properties	✅	4ms
Geoip And Historical Migration.Historical Migration Set In Body	✅	5ms
Geoip And Historical Migration.Historical Migration Absent By Default	✅	5ms

Feature_Flags Tests

✅ 17/17 tests passed

View Details

Test	Status	Duration
Request Payload.Request With Person Properties Device Id	✅	14ms
Request Payload.Flags Request Uses V2 Query Param	✅	15ms
Request Payload.Flags Request Hits Flags Path Not Decide	✅	15ms
Request Payload.Flags Request Omits Authorization Header	✅	14ms
Request Payload.Token In Flags Body Matches Init	✅	14ms
Request Payload.Groups Round Trip	✅	15ms
Request Payload.Groups Default To Empty Object	✅	15ms
Request Payload.Disable Geoip False Propagates As Geoip Disable False	✅	15ms
Request Payload.Disable Geoip Omitted Defaults To False	✅	16ms
Request Payload.Flag Keys To Evaluate Contains Only Requested Key	✅	15ms
Request Lifecycle.No Flags Request On Init Alone	✅	2ms
Request Lifecycle.No Flags Request On Normal Capture	✅	5ms
Request Lifecycle.Two Flag Calls Produce Two Remote Requests	✅	28ms
Request Lifecycle.Mock Response Value Is Returned To Caller	✅	15ms
Retry Behavior.Retries Flags On 502	✅	17ms
Retry Behavior.Retries Flags On 504	✅	17ms
Side Effect Events.Get Feature Flag Captures Feature Flag Called Event	✅	16ms

[posthog]

PostHog Review alpha 🦔 If you find any issues helpful - please reply "valid", "invalid", etc., for evaluation purposes 🙏

[posthog]

PostHog Review

Found 3 must fix, 8 should fix.

[marandaneto]

Live US Cloud validation succeeded for the expanded example: the span was ingested as $ai_generation, and trace/span IDs, provider, model, messages, token counts, latency, base URL, and calculated costs were all populated as expected.

cc @PostHog/team-ai-observability
One caveat for reviewers: gen_ai.response.finish_reasons and gen_ai.response.id remain raw OpenTelemetry properties rather than being normalized to $ai_stop_reason and $ai_generation_id. The Go bridge is exporting the standard GenAI attributes correctly; this is a generic PostHog OTel ingestion normalization gap, not a Go bridge issue.

[marandaneto]

ingested event example

{
  "createdAt": "2026-08-31T09:44:37.958000+00:00",
  "event": "$ai_generation",
  "id": "01a05734-e115-7103-8a19-54b4da519853",
  "properties": {
    "$ai_ingestion_source": "otel",
    "$ai_span_id": "3d0d99fac071acbe",
    "$ai_trace_id": "b46ea40bb3555a8eb0be877729a2013b",
    "$geoip_disable": true,
    "example.run_id": "20260831T094437.958644000Z",
    "gen_ai.response.finish_reasons": [
      "stop"
    ],
    "gen_ai.response.id": "chatcmpl-posthog-go-example",
    "service.name": "posthog-go-otel-example",
    "$ip": "84.115.233.26",
    "$ai_input_tokens": 18,
    "$ai_output_tokens": 11,
    "$ai_model": "gpt-4o-mini-2024-07-18",
    "$ai_provider": "openai",
    "$ai_base_url": "api.openai.com",
    "$ai_span_name": "chat posthog-go OTel example",
    "$ai_latency": 0.051059583,
    "$ai_cache_reporting_exclusive": false,
    "$ai_input_cost_usd": 0.0000027,
    "$ai_output_cost_usd": 0.0000066,
    "$ai_request_cost_usd": 0,
    "$ai_web_search_cost_usd": 0,
    "$ai_total_cost_usd": 0.0000093,
    "$ai_model_cost_used": "openai/gpt-4o-mini-2024-07-18",
    "$ai_cost_model_source": "openrouter",
    "$ai_cost_model_provider": "openai",
    "$ai_input": [
      {
        "role": "system",
        "content": "Answer concisely."
      },
      {
        "role": "user",
        "content": "What is PostHog?"
      }
    ],
    "$ai_output_choices": [
      {
        "role": "assistant",
        "content": "PostHog is an open-source product analytics platform."
      }
    ]
  },
  "sentiment": null
}

[dustinbyrne]

Agent-led review, human-reviewed before posting.