coding-sh (Python SDK)

Send LLM calls, traces and tool spans from any Python service to a coding.sh server (coding.sh serve). Stdlib only, Python 3.9+.

pip install coding-sh
export CODING_SH_URL=https://obs.example.com CODING_SH_INGEST_KEY=csh_ik_...

import coding_sh
from openai import OpenAI

client = coding_sh.observe(OpenAI())   # also AsyncOpenAI, Anthropic, AsyncAnthropic

That's it: every chat.completions.create, responses.create and messages.create / messages.stream call (sync, async and streaming) becomes an llm.call with tokens, model, latency, time to first token, stop reason and tool calls. The server prices it.

Without CODING_SH_URL the SDK does nothing. It never changes return values, errors or stream semantics, and never raises into your app.

Traces, tools and conversations

with coding_sh.trace("support-reply", session_id=conversation_id, user_id=user.id, input=question) as t:
    with coding_sh.span("tool", "lookup_order", input={"id": 42}) as s:
        s.set_output(orders.get(42))
    reply = client.chat.completions.create(model="gpt-4o-mini", messages=msgs)
    t.set_output(reply.choices[0].message.content)

• Each top-level trace() (or untraced call) is its own session. Pass session_id (or wrap code in coding_sh.with_session(id)) to make each trace a turn of one conversation.
• trace and span are context managers (with / async with) and decorators. A decorated span("tool", ...) records the call's arguments as input and its return value as output.
• span("step", name) records a generic step.
• Context follows contextvars, so it crosses await and asyncio tasks. Work handed to a thread pool starts untraced unless you run it with contextvars.copy_context().run.

LLM gateways

Point the client at a gateway and each call records which one it went through, and the request id the gateway's lookup API takes, so coding.sh gateway enrich can add the billed cost and coding.sh sessions --flag gateway:<name> finds the sessions. Detected: Vercel AI Gateway, OpenRouter, LiteLLM, Cloudflare AI Gateway, Portkey, Helicone, Requesty, Ramp Router (router.com), TrueFoundry, Kong, Bifrost and Martian, by base URL (a self-hosted gateway by its URL env var, e.g. LITELLM_PROXY_API_BASE), by response header (cf-aig-log-id, helicone-id, x-litellm-call-id, x-portkey-trace-id, …) or by response id (gen_…, gen-…, rqsty-…). observe() adds a response hook to the client's HTTP client that reads headers only, and only for calls it records. Details.

client = coding_sh.observe(OpenAI(base_url="https://gateway.ai.cloudflare.com/v1/<account>/<gateway>/compat"))

Delivery

One daemon thread sends batches (up to 500 events or 4 MiB, gzip) to POST /v1/ingest. It retries 429, 5xx and network errors (honoring Retry-After, else 1, 2, 4 ... s, at most 5 attempts) with stable event ids, so retries never duplicate. 400/401/403 drop the batch with one warning. A full queue drops new events and counts them. An atexit flush is bounded to 2 s.

coding_sh.flush(timeout=5)     # e.g. at the end of a script, a job or a serverless handler
coding_sh.shutdown()           # flush, then stop
coding_sh.stats()              # {"queued", "sent", "dropped", "failed"}

Config

coding_sh.init(url=, key=, service=, capture=, flush_interval_ms=1000, max_batch=500, max_queue=10000, disabled=, debug=) is optional; arguments override env vars.

OpenTelemetry

Already emitting gen_ai spans? Point OTel at the server instead (no SDK needed), or let the helper do the setup:

from coding_sh.otel import configure, otel_env
configure(service="checkout-api")   # TracerProvider + OTLP/HTTP exporter + Bearer key
env = otel_env()                    # OTEL_EXPORTER_OTLP_* vars for a subprocess

pip install 'coding-sh[otel]'. configure() uses the standard OTEL_EXPORTER_OTLP_* env vars when set, else CODING_SH_URL / CODING_SH_INGEST_KEY, and stamps gen_ai.conversation.id from the current with_session / trace(session_id=...).

Tests

python3 -m unittest discover -s tests -t tests

The provider tests use fake clients and always run. Tests against the real openai, anthropic and OpenTelemetry packages run when those are installed.