tui: tok/s excludes reported reasoning tokens in V2

Description

OpenCode V2's tok/s indicator excludes reported reasoning tokens from the numerator. This understates total generated-token throughput for reasoning-heavy responses.

The footer sums assistant.tokens.output, which is populated from Usage.visibleOutputTokens. That value is the provider's inclusive output count minus its reported reasoning count; reasoning is stored separately in assistant.tokens.reasoning.

The indicator should count both non-reasoning output and reasoning tokens, using a duration that covers the generation of both. For providers such as OpenAI, the raw API output count already includes reasoning, so the calculation must use OpenCode's normalized counters without double-counting the provider total.

Plugins

No plugins are required for the code-level reproduction below.

OpenCode version

V2, upstream source at 2df00955cb933e977427535d2505e50cbc689c69.

Steps to reproduce

Construct a completed assistant message with tokens.output = 20, tokens.reasoning = 80, and a 10-second interval between time.created and time.streamed.
Pass it to turnTokensPerSecond in packages/tui/src/routes/session/rows.ts, with no other assistant messages in the current prompt's history.
The calculation returns 2 tok/s: 20 / 10. Counting all 100 generated tokens over the same interval would give 10 tok/s.

This isolates the missing reasoning count from provider-specific timing behavior. In normal use, it affects the footer whenever a provider reports reasoning tokens separately and the tok/s indicator is enabled.

Screenshot and/or share link

Calculation:

opencode/packages/tui/src/routes/session/rows.ts

Lines 349 to 369 in 2df0095

    
           export function turnTokensPerSecond( 
        
             message: SessionMessageAssistant, 
        
             messages: SessionMessageInfo[], 
        
             position?: number, 
        
           ) { 
        
             const index = position ?? messages.findIndex((item) => item.id === message.id) 
        
             const end = index === -1 ? messages.length : index + 1 
        
             const start = inputIndex(messages, end) 
        
             const steps = messages 
        
               .slice(start + 1, end) 
        
               .filter((item): item is SessionMessageAssistant => item.type === "assistant") 
        
             const durations = steps.flatMap((step) => 
        
               step.time.streamed === undefined ? [] : [Math.max(0, step.time.streamed - step.time.created)], 
        
             ) 
        
             if (steps.length === 0 || durations.length !== steps.length) return 
        
             const output = steps.reduce((total, step) => total + (step.tokens?.output ?? 0), 0) 
        
             const duration = durations.reduce((total, value) => total + value, 0) 
        
             if (output <= 0 || duration <= 0) return 
        
             return output / (duration / 1_000) 
        
           }

Session usage normalization:

opencode/packages/core/src/session/usage.ts

Lines 11 to 19 in 2df0095

    
           export const tokens = (usage: Usage | undefined): TokenUsage.Info => ({ 
        
             input: safe(usage?.nonCachedInputTokens), 
        
             output: safe(usage?.visibleOutputTokens), 
        
             reasoning: safe(usage?.reasoningTokens), 
        
             cache: { 
        
               read: safe(usage?.cacheReadInputTokens), 
        
               write: safe(usage?.cacheWriteInputTokens), 
        
             }, 
        
           })

Operating System

macOS (Darwin 25.6.0, arm64); the calculation is platform-independent.

Terminal

iTerm2

Fixed by #48689 (fix(tui): include reasoning tokens in throughput), merged to v2 as b278ef6 on 2026-09-18. Closing manually because v2 is not the default branch, so the PR did not auto-close this issue.

Navigation Menu

Repository navigation

tui: tok/s excludes reported reasoning tokens in V2 #48705

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	export function turnTokensPerSecond(
	message: SessionMessageAssistant,
	messages: SessionMessageInfo[],
	position?: number,
	) {
	const index = position ?? messages.findIndex((item) => item.id === message.id)
	const end = index === -1 ? messages.length : index + 1
	const start = inputIndex(messages, end)
	const steps = messages
	.slice(start + 1, end)
	.filter((item): item is SessionMessageAssistant => item.type === "assistant")
	const durations = steps.flatMap((step) =>
	step.time.streamed === undefined ? [] : [Math.max(0, step.time.streamed - step.time.created)],
	)
	if (steps.length === 0 \|\| durations.length !== steps.length) return
	const output = steps.reduce((total, step) => total + (step.tokens?.output ?? 0), 0)
	const duration = durations.reduce((total, value) => total + value, 0)
	if (output <= 0 \|\| duration <= 0) return
	return output / (duration / 1_000)
	}

	export const tokens = (usage: Usage \| undefined): TokenUsage.Info => ({
	input: safe(usage?.nonCachedInputTokens),
	output: safe(usage?.visibleOutputTokens),
	reasoning: safe(usage?.reasoningTokens),
	cache: {
	read: safe(usage?.cacheReadInputTokens),
	write: safe(usage?.cacheWriteInputTokens),
	},
	})

tui: tok/s excludes reported reasoning tokens in V2 #48705

Description

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Activity

Hona commented on Sep 22, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions