Bedrock: dynamic max output tokens with empirical AWS probe

- Correct per-model maxTokens in bedrockModels (Opus 4.7 = 128K, Opus 4.6 = 128K,
  Opus 4.5/4.1 = 32K, Sonnet 4.x and Haiku 4.5 = 64K) so the reasoning-budget
  slider is no longer artificially clamped at 16K.
- Add awsModelMaxOutputTokens per-config override and plumb it through
  resolveBedrockModelInfo + AwsBedrockHandler.getModelById.
- New probeBedrockMaxOutputTokens helper that empirically discovers the AWS
  cap for a given model via a 1-token Converse probe (accept-ceiling, parse
  AWS hint, or binary search).
- Wire requestBedrockMaxTokensProbe / bedrockMaxTokensProbe message types
  and handler.
- New webview MaxOutputTokensControl (slider + numeric input + dynamic step)
  and BedrockThinkingBudget wrapper that renders Detect/Reset buttons.
- Bump src/package.json to 3.53.12.

Co-Authored-By: Oz <oz-agent@warp.dev>

Author: allquixotic

packages/types/src/__tests__/bedrock.spec.ts

@@ -9,6 +9,7 @@ import {
 	bedrockModels,
 	expandBedrockTargetsWith1MVariants,
 	hasBedrock1MContextIndicator,
+	resolveBedrockModelInfo,
 	stripBedrock1MContextSuffix,
 } from "../providers/bedrock.js"
 
@@ -46,6 +47,47 @@ describe("Bedrock model catalog", () => {
 		expect(BEDROCK_1M_CONTEXT_DEFAULT_MODEL_IDS.length).toBe(0)
 		expect(BEDROCK_1M_CONTEXT_OPT_IN_MODEL_IDS.length).toBe(BEDROCK_1M_CONTEXT_MODEL_IDS.length)
 	})
+
+	it("matches per-model maxTokens to the documented Bedrock caps for current Anthropic models", () => {
+		// These caps mirror the Anthropic-direct entries in `anthropic.ts`. Bumping them lets the
+		// reasoning-budget slider extend past the legacy 8K cap (the original bug surfaced by Opus 4.7).
+		expect((bedrockModels["anthropic.claude-opus-4-7"] as ModelInfo).maxTokens).toBe(128_000)
+		expect((bedrockModels["anthropic.claude-opus-4-6-v1"] as ModelInfo).maxTokens).toBe(128_000)
+		expect((bedrockModels["anthropic.claude-opus-4-5-20251101-v1:0"] as ModelInfo).maxTokens).toBe(32_000)
+		expect((bedrockModels["anthropic.claude-opus-4-1-20250805-v1:0"] as ModelInfo).maxTokens).toBe(32_000)
+		expect((bedrockModels["anthropic.claude-sonnet-4-6"] as ModelInfo).maxTokens).toBe(64_000)
+		expect((bedrockModels["anthropic.claude-sonnet-4-5-20250929-v1:0"] as ModelInfo).maxTokens).toBe(64_000)
+		expect((bedrockModels["anthropic.claude-haiku-4-5-20251001-v1:0"] as ModelInfo).maxTokens).toBe(64_000)
+	})
+})
+
+describe("resolveBedrockModelInfo", () => {
+	it("prefers the static maxTokens when no override is set", () => {
+		const { info } = resolveBedrockModelInfo({
+			baseModelId: "anthropic.claude-opus-4-7",
+			targetId: "anthropic.claude-opus-4-7",
+		})
+		expect(info.maxTokens).toBe(128_000)
+	})
+
+	it("applies maxOutputTokensOverride above the static cap", () => {
+		const { info } = resolveBedrockModelInfo({
+			baseModelId: "anthropic.claude-opus-4-7",
+			targetId: "anthropic.claude-opus-4-7",
+			maxOutputTokensOverride: 256_000,
+		})
+		expect(info.maxTokens).toBe(256_000)
+	})
+
+	it("lets request-time modelMaxTokens still override (lowering for cost control)", () => {
+		const { info } = resolveBedrockModelInfo({
+			baseModelId: "anthropic.claude-opus-4-7",
+			targetId: "anthropic.claude-opus-4-7",
+			maxOutputTokensOverride: 256_000,
+			modelMaxTokens: 32_000,
+		})
+		expect(info.maxTokens).toBe(32_000)
+	})
 })
 
 describe("expandBedrockTargetsWith1MVariants", () => {

packages/types/src/provider-settings.ts

@@ -234,6 +234,11 @@ const bedrockSchema = apiModelIdProviderModelSchema.extend({
 		.enum(["foundation-model", "system-profile", "application-profile", "custom-arn", "prompt-router", "unknown"])
 		.optional(),
 	awsModelContextWindow: z.number().optional(),
+	// Empirically detected (or manually entered) per-config cap on the model's max output tokens.
+	// Takes precedence over the static `bedrockModels.<id>.maxTokens` table when present and overrides
+	// the static cap inside `resolveBedrockModelInfo`. The user-facing slider also widens to this value
+	// so future Anthropic releases that lift their output ceiling don't require a code change.
+	awsModelMaxOutputTokens: z.number().optional(),
 	awsBedrockEndpointEnabled: z.boolean().optional(),
 	awsBedrockEndpoint: z.string().optional(),
 	awsBedrock1MContext: z.boolean().optional(), // Enable 'context-1m-2025-08-07' beta for 1M context window.

packages/types/src/providers/bedrock.ts

@@ -14,7 +14,8 @@ export const bedrockDefaultPromptRouterModelId: BedrockModelId = "anthropic.clau
 // feature.
 export const bedrockModels = {
 	"anthropic.claude-sonnet-4-5-20250929-v1:0": {
-		maxTokens: 8192,
+		// Mirrors anthropic-direct cap; AWS Bedrock accepts the same upstream maximum.
+		maxTokens: 64_000,
 		contextWindow: 200_000,
 		supportsImages: true,
 		supportsPromptCache: true,
@@ -28,7 +29,8 @@ export const bedrockModels = {
 		cachableFields: ["system", "messages", "tools"],
 	},
 	"anthropic.claude-sonnet-4-6": {
-		maxTokens: 8192,
+		// Mirrors anthropic-direct cap; AWS Bedrock accepts the same upstream maximum.
+		maxTokens: 64_000,
 		contextWindow: 200_000, // Default 200K, extendable to 1M with beta flag 'context-1m-2025-08-07'
 		supportsImages: true,
 		supportsPromptCache: true,
@@ -116,7 +118,8 @@ export const bedrockModels = {
 		cachableFields: ["system"],
 	},
 	"anthropic.claude-sonnet-4-20250514-v1:0": {
-		maxTokens: 8192,
+		// Mirrors anthropic-direct cap; AWS Bedrock accepts the same upstream maximum.
+		maxTokens: 64_000,
 		contextWindow: 200_000,
 		supportsImages: true,
 		supportsPromptCache: true,
@@ -130,7 +133,8 @@ export const bedrockModels = {
 		cachableFields: ["system", "messages", "tools"],
 	},
 	"anthropic.claude-opus-4-1-20250805-v1:0": {
-		maxTokens: 8192,
+		// Mirrors anthropic-direct cap; AWS Bedrock accepts the same upstream maximum.
+		maxTokens: 32_000,
 		contextWindow: 200_000,
 		supportsImages: true,
 		supportsPromptCache: true,
@@ -144,7 +148,8 @@ export const bedrockModels = {
 		cachableFields: ["system", "messages", "tools"],
 	},
 	"anthropic.claude-opus-4-6-v1": {
-		maxTokens: 8192,
+		// Mirrors anthropic-direct cap; AWS Bedrock accepts the same upstream maximum.
+		maxTokens: 128_000,
 		contextWindow: 200_000, // Default 200K, extendable to 1M with beta flag 'context-1m-2025-08-07'
 		supportsImages: true,
 		supportsPromptCache: true,
@@ -168,7 +173,11 @@ export const bedrockModels = {
 		],
 	},
 	"anthropic.claude-opus-4-7": {
-		maxTokens: 8192,
+		// Opus 4.7 ships with a 128K-token max output on Bedrock
+		// (https://builder.aws.com/content/3Cl90CMMnqzCrkk6mXcmnGo1WTG/claude-opus-47-on-amazon-bedrock-apis-features-and-migration-guide).
+		// We default to that ceiling here so the reasoning-budget slider isn't artificially
+		// clamped to the legacy 8K floor used by older Anthropic-on-Bedrock entries.
+		maxTokens: 128_000,
 		// Opus 4.7 natively supports 1M context (no beta flag required) with FLAT $5/$25
 		// pricing at any context length. We still keep a tier entry so the dropdown can
 		// show a "128K" vs "1M" choice - the tier just toggles the context window the UI
@@ -198,7 +207,8 @@ export const bedrockModels = {
 		],
 	},
 	"anthropic.claude-opus-4-5-20251101-v1:0": {
-		maxTokens: 8192,
+		// Mirrors anthropic-direct cap; AWS Bedrock accepts the same upstream maximum.
+		maxTokens: 32_000,
 		contextWindow: 200_000,
 		supportsImages: true,
 		supportsPromptCache: true,
@@ -266,7 +276,8 @@ export const bedrockModels = {
 		cachableFields: ["system", "messages", "tools"],
 	},
 	"anthropic.claude-haiku-4-5-20251001-v1:0": {
-		maxTokens: 8192,
+		// Mirrors anthropic-direct cap; AWS Bedrock accepts the same upstream maximum.
+		maxTokens: 64_000,
 		contextWindow: 200_000,
 		supportsImages: true,
 		supportsPromptCache: true,
@@ -870,12 +881,18 @@ export const resolveBedrockModelInfo = ({
 	optIn1MContext,
 	modelMaxTokens,
 	contextWindowOverride,
+	maxOutputTokensOverride,
 }: {
 	baseModelId?: string
 	targetId?: string
 	optIn1MContext?: boolean
+	// Request-time "how many tokens to ask for" knob (slider value). Mirrors the historic behaviour.
 	modelMaxTokens?: number
 	contextWindowOverride?: number
+	// Static cap override (e.g. empirically detected by the AWS probe). When set, this widens the
+	// effective `info.maxTokens` ceiling that downstream UI and request builders see, even if the
+	// user has not explicitly bumped the slider.
+	maxOutputTokensOverride?: number
 }): { baseModelId: string; info: ModelInfo; uses1MContext: boolean; contextSource: BedrockContextSource } => {
 	const resolvedBaseModelId = parseBedrockBaseModelId(baseModelId || targetId || bedrockDefaultModelId)
 
@@ -906,6 +923,12 @@ export const resolveBedrockModelInfo = ({
 		}
 	}
 
+	// Apply the static-cap override BEFORE the request-time `modelMaxTokens` so users can
+	// explicitly request fewer tokens than the model's headroom (e.g. cost control) without
+	// having the override silently clobber their slider value.
+	if (maxOutputTokensOverride && maxOutputTokensOverride > 0) {
+		info.maxTokens = maxOutputTokensOverride
+	}
 	if (modelMaxTokens && modelMaxTokens > 0) {
 		info.maxTokens = modelMaxTokens
 	}

packages/types/src/vscode-extension-host.ts

@@ -44,6 +44,7 @@ export interface ExtensionMessage {
 		| "listApiConfig"
 		| "routerModels"
 		| "bedrockDiscovery"
+		| "bedrockMaxTokensProbe"
 		| "openAiModels"
 		| "ollamaModels"
 		| "lmStudioModels"
@@ -138,6 +139,13 @@ export interface ExtensionMessage {
 	clineMessage?: ClineMessage
 	routerModels?: RouterModels
 	bedrockDiscovery?: BedrockDiscoveredTarget[]
+	/** Empirically detected cap returned by `requestBedrockMaxTokensProbe`. */
+	bedrockMaxTokensProbe?: {
+		maxOutputTokens: number
+		source: "accepted" | "hint" | "binary-search"
+		attempts: number
+		modelId: string
+	}
 	openAiModels?: string[]
 	ollamaModels?: ModelRecord
 	lmStudioModels?: ModelRecord
@@ -451,6 +459,7 @@ export interface WebviewMessage {
 		| "flushRouterModels"
 		| "requestRouterModels"
 		| "requestBedrockDiscovery"
+		| "requestBedrockMaxTokensProbe"
 		| "requestOpenAiModels"
 		| "requestOllamaModels"
 		| "requestLmStudioModels"

src/api/providers/__tests__/bedrock-max-tokens-probe.spec.ts

@@ -0,0 +1,114 @@
+// npx vitest src/api/providers/__tests__/bedrock-max-tokens-probe.spec.ts
+
+import { describe, expect, it, vi } from "vitest"
+
+import { probeBedrockMaxOutputTokens, BEDROCK_MAX_OUTPUT_PROBE_CEILING } from "../bedrock-discovery"
+
+const baseOptions = {
+	awsRegion: "us-west-2",
+	awsAccessKey: "AKIA",
+	awsSecretKey: "secret",
+} as const
+
+const buildValidationError = (message: string) => {
+	const err = new Error(message)
+	err.name = "ValidationException"
+	;(err as any).$metadata = { httpStatusCode: 400 }
+	return err
+}
+
+describe("probeBedrockMaxOutputTokens", () => {
+	it("returns the ceiling when AWS accepts the first probe", async () => {
+		const runProbe = vi.fn().mockResolvedValue(undefined)
+
+		const result = await probeBedrockMaxOutputTokens({
+			options: baseOptions,
+			modelId: "anthropic.claude-opus-4-7",
+			probeCeiling: 256_000,
+			runProbe,
+		})
+
+		expect(result).toEqual({
+			maxOutputTokens: 256_000,
+			source: "accepted",
+			attempts: 1,
+		})
+		expect(runProbe).toHaveBeenCalledTimes(1)
+		expect(runProbe).toHaveBeenCalledWith(256_000)
+	})
+
+	it("recovers the cap from a parsable AWS error message hint", async () => {
+		const runProbe = vi
+			.fn()
+			.mockRejectedValueOnce(
+				buildValidationError("max_tokens: 256000 must be less than or equal to 128000 for this model"),
+			)
+			.mockResolvedValueOnce(undefined)
+
+		const result = await probeBedrockMaxOutputTokens({
+			options: baseOptions,
+			modelId: "anthropic.claude-opus-4-7",
+			probeCeiling: 256_000,
+			runProbe,
+		})
+
+		expect(result.maxOutputTokens).toBe(128_000)
+		expect(result.source).toBe("hint")
+		expect(result.attempts).toBe(2)
+		expect(runProbe).toHaveBeenNthCalledWith(2, 128_000)
+	})
+
+	it("falls back to binary search when no hint is present", async () => {
+		// Simulate a model that caps at 65_536. AWS rejects anything above with an opaque message.
+		const cap = 65_536
+		const runProbe = vi.fn(async (maxTokens: number) => {
+			if (maxTokens > cap) {
+				throw buildValidationError("max_tokens validation failed: invalid request")
+			}
+			return undefined
+		})
+
+		const result = await probeBedrockMaxOutputTokens({
+			options: baseOptions,
+			modelId: "anthropic.claude-haiku-4-5-20251001-v1:0",
+			probeCeiling: 200_000,
+			runProbe,
+		})
+
+		// Binary search should converge to the exact cap (within 1 token).
+		expect(result.source).toBe("binary-search")
+		expect(result.maxOutputTokens).toBeLessThanOrEqual(cap)
+		expect(result.maxOutputTokens).toBeGreaterThan(cap - 4) // tight bound
+		expect(result.attempts).toBeGreaterThan(2)
+	})
+
+	it("propagates non-validation errors immediately", async () => {
+		const networkError = Object.assign(new Error("connect ETIMEDOUT"), { name: "TimeoutError" })
+		const runProbe = vi.fn().mockRejectedValue(networkError)
+
+		await expect(
+			probeBedrockMaxOutputTokens({
+				options: baseOptions,
+				modelId: "anthropic.claude-opus-4-7",
+				probeCeiling: 256_000,
+				runProbe,
+			}),
+		).rejects.toMatchObject({ message: "connect ETIMEDOUT" })
+		// We never recover from non-validation errors, so only one probe should have been issued.
+		expect(runProbe).toHaveBeenCalledTimes(1)
+	})
+
+	it("rejects when AWS region is missing", async () => {
+		await expect(
+			probeBedrockMaxOutputTokens({
+				options: { ...baseOptions, awsRegion: undefined },
+				modelId: "anthropic.claude-opus-4-7",
+				runProbe: vi.fn(),
+			}),
+		).rejects.toThrow(/region/i)
+	})
+
+	it("respects the documented probe ceiling default", () => {
+		expect(BEDROCK_MAX_OUTPUT_PROBE_CEILING).toBe(1_000_000)
+	})
+})