Integrating a Dify Agent with Dunetrace

Quick Start

pip install dunetrace dify-client-python

import time
from dunetrace import Dunetrace
from dify_client import Client, models

dt = Dunetrace()   # local dev, no API key needed
dify_client = Client(api_key="your-dify-api-key", api_base="https://api.dify.ai/v1")

def chat_with_dify(query: str, user_id: str) -> str:
    with dt.run("my-dify-agent", user_input=query, model="dify-workflow") as run:
        req = models.ChatRequest(query=query, inputs={}, user=user_id, response_mode=models.ResponseMode.BLOCKING)
        run.llm_called("dify-workflow")
        t0 = time.monotonic()
        res = dify_client.chat_messages(req, timeout=60.0)
        run.llm_responded(
            finish_reason="stop",
            output_length=len(res.answer or ""),
            latency_ms=int((time.monotonic() - t0) * 1000),
        )
        run.final_answer()
        return res.answer or ""

chat_with_dify("What is the capital of France?", "user-1")
dt.shutdown()

What this does

Dify agents run on Dify's own server, so there's no local LLM client to auto-instrument — Dunetrace instead wraps the API call itself in dt.run() and treats the whole request/response round-trip as one LLM event. That's enough to catch latency spikes, empty responses, and first-step failures on the Dify side, even though Dify's internal tool executions aren't visible.

Verification

Run your script once, then check the dashboard at http://localhost:3000 — the run should appear within ~15 seconds.

Advanced (optional)

What's captured

Overall latency, success/failure of the API call, response length, and token counts (from Dify's response metadata) — the whole Dify workflow as one step. Not captured: Dify's internal tool executions, unless you parse agent_thoughts/tool_calls from streaming mode yourself and emit run.tool_called()/run.tool_responded().

Most relevant detectors for this integration

SLOW_STEP (the Dify call itself takes too long), EMPTY_LLM_RESPONSE (often means an internal Dify workflow failure), FIRST_STEP_FAILURE (Dify returned an error or empty output early — wrap dify_client.chat_messages() in a try/except to catch network errors as RUN_ERRORED too). If you do manually instrument Dify's internal tools in streaming mode, TOOL_LOOP/RETRY_STORM/CASCADING_TOOL_FAILURE become relevant as well.