fix(ai): fit Claude thinking budgets on Bedrock Converse (#51190)

packages/ai/src/protocols/bedrock-converse.ts

@@ -441,12 +441,32 @@ const isHighReasoningEffort = Schema.is(
   }),
 )
 
+const Options = Schema.Struct({
+  thinking: Schema.optional(Schema.Struct({ type: Schema.Literal("enabled"), budgetTokens: Schema.Number })),
+})
+export type OptionsInput = typeof Options.Type
+const decodeOptions = ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))
+// Claude on Bedrock requires the thinking budget below `maxTokens`, with a minimum of 1,024.
+const MIN_THINKING_BUDGET = 1_024
+
 const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request: LLMRequest) {
   const toolChoice = request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined
   const flattened = ProviderShared.flattenToolRequest(request)
   const generation = request.generation
+  const options = yield* decodeOptions(request.providerOptions ?? {})
   const maxTokens =
     isNova2(request.model) && isHighReasoningEffort(request.http?.body) ? undefined : generation?.maxTokens
+  const thinking =
+    options.thinking === undefined
+      ? undefined
+      : {
+          type: "enabled",
+          budget_tokens: ProviderShared.fitThinkingBudget(
+            options.thinking.budgetTokens,
+            maxTokens,
+            MIN_THINKING_BUDGET,
+          ),
+        }
   // Bedrock-Claude shares Anthropic's 4-breakpoint cap. Spend the budget in
   // tools → system → messages order to favour the highest-impact prefixes.
   const breakpoints = BedrockCache.breakpoints(request.model.id)
@@ -487,9 +507,15 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
     system,
     inferenceConfig,
     toolConfig,
-    // Converse's base inferenceConfig has no topK; Anthropic/Nova accept it
-    // as a model-specific field, so it goes through additionalModelRequestFields.
-    additionalModelRequestFields: generation?.topK === undefined ? undefined : { top_k: generation.topK },
+    // Converse's base inferenceConfig has no topK or thinking; Anthropic/Nova accept them
+    // as model-specific fields, so they go through additionalModelRequestFields.
+    additionalModelRequestFields:
+      generation?.topK === undefined && thinking === undefined
+        ? undefined
+        : {
+            ...(generation?.topK === undefined ? {} : { top_k: generation.topK }),
+            ...(thinking === undefined ? {} : { thinking }),
+          },
   }
 })
 

packages/ai/src/providers/amazon-bedrock.ts

@@ -31,6 +31,7 @@ export interface Settings extends ProviderPackage.Settings {
   readonly profile?: string
   readonly region?: string
   readonly topP?: number
+  readonly thinking?: BedrockConverse.OptionsInput["thinking"]
 }
 export const routes = [BedrockConverse.route]
 
@@ -71,6 +72,7 @@ export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, se
     generation: settings.topP === undefined ? undefined : { topP: settings.topP },
     headers: settings.headers === undefined ? undefined : { ...settings.headers },
     http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.thinking === undefined ? undefined : { thinking: settings.thinking },
     profile: settings.profile,
     region: settings.region,
   }).model(modelID)

packages/ai/test/provider/bedrock-converse.test.ts

@@ -244,6 +244,29 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
+  it.effect("fits a Claude thinking budget below maxTokens", () =>
+    Effect.gen(function* () {
+      const fields = (maxTokens: number, budgetTokens: number, topK?: number) =>
+        compileRequest(
+          LLMRequest.update(baseRequest, {
+            model: AmazonBedrock.model("us.anthropic.claude-haiku-4-5-20251001-v1:0", {
+              baseURL: "https://bedrock-runtime.test",
+              apiKey: "test-bearer",
+              thinking: { type: "enabled", budgetTokens },
+            }),
+            generation: GenerationOptions.make({ maxTokens, topK }),
+          }),
+        ).pipe(Effect.map((prepared) => prepared.body.additionalModelRequestFields))
+
+      expect(yield* fields(64_000, 31_999)).toEqual({ thinking: { type: "enabled", budget_tokens: 31_999 } })
+      expect(yield* fields(20_000, 31_999, 40)).toEqual({
+        top_k: 40,
+        thinking: { type: "enabled", budget_tokens: 10_000 },
+      })
+      expect(yield* fields(1_500, 31_999)).toEqual({ thinking: { type: "enabled", budget_tokens: 1_024 } })
+    }),
+  )
+
   it.effect("omits additionalModelRequestFields when topK is unset", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(baseRequest)

packages/core/src/aisdk-native.ts

@@ -151,8 +151,12 @@ type Overlay = {
 function options(replacement: string, modelID: string | undefined, settings: Legacy): Overlay {
   const converse = replacement === "@opencode/ai/providers/amazon-bedrock" && modelID !== undefined
   const kept = Struct.omit(settings, ["headers", "extraBody", "useCompletionUrls", ...OPENROUTER_KEYS])
+  const thinking = converse ? bedrockThinking(modelID, settings) : undefined
   return {
-    settings: replacement.startsWith("@opencode/ai/providers/amazon-bedrock") ? bedrockSettings(kept, converse) : kept,
+    settings: {
+      ...(replacement.startsWith("@opencode/ai/providers/amazon-bedrock") ? bedrockSettings(kept, converse) : kept),
+      ...(thinking === undefined ? {} : { thinking }),
+    },
     ...(settings.headers === undefined ? {} : { headers: settings.headers }),
     ...(settings.extraBody === undefined ? {} : { body: settings.extraBody }),
     ...(converse ? bedrockRequest(modelID, settings) : {}),
@@ -196,6 +200,13 @@ function bedrockSettings(settings: Legacy, converse: boolean) {
   }
 }
 
+// Claude's enabled budget is a typed setting so the protocol can fit it under the output limit.
+function bedrockThinking(modelID: string | undefined, settings: Legacy) {
+  const reasoning = settings.reasoningConfig
+  if (!modelID?.includes("anthropic") || reasoning?.type !== "enabled" || reasoning.budgetTokens === undefined) return
+  return { type: "enabled", budgetTokens: reasoning.budgetTokens }
+}
+
 function bedrockRequest(modelID: string | undefined, settings: Legacy): Pick<Overlay, "body"> {
   const additional = settings.additionalModelRequestFields ?? {}
   const reasoning = settings.reasoningConfig
@@ -210,9 +221,6 @@ function bedrockRequest(modelID: string | undefined, settings: Legacy): Pick<Ove
   const betas = settings.anthropicBeta ?? []
   const fields = Provider.mergeOverlay(additional, {
     ...(betas.length > 0 ? { anthropic_beta: [...(additional.anthropic_beta ?? []), ...betas] } : {}),
-    ...(anthropic && type === "enabled" && budget !== undefined
-      ? { thinking: { type: "enabled", budget_tokens: budget } }
-      : {}),
     ...(anthropic && type === "adaptive"
       ? { thinking: { type: "adaptive", ...(display === undefined ? {} : { display }) } }
       : {}),

packages/core/src/variant.ts

@@ -398,8 +398,9 @@ const bedrockConverse: Protocol = (model, support) => {
         model,
         support,
         (tokens) =>
+          // Claude's budget is a typed setting so the protocol can fit it under the output limit.
           claude
-            ? fields({ thinking: { type: "enabled", budget_tokens: tokens } })
+            ? { settings: { thinking: { type: "enabled", budgetTokens: tokens } } }
             : fields({ reasoningConfig: { type: "enabled", budgetTokens: tokens } }),
         claude ? ANTHROPIC_OUTPUT_TOKEN_MAX : model.limit.output,
       )

packages/core/test/aisdk-native.test.ts

@@ -247,6 +247,17 @@ describe("AISDKNative", () => {
       },
     })
 
+    expect(
+      map(
+        "@ai-sdk/amazon-bedrock",
+        { reasoningConfig: { type: "enabled", budgetTokens: 12_000 } },
+        "anthropic.claude-sonnet-4-5-20250929-v1:0",
+      ),
+    ).toEqual({
+      package: "@opencode/ai/providers/amazon-bedrock",
+      settings: { thinking: { type: "enabled", budgetTokens: 12_000 } },
+    })
+
     // gpt-oss (Harmony) keeps the flat chat-completions field.
     expect(
       map("@ai-sdk/amazon-bedrock", { reasoningConfig: { maxReasoningEffort: "high" } }, "openai.gpt-oss-120b-1:0")

packages/core/test/variant.test.ts

@@ -196,6 +196,17 @@ test("spells Chat Completions variants for direct providers", () => {
   ])
 })
 
+test("spells Bedrock Converse Claude budgets as a thinking setting", () => {
+  expect(
+    resolve(model("@opencode/ai/providers/amazon-bedrock", "us.anthropic.claude-haiku-4-5-20251001-v1:0", 64_000), [
+      { type: "budget_tokens", min: 1024 },
+    ]),
+  ).toEqual([
+    { id: "high", settings: { thinking: { type: "enabled", budgetTokens: 16_000 } } },
+    { id: "max", settings: { thinking: { type: "enabled", budgetTokens: 31_999 } } },
+  ])
+})
+
 test("spells Bedrock Converse effort for Grok and Nova", () => {
   const supports: Variant.Support[] = [{ type: "effort", values: ["low", "xhigh"] }]
   expect(resolve(model("@opencode/ai/providers/amazon-bedrock", "us.xai.grok-4.6"), supports)).toEqual([