hush-telemetry

External tracing backends for Hush workflows — Langfuse, OpenTelemetry, and local visualization.

Installation

pip install hush-telemetry                 # Langfuse + HushEyes (default)
pip install "hush-telemetry[otel]"         # + OpenTelemetry
pip install "hush-telemetry[all]"          # Everything

Quick Start

from hush.core import Hush, GraphOp, START, END
from hush.telemetry import LangfuseTracer

with GraphOp(name="demo") as graph:
    # ... define ops
    pass

tracer = LangfuseTracer(resource="langfuse:default", tags=["prod"])
engine = Hush(graph, tracer=tracer)
result = await engine.run(inputs={...})
# Traces automatically sent to Langfuse

Supported Backends

Usage Patterns

from hush.telemetry import HushEyesTracer, LangfuseTracer, LangfuseConfig

# Local visualization
tracer = HushEyesTracer(tags=["dev"])

# Langfuse with ResourceHub
tracer = LangfuseTracer(resource="langfuse:default")

# Langfuse with direct config
tracer = LangfuseTracer(config=LangfuseConfig.from_env())

# Multiple tracers at once
engine = Hush(graph, tracer=[HushEyesTracer(), LangfuseTracer(resource="langfuse:default")])

Configuration

# resources.yaml
tracer:
  langfuse:
    public_key: ${LANGFUSE_PUBLIC_KEY}
    secret_key: ${LANGFUSE_SECRET_KEY}
    host: https://cloud.langfuse.com

Features

• Hierarchical tracing — pre-computed TraceNode tree with parent-child relationships
• Token/cost tracking — automatic for LLM ops
• Stream sampling — cap trace items per generator (default: 100)
• Background flushing — non-blocking, ThreadPoolExecutor-based
• No SDK dependency — Langfuse uses custom HTTP client (pure REST API)

Related Packages

License

Apache 2.0