forge-observe 0.1.2

OpenTelemetry instrumentation for qwen-think reasoning sessions.

forge-observe

OpenTelemetry instrumentation for qwen-think reasoning sessions.

Emits traces and metrics for thinking budget consumption, mode switches, token splits, preserve_thinking utilization, and backend flag normalization. Works with any OTel-compatible backend (Jaeger, Grafana Tempo, Datadog, etc.).

Install

pip install forge-observe

# For OTLP export (Jaeger, Tempo, Datadog):
pip install forge-observe[otlp]

Quick start

Auto-instrumentation (recommended)

from forge_observe import instrument
instrument()

# Now use qwen-think as normal -- spans and metrics are emitted automatically.
from qwen_think import ThinkingSession
session = ThinkingSession(client=your_client)
response = session.chat("Implement a binary search tree")

Manual tracing

from forge_observe import ForgeTracer

tracer = ForgeTracer()
with tracer.session(attributes={"model": "Qwen3.6-35B"}) as span:
    with tracer.think_decision(mode="think", reason="coding task"):
        pass
    with tracer.budget_check(
        total_tokens=200000, used_tokens=50000,
        available_tokens=150000, action="ok",
    ):
        pass

Configuration

from forge_observe import configure, instrument

configure(
    service_name="my-app",
    otlp_endpoint="http://localhost:4317",
    otlp_protocol="grpc",
)
instrument()

All settings can be overridden via environment variables:

Variable	Default	Description
FORGE_OBSERVE_SERVICE_NAME	forge-observe	OTel service name
FORGE_OBSERVE_OTLP_ENDPOINT	http://localhost:4317	OTLP collector endpoint
FORGE_OBSERVE_OTLP_PROTOCOL	grpc	grpc or http
FORGE_OBSERVE_ENABLED	true	Kill switch
FORGE_OBSERVE_CONSOLE_EXPORT	false	Print spans to stdout

What it instruments

Spans (auto-instrumented)

forge.session
  |-- forge.session.think_decision    router chose think/instruct, with reason
  |-- forge.session.param_swap        sampling params changed
  |-- forge.session.budget_check      budget evaluated, remaining tokens logged
  |-- forge.backend.normalize         flag normalization applied
  |-- forge.backend.response          response received, thinking/response split
  |-- forge.session.preserve_check    preserve_thinking reuse evaluated

Spans (manual via ForgeTracer)

forge.session.backend_call            wraps a backend API call (ForgeTracer.backend_call)
forge.mtp.config_decision             MTP config selected or disabled, with reason (ForgeTracer.mtp_config)

Metrics

Name	Type	Description
forge.session.budget_remaining	Gauge	Tokens remaining in budget
forge.session.thinking_tokens	Counter	Total thinking tokens emitted
forge.session.response_tokens	Counter	Total response tokens emitted
forge.session.mode_switches	Counter	Think/instruct mode flips
forge.session.preserve_reuse_rate	Gauge	Fraction of turns reusing prior thinking
forge.mtp.acceptance_rate	Gauge	Speculative token acceptance rate

Examples

See examples/ for setups targeting Jaeger, Grafana Tempo, and Datadog.

License

Apache-2.0