[vscode-lm] Add missing model families and correct stale context limits

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

@@ -2,14 +2,24 @@ import { describe, it, expect } from "vitest"
 
 import { vscodeLlmModels, vscodeLlmDefaultModelId } from "../providers/vscode-llm.js"
 
+// The five families added and the two refreshed on 2026-09-10 (VS Code 1.137.0).
+const SCOPED_ROWS = {
+	"gpt-6-astra": { contextWindow: 871793, maxInputTokens: 271783, supportsImages: true },
+	"grok-4.5": { contextWindow: 424794, maxInputTokens: 199783, supportsImages: false },
+	"grok-4.6": { contextWindow: 424794, maxInputTokens: 199784, supportsImages: false },
+	"gemini-3.7-flash": { contextWindow: 935793, maxInputTokens: 935783, supportsImages: true },
+	"gemini-3.8-flash": { contextWindow: 982833, maxInputTokens: 955113, supportsImages: true },
+	"claude-opus-5": { contextWindow: 935793, maxInputTokens: 680456, supportsImages: true },
+	"gemini-3.6-flash": { contextWindow: 935793, maxInputTokens: 935785, supportsImages: true },
+} as const
+
+const REFRESHED_FAMILIES = Object.keys(SCOPED_ROWS) as Array<keyof typeof SCOPED_ROWS>
+
 describe("vscodeLlmModels", () => {
 	it("exposes the opus-4.8 row with its measured maxInputTokens and contextWindow", () => {
-		// The VS Code LM API exposes only maxInputTokens; that is the value the UI reads from this
-		// table (useSelectedModel.ts). For claude-opus-4.8 the two fields intentionally DIVERGE:
-		// maxInputTokens (197897) is the enforced input ceiling, while contextWindow (679560) records
-		// the larger advertised window. The UI reads maxInputTokens, so the divergence is a deliberate
-		// tripwire — assert the actual on-disk literals rather than forcing equality.
-		// See GitHub issue simurg79/Roo-Code#10.
+		// claude-opus-4.8 intentionally DIVERGES: maxInputTokens (197897) is the enforced ceiling the
+		// UI reads, contextWindow (679560) the larger advertised window. Assert the on-disk literals
+		// rather than forcing equality.
 		expect(vscodeLlmModels).toHaveProperty("claude-opus-4.8")
 		expect(vscodeLlmModels["claude-opus-4.8"].contextWindow).toBe(679560)
 		expect(vscodeLlmModels["claude-opus-4.8"].maxInputTokens).toBe(197897)
@@ -23,42 +33,72 @@ describe("vscodeLlmModels", () => {
 	})
 
 	it("includes the 2026-07-14 additions with their measured single-message ceilings", () => {
-		// Measured via single-message binary search on VS Code 1.126.0 (largest input the backend
-		// accepts). claude-sonnet-5 accepts nearly its full advertised window (925449), unlike the
-		// older claude rows that cap at ~197.9K — this divergence is exactly why the values are
-		// measured rather than inferred from a sibling row.
+		// Measured by single-message binary search on VS Code 1.126.0. claude-sonnet-5 accepts nearly
+		// its full window (925449) unlike older claude rows capped at ~197.9K — which is why values
+		// are measured, not inferred from a sibling row.
 		expect(vscodeLlmModels["claude-sonnet-5"].maxInputTokens).toBe(925449)
 		expect(vscodeLlmModels["claude-sonnet-5"].contextWindow).toBe(925449)
 		expect(vscodeLlmModels["gpt-5.6-luna"].maxInputTokens).toBe(199753)
 		expect(vscodeLlmModels["gpt-5.6-sol"].maxInputTokens).toBe(271785)
 		expect(vscodeLlmModels["gpt-5.6-terra"].maxInputTokens).toBe(271785)
 	})
 
-	it("caps the 2026-07-26 model-picker-cache additions at the conservative measured ceiling", () => {
-		// Sourced from `chat.cachedLanguageModels` in User/globalStorage/state.vscdb; not yet
-		// binary-searched, so contextWindow records the advertised window while maxInputTokens holds
-		// the lowest measured enforced ceiling to avoid overflowing on an unverified number.
-		expect(vscodeLlmModels["claude-opus-5"].maxInputTokens).toBe(197897)
+	it("replaces the conservative 197897 placeholders with the 2026-09-10 measured ceilings", () => {
+		// These rows previously carried the placeholder 197897 (a borrowed sibling ceiling, not a
+		// measurement); the 2026-09-10 bracket-derived values supersede it. contextWindow is
+		// unchanged, so a failure here means the ceiling reverted to the placeholder.
+		expect(vscodeLlmModels["claude-opus-5"].maxInputTokens).toBe(680456)
 		expect(vscodeLlmModels["claude-opus-5"].contextWindow).toBe(935793)
-		expect(vscodeLlmModels["gemini-3.6-flash"].maxInputTokens).toBe(197897)
+		expect(vscodeLlmModels["gemini-3.6-flash"].maxInputTokens).toBe(935785)
 		expect(vscodeLlmModels["gemini-3.6-flash"].contextWindow).toBe(935793)
 	})
 
+	it("leaves no 2026-09-10 scoped family pinned to the 197897 placeholder", () => {
+		// Untouched rows (e.g. claude-opus-4.7) legitimately still measure 197897, so this guard is
+		// limited to the families this refresh touched.
+		for (const family of REFRESHED_FAMILIES) {
+			expect(vscodeLlmModels[family].maxInputTokens, `${family} must not be the 197897 placeholder`).not.toBe(
+				197897,
+			)
+		}
+	})
+
+	it("pins every 2026-09-10 row's windows and capability flags", () => {
+		for (const [family, expected] of Object.entries(SCOPED_ROWS)) {
+			const model = vscodeLlmModels[family as keyof typeof vscodeLlmModels]
+
+			expect(model.maxInputTokens, `${family}: maxInputTokens`).toBe(expected.maxInputTokens)
+			expect(model.contextWindow, `${family}: contextWindow`).toBe(expected.contextWindow)
+			expect(model.supportsImages, `${family}: supportsImages`).toBe(expected.supportsImages)
+
+			// Table-wide conventions: prices are 0 because these rows carry no per-token accounting,
+			// and supportsPromptCache is false because no cache is modelled here (not a claim about
+			// the backend). Tool calling is required for Roo to function.
+			expect(model.supportsPromptCache, `${family}: supportsPromptCache`).toBe(false)
+			expect(model.inputPrice, `${family}: inputPrice`).toBe(0)
+			expect(model.outputPrice, `${family}: outputPrice`).toBe(0)
+			expect(model.supportsToolCalling, `${family}: supportsToolCalling`).toBe(true)
+
+			// Provider and gauge look rows up by the live client's family string, so the key, family
+			// and version must agree or lookups silently degrade to the default row.
+			expect(model.family, `${family}: family`).toBe(family)
+			expect(model.version, `${family}: version`).toBe(family)
+		}
+	})
+
 	it("keeps both window fields populated and positive for every row", () => {
-		// NOTE: contextWindow and maxInputTokens are intentionally ALLOWED to differ (claude-opus-4.8
-		// diverges: 679560 vs 197897). The UI reads maxInputTokens, and that divergence is a deliberate
-		// tripwire, so we do NOT assert contextWindow === maxInputTokens here (see simurg79/Roo-Code#10).
-		// The meaningful invariant is that every row carries positive integers for both fields; a
-		// missing/zero value would point to hand-authored drift rather than a real captured row.
+		// The two fields are ALLOWED to differ (claude-opus-4.8: 679560 vs 197897), so equality is
+		// deliberately NOT asserted. The invariant is positive integers on both fields; a missing or
+		// zero value indicates hand-authored drift, not a captured row.
 		for (const [family, model] of Object.entries(vscodeLlmModels)) {
 			expect(model.contextWindow, `${family}: contextWindow must be a positive integer`).toBeGreaterThan(0)
 			expect(model.maxInputTokens, `${family}: maxInputTokens must be a positive integer`).toBeGreaterThan(0)
 		}
 	})
 
 	it("excludes fabricated/internal/alias families and the dropped legacy rows", () => {
-		// Integrity guards: these were never part of the authoritative live capture, or were
-		// removed by the full table REPLACE. Their presence would signal hand-authored drift.
+		// These were never in the live capture, or were removed by the full table REPLACE; their
+		// presence would signal hand-authored drift.
 		expect(vscodeLlmModels).not.toHaveProperty("claude-opus-4.7-high")
 		expect(vscodeLlmModels).not.toHaveProperty("claude-3.5-sonnet")
 		expect(vscodeLlmModels).not.toHaveProperty("claude-4-sonnet")

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

@@ -4,34 +4,11 @@ export type VscodeLlmModelId = keyof typeof vscodeLlmModels
 
 export const vscodeLlmDefaultModelId: VscodeLlmModelId = "claude-sonnet-4.5"
 
-// Rows below were originally enumerated from `vscode.lm.selectChatModels({ vendor: "copilot" })`.
-// The VS Code LM API exposes ONLY `maxInputTokens` (there is no separate context-window field), and
-// that is the single value the runtime/condense gate enforces: getModel() sets
-// contextWindow = Math.max(0, client.maxInputTokens) in src/api/providers/vscode-lm.ts. So for every
-// row `maxInputTokens` IS the enforced context window, and `contextWindow` normally mirrors it
-// (the UI reads maxInputTokens via useSelectedModel.ts). Rows whose advertised window is known to
-// exceed the measured ceiling deliberately diverge: contextWindow records what Copilot advertises
-// while maxInputTokens carries the value the backend actually accepts.
-// These ceilings were measured empirically on 2026-06-18 (VS Code 1.125.0) by binary-searching the
-// single-message "Message exceeds token limit" threshold per model — they are the largest input the
-// backend actually accepts, which for several models is well below the value Copilot advertises:
-//   - claude-opus-4.8:                                   enforced 679560
-//   - claude-opus-4.7 / 4.6, claude-sonnet-4.6,
-//     gemini-3.1-pro-preview, gemini-3.5-flash:          enforced ~197.9K
-//   - gpt-5.5 / gpt-5.4:                                 enforced ~268.4K
-// Additions measured 2026-07-14 (VS Code 1.126.0) with the same single-message binary search:
-//   - claude-sonnet-5:                                   enforced 925449 (near-full window, unlike the older ~197.9K claude rows)
-//   - gpt-5.6-luna:                                      enforced 199753
-//   - gpt-5.6-sol / gpt-5.6-terra:                       enforced 271785
-// Additions sourced 2026-07-26 from the Copilot model-picker cache (`chat.cachedLanguageModels` in
-// VS Code's User/globalStorage/state.vscdb), which persists the metadata `selectChatModels` returns:
-//   - claude-opus-5 / gemini-3.6-flash:                  advertised 935793, NOT yet binary-searched
-// Those two rows keep 935793 in contextWindow as the advertised window, but pin maxInputTokens to
-// 197897 — the lowest enforced ceiling measured on any row here. Trusting an unverified advertised
-// window would overflow the request and hard-fail mid-task, so the conservative floor is used until
-// a binary search establishes the real ceiling (mirrors the claude-opus-4.8 divergence above).
-// Guardrail: these are empirically measured — re-measure (do not hand-tune) if the models change.
-// See GitHub issue simurg79/Roo-Code#10 and myplans/VSCode LM Model Table Integrity/vscode_lm_opus_data_integrity_design.md.
+// contextWindow = advertised window; maxInputTokens = measured single-message accepted ceiling, and
+// it is what the condense gate and the context gauge enforce — so the two deliberately diverge where
+// advertised exceeds measured. Values MUST be re-measured, never hand-tuned or borrowed from a
+// sibling row. Per-row evidence, lower-bound and vendor caveats:
+// myplans/vscode-lm-model-table-integrity/vscode-lm-model-table-integrity-design.md
 export const vscodeLlmModels = {
 	"claude-opus-5": {
 		contextWindow: 935793,
@@ -43,7 +20,7 @@ export const vscodeLlmModels = {
 		version: "claude-opus-5",
 		name: "Claude Opus 5",
 		supportsToolCalling: true,
-		maxInputTokens: 197897,
+		maxInputTokens: 680456,
 	},
 	"claude-opus-4.8": {
 		contextWindow: 679560,
@@ -141,6 +118,18 @@ export const vscodeLlmModels = {
 		supportsToolCalling: true,
 		maxInputTokens: 135790,
 	},
+	"gpt-6-astra": {
+		contextWindow: 871793,
+		supportsImages: true,
+		supportsPromptCache: false,
+		inputPrice: 0,
+		outputPrice: 0,
+		family: "gpt-6-astra",
+		version: "gpt-6-astra",
+		name: "GPT-6 Astra",
+		supportsToolCalling: true,
+		maxInputTokens: 271783,
+	},
 	"gpt-5.6-luna": {
 		contextWindow: 199753,
 		supportsImages: true,
@@ -249,6 +238,58 @@ export const vscodeLlmModels = {
 		supportsToolCalling: true,
 		maxInputTokens: 12078,
 	},
+	"grok-4.6": {
+		contextWindow: 424794,
+		// Grok image requests failed with a nondiagnostic HTTP 400, so support is unknown; false is a
+		// deliberate disable pending evidence, not a measured absence of capability.
+		supportsImages: false,
+		supportsPromptCache: false,
+		inputPrice: 0,
+		outputPrice: 0,
+		family: "grok-4.6",
+		version: "grok-4.6",
+		name: "Grok 4.6",
+		supportsToolCalling: true,
+		maxInputTokens: 199784,
+	},
+	"grok-4.5": {
+		contextWindow: 424794,
+		// See grok-4.6 above: images disabled pending evidence, not proven unsupported.
+		supportsImages: false,
+		supportsPromptCache: false,
+		inputPrice: 0,
+		outputPrice: 0,
+		family: "grok-4.5",
+		version: "grok-4.5",
+		name: "Grok 4.5",
+		supportsToolCalling: true,
+		maxInputTokens: 199783,
+	},
+	"gemini-3.8-flash": {
+		contextWindow: 982833,
+		supportsImages: true,
+		supportsPromptCache: false,
+		inputPrice: 0,
+		outputPrice: 0,
+		family: "gemini-3.8-flash",
+		version: "gemini-3.8-flash",
+		name: "Gemini 3.8 Flash",
+		supportsToolCalling: true,
+		// Largest accepted trial with no rejection observed: a lower bound, not a measured ceiling.
+		maxInputTokens: 955113,
+	},
+	"gemini-3.7-flash": {
+		contextWindow: 935793,
+		supportsImages: true,
+		supportsPromptCache: false,
+		inputPrice: 0,
+		outputPrice: 0,
+		family: "gemini-3.7-flash",
+		version: "gemini-3.7-flash",
+		name: "Gemini 3.7 Flash",
+		supportsToolCalling: true,
+		maxInputTokens: 935783,
+	},
 	"gemini-3.6-flash": {
 		contextWindow: 935793,
 		supportsImages: true,
@@ -259,7 +300,7 @@ export const vscodeLlmModels = {
 		version: "gemini-3.6-flash",
 		name: "Gemini 3.6 Flash",
 		supportsToolCalling: true,
-		maxInputTokens: 197897,
+		maxInputTokens: 935785,
 	},
 	"gemini-3.1-pro-preview": {
 		contextWindow: 197897,

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

@@ -141,6 +141,85 @@ describe("VsCodeLmHandler", () => {
 			noFamilyHandler.dispose()
 		})
 
+		it.each([
+			["gpt-6-astra", 271783],
+			["grok-4.5", 199783],
+			["grok-4.6", 199784],
+			["gemini-3.7-flash", 935783],
+			["gemini-3.8-flash", 955113],
+			["claude-opus-5", 680456],
+			["gemini-3.6-flash", 935785],
+		])("resolves %s to its measured static ceiling when the live client reports it", (family, expected) => {
+			// The live client wins over the selector, and its inflated maxInputTokens (4096 here stands in
+			// for any advertised value) must be ignored in favour of the curated row.
+			const familyHandler = new VsCodeLmHandler({
+				vsCodeLmModelSelector: { vendor: "copilot", family: "claude-sonnet-4.5" },
+			})
+			familyHandler["client"] = {
+				...mockLanguageModelChat,
+				family,
+			} as unknown as vscode.LanguageModelChat
+
+			expect(familyHandler.getCondenseContextWindow()).toBe(expected)
+			expect(familyHandler.getCondenseContextWindow()).toBe(
+				vscodeLlmModels[family as keyof typeof vscodeLlmModels].maxInputTokens,
+			)
+
+			familyHandler.dispose()
+		})
+
+		it("falls back to the claude-sonnet-4.5 row (167790) for a family absent from the table", () => {
+			// Pins the literal so a change to the default row's own window is caught here rather than
+			// silently redefining what "fallback" means for every drifted selector.
+			const unknownHandler = new VsCodeLmHandler({
+				vsCodeLmModelSelector: { vendor: "copilot", family: "gpt-9-nonexistent" },
+			})
+			unknownHandler["client"] = null
+
+			expect(vscodeLlmDefaultModelId).toBe("claude-sonnet-4.5")
+			expect(unknownHandler.getCondenseContextWindow()).toBe(167790)
+
+			unknownHandler.dispose()
+		})
+
+		it("resolves by family alone, independent of the user-selected vendor", () => {
+			// Vendor is a selection/routing concern only; the window lookup is family-keyed, so two
+			// different vendors selecting the same family MUST agree on the condense budget.
+			const copilotHandler = new VsCodeLmHandler({
+				vsCodeLmModelSelector: { vendor: "copilot", family: "gpt-6-astra" },
+			})
+			const otherVendorHandler = new VsCodeLmHandler({
+				vsCodeLmModelSelector: { vendor: "some-other-vendor", family: "gpt-6-astra" },
+			})
+
+			expect(copilotHandler.getCondenseContextWindow()).toBe(271783)
+			expect(otherVendorHandler.getCondenseContextWindow()).toBe(copilotHandler.getCondenseContextWindow())
+
+			copilotHandler.dispose()
+			otherVendorHandler.dispose()
+		})
+
+		it("leaves getModel().contextWindow on the live client window for the new families", () => {
+			// Regression guard for the split: only the condense gate reads the static table. getModel()
+			// must keep reporting the LIVE window, so the refresh is provably behavior-neutral here.
+			const liveWindow = 123456
+			const astraHandler = new VsCodeLmHandler({
+				vsCodeLmModelSelector: { vendor: "copilot", family: "gpt-6-astra" },
+			})
+			astraHandler["client"] = {
+				...mockLanguageModelChat,
+				family: "gpt-6-astra",
+				maxInputTokens: liveWindow,
+			} as unknown as vscode.LanguageModelChat
+
+			expect(astraHandler.getModel().info.contextWindow).toBe(liveWindow)
+			expect(astraHandler.getModel().info.contextWindow).not.toBe(vscodeLlmModels["gpt-6-astra"].maxInputTokens)
+			// supportsImages still comes from the static row.
+			expect(astraHandler.getModel().info.supportsImages).toBe(true)
+
+			astraHandler.dispose()
+		})
+
 		it("falls back to the live window when the static row's maxInputTokens is non-positive", () => {
 			const family = "claude-opus-4.8"
 			const original = vscodeLlmModels[family].maxInputTokens

src/package.json

@@ -3,7 +3,7 @@
 	"displayName": "%extension.displayName%",
 	"description": "%extension.description%",
 	"publisher": "RooVeterinaryInc",
-	"version": "3.53.5",
+	"version": "3.53.6",
 	"icon": "assets/icons/icon.png",
 	"galleryBanner": {
 		"color": "#617A91",