refactor(vscode-lm): default-row condense fallback + shared contextPercent helper

getCondenseContextWindow() resolves the default vscode-lm row for an unknown/absent family (catalog drift) instead of the inflated live window; only a non-positive static maxInputTokens still falls back to it. Extract the duplicated contextPercent math shared by willManageContext and manageContext into computeContextPercent so the two stay in lockstep. Addresses PR review feedback.

Author: Bertan Ari

src/api/providers/__tests__/vscode-lm.spec.ts

@@ -63,7 +63,7 @@ import * as vscode from "vscode"
 import { VsCodeLmHandler } from "../vscode-lm"
 import type { ApiHandlerOptions } from "../../../shared/api"
 import type { Anthropic } from "@anthropic-ai/sdk"
-import { openAiModelInfoSaneDefaults, vscodeLlmModels } from "@roo-code/types"
+import { openAiModelInfoSaneDefaults, vscodeLlmDefaultModelId, vscodeLlmModels } from "@roo-code/types"
 
 const mockLanguageModelChat = {
 	id: "test-model",
@@ -483,20 +483,21 @@ describe("VsCodeLmHandler", () => {
 			opusHandler.dispose()
 		})
 
-		it("falls back to the live model context window for families not in the static table", () => {
-			// Not a curated row, so the gate uses the live runtime window.
+		it("falls back to the default-row maxInputTokens for an unknown family (catalog drift)", () => {
+			// `test-family` isn't a curated row (e.g. a selector left over from a dropped model), so the
+			// gate resolves the default row instead of the inflated live window.
 			handler["client"] = mockLanguageModelChat as unknown as vscode.LanguageModelChat
-			expect(handler.getCondenseContextWindow()).toBe(handler.getModel().info.contextWindow)
-			expect(handler.getCondenseContextWindow()).toBe(mockLanguageModelChat.maxInputTokens)
+			expect(handler.getCondenseContextWindow()).toBe(vscodeLlmModels[vscodeLlmDefaultModelId].maxInputTokens)
 		})
 
-		it("falls back to the live window when no family is resolvable (no client, no selector family)", () => {
-			// No client and no selector family means `family` is undefined, so the gate skips the
-			// static lookup and uses getModel().info.contextWindow.
+		it("falls back to the default-row maxInputTokens when no family is resolvable (no client, no selector family)", () => {
+			// No client and no selector family means `family` is undefined, so the gate uses the default
+			// row's maxInputTokens rather than the live getModel().info.contextWindow.
 			const noFamilyHandler = new VsCodeLmHandler({ vsCodeLmModelSelector: { vendor: "copilot" } })
 			noFamilyHandler["client"] = null
-			expect(noFamilyHandler.getCondenseContextWindow()).toBe(noFamilyHandler.getModel().info.contextWindow)
-			expect(noFamilyHandler.getCondenseContextWindow()).toBe(openAiModelInfoSaneDefaults.contextWindow)
+			expect(noFamilyHandler.getCondenseContextWindow()).toBe(
+				vscodeLlmModels[vscodeLlmDefaultModelId].maxInputTokens,
+			)
 			noFamilyHandler.dispose()
 		})
 

src/api/providers/vscode-lm.ts

@@ -2,7 +2,7 @@ import { Anthropic } from "@anthropic-ai/sdk"
 import * as vscode from "vscode"
 import OpenAI from "openai"
 
-import { type ModelInfo, openAiModelInfoSaneDefaults, vscodeLlmModels } from "@roo-code/types"
+import { type ModelInfo, openAiModelInfoSaneDefaults, vscodeLlmDefaultModelId, vscodeLlmModels } from "@roo-code/types"
 
 import type { ApiHandlerOptions } from "../../shared/api"
 import { SELECTOR_SEPARATOR, stringifyVsCodeLmModelSelector } from "../../shared/vsCodeSelectorUtils"
@@ -565,12 +565,15 @@ export class VsCodeLmHandler extends BaseProvider implements SingleCompletionHan
 	/**
 	 * Context window for auto-condense. The API's advertised `client.maxInputTokens` is far larger
 	 * than usable, so relying on it stops auto-condense from firing; measure against the curated
-	 * static table's `maxInputTokens` instead (the same value the bar uses). Fall back to the live
-	 * window when the model isn't in the table.
+	 * static table's `maxInputTokens` instead (the same value the bar uses). An unknown family (e.g.
+	 * a selector left over from a model dropped from the catalog) resolves to the default row rather
+	 * than the inflated live window; only a non-positive static `maxInputTokens` falls back to it.
 	 */
 	getCondenseContextWindow(): number {
 		const family = this.client?.family ?? this.options.vsCodeLmModelSelector?.family
-		const staticModel = family ? vscodeLlmModels[family as keyof typeof vscodeLlmModels] : undefined
+		const staticModel = family
+			? (vscodeLlmModels[family as keyof typeof vscodeLlmModels] ?? vscodeLlmModels[vscodeLlmDefaultModelId])
+			: vscodeLlmModels[vscodeLlmDefaultModelId]
 
 		if (staticModel && typeof staticModel.maxInputTokens === "number" && staticModel.maxInputTokens > 0) {
 			return staticModel.maxInputTokens

src/core/context-management/index.ts

@@ -40,6 +40,32 @@ export async function estimateTokenCount(
 	return apiHandler.countTokens(content)
 }
 
+/**
+ * Computes the percentage of the context budget consumed by the prior context.
+ *
+ * Default: divide by the full context window. Opt-in (vscode-lm) divides by available input
+ * (window minus reserved output); an unknown/unlimited reserve (maxTokens -1) falls back to the
+ * full window. Shared by `willManageContext` and `manageContext` so the two stay in lockstep.
+ */
+function computeContextPercent({
+	prevContextTokens,
+	contextWindow,
+	maxTokens,
+	useAvailableInputForContextPercent,
+}: {
+	prevContextTokens: number
+	contextWindow: number
+	maxTokens?: number | null
+	useAvailableInputForContextPercent?: boolean
+}): number {
+	if (!useAvailableInputForContextPercent) {
+		return (100 * prevContextTokens) / contextWindow
+	}
+	const reservedForOutput = maxTokens && maxTokens > 0 ? maxTokens : 0
+	const availableInputTokens = contextWindow - reservedForOutput
+	return availableInputTokens > 0 ? (100 * prevContextTokens) / availableInputTokens : 100
+}
+
 /**
  * Result of truncation operation, includes the truncation ID for UI events.
  */
@@ -200,16 +226,12 @@ export function willManageContext({
 		// Invalid values fall back to global setting (effectiveThreshold already set)
 	}
 
-	// Default: divide by the full context window. Opt-in (vscode-lm) divides by available input
-	// (window minus reserved output); an unknown/unlimited reserve (-1) falls back to the full window.
-	let contextPercent: number
-	if (useAvailableInputForContextPercent) {
-		const reservedForOutput = maxTokens && maxTokens > 0 ? maxTokens : 0
-		const availableInputTokens = contextWindow - reservedForOutput
-		contextPercent = availableInputTokens > 0 ? (100 * prevContextTokens) / availableInputTokens : 100
-	} else {
-		contextPercent = (100 * prevContextTokens) / contextWindow
-	}
+	const contextPercent = computeContextPercent({
+		prevContextTokens,
+		contextWindow,
+		maxTokens,
+		useAvailableInputForContextPercent,
+	})
 	return contextPercent >= effectiveThreshold || prevContextTokens > allowedTokens
 }
 
@@ -328,16 +350,12 @@ export async function manageContext({
 	// If no specific threshold is found for the profile, fall back to global setting
 
 	if (autoCondenseContext) {
-		// Default: divide by the full context window. Opt-in (vscode-lm) divides by available input
-		// (window minus reserved output); an unknown/unlimited reserve (-1) falls back to the full window.
-		let contextPercent: number
-		if (useAvailableInputForContextPercent) {
-			const reservedForOutput = maxTokens && maxTokens > 0 ? maxTokens : 0
-			const availableInputTokens = contextWindow - reservedForOutput
-			contextPercent = availableInputTokens > 0 ? (100 * prevContextTokens) / availableInputTokens : 100
-		} else {
-			contextPercent = (100 * prevContextTokens) / contextWindow
-		}
+		const contextPercent = computeContextPercent({
+			prevContextTokens,
+			contextWindow,
+			maxTokens,
+			useAvailableInputForContextPercent,
+		})
 		if (contextPercent >= effectiveThreshold || prevContextTokens > allowedTokens) {
 			// Attempt to intelligently condense the context
 			const result = await summarizeConversation({