scriptoriumadmin/Roo-Code
1
0
Fork 0
mirror of https://github.com/RooVetGit/Roo-Code.git synced 2026-09-05 08:10:14 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

refactor: move applyCacheBreakpoints into providers, reduce maxBreakpoints to 1

Browse source 
- Move applyCacheBreakpoints() from Task.ts into each provider's createMessage()
  (anthropic, bedrock, openrouter, requesty, anthropic-vertex, roo)
- Aligns with AI SDK best practice: cache control is a call-site concern
- Remove stripCacheProviderOptions() from saveRooMessages() save path
  (no longer needed since cache markers don't contaminate persistent history)
- Change maxBreakpoints default from 2 to 1 (3-point strategy: 1 system, 1 tool, 1 message)
This commit is contained in:
Hannes Rudolph 2026-02-13 09:25:12 -07:00
parent f656fb1709
commit b6fd74b3ec
10 changed files with 48 additions and 33 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
src/api/providers/anthropic-vertex.ts
View file

@ -25,7 +25,7 @@ import {
handleAiSdkError,
yieldResponseMessage,
} from "../transform/ai-sdk"
import { applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { applyCacheBreakpoints, applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { calculateApiCostAnthropic } from "../../shared/cost"
import { DEFAULT_HEADERS } from "./constants"
@ -127,6 +127,8 @@ export class AnthropicVertexHandler extends BaseProvider implements SingleComple
metadata?.systemProviderOptions,
)
applyCacheBreakpoints(aiSdkMessages)
// Build streamText request
// Cast providerOptions to any to bypass strict JSONObject typing — the AI SDK accepts the correct runtime values
const requestOptions: Parameters<typeof streamText>[0] = {

4
src/api/providers/anthropic.ts
View file

@ -23,7 +23,7 @@ import {
handleAiSdkError,
yieldResponseMessage,
} from "../transform/ai-sdk"
import { applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { applyCacheBreakpoints, applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { sanitizeMessagesForProvider } from "../transform/sanitize-messages"
import { calculateApiCostAnthropic } from "../../shared/cost"
@ -116,6 +116,8 @@ export class AnthropicHandler extends BaseProvider implements SingleCompletionHa
metadata?.systemProviderOptions,
)
applyCacheBreakpoints(aiSdkMessages)
// Build streamText request
// Cast providerOptions to any to bypass strict JSONObject typing — the AI SDK accepts the correct runtime values
const requestOptions: Parameters<typeof streamText>[0] = {

4
src/api/providers/bedrock.ts
View file

@ -31,7 +31,7 @@ import {
handleAiSdkError,
yieldResponseMessage,
} from "../transform/ai-sdk"
import { applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { applyCacheBreakpoints, applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { getModelParams } from "../transform/model-params"
import { sanitizeMessagesForProvider } from "../transform/sanitize-messages"
import { shouldUseReasoningBudget } from "../../shared/api"
@ -252,6 +252,8 @@ export class AwsBedrockHandler extends BaseProvider implements SingleCompletionH
? applySystemPromptCaching(systemPrompt, aiSdkMessages, metadata?.systemProviderOptions)
: systemPrompt || undefined
applyCacheBreakpoints(aiSdkMessages)
// Strip non-Bedrock cache annotations from messages when caching is disabled,
// and strip Bedrock-specific annotations when caching is disabled.
if (!usePromptCache) {

4
src/api/providers/openrouter.ts
View file

@ -18,7 +18,7 @@ import { calculateApiCostOpenAI } from "../../shared/cost"
import { getModelParams } from "../transform/model-params"
import { convertToolsForAiSdk, processAiSdkStreamPart, yieldResponseMessage } from "../transform/ai-sdk"
import { applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { applyCacheBreakpoints, applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { BaseProvider } from "./base-provider"
import { getModels, getModelsFromCache } from "./fetchers/modelCache"
@ -182,6 +182,8 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
metadata?.systemProviderOptions,
)
applyCacheBreakpoints(aiSdkMessages)
try {
const result = streamText({
model: openrouter.chat(modelId),

4
src/api/providers/requesty.ts
View file

@ -8,7 +8,7 @@ import type { ApiHandlerOptions } from "../../shared/api"
import { calculateApiCostOpenAI } from "../../shared/cost"
import { convertToolsForAiSdk, consumeAiSdkStream, mapToolChoice, handleAiSdkError } from "../transform/ai-sdk"
import { applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { applyCacheBreakpoints, applyToolCacheOptions, applySystemPromptCaching } from "../transform/cache-breakpoints"
import { ApiStream, ApiStreamUsageChunk } from "../transform/stream"
import { getModelParams } from "../transform/model-params"
@ -205,6 +205,8 @@ export class RequestyHandler extends BaseProvider implements SingleCompletionHan
metadata?.systemProviderOptions,
)
applyCacheBreakpoints(aiSdkMessages)
const requestOptions: Parameters<typeof streamText>[0] = {
model: languageModel,
system: effectiveSystemPrompt,

4
src/api/providers/roo.ts
View file

@ -17,7 +17,7 @@ import {
mapToolChoice,
yieldResponseMessage,
} from "../transform/ai-sdk"
import { applyToolCacheOptions } from "../transform/cache-breakpoints"
import { applyCacheBreakpoints, applyToolCacheOptions } from "../transform/cache-breakpoints"
import type { RooReasoningParams } from "../transform/reasoning"
import { getRooReasoning } from "../transform/reasoning"
@ -161,6 +161,8 @@ export class RooHandler extends BaseProvider implements SingleCompletionHandler
const tools = convertToolsForAiSdk(this.convertToolsForOpenAI(metadata?.tools))
applyToolCacheOptions(tools as Parameters<typeof applyToolCacheOptions>[0], metadata?.toolProviderOptions)
applyCacheBreakpoints(aiSdkMessages)
let lastStreamError: string | undefined
try {

39
src/api/transform/__tests__/cache-breakpoints.spec.ts
View file

@ -44,21 +44,28 @@ describe("applyCacheBreakpoints", () => {
expect(messages[0].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
})
it("places 2 breakpoints on 2 user messages", () => {
it("places 1 breakpoint on the last of 2 user messages (default maxBreakpoints=1)", () => {
const messages: TestMessage[] = [{ role: "user" }, { role: "user" }]
applyCacheBreakpoints(messages)
expect(messages[0].providerOptions).toBeUndefined()
expect(messages[1].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
})
it("places 2 breakpoints on 2 user messages when maxBreakpoints=2", () => {
const messages: TestMessage[] = [{ role: "user" }, { role: "user" }]
applyCacheBreakpoints(messages, { maxBreakpoints: 2 })
expect(messages[0].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
expect(messages[1].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
})
it("places 2 breakpoints on 2 tool messages", () => {
it("places 1 breakpoint on the last of 2 tool messages (default maxBreakpoints=1)", () => {
const messages: TestMessage[] = [{ role: "tool" }, { role: "tool" }]
applyCacheBreakpoints(messages)
expect(messages[0].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
expect(messages[0].providerOptions).toBeUndefined()
expect(messages[1].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
})
it("targets last 2 non-assistant messages in a mixed conversation", () => {
it("targets last non-assistant message in a mixed conversation", () => {
const messages: TestMessage[] = [
{ role: "user" },
{ role: "assistant" },
@ -67,15 +74,15 @@ describe("applyCacheBreakpoints", () => {
{ role: "tool" },
]
applyCacheBreakpoints(messages)
// Last 2 non-assistant: index 2 (user) and index 4 (tool)
// Last 1 non-assistant: index 4 (tool)
expect(messages[0].providerOptions).toBeUndefined()
expect(messages[1].providerOptions).toBeUndefined()
expect(messages[2].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
expect(messages[2].providerOptions).toBeUndefined()
expect(messages[3].providerOptions).toBeUndefined()
expect(messages[4].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
})
it("targets indices 3 and 5 in [user, assistant, tool, user, assistant, tool]", () => {
it("targets only index 5 in [user, assistant, tool, user, assistant, tool]", () => {
const messages: TestMessage[] = [
{ role: "user" },
{ role: "assistant" },
@ -88,7 +95,7 @@ describe("applyCacheBreakpoints", () => {
expect(messages[0].providerOptions).toBeUndefined()
expect(messages[1].providerOptions).toBeUndefined()
expect(messages[2].providerOptions).toBeUndefined()
expect(messages[3].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
expect(messages[3].providerOptions).toBeUndefined()
expect(messages[4].providerOptions).toBeUndefined()
expect(messages[5].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
})
@ -97,7 +104,7 @@ describe("applyCacheBreakpoints", () => {
const messages: TestMessage[] = [{ role: "system" }, { role: "user" }, { role: "assistant" }, { role: "user" }]
applyCacheBreakpoints(messages)
expect(messages[0].providerOptions).toBeUndefined()
expect(messages[1].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
expect(messages[1].providerOptions).toBeUndefined()
expect(messages[3].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
})
@ -132,7 +139,7 @@ describe("applyCacheBreakpoints", () => {
it("adds anchor breakpoint at ~1/3 with useAnchor and enough messages", () => {
// 6 non-assistant messages (indices 0-5 in nonAssistantIndices)
// Anchor at floor(6/3) = index 2 in nonAssistantIndices -> messages index 4
// Last 2: indices 10 and 8
// Last 1: index 10
const messages: TestMessage[] = [
{ role: "user" }, // 0 - nonAssistant[0]
{ role: "assistant" }, // 1
@ -142,21 +149,21 @@ describe("applyCacheBreakpoints", () => {
{ role: "assistant" }, // 5
{ role: "user" }, // 6 - nonAssistant[3]
{ role: "assistant" }, // 7
{ role: "user" }, // 8 - nonAssistant[4] <- last 2
{ role: "user" }, // 8 - nonAssistant[4]
{ role: "assistant" }, // 9
{ role: "user" }, // 10 - nonAssistant[5] <- last 2
{ role: "user" }, // 10 - nonAssistant[5] <- last 1
]
applyCacheBreakpoints(messages, { useAnchor: true })
// Should have 3 breakpoints: indices 4, 8, 10
// Should have 2 breakpoints: indices 4 (anchor) and 10 (last 1)
expect(messages[4].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
expect(messages[8].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
expect(messages[10].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
// Others should NOT have breakpoints
expect(messages[0].providerOptions).toBeUndefined()
expect(messages[2].providerOptions).toBeUndefined()
expect(messages[6].providerOptions).toBeUndefined()
expect(messages[8].providerOptions).toBeUndefined()
})
it("does not add anchor when below anchorThreshold", () => {
@ -170,9 +177,9 @@ describe("applyCacheBreakpoints", () => {
// 3 non-assistant messages, below default threshold of 5
applyCacheBreakpoints(messages, { useAnchor: true })
// Last 2 only: indices 2 and 4
// Last 1 only: index 4
expect(messages[0].providerOptions).toBeUndefined()
expect(messages[2].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
expect(messages[2].providerOptions).toBeUndefined()
expect(messages[4].providerOptions).toEqual(UNIVERSAL_CACHE_OPTIONS)
})

10
src/api/transform/cache-breakpoints.ts
View file

@ -12,7 +12,7 @@ export const UNIVERSAL_CACHE_OPTIONS: Record<string, Record<string, unknown>> =
* Optional targeting configuration for cache breakpoint placement.
*/
export interface CacheBreakpointTargeting {
/** Maximum number of message breakpoints to place. Default: 2 */
/** Maximum number of message breakpoints to place. Default: 1 */
maxBreakpoints?: number
/** Whether to add an anchor breakpoint at ~1/3 through the conversation. Default: false */
useAnchor?: boolean
@ -23,19 +23,19 @@ export interface CacheBreakpointTargeting {
/**
* Apply cache breakpoints to AI SDK messages with ALL provider namespaces.
*
* 4-breakpoint strategy:
* 3-breakpoint strategy:
* 1. System prompt — passed as first message in messages[] with providerOptions
* 2. Tool definitions — handled externally via `toolProviderOptions` in `streamText()`
* 3-4. Last 2 non-assistant messages — this function handles these
* 3. Last non-assistant message — this function handles this
*
* @param messages - The AI SDK message array (mutated in place)
* @param targeting - Optional targeting options (defaults: 2 breakpoints, no anchor)
* @param targeting - Optional targeting options (defaults: 1 breakpoint, no anchor)
*/
export function applyCacheBreakpoints(
messages: { role: string; providerOptions?: Record<string, Record<string, unknown>> }[],
targeting: CacheBreakpointTargeting = {},
): void {
const { maxBreakpoints = 2, useAnchor = false, anchorThreshold = 5 } = targeting
const { maxBreakpoints = 1, useAnchor = false, anchorThreshold = 5 } = targeting
// 1. Collect non-assistant message indices (user | tool roles)
const nonAssistantIndices: number[] = []

3
src/core/task-persistence/apiMessages.ts
View file

@ -282,10 +282,9 @@ export async function saveRooMessages({
try {
const taskDir = await getTaskDirectoryPath(globalStoragePath, taskId)
const filePath = path.join(taskDir, GlobalFileNames.apiConversationHistory)
const strippedMessages = stripCacheProviderOptions(messages)
const envelope: RooMessageHistory = {
version: ROO_MESSAGE_VERSION,
messages: strippedMessages,
messages,
}
await safeWriteJson(filePath, envelope)
return true

5
src/core/task/Task.ts
View file

@ -63,7 +63,7 @@ import { ApiHandler, ApiHandlerCreateMessageMetadata, buildApiHandler } from "..
import type { AssistantModelMessage } from "ai"
import { ApiStream, GroundingSource } from "../../api/transform/stream"
import { maybeRemoveImageBlocks } from "../../api/transform/image-cleaning"
import { applyCacheBreakpoints, UNIVERSAL_CACHE_OPTIONS } from "../../api/transform/cache-breakpoints"
import { UNIVERSAL_CACHE_OPTIONS } from "../../api/transform/cache-breakpoints"
// shared
import { findLastIndex } from "../../shared/array"
@ -4387,9 +4387,6 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
const mergedForApi = mergeConsecutiveApiMessages(messagesSinceLastSummary, { roles: ["user"] })
const messagesWithoutImages = maybeRemoveImageBlocks(mergedForApi, this.api)
// Breakpoints 3-4: Apply cache breakpoints to the last 2 non-assistant messages
applyCacheBreakpoints(messagesWithoutImages.filter(isRooRoleMessage))
// Check auto-approval limits
const approvalResult = await this.autoApprovalHandler.checkAutoApprovalLimits(
state,