scriptoriumadmin/Roo-Code
1
0
Fork 0
mirror of https://github.com/RooVetGit/Roo-Code.git synced 2026-10-06 02:47:56 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Update the max_tokens fallback logic in the sliding window (#5993)

Browse source
This commit is contained in:
Matt Rubens 2025-07-21 22:26:39 -04:00 • committed by GitHub
parent 0500894b34
commit 8334f0869f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 8 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

12
src/core/sliding-window/__tests__/sliding-window.spec.ts
View file

@ -1103,9 +1103,9 @@ describe("Sliding Window", () => {
expect(result2.prevContextTokens).toBe(50001)
})
it("should use 20% of context window as buffer when maxTokens is undefined", async () => {
it("should use ANTHROPIC_DEFAULT_MAX_TOKENS as buffer when maxTokens is undefined", async () => {
const modelInfo = createModelInfo(100000, undefined)
// Max tokens = 100000 - (100000 * 0.2) = 80000
// Max tokens = 100000 - ANTHROPIC_DEFAULT_MAX_TOKENS = 100000 - 8192 = 91808
// Create messages with very small content in the last one to avoid token overflow
const messagesWithSmallContent = [
@ -1117,7 +1117,7 @@ describe("Sliding Window", () => {
// Below max tokens and buffer - no truncation
const result1 = await truncateConversationIfNeeded({
messages: messagesWithSmallContent,
totalTokens: 69999, // Well below threshold + dynamic buffer
totalTokens: 81807, // Well below threshold + dynamic buffer (91808 - 10000 = 81808)
contextWindow: modelInfo.contextWindow,
maxTokens: modelInfo.maxTokens,
apiHandler: mockApiHandler,
@ -1132,13 +1132,13 @@ describe("Sliding Window", () => {
messages: messagesWithSmallContent,
summary: "",
cost: 0,
prevContextTokens: 69999,
prevContextTokens: 81807,
})
// Above max tokens - truncate
const result2 = await truncateConversationIfNeeded({
messages: messagesWithSmallContent,
totalTokens: 80001, // Above threshold
totalTokens: 81809, // Above threshold (81808)
contextWindow: modelInfo.contextWindow,
maxTokens: modelInfo.maxTokens,
apiHandler: mockApiHandler,
@ -1153,7 +1153,7 @@ describe("Sliding Window", () => {
expect(result2.messages.length).toBe(3) // Truncated with 0.5 fraction
expect(result2.summary).toBe("")
expect(result2.cost).toBe(0)
expect(result2.prevContextTokens).toBe(80001)
expect(result2.prevContextTokens).toBe(81809)
})
it("should handle small context windows appropriately", async () => {

3
src/core/sliding-window/index.ts
View file

@ -5,6 +5,7 @@ import { TelemetryService } from "@roo-code/telemetry"
import { ApiHandler } from "../../api"
import { MAX_CONDENSE_THRESHOLD, MIN_CONDENSE_THRESHOLD, summarizeConversation, SummarizeResponse } from "../condense"
import { ApiMessage } from "../task-persistence/apiMessages"
import { ANTHROPIC_DEFAULT_MAX_TOKENS } from "@roo-code/types"
/**
* Default percentage of the context window to use as a buffer when deciding when to truncate
@ -105,7 +106,7 @@ export async function truncateConversationIfNeeded({
let error: string | undefined
let cost = 0
// Calculate the maximum tokens reserved for response
const reservedTokens = maxTokens || contextWindow * 0.2
const reservedTokens = maxTokens || ANTHROPIC_DEFAULT_MAX_TOKENS
// Estimate tokens for the last message (which is always a user message)
const lastMessage = messages[messages.length - 1]