[pull] main from Kilo-Org:main

.changeset/loud-lights-build.md

@@ -0,0 +1,5 @@
+---
+"kilo-code": minor
+---
+
+Added gemini-3-flash-preview model

packages/types/src/providers/gemini.ts

@@ -3,7 +3,7 @@ import type { ModelInfo } from "../model.js"
 // https://ai.google.dev/gemini-api/docs/models/gemini
 export type GeminiModelId = keyof typeof geminiModels
 
-export const geminiDefaultModelId: GeminiModelId = "gemini-3-pro-preview" // kilocode_change
+export const geminiDefaultModelId: GeminiModelId = "gemini-3-pro-preview"
 
 export const geminiModels = {
 	"gemini-3-pro-preview": {
@@ -31,6 +31,22 @@ export const geminiModels = {
 			},
 		],
 	},
+	"gemini-3-flash-preview": {
+		maxTokens: 65_536,
+		contextWindow: 1_048_576,
+		supportsImages: true,
+		supportsNativeTools: true,
+		defaultToolProtocol: "native",
+		supportsPromptCache: true,
+		supportsReasoningEffort: ["minimal", "low", "medium", "high"],
+		reasoningEffort: "medium",
+		supportsTemperature: true,
+		defaultTemperature: 1,
+		inputPrice: 0.3,
+		outputPrice: 2.5,
+		cacheReadsPrice: 0.075,
+		cacheWritesPrice: 1.0,
+	},
 	// 2.5 Pro models
 	"gemini-2.5-pro": {
 		maxTokens: 64_000,

packages/types/src/providers/vertex.ts

@@ -6,19 +6,6 @@ export type VertexModelId = keyof typeof vertexModels
 export const vertexDefaultModelId: VertexModelId = "claude-sonnet-4-5@20250929"
 
 export const vertexModels = {
-	"gemini-3-flash-preview": {
-		maxTokens: 65_536,
-		contextWindow: 1_048_576,
-		supportsImages: true,
-		supportsNativeTools: true,
-		supportsPromptCache: true,
-		supportsReasoningEffort: ["minimal", "low", "medium", "high"],
-		reasoningEffort: "medium",
-		supportsTemperature: true,
-		defaultTemperature: 1,
-		inputPrice: 0.5,
-		outputPrice: 3.0,
-	},
 	"gemini-3-pro-preview": {
 		maxTokens: 65_536,
 		contextWindow: 1_048_576,
@@ -44,6 +31,22 @@ export const vertexModels = {
 			},
 		],
 	},
+	"gemini-3-flash-preview": {
+		maxTokens: 65_536,
+		contextWindow: 1_048_576,
+		supportsImages: true,
+		supportsNativeTools: true,
+		defaultToolProtocol: "native",
+		supportsPromptCache: true,
+		supportsReasoningEffort: ["minimal", "low", "medium", "high"],
+		reasoningEffort: "medium",
+		supportsTemperature: true,
+		defaultTemperature: 1,
+		inputPrice: 0.3,
+		outputPrice: 2.5,
+		cacheReadsPrice: 0.075,
+		cacheWritesPrice: 1.0,
+	},
 	"gemini-2.5-flash-preview-05-20:thinking": {
 		maxTokens: 65_535,
 		contextWindow: 1_048_576,

src/services/continuedev/core/autocomplete/postprocessing/__tests__/removePrefixOverlap.test.ts

@@ -0,0 +1,33 @@
+import { describe, it, expect } from "vitest"
+import { removePrefixOverlap } from "../removePrefixOverlap.js"
+
+describe("removePrefixOverlap", () => {
+	it("should remove full prefix match", () => {
+		expect(removePrefixOverlap("hello world", "hello ")).toBe("world")
+	})
+
+	it("should remove trimmed prefix when prefix has extra whitespace", () => {
+		expect(removePrefixOverlap("hello world", "  hello  ")).toBe(" world")
+	})
+
+	it("should remove partial word overlap", () => {
+		expect(removePrefixOverlap("hello world", "hel")).toBe("lo world")
+	})
+
+	it("should return completion as-is when no overlap", () => {
+		expect(removePrefixOverlap("world", "hello")).toBe("world")
+	})
+
+	it("should use last line of multi-line prefix", () => {
+		expect(removePrefixOverlap("hello world", "line 1\nline 2\nhello ")).toBe("world")
+	})
+
+	it("should handle empty strings", () => {
+		expect(removePrefixOverlap("hello world", "")).toBe("hello world")
+		expect(removePrefixOverlap("", "hello")).toBe("")
+	})
+
+	it("should be case-sensitive", () => {
+		expect(removePrefixOverlap("hello world", "HELLO ")).toBe("hello world")
+	})
+})

src/services/continuedev/core/autocomplete/postprocessing/index.ts

@@ -1,5 +1,6 @@
 import { longestCommonSubsequence } from "../../util/lcs.js"
 import { lineIsRepeated } from "../util/textSimilarity.js"
+import { removePrefixOverlap } from "./removePrefixOverlap.js"
 
 function rewritesLineAbove(completion: string, prefix: string): boolean {
 	const lineAbove = prefix
@@ -143,21 +144,7 @@ export function postprocessCompletion({
 	}
 
 	if (llm.model.includes("mercury") || llm.model.includes("granite")) {
-		// Granite tends to repeat the start of the line in the completion output
-		const prefixEnd = prefix.split("\n").pop()
-		if (prefixEnd) {
-			if (completion.startsWith(prefixEnd)) {
-				completion = completion.slice(prefixEnd.length)
-			} else {
-				const trimmedPrefix = prefixEnd.trim()
-				const lastWord = trimmedPrefix.split(/\s+/).pop()
-				if (lastWord && completion.startsWith(lastWord)) {
-					completion = completion.slice(lastWord.length)
-				} else if (completion.startsWith(trimmedPrefix)) {
-					completion = completion.slice(trimmedPrefix.length)
-				}
-			}
-		}
+		completion = removePrefixOverlap(completion, prefix)
 	}
 
 	// // If completion starts with multiple whitespaces, but the cursor is at the end of the line

src/services/continuedev/core/autocomplete/postprocessing/removePrefixOverlap.ts

@@ -0,0 +1,17 @@
+export function removePrefixOverlap(completion: string, prefix: string): string {
+	const prefixEnd = prefix.split("\n").pop()
+	if (prefixEnd) {
+		if (completion.startsWith(prefixEnd)) {
+			completion = completion.slice(prefixEnd.length)
+		} else {
+			const trimmedPrefix = prefixEnd.trim()
+			const lastWord = trimmedPrefix.split(/\s+/).pop()
+			if (lastWord && completion.startsWith(lastWord)) {
+				completion = completion.slice(lastWord.length)
+			} else if (completion.startsWith(trimmedPrefix)) {
+				completion = completion.slice(trimmedPrefix.length)
+			}
+		}
+	}
+	return completion
+}

src/services/ghost/chat-autocomplete/ChatTextAreaAutocomplete.ts

@@ -3,6 +3,7 @@ import { GhostModel } from "../GhostModel"
 import { ProviderSettingsManager } from "../../../core/config/ProviderSettingsManager"
 import { VisibleCodeContext } from "../types"
 import { ApiStreamChunk } from "../../../api/transform/stream"
+import { removePrefixOverlap } from "../../continuedev/core/autocomplete/postprocessing/removePrefixOverlap.js"
 
 /**
  * Service for providing FIM-based autocomplete suggestions in ChatTextArea
@@ -64,6 +65,9 @@ export class ChatTextAreaAutocomplete {
 		}
 
 		const cleanedSuggestion = this.cleanSuggestion(response, userText)
+		console.log(
+			`[ChatAutocomplete] prefix: ${JSON.stringify(userText)} | response: ${JSON.stringify(response)} | cleanedSuggestion: ${JSON.stringify(cleanedSuggestion)}`,
+		)
 
 		return { suggestion: cleanedSuggestion }
 	}
@@ -81,6 +85,8 @@ export class ChatTextAreaAutocomplete {
 - Use context from visible code if relevant
 - NEVER repeat what the user already typed
 - NEVER start with comments (//, /*, #)
+- If the user is in the middle of typing a word (e.g., "hel"), include the COMPLETE word in your response (e.g., "hello world" not just "lo world")
+- This allows proper prefix matching to remove the overlap correctly
 - Return ONLY the completion text, no explanations or formatting`
 	}
 
@@ -90,7 +96,9 @@ export class ChatTextAreaAutocomplete {
 	private getChatUserPrompt(prefix: string): string {
 		return `${prefix}
 
-TASK: Complete the user's message naturally. Return ONLY the completion text (what comes next), no explanations.`
+TASK: Complete the user's message naturally. 
+- If the user is mid-word (e.g., typed "hel"), return the COMPLETE word (e.g., "hello world") so prefix matching can work correctly
+- Return ONLY the completion text (what comes next), no explanations.`
 	}
 
 	/**
@@ -146,18 +154,15 @@ TASK: Complete the user's message naturally. Return ONLY the completion text (wh
 	 * and filtering out unwanted patterns like comments
 	 */
 	private cleanSuggestion(suggestion: string, userText: string): string {
-		let cleaned = suggestion.trim()
+		let cleaned = suggestion
 
-		if (cleaned.startsWith(userText)) {
-			cleaned = cleaned.substring(userText.length)
-		}
+		cleaned = removePrefixOverlap(cleaned, userText)
 
 		const firstNewline = cleaned.indexOf("\n")
 		if (firstNewline !== -1) {
 			cleaned = cleaned.substring(0, firstNewline)
 		}
-
-		cleaned = cleaned.trimStart()
+		cleaned = cleaned.trimEnd() // Do NOT trim the end of the suggestion
 
 		// Filter out suggestions that start with comment patterns
 		// This happens because the context uses // prefixes for labels