scriptoriumadmin/Roo-Code
1
0
Fork 0
mirror of https://github.com/RooVetGit/Roo-Code.git synced 2026-10-09 03:17:58 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Fix max tokens in the task header (#3893)

Browse source
This commit is contained in:
Chris Estreich 2025-05-23 02:32:21 -07:00 • committed by GitHub
parent fa1e7b415c
commit 872902754a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 85 additions and 18 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
.changeset/ten-bags-hang.md Normal file
View file

@ -0,0 +1,5 @@
---
"roo-cline": patch
---
Fix max tokens in task header

12
src/api/transform/model-params.ts
View file

@ -87,16 +87,20 @@ export function getModelParams({
reasoningEffort = customReasoningEffort ?? model.reasoningEffort
}
// TODO: We should consolidate this logic to compute `maxTokens` with
// `getModelMaxOutputTokens` in order to maintain a single source of truth.
const isAnthropic = format === "anthropic" || (format === "openrouter" && modelId.startsWith("anthropic/"))
// For "Hybrid" reasoning models, we should discard the model's actual
// `maxTokens` value if we're not using reasoning.
if (model.supportsReasoningBudget && !reasoningBudget) {
// `maxTokens` value if we're not using reasoning. We do this for Anthropic
// models only for now. Should we do this for Gemini too?
if (model.supportsReasoningBudget && !reasoningBudget && isAnthropic) {
maxTokens = ANTHROPIC_DEFAULT_MAX_TOKENS
}
// For Anthropic models we should always make sure a `maxTokens` value is
// set.
const isAnthropic = format === "anthropic" || (format === "openrouter" && modelId.startsWith("anthropic/"))
if (!maxTokens && isAnthropic) {
maxTokens = ANTHROPIC_DEFAULT_MAX_TOKENS
}

61
src/shared/__tests__/api.test.ts
View file

@ -7,8 +7,11 @@ import {
shouldUseReasoningBudget,
shouldUseReasoningEffort,
} from "../api"
import { ANTHROPIC_DEFAULT_MAX_TOKENS } from "../../api/providers/constants"
describe("getMaxTokensForModel", () => {
const modelId = "test"
/**
* Testing the specific fix in commit cc79178f:
* For thinking models, use apiConfig.modelMaxTokens if available,
@ -27,7 +30,7 @@ describe("getMaxTokensForModel", () => {
modelMaxTokens: 4000,
}
expect(getModelMaxOutputTokens({ model, settings })).toBe(4000)
expect(getModelMaxOutputTokens({ modelId, model, settings })).toBe(4000)
})
it("should return 16_384 for thinking models when modelMaxTokens not provided", () => {
@ -40,7 +43,7 @@ describe("getMaxTokensForModel", () => {
const settings = {}
expect(getModelMaxOutputTokens({ model, settings })).toBe(16_384)
expect(getModelMaxOutputTokens({ modelId, model, settings })).toBe(16_384)
})
it("should return 16_384 for thinking models when apiConfig is undefined", () => {
@ -51,7 +54,7 @@ describe("getMaxTokensForModel", () => {
maxTokens: 8000,
}
expect(getModelMaxOutputTokens({ model, settings: undefined })).toBe(16_384)
expect(getModelMaxOutputTokens({ modelId, model, settings: undefined })).toBe(16_384)
})
it("should return modelInfo.maxTokens for non-thinking models", () => {
@ -65,7 +68,7 @@ describe("getMaxTokensForModel", () => {
modelMaxTokens: 4000,
}
expect(getModelMaxOutputTokens({ model, settings })).toBe(8000)
expect(getModelMaxOutputTokens({ modelId, model, settings })).toBe(8000)
})
it("should return undefined for non-thinking models with undefined maxTokens", () => {
@ -78,7 +81,7 @@ describe("getMaxTokensForModel", () => {
modelMaxTokens: 4000,
}
expect(getModelMaxOutputTokens({ model, settings })).toBeUndefined()
expect(getModelMaxOutputTokens({ modelId, model, settings })).toBeUndefined()
})
test("should return maxTokens from modelInfo when thinking is false", () => {
@ -92,7 +95,7 @@ describe("getMaxTokensForModel", () => {
modelMaxTokens: 4096,
}
const result = getModelMaxOutputTokens({ model, settings })
const result = getModelMaxOutputTokens({ modelId, model, settings })
expect(result).toBe(2048)
})
@ -108,7 +111,7 @@ describe("getMaxTokensForModel", () => {
modelMaxTokens: 4096,
}
const result = getModelMaxOutputTokens({ model, settings })
const result = getModelMaxOutputTokens({ modelId, model, settings })
expect(result).toBe(4096)
})
@ -122,7 +125,7 @@ describe("getMaxTokensForModel", () => {
const settings: ProviderSettings = {}
const result = getModelMaxOutputTokens({ model, settings: undefined })
const result = getModelMaxOutputTokens({ modelId, model, settings: undefined })
expect(result).toBe(16_384)
})
@ -133,7 +136,7 @@ describe("getMaxTokensForModel", () => {
maxTokens: 2048,
}
expect(getModelMaxOutputTokens({ model: modelInfoOnly, settings: undefined })).toBe(2048)
expect(getModelMaxOutputTokens({ modelId, model: modelInfoOnly, settings: undefined })).toBe(2048)
})
test("should handle missing properties gracefully", () => {
@ -147,7 +150,7 @@ describe("getMaxTokensForModel", () => {
modelMaxTokens: 4096,
}
expect(getModelMaxOutputTokens({ model: modelInfoWithoutMaxTokens, settings })).toBe(4096)
expect(getModelMaxOutputTokens({ modelId, model: modelInfoWithoutMaxTokens, settings })).toBe(4096)
const modelInfoWithoutThinking: ModelInfo = {
contextWindow: 200_000,
@ -155,7 +158,43 @@ describe("getMaxTokensForModel", () => {
maxTokens: 2048,
}
expect(getModelMaxOutputTokens({ model: modelInfoWithoutThinking, settings: undefined })).toBe(2048)
expect(getModelMaxOutputTokens({ modelId, model: modelInfoWithoutThinking, settings: undefined })).toBe(2048)
})
test("should return ANTHROPIC_DEFAULT_MAX_TOKENS for Anthropic models that support reasoning budget but aren't using it", () => {
// Test case for models that support reasoning budget but enableReasoningEffort is false
const anthropicModelId = "claude-sonnet-4-20250514"
const model: ModelInfo = {
contextWindow: 200_000,
supportsPromptCache: true,
supportsReasoningBudget: true,
maxTokens: 64_000, // This should be ignored
}
const settings: ProviderSettings = {
enableReasoningEffort: false, // Not using reasoning
}
const result = getModelMaxOutputTokens({ modelId: anthropicModelId, model, settings })
expect(result).toBe(ANTHROPIC_DEFAULT_MAX_TOKENS) // Should be 8192, not 64_000
})
test("should return model.maxTokens for non-Anthropic models that support reasoning budget but aren't using it", () => {
// Test case for non-Anthropic models - should still use model.maxTokens
const geminiModelId = "gemini-2.5-flash-preview-04-17"
const model: ModelInfo = {
contextWindow: 1_048_576,
supportsPromptCache: false,
supportsReasoningBudget: true,
maxTokens: 65_535,
}
const settings: ProviderSettings = {
enableReasoningEffort: false, // Not using reasoning
}
const result = getModelMaxOutputTokens({ modelId: geminiModelId, model, settings })
expect(result).toBe(65_535) // Should use model.maxTokens, not ANTHROPIC_DEFAULT_MAX_TOKENS
})
})

12
src/shared/api.ts
View file

@ -1,3 +1,4 @@
import { ANTHROPIC_DEFAULT_MAX_TOKENS } from "../api/providers/constants"
import { ModelInfo, ProviderName, ProviderSettings } from "../schemas"
export type { ModelInfo, ProviderName, ProviderSettings }
@ -1936,9 +1937,11 @@ export const DEFAULT_HYBRID_REASONING_MODEL_MAX_TOKENS = 16_384
export const DEFAULT_HYBRID_REASONING_MODEL_THINKING_TOKENS = 8_192
export const getModelMaxOutputTokens = ({
modelId,
model,
settings,
}: {
modelId: string
model: ModelInfo
settings?: ProviderSettings
}): number | undefined => {
@ -1946,5 +1949,14 @@ export const getModelMaxOutputTokens = ({
return settings?.modelMaxTokens || DEFAULT_HYBRID_REASONING_MODEL_MAX_TOKENS
}
const isAnthropicModel = modelId.includes("claude")
// For "Hybrid" reasoning models, we should discard the model's actual
// `maxTokens` value if we're not using reasoning. We do this for Anthropic
// models only for now. Should we do this for Gemini too?
if (model.supportsReasoningBudget && isAnthropicModel) {
return ANTHROPIC_DEFAULT_MAX_TOKENS
}
return model.maxTokens ?? undefined
}

1
webview-ui/src/__tests__/ContextWindowProgress.test.tsx
View file

@ -38,6 +38,7 @@ jest.mock("@src/components/chat/TaskHeader", () => {
// Mock useSelectedModel hook
jest.mock("@src/components/ui/hooks/useSelectedModel", () => ({
useSelectedModel: jest.fn(() => ({
id: "test",
info: { contextWindow: 4000 },
})),
}))

12
webview-ui/src/components/chat/TaskHeader.tsx
View file

@ -48,7 +48,7 @@ const TaskHeader = ({
}: TaskHeaderProps) => {
const { t } = useTranslation()
const { apiConfiguration, currentTaskItem } = useExtensionState()
const { info: model } = useSelectedModel(apiConfiguration)
const { id: modelId, info: model } = useSelectedModel(apiConfiguration)
const [isTaskExpanded, setIsTaskExpanded] = useState(false)
const textContainerRef = useRef<HTMLDivElement>(null)
@ -101,7 +101,9 @@ const TaskHeader = ({
contextWindow={contextWindow}
contextTokens={contextTokens || 0}
maxTokens={
model ? getModelMaxOutputTokens({ model, settings: apiConfiguration }) : undefined
model
? getModelMaxOutputTokens({ modelId, model, settings: apiConfiguration })
: undefined
}
/>
{!!totalCost && <VSCodeBadge>${totalCost.toFixed(2)}</VSCodeBadge>}
@ -140,7 +142,11 @@ const TaskHeader = ({
contextTokens={contextTokens || 0}
maxTokens={
model
? getModelMaxOutputTokens({ model, settings: apiConfiguration })
? getModelMaxOutputTokens({
modelId,
model,
settings: apiConfiguration,
})
: undefined
}
/>