Skip to main content
Halios accepts OpenTelemetry traces from standard OTel SDKs and common AI instrumentation libraries. If you already emit compatible telemetry, point your exporter at Halios. If not, your coding agent can add or adapt instrumentation.

​
Compatible instrumentation

Halios works with standard OpenTelemetry-compatible traces and auto-instrumentors:

​
Configure the OTLP exporter

Point your standard OpenTelemetry exporter to Halios using standard environment variables:

​
GenAI semantics Halios understands

Halios recognizes common GenAI operations such as:
  • Model calls (input prompts, output completions, token usage)
  • Tool execution (tool name, arguments, output results or errors)
  • Retrieval (search queries, returned document chunks and IDs)
  • Agent / Sub-agent delegation (multi-agent workflows, parent-child context)
Where available, Halios uses OpenTelemetry GenAI semantic convention attributes such as gen_ai.operation.name, model/provider metadata, tool names, messages, usage, and related execution context.
For effective evaluations and debugging, aim to send:
  • One trace per agent interaction or run
  • Parent-child span relationships preserved across async calls
  • Model and tool operation names
  • Tool arguments and results where available
  • Input and output messages if evaluations inspect them
  • Service and environment metadata (service.name, deployment.environment.name, service.version)
Halios can ingest partial traces, but richer traces produce better debugging and evaluation data.

​
Redacting sensitive data

Trace payloads may contain prompts, user data, and tool arguments. Use your OpenTelemetry processor or exporter to redact or drop sensitive fields before export when required.

​
Rate limits and retries

Rate-limited OTLP responses include the structured RateLimit-Policy and RateLimit HTTP fields. The policy field describes the applicable request and content-byte windows; the current field reports the remaining quota and effective reset time. For example:
When Halios returns 429 Too Many Requests, it also returns Retry-After as a non-negative number of seconds. Exporters and coding agents should pause for at least that duration before retrying and should use exponential backoff with jitter for subsequent failures. Capacity-backpressure responses can also include Retry-After without a quota policy. Treat that value as authoritative for the affected request.

​
Verify traces with the CLI

You can inspect and verify stored traces using the Halios CLI:
Verification confirms that traces have valid W3C IDs, correct parent-child relationships, ended statuses, and captured operation metadata.

CLI reference

Command reference for automation and CI pipelines.

Evaluation concepts

Core mental model and key evaluation terms.