forgesight-datadog

The Datadog exporter for ForgeSight. Surfaces agent telemetry in Datadog APM with the unified service / env / version tags, LLM / tool / MCP calls as child spans, and the SDK's computed cost as the monitorable DD metric forgesight.cost_usd — the same number every other backend reports.

pip install forgesight-datadog

import forgesight
from forgesight_datadog import DatadogExporter

forgesight.configure(exporters=[
    DatadogExporter(api_key="...", site="datadoghq.com",
                    service="issue-classifier", env="prod"),
])

Or by name: exporters: [{name: datadog, config: {api_key: "${DD_API_KEY}", service: ...}}].

Two transports

• agent (default) — maps each record to a DD APM span via ddtrace and writes it to a local DD Agent (agent_endpoint: http://datadog-agent:8126), plus emits cost/token DD metrics. Direct intake (no agent_endpoint) requires api_key.
• otlp — sends OTLP/HTTP to the DD Agent's OTLP port (agent_endpoint: http://datadog-agent:4318) with the DD unified tags applied as resource attributes. Reuses forgesight-otel.

A DD Agent / intake outage makes export() return FAILURE (counted, never raised — P6); it never blocks the agent. Prompt/response content is attached only with capture_content=True (off by default, P7).

OTLP-native backends need no package

Because the domain model maps cleanly onto the OTel GenAI conventions, anything that ingests OTLP works through forgesight-otel with no dedicated package — point it at the backend and you're done:

Datadog earns a package only because its richest path (DD-native APM tagging + cost-as-DD-metric) is DD-specific. A team that only wants generic spans in Datadog can use the OTLP path and skip this package entirely.

Configuration

Constructor kwargs win over env (FR-12).

License

Apache-2.0