fix(vscode-lm): complete Copilot model port with safe input ceilings

packages/types/src/__tests__/vscode-llm.spec.ts

@@ -16,15 +16,15 @@ describe("vscodeLlmModels", () => {
 		expect(vscodeLlmModels["gemini-2.5-pro"].maxInputTokens).toBe(108594)
 	})
 	it.each([
-		["claude-opus-5", "Claude Opus 5", 935793, "advertised, unverified"],
-		["claude-sonnet-5", "Claude Sonnet 5", 925449, "measured"],
-		["gpt-5.6-luna", "GPT-5.6 Luna", 199753, "measured"],
-		["gpt-5.6-sol", "GPT-5.6 Sol", 271785, "measured"],
-		["gpt-5.6-terra", "GPT-5.6 Terra", 271785, "measured"],
-		["gemini-3.6-flash", "Gemini 3.6 Flash", 935793, "advertised, unverified"],
-	] as const)("preserves PR #13 row %s (%s), limit %i (%s)", (...[family, name, limit]) => {
+		["claude-opus-5", "Claude Opus 5", 935793, 197897],
+		["claude-sonnet-5", "Claude Sonnet 5", 925449, 925449],
+		["gpt-5.6-luna", "GPT-5.6 Luna", 199753, 199753],
+		["gpt-5.6-sol", "GPT-5.6 Sol", 271785, 271785],
+		["gpt-5.6-terra", "GPT-5.6 Terra", 271785, 271785],
+		["gemini-3.6-flash", "Gemini 3.6 Flash", 935793, 197897],
+	] as const)("preserves PR #13 row %s (%s) with PR #14 ceilings", (family, name, contextWindow, maxInputTokens) => {
 		expect(vscodeLlmModels[family]).toEqual({
-			contextWindow: limit,
+			contextWindow,
 			supportsImages: true,
 			supportsPromptCache: false,
 			inputPrice: 0,
@@ -33,7 +33,7 @@ describe("vscodeLlmModels", () => {
 			version: family,
 			name,
 			supportsToolCalling: true,
-			maxInputTokens: limit,
+			maxInputTokens,
 		})
 	})
 

packages/types/src/providers/vscode-llm.ts

@@ -14,12 +14,11 @@ export type VscodeLlmModelId = keyof typeof vscodeLlmModels
 
 export const vscodeLlmDefaultModelId: VscodeLlmModelId = "claude-sonnet-4.5"
 
-// Curated VS Code LM (GitHub Copilot) model catalog.
-// The API exposes only `maxInputTokens`; the UI and condense gate read that. `contextWindow` is
-// the advertised window, kept for rows where it diverges from the ceiling (e.g. claude-opus-4.8).
-// https://github.com/simurg79/Roo-Code/pull/13: sonnet-5 and gpt-5.6 limits were measured
-// by single-message binary search on VS Code 1.126.0 (2026-07-14). Opus-5 and gemini-3.6-flash
-// retain advertised, unverified limits from the Copilot model-picker cache (2026-07-26).
+// The UI and condense gate use maxInputTokens; contextWindow preserves the advertised window.
+// https://github.com/simurg79/Roo-Code/pull/13: sonnet-5 and gpt-5.6 were binary-searched.
+// https://github.com/simurg79/Roo-Code/pull/14: opus-5 and gemini-3.6-flash retain advertised
+// 935793 windows but conservatively reuse the 197897 ceiling measured on other models to avoid
+// request overflow; neither model's actual input ceiling has been measured.
 export const vscodeLlmModels = {
 	"claude-opus-5": {
 		contextWindow: 935793,
@@ -31,7 +30,7 @@ export const vscodeLlmModels = {
 		version: "claude-opus-5",
 		name: "Claude Opus 5",
 		supportsToolCalling: true,
-		maxInputTokens: 935793,
+		maxInputTokens: 197897,
 	},
 	"claude-opus-4.8": {
 		contextWindow: 679560,
@@ -247,7 +246,7 @@ export const vscodeLlmModels = {
 		version: "gemini-3.6-flash",
 		name: "Gemini 3.6 Flash",
 		supportsToolCalling: true,
-		maxInputTokens: 935793,
+		maxInputTokens: 197897,
 	},
 	"gemini-3.1-pro-preview": {
 		contextWindow: 197897,

webview-ui/src/components/ui/hooks/__tests__/useSelectedModel.spec.ts

@@ -1422,24 +1422,25 @@ describe("useSelectedModel", () => {
 			expect(result.current.info?.supportsImages).toBe(false)
 		})
 
-		it("pins a divergent family's contextWindow to maxInputTokens, not its advertised window", () => {
-			// claude-opus-4.8 is the row where contextWindow and maxInputTokens differ; a field swap to
-			// the advertised window would be caught here.
-			const family = "claude-opus-4.8"
-			const apiConfiguration: ProviderSettings = {
-				apiProvider: providerIdentifiers.vscodeLm,
-				vsCodeLmModelSelector: { vendor: "copilot", family },
-			}
+		it.each(["claude-opus-4.8", "claude-opus-5", "gemini-3.6-flash"] as const)(
+			"pins %s to the 197897 input ceiling, not its advertised window",
+			(family) => {
+				const apiConfiguration: ProviderSettings = {
+					apiProvider: providerIdentifiers.vscodeLm,
+					vsCodeLmModelSelector: { vendor: "copilot", family },
+				}
 
-			const wrapper = createWrapper()
-			const { result } = renderHook(() => useSelectedModel(apiConfiguration), { wrapper })
+				const wrapper = createWrapper()
+				const { result } = renderHook(() => useSelectedModel(apiConfiguration), { wrapper })
 
-			expect(result.current.provider).toBe(providerIdentifiers.vscodeLm)
-			expect(result.current.id).toBe(`copilot/${family}`)
-			expect(result.current.info?.contextWindow).toBe(vscodeLlmModels[family].maxInputTokens) // 197897
-			expect(result.current.info?.contextWindow).not.toBe(vscodeLlmModels[family].contextWindow) // NOT 679560
-			expect(result.current.info?.supportsImages).toBe(false)
-		})
+				expect(result.current.provider).toBe(providerIdentifiers.vscodeLm)
+				expect(result.current.id).toBe(`copilot/${family}`)
+				expect(result.current.info?.contextWindow).toBe(197897)
+				expect(result.current.info?.contextWindow).toBe(vscodeLlmModels[family].maxInputTokens)
+				expect(result.current.info?.contextWindow).not.toBe(vscodeLlmModels[family].contextWindow)
+				expect(result.current.info?.supportsImages).toBe(false)
+			},
+		)
 
 		it("falls back to the default model's window for an unlisted family (NOT 128000)", () => {
 			const apiConfiguration: ProviderSettings = {