Traces show input but no output; generations show output but no input

[DerJacques]

In the Langfuse UI, traces produced by this plugin only ever show input; they never show output. Conversely, the nested opencode.generation observations have output but no input. As a result, no single observation (trace or generation) has a paired input + output, which makes it impractical to build LLM-as-judge evaluators on these traces.

Here are some example of what that looks like:

Please note that this is my first time using Langfuse, but I would have expected a different behavior based on the short video in https://langfuse.com/docs/observability/get-started#see-your-trace-in-langfuse.

Reproduction

1. Enable the plugin per the README:

   {
     "experimental": { "openTelemetry": true },
     "plugin": ["@langfuse/opencode-observability-plugin@latest"]
   }

2. Run any prompt with opencode run "say hello" (or via the TUI).
3. Open the resulting trace in Langfuse.

Observed:

• The trace pane shows the user prompt as input, but output is empty.
• Click into the trace to see observations:
  • opencode.turn (root, drives trace-level I/O) — has input, no output.
  • opencode.generation — has output, no input.

Expected: at least one of the observations should carry both the input and the output of the LLM call, so users can run trace-level or observation-level LLM judges without manually joining sibling spans.

Root cause (from reading the source)

In src/langfuse.ts:

• traceUserMessage creates the opencode.turn root span and sets langfuse.observation.input to the formatted user message. It never sets langfuse.observation.output on this span — neither at creation nor later when the turn finishes.
• traceGeneration / startActiveGenerationStep create the nested opencode.generation span and set langfuse.observation.output (assistant text), model.name, usage_details, cost_details. They never set langfuse.observation.input.

Since Langfuse derives trace-level input/output from the root observation's attributes, the missing output on opencode.turn is what surfaces in the UI as the empty trace output column.

Environment

• Plugin: @langfuse/opencode-observability-plugin (latest as of filing)
• OpenCode: 1.17.9
• Configuration: experimental.openTelemetry: true, plugin loaded from opencode.json

[bezbac]

Hi @DerJacques,

The way I built this plugin, a single opencode.turn can contain multiple opencode.generation events, each with its own output. Because of that, I don't think trace-level input/output fields map well to the underlying data model - there isn't necessarily a single canonical output for a turn.

This is also reflected in the Langfuse UI. In Fast Preview, there is no dedicated trace-level view; users primarily interact with observation/event-level data. Given that, surfacing inputs and outputs on the individual generation observations seems more accurate and useful than trying to aggregate them onto the trace.

I'd love to better understand the use case you're trying to support here. Could you share a bit more about how you plan to use the data you are ingesting into Langfuse with this plugin?

[LemurPwned]

@bezbac I'm also interested in this.
In my case, I'd like to trace the context and tools that are available to the open code. Right now, neither is visible with the plugin.
In addition, I think the turns are not properly assembled into a single span?
I remember that I was able to make it work with Comet Opik with minimal config.

I also tested this plugin and it seems to work really good, having proper input/output and system prompt in the turn: https://github.com/omercnet/opencode-plugin-langfuse