refactor(llm): normalize Usage as inclusive total + non-overlapping breakdown (#26735)

This commit is contained in:
Kit Langton
2026-05-10 21:52:50 -04:00
committed by GitHub
parent 38e4540119
commit 2703eff2e2
12 changed files with 275 additions and 68 deletions
+29 -1
View File
@@ -1,6 +1,7 @@
import { describe, expect, test } from "bun:test"
import { Schema } from "effect"
import { ContentPart, LLMEvent, LLMRequest, ModelID, ModelLimits, ModelRef, ProviderID } from "../src/schema"
import { ContentPart, LLMEvent, LLMRequest, ModelID, ModelLimits, ModelRef, ProviderID, Usage } from "../src/schema"
import { ProviderShared } from "../src/protocols/shared"
const model = new ModelRef({
id: ModelID.make("fake-model"),
@@ -48,3 +49,30 @@ describe("llm schema", () => {
expect(ContentPart.guards.media({ type: "text", text: "hi" })).toBe(false)
})
})
describe("LLM.Usage", () => {
test("subtractTokens clamps non-sensical breakdowns to zero", () => {
// Defense against a provider reporting cached_tokens > prompt_tokens or
// reasoning_tokens > completion_tokens — the negative would otherwise
// round-trip through the pipeline and crash strict downstream schemas.
expect(ProviderShared.subtractTokens(5, 3)).toBe(2)
expect(ProviderShared.subtractTokens(5, 10)).toBe(0)
expect(ProviderShared.subtractTokens(5, undefined)).toBe(5)
expect(ProviderShared.subtractTokens(undefined, 3)).toBeUndefined()
expect(ProviderShared.subtractTokens(undefined, undefined)).toBeUndefined()
})
test("sumTokens returns undefined only when every input is undefined", () => {
expect(ProviderShared.sumTokens(1, 2, 3)).toBe(6)
expect(ProviderShared.sumTokens(1, undefined, 3)).toBe(4)
expect(ProviderShared.sumTokens(undefined, undefined, undefined)).toBeUndefined()
expect(ProviderShared.sumTokens()).toBeUndefined()
})
test("visibleOutputTokens clamps reasoning > output to zero", () => {
expect(new Usage({ outputTokens: 10, reasoningTokens: 4 }).visibleOutputTokens).toBe(6)
expect(new Usage({ outputTokens: 10 }).visibleOutputTokens).toBe(10)
expect(new Usage({ outputTokens: 4, reasoningTokens: 10 }).visibleOutputTokens).toBe(0)
expect(new Usage({}).visibleOutputTokens).toBe(0)
})
})