Feature Request: Add LLM API Call Duration Logging

Summary

OpenCode currently tracks LLM call timing internally but does not log the total duration of LLM API calls. This makes it difficult to monitor and analyze API latency.

Problem

When debugging performance issues or monitoring API costs, users need visibility into:

• How long each LLM API call takes
• Which models/providers have high latency
• Overall time spent waiting for LLM responses

Currently:

• service=llm logs only show when streaming starts
• No duration/end time is logged for LLM calls
• Internal timing exists in MessageV2 Parts but is not exposed in logs

Proposed Solution

Add duration logging for LLM API calls, similar to how service=session.prompt logs duration=xxx.

Examples

Current logging:

INFO 2026-01-17T06:30:36 +0ms service=llm providerID=zhipuai-coding-plan modelID=glm-4.7 sessionID=ses_4355b9979ffeLoxgoHRHr6IAnb small=false agent=plan stream

Proposed logging:

INFO 2026-01-17T06:30:36 +0ms service=llm providerID=zhipuai-coding-plan modelID=glm-4.7 sessionID=ses_4355b9979ffeLoxgoHRHr6IAnb small=false agent=plan stream
INFO 2026-01-17T06:30:50 +14321ms service=llm providerID=zhipuai-coding-plan modelID=glm-4.7 sessionID=ses_4355b9979ffeLoxgoHRHr6IAnb status=completed duration=14321

Implementation Details

The timing information can be captured in packages/opencode/src/session/llm.ts:

// At the start of stream()
const startTime = Date.now()
l.info("stream", { modelID: input.model.id, providerID: input.model.providerID })

// Track stream completion and log duration
const streamResult = streamText({...})

// Use for-await-of to track when the stream completes
let streamCompleted = false
const wrappedStream = {
  ...streamResult,
  fullStream: (async function* () {
    for await (const chunk of streamResult.fullStream) {
      yield chunk
    }
    // Log completion when stream finishes
    if (!streamCompleted) {
      streamCompleted = true
      const duration = Date.now() - startTime
      l.info("stream completed", { 
        modelID: input.model.id, 
        providerID: input.model.providerID,
        duration 
      })
    }
  })()
}
return wrappedStream

Benefits

• Better performance monitoring
• Easier debugging of latency issues
• Improved cost analysis (time × rate)
• Consistent with existing logging patterns in OpenCode

Alternatives Considered

• Using OpenTelemetry (already supported via experimental_telemetry config) - but logs are more accessible for most users
• Manual calculation from MessageV2 timestamps - requires querying internal storage

Additional Context

OpenCode already tracks timing for:

• Tool execution (tool.call to tool.result)
• Text generation (time.start to time.end in TextPart)
• Message completion (time.completed)

Adding LLM stream duration would complete the picture.