integration: validate OTLP export against Telemetry.dev #241

Goal

Run one real compatibility experiment between AgentTrace's OTLP/HTTP JSON exporter and Telemetry.dev, then document the integration only if the received trace preserves the evidence needed for review.

This is the distribution experiment proposed in the Telemetry.dev learning, not a dependency on a hosted backend.

Experiment

1. Create or select a non-sensitive representative session containing:
   • parent/child span structure;
   • provider and model attribution;
   • tool calls and results;
   • a failure and retry;
   • token and estimated-cost fields;
   • redacted content; and
   • at least one known capture gap.
2. Export it through the current OTLP/HTTP JSON path to Telemetry.dev's documented /v1/traces endpoint.
3. Inspect the received trace and record:
   • hierarchy preservation;
   • mapped, renamed, dropped, or rejected attributes;
   • model/provider/tool representation;
   • token, cost, error, and status behavior;
   • redaction behavior; and
   • whether AgentTrace evidence-health or provenance fields survive.
4. Compare otlp and otlp-genai where useful.
5. Publish a backend guide only if the mapping is accurate and reproducible.
6. If it passes, propose an AgentTrace listing in Telemetry.dev's coding-agent integration directory.

Safety and boundaries

• Use synthetic or explicitly approved non-sensitive data.
• Do not send a customer session or repository content.
• Do not add Telemetry.dev as a required dependency or default endpoint.
• Do not claim compatibility based only on an HTTP 2xx response.
• Document limitations prominently if the backend cannot preserve review-critical semantics.

Acceptance criteria

• A real export reaches a Telemetry.dev project.
• The exact AgentTrace and backend versions and test date are recorded.
• The sent payload and received mapping can be compared without secrets.
• Dropped, renamed, or flattened attributes are listed.
• Hierarchy, errors, tokens, cost, redaction, provenance, and gap semantics are explicitly evaluated.
• A pass, partial, or fail decision is recorded.
• Documentation and directory outreach happen only after a pass or clearly bounded partial result.