Files
orca/src/shared/native-chat/native-chat-context-usage.test.ts
T

86 lines
2.7 KiB
TypeScript

import { describe, expect, it } from 'vitest'
import { deriveNativeChatContextUsage } from './native-chat-context-usage'
import type { AgentSessionTokenUsage } from '../agent-session/agent-session-context-usage'
import type { NativeChatMessage } from './native-chat-types'
function assistant(
id: string,
usage?: AgentSessionTokenUsage,
model = 'gpt-5.5'
): NativeChatMessage {
return {
id,
role: 'assistant',
blocks: [{ type: 'text', text: id }],
timestamp: 100,
source: 'transcript',
model,
provider: 'openai-codex',
...(usage ? { usage } : {})
}
}
function usage(inputTokens: number, cacheRead = 0, cacheWrite = 0): AgentSessionTokenUsage {
return {
inputTokens,
cacheCreationInputTokens: cacheWrite,
cacheReadInputTokens: cacheRead,
outputTokens: 16
}
}
const compaction: NativeChatMessage = {
id: 'compacted',
role: 'system',
blocks: [{ type: 'text', text: 'Context compacted', presentation: 'compaction' }],
timestamp: 150,
source: 'transcript'
}
const noWindow = (): number | null => null
describe('deriveNativeChatContextUsage', () => {
it('sums input and both cache counts of the newest response against its window', () => {
const windows: NativeChatMessage[] = []
const derived = deriveNativeChatContextUsage(
[assistant('older', usage(5)), assistant('newest', usage(680, 20_992, 8))],
(message) => {
windows.push(message)
return 272_000
}
)
expect(derived).toEqual({
usedTokens: 21_680,
windowTokens: 272_000,
percentage: 8
})
// The window is asked of the response that was measured, not the session's picker.
expect(windows.map(({ id }) => id)).toEqual(['newest'])
})
it('reports the used figure alone when the window is unknown', () => {
expect(deriveNativeChatContextUsage([assistant('a', usage(450_000))], noWindow)).toEqual({
usedTokens: 450_000,
windowTokens: null,
percentage: null
})
expect(
deriveNativeChatContextUsage([assistant('a', usage(10))], () => 0)?.windowTokens
).toBeNull()
})
it('skips responses whose accounting does not reflect the prompt', () => {
const messages = [assistant('measured', usage(100)), assistant('aborted')]
expect(deriveNativeChatContextUsage(messages, noWindow)?.usedTokens).toBe(100)
})
it('knows nothing after a compaction until the next response lands', () => {
const before = [assistant('before', usage(200_000)), compaction]
expect(deriveNativeChatContextUsage(before, noWindow)).toBeNull()
expect(
deriveNativeChatContextUsage([...before, assistant('after', usage(30_000))], noWindow)
?.usedTokens
).toBe(30_000)
})
})