Sepurux Python SDK

Record agent traces, run reliability campaigns, and instrument OpenAI or LangChain workflows — all from Python.

Install

pip install -e sdk

Quick start

import os

from sepurux import SepuruxClient

os.environ["SEPURUX_API_BASE_URL"] = "http://localhost:8000"
os.environ["SEPURUX_API_KEY"] = "sepurux-dev-key"
os.environ["SEPURUX_PROJECT_ID"] = "22222222-2222-2222-2222-222222222222"

with SepuruxClient.from_env() as client:
    with client.trace("example_task", {"user_id": "u-123"}) as trace:
        trace.model_step("plan", {"goal": "create issue"}, output={"ok": True})
        trace.tool_call("jira.create_issue(commit)", {"summary": "SDK test"})
        trace.tool_result("jira.create_issue(commit)", {"issue_id": "OPS-123"})

    print("trace_id:", trace.trace_id)

This is the intended ergonomic path:

• SepuruxClient.from_env() reads SEPURUX_API_BASE_URL, SEPURUX_API_KEY, and SEPURUX_PROJECT_ID
• client.trace(...) records and uploads automatically on exit
• trace.trace_id and trace.run_id are available after the context closes

LangSmith-style ergonomics

Auto-upload trace context

from sepurux import SepuruxClient

client = SepuruxClient.from_env()

with client.trace(
    "checkout_refund",
    {"ticket_id": "t-123"},
    campaign_id="refund_reliability",
    mutation_pack="sepurux.core.reliability",
) as trace:
    trace.model_step("triage", {"text": "Refund the duplicate charge"})
    trace.tool_call("payments.refund", {"payment_id": "pay_123", "amount": 4200})
    trace.tool_result("payments.refund", {"refund_id": "rf_123", "status": "queued"})

print(trace.trace_id)
print(trace.run_id)

@traceable decorator

Works with both sync and async def functions:

from sepurux import SepuruxClient

client = SepuruxClient.from_env()

# Sync function
@client.traceable(campaign_id="refund_reliability")
def score_refund_risk(amount: int, currency: str) -> dict:
    return {"risk": "medium", "approve": amount < 10000, "currency": currency}

result = score_refund_risk(4200, "usd")

# Async function — same decorator, awaitable result
@client.traceable(campaign_id="refund_reliability")
async def fetch_risk_score(payment_id: str) -> dict:
    data = await some_async_api(payment_id)
    return {"risk": data["score"]}

result = await fetch_risk_score("pay_123")

Global configuration

from sepurux import configure, trace, traceable

configure()

@traceable()
def classify_ticket(text: str) -> dict:
    return {"label": "bug", "priority": "p2"}

with trace("support_session", {"channel": "email"}) as rec:
    rec.model_step("classify", {"text": "checkout timed out"})

API

SepuruxClient

SepuruxClient(base_url, api_key=None, project_id=None, timeout=30, sdk_header=None)
SepuruxClient.from_env(...)

The SDK automatically sends X-Sepurux-SDK: py/<version> on requests. Use sdk_header only if you need to override it manually.

Methods:

• trace(task_name, task_input=None, campaign_id=None, mutation_pack=None) -> recorder
• traceable(name=None, campaign_id=None, mutation_pack=None) -> decorator
• upload_trace(trace: dict) -> str
• list_traces(limit=50, offset=0) -> list[dict]
• get_trace(trace_id, include_payload=False) -> dict
• get_trace_timeline(trace_id) -> list[dict]
• bootstrap_trace(trace_id) -> dict
• create_campaign(name, mutation_set, eval_set, mutation_pack_id=None) -> str
• preflight_run_inputs(trace_id, campaign_id) -> None
• start_run(trace_id, campaign_id, thresholds=None, preflight=False) -> str
• run_pack(campaign_id, mutation_pack) -> str
• get_run(run_id) -> dict
• get_ci_run(run_id) -> dict
• get_campaign_failures(campaign_id, run_id=None) -> list[dict]
• get_campaign_reliability(campaign_id, run_id=None) -> dict

TraceBuilder

TraceBuilder outputs backend-compatible traces:

{
  "trace_version": "0.1",
  "source": "sdk",
  "task": {"name": "...", "input": {}},
  "events": []
}

Recommended tracing flow

Use client.trace(...), client.traceable(...), or the global trace(...) / traceable(...) helpers as the default integration path. Legacy helpers like sepurux_trace(...) and record_trace(...) remain available for compatibility, but they are deprecated and no longer the documented path.

Server-side pack execution

result = client.run_campaign("demo_campaign", mutation_pack="core.reliability")
print(result["run_id"])
print(result["reliability_score"])
print([failure["type"] for failure in result["failures"]])

Integrations

OpenAI

instrument_openai wraps any openai.OpenAI (or AsyncOpenAI) client and automatically records every chat.completions.create call as an llm_call event. Tool calls present in the response are also recorded. No manual instrumentation required.

pip install sepurux openai

import openai
from sepurux import SepuruxClient
from sepurux.integrations.openai import instrument_openai

client = SepuruxClient.from_env()
openai_client = openai.OpenAI()

with client.trace(
    "customer_refund_flow",
    {"ticket_id": "t-101"},
    campaign_id=os.environ["SEPURUX_CAMPAIGN_ID"],
) as trace:
    ai = instrument_openai(openai_client, recorder=trace)
    response = ai.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": "Classify this support ticket"}],
    )

print(trace.trace_id)
print(trace.run_id)

What gets recorded automatically:

• llm_call event with model name, messages input, full response output, and latency
• tool_call events for any function calls returned by the model
• failure event if the API call raises an exception

For async usage, call await ai.chat.completions.acreate(...) — the wrapper detects the async path automatically.

LangChain

SepuruxCallbackHandler automatically records LLM calls, tool calls, and agent outputs from any LangChain chain or agent. Requires langchain-core:

pip install sepurux langchain-core

import os
from langchain_openai import ChatOpenAI
from langchain.agents import AgentExecutor, create_tool_calling_agent
from langchain_core.prompts import ChatPromptTemplate
from sepurux.integrations.langchain import SepuruxCallbackHandler

handler = SepuruxCallbackHandler(
    "customer_refund_flow",
    {"ticket_id": "t-101"},
    campaign_id=os.environ["SEPURUX_CAMPAIGN_ID"],
)

# Works with chains, agents, or any LangChain runnable
result = agent_executor.invoke(
    {"input": "Process refund for ticket t-101"},
    config={"callbacks": [handler]},
)

# Upload trace + start reliability run
handler.finish()
print(handler.trace_id)
print(handler.run_id)

Or use it as a context manager — finish() is called automatically on exit:

with SepuruxCallbackHandler("refund_flow", campaign_id="...") as handler:
    result = chain.invoke(inputs, config={"callbacks": [handler]})

print(handler.trace_id)

What gets recorded automatically:

• LLM calls with model name, prompt input, generated output, and latency
• Tool calls with name and parsed arguments
• Tool results
• Agent final output

Optional environment variables:

• SEPURUX_API_BASE_URL (default http://localhost:8000)
• SEPURUX_UI_BASE_URL (default http://localhost:3000)
• SEPURUX_API_KEY
• SEPURUX_PROJECT_ID
• SEPURUX_CAMPAIGN_ID (if provided, script starts a run)