Problem
OpenCode skill and task tool calls are exported as Langfuse TOOL observations named only skill and task.
The semantic values are preserved only inside the observation input:
- Skill name:
input.name
- Subagent type:
input.subagent_type
Langfuse dashboards can group tool observations by observation name, but cannot group or filter metrics by fields nested inside observation input. As a result, dashboards can show only total skill invocations and total subagent spawns. They cannot break those calls down by skill name or subagent type.
Current behavior
A skill invocation such as:
{
"name": "resolve-dependencies"
}is exported as an observation named skill.
A subagent spawn such as:
{
"subagent_type": "developer"
}is exported as an observation named task.
On current main, traceToolStart creates the span with startSpan(input.tool, ...), so these semantic names are not exposed through the observation name. I reproduced the behavior with plugin 0.3.0, and the same naming behavior remains in the current 0.4.0 source.
Expected behavior
Expose the semantic skill or subagent name through a field that Langfuse metrics can group and filter by.
Possible approaches:
- Name observations
skill:<name> and task:<subagent_type>.
- Add stable Langfuse attributes or metadata fields that are supported as dashboard dimensions.
The original tool input should remain available unchanged.
Why this matters
This would allow native Langfuse dashboards to report:
- Skill invocation counts by skill name
- Subagent spawn counts by subagent type
- Per-trace skill and subagent usage
Without this, users must build a separate pipeline that parses observation inputs and writes the extracted values back into Langfuse.
Problem
OpenCode
skillandtasktool calls are exported as Langfuse TOOL observations named onlyskillandtask.The semantic values are preserved only inside the observation input:
input.nameinput.subagent_typeLangfuse dashboards can group tool observations by observation
name, but cannot group or filter metrics by fields nested inside observationinput. As a result, dashboards can show only total skill invocations and total subagent spawns. They cannot break those calls down by skill name or subagent type.Current behavior
A skill invocation such as:
{ "name": "resolve-dependencies" }is exported as an observation named
skill.A subagent spawn such as:
{ "subagent_type": "developer" }is exported as an observation named
task.On current
main,traceToolStartcreates the span withstartSpan(input.tool, ...), so these semantic names are not exposed through the observation name. I reproduced the behavior with plugin0.3.0, and the same naming behavior remains in the current0.4.0source.Expected behavior
Expose the semantic skill or subagent name through a field that Langfuse metrics can group and filter by.
Possible approaches:
skill:<name>andtask:<subagent_type>.The original tool input should remain available unchanged.
Why this matters
This would allow native Langfuse dashboards to report:
Without this, users must build a separate pipeline that parses observation inputs and writes the extracted values back into Langfuse.