analyzing-llm-rationale: Skill for Claude Code

OTel FastAPI Style

Use native FastAPI instrumentation. Do not replace request handling with manual middleware just to create spans.

from fastapi import FastAPI
from opentelemetry.instrumentation.fastapi import FastAPIInstrumentor

def init_observability(app: FastAPI) -> bool:
    ...
    FastAPIInstrumentor.instrument_app(app)
    return True

Keep FastAPIInstrumentor.instrument_app(app) with the rest of the observability setup so the bootstrap is easy to reason about. If the generated token is not claimed yet because signup is still in progress, it is still okay to initialize providers; ingest may reject exports until the browser flow finishes.

Entrypoint

Initialize before serving user traffic.

app = FastAPI(title="Mugline API")
init_observability(app)

Bounded Work

Use module-scope OTel objects and decorators for helpers that auto-instrumented HTTP spans cannot see.

tracer = trace.get_tracer("mugline.api")
meter = metrics.get_meter("mugline.api")

@tracer.start_as_current_span("mug.recommend")
async def recommend_mug(*, tenant_id: str, preference: str) -> dict[str, str]:
    span = trace.get_current_span()
    span.set_attribute("tenant.id", tenant_id)
    ...

Logs

For OTLP-forwarded stdlib logs, configure all of these:

• LoggerProvider
• set_logger_provider(logger_provider)
• OTLPLogExporter
• LoggingHandler
• LoggingInstrumentor().instrument(...)

LLM Calls

LLM routes need token coverage:

• llm.tokens.input
• llm.tokens.output

Tag explicit token counters with tenant, provider, model, use case, call site, and outcome only when provider instrumentation cannot capture token usage. Do not add app-side LLM cost metrics or pricing tables; Superlog estimates cost centrally from provider/model/token data.