Feature Request: Add Langfuse Tracing

[tito]

Summary

Add tracing/observability to OpenCode using Langfuse.

Why

• Currently no way to trace execution or monitor performance
• Difficult to debug issues in production
• Missing insights on usage patterns and bottlenecks

Proposal

Integrate Langfuse (https://langfuse.com/) for tracing, which is already supported by Goose (https://block.github.io/goose/) per their documentation.

Benefits

• Simple debugging
• Performance monitoring
• Usage analytics
• Low implementation effort

[ColeMurray]

langfuse supports Otel, so preferably we could instrument with Otel and have langfuse as one of N many different OtelProcessors

https://langfuse.com/docs/opentelemetry/get-started

[adrielp]

👍🏼 to this. Would love to be able to opt-in to emitting native instrumentation for as much of the stack as possible simply by adding the OTEL exporter env vars + leveraging the GenAI semantic conventions I'd be happy to assist on this as well.

[mattzhango]

+1 to having tracing/observability, especially with the release of the SDK

[noamzbr]

Following #4978, Langfuse tracing is now possible with a minimal plugin that can look something like

/**
 * Langfuse OpenTelemetry Plugin for OpenCode
 * 
 * Requires: bun add @langfuse/otel @opentelemetry/sdk-node
 * Env vars: LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_HOST
 */
import { NodeSDK } from "@opentelemetry/sdk-node"
import { LangfuseSpanProcessor } from "@langfuse/otel"

let initialized = false

export const LangfusePlugin = async () => ({
  async config(config) {
    if (!config.open_telemetry || initialized) return
    initialized = true

    const processor = new LangfuseSpanProcessor()
    new NodeSDK({ spanProcessors: [processor] }).start()

    // // Flush frequently - Required in headless mode to make sure flushes are made before process exit
    // setInterval(() => processor.forceFlush(), 500)
  },
})

(Probably worth finding a better solution for headless mode)

[modelnova-ai]

Following #4978, Langfuse tracing is now possible with a minimal plugin that can look something like

/**
 * Langfuse OpenTelemetry Plugin for OpenCode
 * 
 * Requires: bun add @langfuse/otel @opentelemetry/sdk-node
 * Env vars: LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_HOST
 */
import { NodeSDK } from "@opentelemetry/sdk-node"
import { LangfuseSpanProcessor } from "@langfuse/otel"

let initialized = false

export const LangfusePlugin = async () => ({
  async config(config) {
    if (!config.open_telemetry || initialized) return
    initialized = true

    const processor = new LangfuseSpanProcessor()
    new NodeSDK({ spanProcessors: [processor] }).start()

    // // Flush frequently - Required in headless mode to make sure flushes are made before process exit
    // setInterval(() => processor.forceFlush(), 500)
  },
})

(Probably worth finding a better solution for headless mode)

Is it possible to record telemetry on a per user level with this plugin? if not how to trace user level tracing in langfuse

[omercnet]

https://github.com/omercnet/opencode-plugin-langfuse

[pai4451]

@omercnet Do you know what opentelemetry metrics does opencode emits? I couldn't find any telemetry metrics and logs even I configure

{
  "experimental": {
    "openTelemetry": true
  },
  "plugin": ["opencode-plugin-langfuse"]
}

[omercnet]

@pai4451 make sure you set env vars properly, try in terminal first with opencode --print-logs --log-level WARN run test to see any warning logs

and kindly open issues on the plugin repo so we can handle any bugs