fix(ai): guard Sonnet 5.5 Bedrock Converse requests

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

@@ -6,6 +6,7 @@ import {
   AIError,
   LLMEvent,
   Usage,
+  mergeJsonRecords,
   type CacheHint,
   type FinishReason,
   type FinishReasonDetails,
@@ -23,7 +24,7 @@ import { JsonObject, optionalArray, ProviderShared } from "./shared.js"
 import { BedrockAuth } from "./utils/bedrock-auth.js"
 import { BedrockCache } from "./utils/bedrock-cache.js"
 import { BedrockMedia } from "./utils/bedrock-media.js"
-import { supportsThinkingBlockBinding, THINKING_BINDING_BETA } from "./utils/claude-model.js"
+import { claudeVersion, supportsThinkingBlockBinding, THINKING_BINDING_BETA } from "./utils/claude-model.js"
 import { Lifecycle } from "./utils/lifecycle.js"
 import { MistralToolID } from "./utils/mistral-tool-id.js"
 import { ToolStream } from "./utils/tool-stream.js"
@@ -428,6 +429,10 @@ const lowerSystem = (breakpoints: BedrockCache.Breakpoints, system: ReadonlyArra
 // Nova 2 rejects `maxTokens` at high reasoning effort, where its output can exceed the field's maximum. Other models
 // that take `reasoningConfig`, such as Grok on Bedrock, accept it.
 const isNova2 = (model: LanguageModel) => /\bamazon\.nova-2-/.test(model.id)
+const isSonnet55 = (model: LanguageModel) => {
+  const version = claudeVersion(model.id)
+  return version?.family === "sonnet" && version.major === 5 && version.minor === 5
+}
 const isHighReasoningEffort = Schema.is(
   Schema.Struct({
     additionalModelRequestFields: Schema.Struct({
@@ -437,23 +442,31 @@ const isHighReasoningEffort = Schema.is(
 )
 
 const Options = Schema.Struct({
-  thinking: Schema.optional(Schema.Struct({ type: Schema.Literal("enabled"), budgetTokens: Schema.Number })),
+  thinking: Schema.optional(
+    Schema.Union([
+      Schema.Struct({ type: Schema.Literal("enabled"), budgetTokens: Schema.Number }),
+      Schema.Struct({ type: Schema.Literal("between_tools") }),
+    ]),
+  ),
 })
 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 isThinkingDisabled = Schema.is(
+const isThinkingWithoutBinding = Schema.is(
   Schema.Struct({
-    additionalModelRequestFields: Schema.Struct({ thinking: Schema.Struct({ type: Schema.Literal("disabled") }) }),
+    additionalModelRequestFields: Schema.Struct({
+      thinking: Schema.Struct({ type: Schema.Literals(["disabled", "between_tools"]) }),
+    }),
   }),
 )
 
 // Claude 5.1+ binds each thinking signature to the prefix above it. Ask Bedrock to drop the affected blocks instead of
-// failing when that prefix changes. `http.body` overlays this field by field, so callers can still override it.
+// failing when that prefix changes. `http.body` overlays this field by field; between_tools cannot carry binding.
 const applyThinkingBindingDefault = (request: LLMRequest, thinking: Readonly<Record<string, unknown>> | undefined) => {
-  if (isThinkingDisabled(request.http?.body)) return thinking
+  if (thinking?.type === "between_tools") return thinking
+  if (isThinkingWithoutBinding(request.http?.body)) return thinking
   if (!supportsThinkingBlockBinding(request.model)) return thinking
   return {
     ...(thinking ?? { type: "adaptive" as const }),
@@ -466,20 +479,24 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
   const flattened = ProviderShared.flattenToolRequest(request)
   const generation = request.generation
   const options = yield* decodeOptions(request.providerOptions ?? {})
+  if (options.thinking?.type === "between_tools" && !isSonnet55(request.model))
+    return yield* ProviderShared.invalidRequest("Bedrock Converse between_tools thinking requires Claude Sonnet 5.5")
   const maxTokens =
     isNova2(request.model) && isHighReasoningEffort(request.http?.body) ? undefined : generation?.maxTokens
   const thinking = applyThinkingBindingDefault(
     request,
     options.thinking === undefined
       ? undefined
-      : {
-          type: "enabled",
-          budget_tokens: ProviderShared.fitThinkingBudget(
-            options.thinking.budgetTokens,
-            maxTokens,
-            MIN_THINKING_BUDGET,
-          ),
-        },
+      : options.thinking.type === "between_tools"
+        ? { type: "between_tools" }
+        : {
+            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.
@@ -515,7 +532,7 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
       stopSequences: generation?.stop,
     }
   })()
-  return {
+  const body = {
     modelId: request.model.id,
     messages,
     system,
@@ -533,6 +550,69 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
             ...(thinking?.block_binding === undefined ? {} : { anthropic_beta: [THINKING_BINDING_BETA] }),
           },
   }
+  if (isSonnet55(request.model)) yield* validateSonnet55Body(request, body)
+  return body
+})
+
+const validateSonnet55Body = Effect.fn("BedrockConverse.validateSonnet55Body")(function* (
+  request: LLMRequest,
+  body: Record<string, unknown>,
+) {
+  // The HTTP body overlay is merged after protocol lowering; check the request that will actually be sent.
+  const effective = mergeJsonRecords(body, request.http?.body)
+  const config = effective?.toolConfig
+  const choice = ProviderShared.isRecord(config) ? config.toolChoice : undefined
+  if (ProviderShared.isRecord(choice) && ("any" in choice || "tool" in choice))
+    return yield* ProviderShared.invalidRequest(
+      "Bedrock Converse Claude Sonnet 5.5 does not support forced tool choice; use auto instead",
+    )
+  const additional = effective?.additionalModelRequestFields
+  const thinking = ProviderShared.isRecord(additional) ? additional.thinking : undefined
+  if (ProviderShared.isRecord(thinking) && thinking.type === "enabled")
+    return yield* ProviderShared.invalidRequest(
+      "Bedrock Converse Claude Sonnet 5.5 does not support manual thinking budgets; use adaptive thinking instead",
+    )
+  if (ProviderShared.isRecord(thinking) && thinking.type === "disabled")
+    return yield* ProviderShared.invalidRequest(
+      "Bedrock Converse Claude Sonnet 5.5 does not support disabled thinking; use between_tools instead",
+    )
+  if (
+    ProviderShared.isRecord(thinking) &&
+    thinking.type === "between_tools" &&
+    ("budget_tokens" in thinking || "block_binding" in thinking || "display" in thinking)
+  )
+    return yield* ProviderShared.invalidRequest(
+      "Bedrock Converse Claude Sonnet 5.5 between_tools does not accept a budget, block binding, or display",
+    )
+  const output = ProviderShared.isRecord(additional) ? additional.output_config : undefined
+  if (
+    ProviderShared.isRecord(thinking) &&
+    thinking.type === "between_tools" &&
+    ProviderShared.isRecord(output) &&
+    (output.effort === "xhigh" || output.effort === "max")
+  )
+    return yield* ProviderShared.invalidRequest(
+      "Bedrock Converse Claude Sonnet 5.5 between_tools does not support xhigh or max effort; use adaptive thinking",
+    )
+  const inference = effective?.inferenceConfig
+  if (ProviderShared.isRecord(inference) && inference.temperature !== undefined && inference.temperature !== 1)
+    return yield* ProviderShared.invalidRequest(
+      "Bedrock Converse Claude Sonnet 5.5 does not support non-default temperature; omit temperature instead",
+    )
+  // AWS's Claude parameter reference documents 0.999 for top_p, while Anthropic's API uses 1.
+  if (
+    ProviderShared.isRecord(inference) &&
+    inference.topP !== undefined &&
+    inference.topP !== 1 &&
+    inference.topP !== 0.999
+  )
+    return yield* ProviderShared.invalidRequest(
+      "Bedrock Converse Claude Sonnet 5.5 does not support non-default topP; omit topP instead",
+    )
+  if (ProviderShared.isRecord(additional) && additional.top_k !== undefined && additional.top_k !== 0)
+    return yield* ProviderShared.invalidRequest(
+      "Bedrock Converse Claude Sonnet 5.5 does not support non-default top_k; omit topK instead",
+    )
 })
 
 // =============================================================================

packages/ai/test/provider/bedrock-converse-sonnet55.test.ts

@@ -0,0 +1,223 @@
+import { EventStreamCodec } from "@smithy/eventstream-codec"
+import { fromUtf8, toUtf8 } from "@smithy/util-utf8"
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { HttpClientRequest } from "effect/unstable/http"
+import { LLM, LLMRequest, ToolChoice, ToolDefinition } from "../../src/index.js"
+import { AmazonBedrock } from "../../src/providers.js"
+import { LLMClient } from "../../src/route.js"
+import { compileRequest } from "../../src/route/client.js"
+import { it } from "../lib/effect.js"
+import { dynamicResponse } from "../lib/http.js"
+
+const bedrock = (id: string, settings: Parameters<typeof AmazonBedrock.model>[1] = {}) =>
+  AmazonBedrock.model(id, { baseURL: "https://bedrock-runtime.test", apiKey: "test-bearer", ...settings })
+
+const request = LLM.request({
+  model: bedrock("global.anthropic.claude-sonnet-5-5"),
+  prompt: "What is the weather?",
+  cache: "none",
+  tools: [ToolDefinition.make({ name: "lookup", description: "Weather lookup", inputSchema: { type: "object" } })],
+})
+
+describe("Bedrock Converse Claude Sonnet 5.5", () => {
+  it.effect("rejects required tool choice rather than silently allowing a different tool policy", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLMRequest.update(request, { toolChoice: ToolChoice.make({ type: "required" }) }),
+      ).pipe(Effect.flip)
+
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toContain("tool choice")
+    }),
+  )
+
+  it.effect("rejects named tool choice supplied by a body overlay", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLMRequest.update(request, { http: { body: { toolConfig: { toolChoice: { tool: { name: "lookup" } } } } } }),
+      ).pipe(Effect.flip)
+
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toContain("tool choice")
+    }),
+  )
+
+  it.effect("rejects manual thinking budgets instead of silently dropping the budget", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLMRequest.update(request, {
+          model: bedrock("global.anthropic.claude-sonnet-5-5", {
+            thinking: { type: "enabled", budgetTokens: 4_000 },
+          }),
+        }),
+      ).pipe(Effect.flip)
+
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toContain("manual thinking")
+    }),
+  )
+
+  it.effect("rejects explicitly disabled thinking with an actionable alternative", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLMRequest.update(request, {
+          http: { body: { additionalModelRequestFields: { thinking: { type: "disabled" } } } },
+        }),
+      ).pipe(Effect.flip)
+
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toContain("between_tools")
+    }),
+  )
+
+  it.effect("rejects non-default temperature without discarding the caller's sampling preference", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(LLMRequest.update(request, { generation: { temperature: 0 } })).pipe(
+        Effect.flip,
+      )
+
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toContain("temperature")
+    }),
+  )
+
+  it.effect("rejects non-default topP configured through the Bedrock provider", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLMRequest.update(request, { model: bedrock("global.anthropic.claude-sonnet-5-5", { topP: 0.8 }) }),
+      ).pipe(Effect.flip)
+
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toContain("topP")
+    }),
+  )
+
+  it.effect("rejects non-default topK in additional model request fields", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(LLMRequest.update(request, { generation: { topK: 40 } })).pipe(Effect.flip)
+
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toContain("top_k")
+    }),
+  )
+
+  it.effect("does not add adaptive thinking binding to an explicit between_tools request", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLMRequest.update(request, {
+          http: { body: { additionalModelRequestFields: { thinking: { type: "between_tools" } } } },
+        }),
+      )
+
+      expect(prepared.body.additionalModelRequestFields).toBeUndefined()
+    }),
+  )
+
+  it.effect("accepts between_tools as a typed Bedrock model setting", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLMRequest.update(request, {
+          model: bedrock("global.anthropic.claude-sonnet-5-5", { thinking: { type: "between_tools" } }),
+        }),
+      )
+
+      expect(prepared.body.additionalModelRequestFields).toEqual({ thinking: { type: "between_tools" } })
+    }),
+  )
+
+  it.effect("rejects a between_tools overlay that retains a manual budget", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLMRequest.update(request, {
+          model: bedrock("global.anthropic.claude-sonnet-5-5", {
+            thinking: { type: "enabled", budgetTokens: 4_000 },
+          }),
+          http: { body: { additionalModelRequestFields: { thinking: { type: "between_tools" } } } },
+        }),
+      ).pipe(Effect.flip)
+
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toContain("between_tools")
+    }),
+  )
+
+  it.effect("rejects between_tools at unsupported max effort", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLMRequest.update(request, {
+          model: bedrock("global.anthropic.claude-sonnet-5-5", { thinking: { type: "between_tools" } }),
+          http: { body: { additionalModelRequestFields: { output_config: { effort: "max" } } } },
+        }),
+      ).pipe(Effect.flip)
+
+      expect(error.reason._tag).toBe("InvalidRequest")
+      expect(error.message).toContain("max")
+    }),
+  )
+
+  it.effect("sends between_tools without block binding and preserves auto tool choice on the wire", () =>
+    Effect.gen(function* () {
+      const bodies: Array<Record<string, unknown>> = []
+      const response = new EventStreamCodec(toUtf8, fromUtf8).encode({
+        headers: {
+          ":message-type": { type: "string", value: "event" },
+          ":event-type": { type: "string", value: "messageStop" },
+          ":content-type": { type: "string", value: "application/json" },
+        },
+        body: new TextEncoder().encode('{"stopReason":"end_turn"}'),
+      })
+      yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          toolChoice: ToolChoice.make({ type: "auto" }),
+          http: { body: { additionalModelRequestFields: { thinking: { type: "between_tools" } } } },
+        }),
+      ).pipe(
+        Effect.provide(
+          dynamicResponse((input) =>
+            Effect.gen(function* () {
+              const sent = yield* HttpClientRequest.toWeb(input.request)
+              bodies.push(JSON.parse(yield* Effect.promise(() => sent.text())))
+              return input.respond(response, { headers: { "content-type": "application/vnd.amazon.eventstream" } })
+            }),
+          ),
+        ),
+      )
+
+      expect(bodies[0]?.additionalModelRequestFields).toEqual({ thinking: { type: "between_tools" } })
+      expect(bodies[0]?.toolConfig).toMatchObject({ toolChoice: { auto: {} } })
+    }),
+  )
+
+  it.effect("keeps Sonnet 5.5 defaults and supported sampling defaults", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLMRequest.update(request, {
+          toolChoice: ToolChoice.make({ type: "none" }),
+          generation: { temperature: 1, topP: 0.999, topK: 0 },
+        }),
+      )
+
+      expect(prepared.body.toolConfig?.toolChoice).toBeUndefined()
+      expect(prepared.body.inferenceConfig).toMatchObject({ temperature: 1, topP: 0.999 })
+      expect(prepared.body.additionalModelRequestFields).toMatchObject({ top_k: 0, thinking: { type: "adaptive" } })
+    }),
+  )
+
+  it.effect("keeps earlier Claude and other Bedrock models' request behavior", () =>
+    Effect.gen(function* () {
+      for (const id of ["us.anthropic.claude-sonnet-4-6", "us.amazon.nova-2-lite-v1:0"]) {
+        const prepared = yield* compileRequest(
+          LLMRequest.update(request, {
+            model: bedrock(id),
+            toolChoice: ToolChoice.make({ type: "required" }),
+            generation: { temperature: 0 },
+          }),
+        )
+
+        expect(prepared.body.toolConfig?.toolChoice).toEqual({ any: {} })
+        expect(prepared.body.inferenceConfig).toEqual({ temperature: 0 })
+      }
+    }),
+  )
+})