@warlock.js/ai-bedrock 4.2.10 → 4.3.0

package/CHANGELOG.md

@@ -6,6 +6,15 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+## 4.3.0 - 2026-06-21
+
+### Added
+
+- **Cost-truth contract wiring.** Capabilities now report `reasoning`, `promptCaching`, `pdf`, and `audio` truthfully per model family (inferred from the model id, overridable via `bedrock.model({ name, reasoning?, promptCaching?, pdf?, audio? })`): `reasoning` for Claude 3.7 + Claude 4, `promptCaching` for Claude 3.5+/3.7/4 and Nova, `pdf` for Claude 3+ and Nova, `audio` defaulting to `false`.
+- **Reasoning / extended thinking.** `ModelCallOptions.reasoning` maps to Converse `additionalModelRequestFields.thinking = { type: "enabled", budget_tokens }` for reasoning-capable models — `reasoning.maxTokens` is the explicit budget, `reasoning.effort` (low/medium/high) maps to a conventional budget. No-ops for non-reasoning models so unsupported params never reach the wire.
+- **Prompt-cache write breakpoints.** `ModelCallOptions.cacheControl.breakpoints` appends a Converse `cachePoint` block to the last message for `promptCaching`-capable models.
+- **`Usage.cacheWriteTokens`** populated from Converse `usage.cacheWriteInputTokens` on both `complete()` and `stream()` (alongside the existing `cachedTokens` from `cacheReadInputTokens`). `Usage.reasoningTokens` is intentionally left unset — Bedrock's Converse `TokenUsage` reports no reasoning-token channel.
+
 ## 4.1.15
 
 - Baseline — per-package changelog tracking starts at this version.

package/cjs/index.cjs

@@ -402,6 +402,112 @@ var BedrockEmbedder = class {
 	}
 };
 
+//#endregion
+//#region ../@warlock.js/ai-bedrock/src/known-capabilities.ts
+/**
+* Cost-truth capability inference for Bedrock Converse model ids.
+*
+* Bedrock model ids are provider-prefixed and version-suffixed
+* (`anthropic.claude-3-7-sonnet-20250219-v1:0`, `us.amazon.nova-pro-v1:0`),
+* so — exactly like `known-vision-models.ts` — a lowercase substring scan
+* is the only robust check across cross-region inference-profile prefixes
+* (`us.`, `eu.`, `apac.`) and date/version tags.
+*
+* Each predicate answers a single `ModelCapabilities` flag the agent reads
+* to decide whether to forward a cost-truth option (`reasoning`,
+* `cacheControl`) or up-front-reject an attachment (`pdf`). Unknown ids
+* default to `false` so an unsupported request fails fast with a clear
+* capability error instead of an opaque Bedrock `ValidationException`.
+* Every inference is overridable per-model via `bedrock.model({ name, … })`.
+*/
+/**
+* Families that expose Anthropic-style extended thinking on Bedrock
+* Converse via `additionalModelRequestFields.thinking`. Only Claude 3.7
+* and the Claude 4 line (Sonnet / Opus / Haiku) support a configurable
+* thinking budget; earlier Claude, Nova, Llama, Mistral and Cohere do
+* not, so they are intentionally absent.
+*/
+const REASONING_CAPABLE_SUBSTRINGS = [
+	"claude-3-7",
+	"claude-sonnet-4",
+	"claude-opus-4",
+	"claude-haiku-4"
+];
+/**
+* Families that honor Converse `cachePoint` prompt-cache breakpoints.
+* Anthropic Claude 3.5+ / 3.7 / 4 and the Amazon Nova line support
+* cache points; text-only legacy families do not.
+*/
+const PROMPT_CACHING_CAPABLE_SUBSTRINGS = [
+	"claude-3-5",
+	"claude-3-7",
+	"claude-sonnet-4",
+	"claude-opus-4",
+	"claude-haiku-4",
+	"nova-lite",
+	"nova-pro",
+	"nova-premier",
+	"nova-micro"
+];
+/**
+* Families that accept Converse `document` content blocks (PDF / docx /
+* txt input). The multimodal Claude 3+ and Nova families support
+* document blocks; the substring set mirrors the vision-capable list
+* minus the image-only Llama entries (Llama on Bedrock takes images but
+* not document blocks via Converse).
+*/
+const PDF_CAPABLE_SUBSTRINGS = [
+	"claude-3",
+	"claude-sonnet-4",
+	"claude-opus-4",
+	"claude-haiku-4",
+	"nova-lite",
+	"nova-pro",
+	"nova-premier"
+];
+function matchesAny(modelId, fragments) {
+	const normalized = modelId.toLowerCase();
+	return fragments.some((fragment) => normalized.includes(fragment));
+}
+/**
+* Infer whether a Bedrock model id exposes extended-thinking / reasoning
+* (Claude 3.7 + Claude 4). When true the adapter forwards
+* `ModelCallOptions.reasoning` as Converse
+* `additionalModelRequestFields.thinking`.
+*
+* @example
+* inferReasoningCapability("anthropic.claude-3-7-sonnet-20250219-v1:0"); // → true
+* inferReasoningCapability("us.amazon.nova-pro-v1:0");                    // → false
+*/
+function inferReasoningCapability(modelId) {
+	return matchesAny(modelId, REASONING_CAPABLE_SUBSTRINGS);
+}
+/**
+* Infer whether a Bedrock model id honors Converse `cachePoint`
+* breakpoints (Claude 3.5+ / Nova). When true the adapter both maps
+* `ModelCallOptions.cacheControl` write breakpoints to cache points and
+* reports `Usage.cachedTokens` / `Usage.cacheWriteTokens`.
+*
+* @example
+* inferPromptCachingCapability("us.amazon.nova-pro-v1:0");                       // → true
+* inferPromptCachingCapability("meta.llama3-1-8b-instruct-v1:0");                // → false
+*/
+function inferPromptCachingCapability(modelId) {
+	return matchesAny(modelId, PROMPT_CACHING_CAPABLE_SUBSTRINGS);
+}
+/**
+* Infer whether a Bedrock model id accepts Converse `document` content
+* blocks (PDF / document input — Claude 3+ / Nova). When false the agent
+* rejects a PDF attachment up front instead of dropping it at the wire.
+*
+* @example
+* inferPdfCapability("anthropic.claude-3-5-sonnet-20240620-v1:0"); // → true
+* inferPdfCapability("meta.llama3-2-90b-instruct-v1:0");           // → false
+*/
+function inferPdfCapability(modelId) {
+	return matchesAny(modelId, PDF_CAPABLE_SUBSTRINGS);
+}
+
 //#endregion
 //#region ../@warlock.js/ai-bedrock/src/known-vision-models.ts
 /**
@@ -454,6 +560,18 @@ function inferVisionCapability(modelId) {
 //#region ../@warlock.js/ai-bedrock/src/model.ts
 const LOG_MODULE = "ai.bedrock";
 /**
+* Conventional extended-thinking token budgets for the neutral
+* `reasoning.effort` levels, used when the caller asks for an effort
+* tier without naming an explicit `reasoning.maxTokens` budget. Mirrors
+* the low / medium / high spread other reasoning adapters expose so the
+* vendor-neutral option behaves consistently across providers.
+*/
+const EFFORT_THINKING_BUDGET = {
+	low: 1024,
+	medium: 4096,
+	high: 16384
+};
+/**
 * Bedrock-backed implementation of `ModelContract`.
 *
 * **Role.** The provider-facing bridge between the vendor-neutral
@@ -498,7 +616,11 @@ var BedrockModel = class {
 		this.pricing = config.pricing;
 		this.capabilities = {
 			structuredOutput: config.structuredOutput ?? true,
-			vision: config.vision ?? inferVisionCapability(config.name)
+			vision: config.vision ?? inferVisionCapability(config.name),
+			reasoning: config.reasoning ?? inferReasoningCapability(config.name),
+			promptCaching: config.promptCaching ?? inferPromptCachingCapability(config.name),
+			pdf: config.pdf ?? inferPdfCapability(config.name),
+			audio: config.audio ?? false
 		};
 	}
 	/**
@@ -604,6 +726,7 @@ var BedrockModel = class {
 					usage.output = raw.outputTokens ?? 0;
 					usage.total = raw.totalTokens ?? usage.input + usage.output;
 					if (raw.cacheReadInputTokens && raw.cacheReadInputTokens > 0) usage.cachedTokens = raw.cacheReadInputTokens;
+					if (raw.cacheWriteInputTokens && raw.cacheWriteInputTokens > 0) usage.cacheWriteTokens = raw.cacheWriteInputTokens;
 				}
 			}
 		} catch (thrown) {
@@ -630,19 +753,66 @@ var BedrockModel = class {
 		const { system, messages: bedrockMessages } = toBedrockMessages(messages);
 		const maxTokens = options?.maxTokens ?? this.config.maxTokens;
 		const temperature = options?.temperature ?? this.config.temperature;
+		const cachedMessages = this.applyCacheBreakpoints(bedrockMessages, options?.cacheControl);
 		return {
 			modelId: this.name,
-			messages: bedrockMessages,
+			messages: cachedMessages,
 			...system ? { system } : {},
 			inferenceConfig: {
 				...maxTokens !== void 0 ? { maxTokens } : {},
 				...temperature !== void 0 ? { temperature } : {}
 			},
 			...this.buildToolConfig(options?.tools),
-			...this.buildOutputConfig(options?.responseSchema)
+			...this.buildOutputConfig(options?.responseSchema),
+			...this.buildReasoningConfig(options?.reasoning)
 		};
 	}
 	/**
+	* Append a Converse `cachePoint` block to the LAST message when the
+	* caller supplies a `cacheControl` write breakpoint and the model is
+	* `promptCaching`-capable. A cache point tells Bedrock to cache the
+	* whole prefix up to that block, so subsequent calls reusing the same
+	* prefix bill the cached portion at the discounted read rate
+	* (surfaced as `Usage.cachedTokens`). No-ops gracefully when caching
+	* is unsupported, no breakpoint was requested, or there are no
+	* messages to mark — Bedrock then prices the call normally.
+	*
+	* Bedrock only honors `CachePointType.DEFAULT`; the neutral
+	* `breakpoints` count is a presence hint (one trailing breakpoint is
+	* the only placement Converse supports without manual block surgery),
+	* so any positive value marks the trailing message.
+	*/
+	applyCacheBreakpoints(messages, cacheControl) {
+		const breakpoints = cacheControl?.breakpoints ?? 0;
+		if (!this.capabilities.promptCaching || breakpoints <= 0 || !messages || messages.length === 0) return messages;
+		const last = messages.length - 1;
+		const lastMessage = messages[last];
+		return [...messages.slice(0, last), {
+			...lastMessage,
+			content: [...lastMessage.content ?? [], { cachePoint: { type: "default" } }]
+		}];
+	}
+	/**
+	* Translate the neutral `reasoning` option into Claude-on-Bedrock's
+	* extended-thinking control, carried in Converse's escape hatch
+	* `additionalModelRequestFields.thinking`. Emitted only when the model
+	* is `reasoning`-capable and a budget can be resolved — `maxTokens`
+	* (explicit thinking budget) wins, otherwise `effort` maps to a
+	* conventional token budget so callers can opt in without picking a
+	* number. Returns an empty object (no-op) for non-reasoning models or
+	* when no reasoning option was supplied, so unsupported params never
+	* reach the wire.
+	*/
+	buildReasoningConfig(reasoning) {
+		if (!this.capabilities.reasoning || !reasoning) return {};
+		const budgetTokens = reasoning.maxTokens ?? EFFORT_THINKING_BUDGET[reasoning.effort ?? ""];
+		if (budgetTokens === void 0) return {};
+		return { additionalModelRequestFields: { thinking: {
+			type: "enabled",
+			budget_tokens: budgetTokens
+		} } };
+	}
+	/**
 	* Spread-friendly tool fragment. Returns an empty object when no
 	* tools were supplied (Bedrock rejects an empty `tools` array).
 	*/
@@ -694,8 +864,12 @@ var BedrockModel = class {
 	}
 	/**
 	* Normalize Converse's `TokenUsage` into the neutral `Usage` shape.
-	* Bedrock supplies a pre-summed `totalTokens`; cache-read tokens are
-	* surfaced as `cachedTokens` only when non-zero.
+	* Bedrock supplies a pre-summed `totalTokens`; cache-read and
+	* cache-write tokens are surfaced as `cachedTokens` /
+	* `cacheWriteTokens` only when non-zero so callers can price the
+	* discounted read rate and the one-time write cost separately.
+	* Bedrock's Converse `TokenUsage` carries no reasoning-token channel,
+	* so `Usage.reasoningTokens` is intentionally left unset here.
 	*/
 	extractUsage(raw) {
 		if (!raw) return {
@@ -706,11 +880,13 @@ var BedrockModel = class {
 		const input = raw.inputTokens ?? 0;
 		const output = raw.outputTokens ?? 0;
 		const cached = raw.cacheReadInputTokens;
+		const cacheWrite = raw.cacheWriteInputTokens;
 		return {
 			input,
 			output,
 			total: raw.totalTokens ?? input + output,
-			...cached && cached > 0 ? { cachedTokens: cached } : {}
+			...cached && cached > 0 ? { cachedTokens: cached } : {},
+			...cacheWrite && cacheWrite > 0 ? { cacheWriteTokens: cacheWrite } : {}
 		};
 	}
 	/**

package/cjs/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","names":["InvalidRequestError","AIError","ProviderTimeoutError","ProviderAuthError","QuotaExceededError","ProviderRateLimitError","ContextLengthExceededError","InvalidRequestError","ProviderError","LOG_MODULE","log","InvokeModelCommand","log","ConverseCommand","ConverseStreamCommand","BedrockRuntimeClient"],"sources":["../../../../../../@warlock.js/ai-bedrock/src/utils/map-stop-reason.ts","../../../../../../@warlock.js/ai-bedrock/src/utils/to-bedrock-messages.ts","../../../../../../@warlock.js/ai-bedrock/src/utils/to-bedrock-tools.ts","../../../../../../@warlock.js/ai-bedrock/src/utils/wrap-bedrock-error.ts","../../../../../../@warlock.js/ai-bedrock/src/embedder.ts","../../../../../../@warlock.js/ai-bedrock/src/known-vision-models.ts","../../../../../../@warlock.js/ai-bedrock/src/model.ts","../../../../../../@warlock.js/ai-bedrock/src/sdk.ts"],"sourcesContent":["import type { FinishReason } from \"@warlock.js/ai\";\n\nconst stopReasonMap: Record<string, FinishReason> = {\n  end_turn: \"stop\",\n  stop_sequence: \"stop\",\n  max_tokens: \"length\",\n  tool_use: \"tool_calls\",\n};\n\n/**\n * Map Bedrock Converse's `stopReason` to the normalized `FinishReason`\n * union.\n *\n * `end_turn` / `stop_sequence` are natural stops. `max_tokens` maps to\n * `length`. `tool_use` maps to `tool_calls`. Everything else —\n * `content_filtered`, `guardrail_intervened`, `malformed_tool_use`,\n * `malformed_model_output`, `model_context_window_exceeded`, `null`,\n * or any future value — falls through to `\"error\"`: none produced a\n * clean terminal answer, so the agent must not treat them as success.\n *\n * @example\n * mapStopReason(\"end_turn\");              // \"stop\"\n * mapStopReason(\"tool_use\");              // \"tool_calls\"\n * mapStopReason(\"guardrail_intervened\");  // \"error\"\n * mapStopReason(undefined);               // \"error\"\n */\nexport function mapStopReason(raw: string | null | undefined): FinishReason {\n  return stopReasonMap[raw ?? \"\"] ?? \"error\";\n}\n","import { InvalidRequestError, type ContentPart, type Message } from \"@warlock.js/ai\";\nimport type {\n  ContentBlock,\n  ImageFormat,\n  Message as BedrockMessage,\n  SystemContentBlock,\n} from \"@aws-sdk/client-bedrock-runtime\";\n\n/**\n * Result of splitting a vendor-neutral `Message[]` for the Bedrock\n * Converse API: system prompts hoist to a separate `SystemContentBlock[]`\n * (Converse has no `\"system\"` role inside `messages`), and the\n * remaining turns map to Bedrock `Message[]`.\n */\nexport type BedrockMessages = {\n  system: SystemContentBlock[] | undefined;\n  messages: BedrockMessage[];\n};\n\nconst MEDIA_TYPE_TO_FORMAT: Record<string, ImageFormat> = {\n  \"image/jpeg\": \"jpeg\",\n  \"image/png\": \"png\",\n  \"image/gif\": \"gif\",\n  \"image/webp\": \"webp\",\n};\n\n/**\n * Convert vendor-neutral `Message[]` into Bedrock Converse's request\n * shape.\n *\n * Converse differs from the OpenAI Chat protocol in three ways this\n * function absorbs:\n *\n * 1. **No `system` role.** System messages become a separate\n *    `SystemContentBlock[]` (one `{ text }` block each).\n * 2. **Tool results are `user` turns.** A neutral `tool` message\n *    becomes a `user` message carrying a single `toolResult` block.\n * 3. **Tool calls are `toolUse` content blocks.** An assistant message\n *    with `toolCalls` becomes an `assistant` message: an optional\n *    leading `text` block followed by one `toolUse` block per call.\n *\n * @example\n * const { system, messages } = toBedrockMessages([\n *   { role: \"system\", content: \"Be concise.\" },\n *   { role: \"user\", content: \"Hi\" },\n * ]);\n */\nexport function toBedrockMessages(messages: Message[]): BedrockMessages {\n  const system: SystemContentBlock[] = [];\n  const mapped: BedrockMessage[] = [];\n\n  for (const message of messages) {\n    if (message.role === \"system\") {\n      system.push({ text: stringifyContent(message.content) });\n\n      continue;\n    }\n\n    if (message.role === \"tool\") {\n      mapped.push({\n        role: \"user\",\n        content: [\n          {\n            toolResult: {\n              toolUseId: message.toolCallId ?? \"\",\n              content: [{ text: stringifyContent(message.content) }],\n            },\n          },\n        ],\n      });\n\n      continue;\n    }\n\n    if (message.role === \"assistant\" && message.toolCalls && message.toolCalls.length > 0) {\n      const blocks: ContentBlock[] = [];\n      const text = stringifyContent(message.content);\n\n      if (text) {\n        blocks.push({ text });\n      }\n\n      for (const toolCall of message.toolCalls) {\n        blocks.push({\n          toolUse: {\n            toolUseId: toolCall.id,\n            name: toolCall.name,\n            input: toolCall.input ?? {},\n          },\n        } as ContentBlock);\n      }\n\n      mapped.push({ role: \"assistant\", content: blocks });\n\n      continue;\n    }\n\n    if (message.role === \"user\" && Array.isArray(message.content)) {\n      mapped.push({\n        role: \"user\",\n        content: message.content.map(toBedrockContentBlock),\n      });\n\n      continue;\n    }\n\n    mapped.push({\n      role: message.role === \"assistant\" ? \"assistant\" : \"user\",\n      content: [{ text: stringifyContent(message.content) }],\n    });\n  }\n\n  return {\n    system: system.length > 0 ? system : undefined,\n    messages: mapped,\n  };\n}\n\n/**\n * Multipart content is only meaningful on user messages — for any other\n * role collapse a `ContentPart[]` to its concatenated text. Plain\n * strings pass through unchanged.\n */\nfunction stringifyContent(content: string | ContentPart[]): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n\n  return content\n    .filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n    .map((part) => part.text)\n    .join(\"\");\n}\n\n/**\n * Map a resolved `ContentPart` to a Bedrock `ContentBlock`. Bedrock's\n * `ImageSource` only accepts raw bytes or an S3 location — there is no\n * remote-URL source. A neutral `{ url }` image therefore cannot be\n * sent and surfaces a typed `InvalidRequestError` upfront rather than\n * a downstream Bedrock validation fault. The agent has already\n * resolved attachments, so this never fetches or reads anything.\n */\nfunction toBedrockContentBlock(part: ContentPart): ContentBlock {\n  if (part.type === \"text\") {\n    return { text: part.text };\n  }\n\n  if (\"url\" in part.source) {\n    throw new InvalidRequestError(\n      \"Bedrock Converse does not support remote-URL image sources; supply base64 image bytes instead.\",\n    );\n  }\n\n  const format = MEDIA_TYPE_TO_FORMAT[part.source.mediaType];\n\n  if (!format) {\n    throw new InvalidRequestError(\n      `Unsupported image media type for Bedrock: \"${part.source.mediaType}\" (expected image/jpeg, image/png, image/gif, or image/webp).`,\n    );\n  }\n\n  return {\n    image: {\n      format,\n      source: { bytes: Buffer.from(part.source.base64, \"base64\") },\n    },\n  };\n}\n","import { extractJsonSchema, type ToolConfig } from \"@warlock.js/ai\";\nimport type { Tool, ToolConfiguration, ToolInputSchema } from \"@aws-sdk/client-bedrock-runtime\";\n\n/**\n * Convert vendor-neutral `ToolConfig[]` into Bedrock Converse's\n * `ToolConfiguration`. Each tool becomes a `toolSpec` with a JSON\n * `inputSchema`. Bedrock requires the schema root to be an object —\n * a non-object extraction degrades to a parameterless object schema\n * so registration never fails.\n *\n * Returns `undefined` when there are no tools so the caller can omit\n * `toolConfig` from the request entirely (Bedrock rejects an empty\n * `tools` array).\n *\n * @example\n * const toolConfig = toBedrockToolConfig([weatherTool]);\n * await client.send(new ConverseCommand({ modelId, messages, toolConfig }));\n */\nexport function toBedrockToolConfig(\n  tools: ToolConfig<unknown, unknown>[] | undefined,\n): ToolConfiguration | undefined {\n  if (!tools || tools.length === 0) {\n    return undefined;\n  }\n\n  return {\n    tools: tools.map(\n      (tool): Tool => ({\n        toolSpec: {\n          name: tool.name,\n          description: tool.description,\n          inputSchema: { json: toJsonSchema(tool.input) } as ToolInputSchema,\n        },\n      }),\n    ),\n  };\n}\n\n/**\n * Resolve a tool's input schema to a JSON-Schema object. Bedrock's\n * `ToolInputSchema.json` requires an object root; anything else (or a\n * failed extraction) degrades to a parameterless object so the tool\n * still registers.\n */\nfunction toJsonSchema(input: ToolConfig<unknown, unknown>[\"input\"]): Record<string, unknown> {\n  const schema = extractJsonSchema(input);\n\n  if (schema && schema.type === \"object\") {\n    return schema;\n  }\n\n  return { type: \"object\" };\n}\n","import {\n  AIError,\n  ContextLengthExceededError,\n  InvalidRequestError,\n  ProviderAuthError,\n  ProviderError,\n  ProviderRateLimitError,\n  ProviderTimeoutError,\n  QuotaExceededError,\n} from \"@warlock.js/ai\";\n\n/**\n * Raw-error fields the wrapper reads off an AWS SDK exception. Every\n * Bedrock error is a Smithy `__BaseException` with a stable `name`\n * (`\"ThrottlingException\"`, `\"ValidationException\"`, …) and a\n * `$metadata` carrying `httpStatusCode` + `requestId`. We duck-type\n * because retries and proxies sometimes flatten the prototype chain.\n */\ntype BedrockErrorShape = {\n  name?: string;\n  message?: string;\n  httpStatusCode?: number;\n  requestId?: string;\n  code?: string;\n};\n\nconst TIMEOUT_NAMES = new Set([\n  \"ModelTimeoutException\",\n  \"TimeoutError\",\n  \"RequestTimeout\",\n  \"RequestTimeoutException\",\n]);\n\n/**\n * Wrap any thrown value caught inside the Bedrock adapter into the\n * appropriate `@warlock.js/ai` `AIError` subclass.\n *\n * **Dispatch strategy.** AWS errors carry no provider machine `code`;\n * the stable identifier is the Smithy exception `name`. Dispatch keys\n * on `name`, falls back to `$metadata.httpStatusCode` when the name is\n * missing (flattened/proxied errors). `ValidationException` is split:\n * the \"input is too long / exceeds context window\" phrasing maps to\n * `ContextLengthExceededError`, everything else to\n * `InvalidRequestError`.\n *\n * `AIError` instances pass through unchanged so `catch/throw wrap(e)`\n * pipelines never double-wrap.\n *\n * @example\n * try {\n *   return await this.client.send(new ConverseCommand(...));\n * } catch (thrown) {\n *   throw wrapBedrockError(thrown);\n * }\n */\nexport function wrapBedrockError(thrown: unknown): AIError {\n  if (thrown instanceof AIError) {\n    return thrown;\n  }\n\n  const shape = toShape(thrown);\n  const context = buildContext(shape);\n  const message = shape.message ?? (thrown instanceof Error ? thrown.message : String(thrown));\n\n  if (isTimeout(shape)) {\n    return new ProviderTimeoutError(message, { cause: thrown, context });\n  }\n\n  if (shape.name === \"AccessDeniedException\" || shape.httpStatusCode === 403) {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.httpStatusCode === 401) {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.name === \"ServiceQuotaExceededException\") {\n    return new QuotaExceededError(message, { cause: thrown, context });\n  }\n\n  if (shape.name === \"ThrottlingException\" || shape.httpStatusCode === 429) {\n    return new ProviderRateLimitError(message, { cause: thrown, context });\n  }\n\n  if (shape.name === \"ValidationException\") {\n    if (/too long|context window|maximum context|exceeds the maximum/i.test(message)) {\n      return new ContextLengthExceededError(message, { cause: thrown, context });\n    }\n\n    return new InvalidRequestError(message, { cause: thrown, context });\n  }\n\n  if (\n    shape.name === \"ResourceNotFoundException\" ||\n    shape.name === \"ConflictException\" ||\n    isClientStatus(shape.httpStatusCode)\n  ) {\n    return new InvalidRequestError(message, { cause: thrown, context });\n  }\n\n  return new ProviderError(message, { cause: thrown, context });\n}\n\n/**\n * Read the raw error shape without depending on `instanceof`. AWS\n * exceptions expose `$metadata`; plain/proxied errors may carry\n * `status` / `code` instead.\n */\nfunction toShape(thrown: unknown): BedrockErrorShape {\n  if (typeof thrown !== \"object\" || thrown === null) {\n    return {};\n  }\n\n  const raw = thrown as Record<string, unknown>;\n  const metadata = raw.$metadata as { httpStatusCode?: number; requestId?: string } | undefined;\n\n  return {\n    name: typeof raw.name === \"string\" ? raw.name : undefined,\n    message: typeof raw.message === \"string\" ? raw.message : undefined,\n    httpStatusCode:\n      metadata && typeof metadata.httpStatusCode === \"number\"\n        ? metadata.httpStatusCode\n        : typeof raw.status === \"number\"\n          ? (raw.status as number)\n          : undefined,\n    requestId: metadata && typeof metadata.requestId === \"string\" ? metadata.requestId : undefined,\n    code: typeof raw.code === \"string\" ? raw.code : undefined,\n  };\n}\n\n/**\n * Decide whether the error is a timeout. Bedrock surfaces\n * `ModelTimeoutException`; the AWS transport layer surfaces\n * `TimeoutError` / `ETIMEDOUT` / `ECONNABORTED`.\n */\nfunction isTimeout(shape: BedrockErrorShape): boolean {\n  if (shape.name && TIMEOUT_NAMES.has(shape.name)) {\n    return true;\n  }\n\n  return shape.code === \"ETIMEDOUT\" || shape.code === \"ECONNABORTED\";\n}\n\n/** True for HTTP 4xx — a client-side request problem, not a server fault. */\nfunction isClientStatus(status: number | undefined): boolean {\n  return typeof status === \"number\" && status >= 400 && status < 500;\n}\n\n/**\n * Attach the raw diagnostic fields to `error.context`. The Smithy\n * exception `name` is the closest thing Bedrock has to a stable error\n * code, so it lands on `context.code`.\n */\nfunction buildContext(shape: BedrockErrorShape): Record<string, unknown> {\n  const context: Record<string, unknown> = {};\n\n  if (shape.httpStatusCode !== undefined) {\n    context.status = shape.httpStatusCode;\n  }\n\n  if (shape.name) {\n    context.code = shape.name;\n  }\n\n  if (shape.requestId) {\n    context.requestId = shape.requestId;\n  }\n\n  return context;\n}\n","import {\n  type EmbeddingBatchResult,\n  type EmbeddingResult,\n  type EmbeddingUsage,\n  type EmbedderContract,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport { InvokeModelCommand, type BedrockRuntimeClient } from \"@aws-sdk/client-bedrock-runtime\";\nimport type { BedrockEmbedderConfig } from \"./config.type\";\nimport { wrapBedrockError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.bedrock\";\n\n/** Shape of the Amazon Titan Text Embeddings response body. */\ntype TitanEmbeddingResponse = {\n  embedding: number[];\n  inputTextTokenCount: number;\n};\n\n/**\n * Bedrock-backed implementation of `EmbedderContract`, targeting the\n * Amazon Titan Text Embeddings family\n * (`amazon.titan-embed-text-v2:0` / v1) via `InvokeModel`.\n *\n * **Role.** Converts text into floating-point vectors. Standalone\n * primitive — unrelated to Converse / tools / the agent loop.\n *\n * **Single-input only upstream.** Titan's `InvokeModel` body accepts\n * one `inputText` per call — there is no batch endpoint. `embedMany`\n * therefore issues one request per input sequentially and aggregates\n * token usage. This is a deliberate, documented trade-off: a real\n * batch API does not exist for Titan on Bedrock, so the alternative\n * (failing `embedMany`) would be worse. Cohere embeddings on Bedrock\n * *do* batch but use an incompatible body shape — out of scope; use\n * the OpenAI adapter or a future Cohere adapter when batch throughput\n * matters.\n *\n * **Dimensions.** When no `dimensions` override is given,\n * `this.dimensions` starts at `0` and is populated from the first\n * response's vector length, then cached. Passing `dimensions` forwards\n * Titan v2's truncation hint (256 / 512 / 1024) and sets the initial\n * value immediately.\n *\n * @example\n * const embedder = new BedrockEmbedder(client, { name: \"amazon.titan-embed-text-v2:0\" });\n * const { vector } = await embedder.embed(\"Hello world\");\n * const { vectors } = await embedder.embedMany([\"doc 1\", \"doc 2\"]);\n */\nexport class BedrockEmbedder implements EmbedderContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public dimensions: number;\n\n  private readonly client: BedrockRuntimeClient;\n  private readonly configuredDimensions: number | undefined;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: BedrockRuntimeClient,\n    config: BedrockEmbedderConfig,\n    provider: string = \"bedrock\",\n  ) {\n    this.client = client;\n    this.name = config.name;\n    this.provider = provider;\n    this.configuredDimensions = config.dimensions;\n    this.dimensions = config.dimensions ?? 0;\n  }\n\n  public async embed(input: string): Promise<EmbeddingResult> {\n    const { vector, tokens } = await this.invoke(input);\n\n    return {\n      vector,\n      dimensions: this.dimensions,\n      usage: { promptTokens: tokens, totalTokens: tokens },\n    };\n  }\n\n  public async embedMany(inputs: string[]): Promise<EmbeddingBatchResult> {\n    const vectors: number[][] = [];\n    let tokens = 0;\n\n    for (const input of inputs) {\n      const result = await this.invoke(input);\n\n      vectors.push(result.vector);\n      tokens += result.tokens;\n    }\n\n    const usage: EmbeddingUsage = { promptTokens: tokens, totalTokens: tokens };\n\n    return { vectors, dimensions: this.dimensions, usage };\n  }\n\n  /**\n   * Issue a single Titan `InvokeModel` embedding request: encode the\n   * JSON body, send, wrap provider errors, decode the response, and\n   * cache `dimensions` on the first successful call.\n   */\n  private async invoke(input: string): Promise<{ vector: number[]; tokens: number }> {\n    this.logger.debug(LOG_MODULE, \"embedder.request\", \"InvokeModel embeddings\", {\n      model: this.name,\n    });\n\n    const body = JSON.stringify({\n      inputText: input,\n      ...(this.configuredDimensions !== undefined\n        ? { dimensions: this.configuredDimensions }\n        : {}),\n    });\n\n    let raw;\n\n    try {\n      raw = await this.client.send(\n        new InvokeModelCommand({\n          modelId: this.name,\n          contentType: \"application/json\",\n          accept: \"application/json\",\n          body: new TextEncoder().encode(body),\n        }),\n      );\n    } catch (thrown) {\n      const wrapped = wrapBedrockError(thrown);\n\n      this.logger.error(LOG_MODULE, \"embedder.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    const decoded = JSON.parse(new TextDecoder().decode(raw.body)) as TitanEmbeddingResponse;\n\n    if (this.dimensions === 0) {\n      this.dimensions = decoded.embedding.length;\n    }\n\n    this.logger.debug(LOG_MODULE, \"embedder.response\", \"InvokeModel embeddings returned\", {\n      dimensions: decoded.embedding.length,\n      tokens: decoded.inputTextTokenCount,\n    });\n\n    return { vector: decoded.embedding, tokens: decoded.inputTextTokenCount };\n  }\n}\n","/**\n * Substrings that identify Bedrock model ids whose family accepts image\n * input on the Converse API.\n *\n * Bedrock model ids are provider-prefixed and version-suffixed\n * (`anthropic.claude-3-5-sonnet-20240620-v1:0`, `us.amazon.nova-pro-v1:0`,\n * `meta.llama3-2-90b-instruct-v1:0`), so a substring match is the only\n * robust check across the cross-region inference-profile prefixes\n * (`us.`, `eu.`, `apac.`) and date/version tags.\n *\n * Multimodal families covered: Anthropic Claude 3 / 3.5 / 3.7 / 4,\n * Amazon Nova Lite/Pro/Premier, Meta Llama 3.2 (11B/90B) and Llama 4.\n * Text-only families (Llama 3/3.1, Titan Text, Mistral 7B, Cohere\n * Command) are intentionally absent. Override per-model via\n * `bedrock.model({ name, vision: true | false })`.\n */\nconst VISION_CAPABLE_SUBSTRINGS = [\n  \"claude-3\",\n  \"claude-sonnet-4\",\n  \"claude-opus-4\",\n  \"claude-haiku-4\",\n  \"nova-lite\",\n  \"nova-pro\",\n  \"nova-premier\",\n  \"llama3-2-11b\",\n  \"llama3-2-90b\",\n  \"llama4\",\n];\n\n/**\n * Infer whether a Bedrock model id supports vision based on the known\n * multimodal-family substrings. Unknown ids default to `false` so that\n * passing an image attachment to an unsupported model surfaces a clear,\n * agent-side capability error instead of an opaque Bedrock validation\n * fault.\n *\n * @example\n * inferVisionCapability(\"anthropic.claude-3-5-sonnet-20240620-v1:0\"); // → true\n * inferVisionCapability(\"us.amazon.nova-pro-v1:0\");                    // → true\n * inferVisionCapability(\"meta.llama3-1-8b-instruct-v1:0\");             // → false\n * inferVisionCapability(\"amazon.titan-text-express-v1\");               // → false\n */\nexport function inferVisionCapability(modelId: string): boolean {\n  const normalized = modelId.toLowerCase();\n\n  return VISION_CAPABLE_SUBSTRINGS.some((fragment) => normalized.includes(fragment));\n}\n","import {\n  safeJsonParse,\n  type Message,\n  type ModelCallOptions,\n  type ModelCapabilities,\n  type ModelContract,\n  type ModelPricing,\n  type ModelResponse,\n  type ModelStreamChunk,\n  type ModelToolCallRequest,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport {\n  ConverseCommand,\n  ConverseStreamCommand,\n  type BedrockRuntimeClient,\n  type ContentBlock,\n  type ConverseRequest,\n  type TokenUsage,\n} from \"@aws-sdk/client-bedrock-runtime\";\nimport type { BedrockModelConfig } from \"./config.type\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapStopReason, toBedrockMessages, toBedrockToolConfig, wrapBedrockError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.bedrock\";\n\n/**\n * Bedrock-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and AWS Bedrock's Converse /\n * ConverseStream API. Converse is the model-agnostic surface — one\n * wire mapping covers every Bedrock-hosted family (Anthropic Claude,\n * Amazon Nova, Meta Llama, Mistral, Cohere) instead of per-family\n * `InvokeModel` body shapes.\n *\n * **Responsibility.**\n * - Owns: a long-lived `BedrockRuntimeClient` + frozen `ModelConfig`\n *   (modelId, temperature, maxTokens) used as per-call defaults.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Converse shapes (system hoisting, `toolUse` / `toolResult` blocks,\n *   image bytes) on the way out, and Converse's content-block response\n *   (text, tool calls, stop reason, token usage) back into the neutral\n *   shapes on the way in.\n * - Does NOT own: dispatching tools, looping, history, retries — those\n *   are agent concerns. The model is a per-call protocol adapter.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across calls\"): the AWS client is heavy to construct and reused for\n * the SDK's lifetime.\n *\n * @example\n * import { BedrockRuntimeClient } from \"@aws-sdk/client-bedrock-runtime\";\n * const client = new BedrockRuntimeClient({ region: \"us-east-1\" });\n * const model = new BedrockModel(client, {\n *   name: \"anthropic.claude-sonnet-4-5-20250929-v1:0\",\n * });\n *\n * const myAgent = agent({ model, tools: [searchTool] });\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class BedrockModel implements ModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly capabilities: ModelCapabilities;\n  public readonly pricing?: ModelPricing;\n\n  private readonly client: BedrockRuntimeClient;\n  private readonly config: BedrockModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: BedrockRuntimeClient,\n    config: BedrockModelConfig,\n    provider: string = \"bedrock\",\n  ) {\n    this.client = client;\n    this.config = config;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    this.capabilities = {\n      structuredOutput: config.structuredOutput ?? true,\n      vision: config.vision ?? inferVisionCapability(config.name),\n    };\n  }\n\n  /**\n   * Single-shot completion via the Converse API. Sends the full\n   * message list, waits for the terminal response, and reshapes it\n   * into a vendor-neutral `ModelResponse`. Per-call `options` override\n   * the instance defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting Converse call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response;\n\n    try {\n      response = await this.client.send(\n        new ConverseCommand(this.buildRequest(messages, options)),\n        options?.signal ? { abortSignal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const blocks = response.output?.message?.content ?? [];\n    const finishReason = mapStopReason(response.stopReason);\n    const usage = this.extractUsage(response.usage);\n    const toolCalls = this.extractToolCalls(blocks);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Converse call succeeded\", { finishReason, usage });\n\n    return {\n      content: this.extractText(blocks),\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion via ConverseStream. Yields neutral\n   * `ModelStreamChunk`s — `delta` for text, `tool-call` once a\n   * `toolUse` block's accumulated input JSON is complete, and a\n   * terminal `done` with the final finish reason + usage totals.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting ConverseStream call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response;\n\n    try {\n      response = await this.client.send(\n        new ConverseStreamCommand(this.buildRequest(messages, options)),\n        options?.signal ? { abortSignal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawStopReason: string | undefined;\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n    const toolBlocks = new Map<number, { id: string; name: string; json: string }>();\n\n    try {\n      for await (const event of response.stream ?? []) {\n        if (event.contentBlockStart?.start?.toolUse) {\n          const start = event.contentBlockStart.start.toolUse;\n\n          toolBlocks.set(event.contentBlockStart.contentBlockIndex ?? 0, {\n            id: start.toolUseId ?? \"\",\n            name: start.name ?? \"\",\n            json: \"\",\n          });\n\n          continue;\n        }\n\n        if (event.contentBlockDelta?.delta) {\n          const delta = event.contentBlockDelta.delta;\n\n          if (delta.text) {\n            yield { type: \"delta\", content: delta.text };\n          } else if (delta.toolUse) {\n            const accumulator = toolBlocks.get(event.contentBlockDelta.contentBlockIndex ?? 0);\n\n            if (accumulator) {\n              accumulator.json += delta.toolUse.input ?? \"\";\n            }\n          }\n\n          continue;\n        }\n\n        if (event.contentBlockStop) {\n          const accumulator = toolBlocks.get(event.contentBlockStop.contentBlockIndex ?? 0);\n\n          if (accumulator) {\n            yield {\n              type: \"tool-call\",\n              id: accumulator.id,\n              name: accumulator.name,\n              input: safeJsonParse<Record<string, unknown>>(accumulator.json, {}),\n            };\n\n            toolBlocks.delete(event.contentBlockStop.contentBlockIndex ?? 0);\n          }\n\n          continue;\n        }\n\n        if (event.messageStop) {\n          rawStopReason = event.messageStop.stopReason;\n        }\n\n        if (event.metadata?.usage) {\n          const raw = event.metadata.usage;\n\n          usage.input = raw.inputTokens ?? 0;\n          usage.output = raw.outputTokens ?? 0;\n          usage.total = raw.totalTokens ?? usage.input + usage.output;\n\n          if (raw.cacheReadInputTokens && raw.cacheReadInputTokens > 0) {\n            usage.cachedTokens = raw.cacheReadInputTokens;\n          }\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = mapStopReason(rawStopReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"ConverseStream call succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the Converse request shared by `complete()` and\n   * `stream()` (both command shapes take the same input). Hoists the\n   * system prompt, maps inference params, and conditionally attaches\n   * tools and native structured output.\n   */\n  private buildRequest(\n    messages: Message[],\n    options: ModelCallOptions | undefined,\n  ): ConverseRequest {\n    const { system, messages: bedrockMessages } = toBedrockMessages(messages);\n    const maxTokens = options?.maxTokens ?? this.config.maxTokens;\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      modelId: this.name,\n      messages: bedrockMessages,\n      ...(system ? { system } : {}),\n      inferenceConfig: {\n        ...(maxTokens !== undefined ? { maxTokens } : {}),\n        ...(temperature !== undefined ? { temperature } : {}),\n      },\n      ...this.buildToolConfig(options?.tools),\n      ...this.buildOutputConfig(options?.responseSchema),\n    };\n  }\n\n  /**\n   * Spread-friendly tool fragment. Returns an empty object when no\n   * tools were supplied (Bedrock rejects an empty `tools` array).\n   */\n  private buildToolConfig(tools: ModelCallOptions[\"tools\"]): Pick<ConverseRequest, \"toolConfig\"> {\n    const toolConfig = toBedrockToolConfig(tools);\n\n    return toolConfig ? { toolConfig } : {};\n  }\n\n  /**\n   * Translate the neutral `responseSchema` into Converse's native\n   * `outputConfig.textFormat` (JSON-schema structured output). Bedrock\n   * requires the schema as a stringified JSON document and only\n   * accepts an object root. Emitted only when the model is\n   * `structuredOutput`-capable and the schema is an object — otherwise\n   * the agent's soft system-prompt hint + client-side `validate()`\n   * carry shape (same degradation philosophy as the OpenAI adapter).\n   */\n  private buildOutputConfig(\n    responseSchema: Record<string, unknown> | undefined,\n  ): Pick<ConverseRequest, \"outputConfig\"> {\n    if (!responseSchema || !this.capabilities.structuredOutput) {\n      return {};\n    }\n\n    if (responseSchema.type !== \"object\" || typeof responseSchema.properties !== \"object\") {\n      return {};\n    }\n\n    return {\n      outputConfig: {\n        textFormat: {\n          type: \"json_schema\",\n          structure: {\n            jsonSchema: { name: \"response\", schema: JSON.stringify(responseSchema) },\n          },\n        },\n      },\n    };\n  }\n\n  /**\n   * Concatenate every `text` content block into the single neutral\n   * `content` string. `toolUse` and other block types are surfaced\n   * separately via `extractToolCalls`.\n   */\n  private extractText(blocks: ContentBlock[]): string {\n    return blocks\n      .map((block) => (\"text\" in block && typeof block.text === \"string\" ? block.text : \"\"))\n      .join(\"\");\n  }\n\n  /**\n   * Reshape Converse `toolUse` content blocks into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when no tools were\n   * requested so callers can branch on presence.\n   */\n  private extractToolCalls(blocks: ContentBlock[]): ModelToolCallRequest[] | undefined {\n    const toolCalls: ModelToolCallRequest[] = [];\n\n    for (const block of blocks) {\n      if (\"toolUse\" in block && block.toolUse) {\n        toolCalls.push({\n          id: block.toolUse.toolUseId ?? \"\",\n          name: block.toolUse.name ?? \"\",\n          input: (block.toolUse.input ?? {}) as Record<string, unknown>,\n        });\n      }\n    }\n\n    return toolCalls.length > 0 ? toolCalls : undefined;\n  }\n\n  /**\n   * Normalize Converse's `TokenUsage` into the neutral `Usage` shape.\n   * Bedrock supplies a pre-summed `totalTokens`; cache-read tokens are\n   * surfaced as `cachedTokens` only when non-zero.\n   */\n  private extractUsage(raw: TokenUsage | undefined): Usage {\n    if (!raw) {\n      return { input: 0, output: 0, total: 0 };\n    }\n\n    const input = raw.inputTokens ?? 0;\n    const output = raw.outputTokens ?? 0;\n    const cached = raw.cacheReadInputTokens;\n\n    return {\n      input,\n      output,\n      total: raw.totalTokens ?? input + output,\n      ...(cached && cached > 0 ? { cachedTokens: cached } : {}),\n    };\n  }\n\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy\n   * and emit the standard error log line before it propagates. Shared\n   * by every catch site so the log shape stays identical.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapBedrockError(thrown);\n\n    this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n      code: wrapped.code,\n      context: wrapped.context,\n    });\n\n    return wrapped;\n  }\n}\n","import { BedrockRuntimeClient } from \"@aws-sdk/client-bedrock-runtime\";\nimport type {\n  EmbedderContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n} from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type {\n  BedrockEmbedderConfig,\n  BedrockModelConfig,\n  BedrockSDKConfig,\n} from \"./config.type\";\nimport { BedrockEmbedder } from \"./embedder\";\nimport { BedrockModel } from \"./model\";\n\n/**\n * AWS Bedrock-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for any Bedrock-hosted model via\n * the Converse API. A single `BedrockSDK` holds one live\n * `BedrockRuntimeClient`, shared by every `ModelContract` and\n * `EmbedderContract` it produces. Construct one SDK per AWS\n * account/region and reuse it everywhere.\n *\n * **Responsibility.**\n * - Owns: a long-lived `BedrockRuntimeClient` (region, credential\n *   chain) and its lifetime. Factory for `BedrockModel` /\n *   `BedrockEmbedder` instances sharing that client.\n * - Does NOT own: anything per-call — those live in `BedrockModel` /\n *   `BedrockEmbedder` and the agent runtime.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"): the AWS client is heavy to construct and\n * designed for reuse; keeping it on `this` aligns with the\n * `new BedrockRuntimeClient(...)` upstream convention.\n *\n * @example\n * const bedrock = new BedrockSDK({ region: \"us-east-1\" });\n * const model = bedrock.model({ name: \"anthropic.claude-sonnet-4-5-20250929-v1:0\" });\n * const embedder = bedrock.embedder({ name: \"amazon.titan-embed-text-v2:0\" });\n */\nexport class BedrockSDK implements SDKAdapterContract {\n  private readonly client: BedrockRuntimeClient;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: BedrockSDKConfig) {\n    const { provider, pricing, ...clientConfig } = config;\n\n    this.client = new BedrockRuntimeClient(clientConfig);\n    this.provider = provider ?? \"bedrock\";\n    this.pricing = pricing;\n  }\n\n  /**\n   * Build a `BedrockModel` bound to this SDK's client. Each call\n   * returns a fresh instance; all instances share the underlying AWS\n   * client so connection pools, credential refresh, and retry config\n   * stay unified. The SDK's `provider` label is forwarded.\n   *\n   * Pricing resolution: per-model `config.pricing` wins; otherwise the\n   * SDK-level registry entry keyed by `config.name`; otherwise\n   * `undefined` (no cost computed).\n   */\n  public model(config: BedrockModelConfig): ModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: BedrockModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new BedrockModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate. Uses the character-heuristic\n   * (`approximateTokenCount`) from the core package — Bedrock has no\n   * offline tokenizer and the per-model tokenizers differ; good enough\n   * for budgeting and quota guards, not for billing.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n\n  /**\n   * Build a `BedrockEmbedder` (Amazon Titan Text Embeddings) bound to\n   * this SDK's client.\n   *\n   * @example\n   * const embedder = bedrock.embedder({ name: \"amazon.titan-embed-text-v2:0\" });\n   * const { vector } = await embedder.embed(\"Hello world\");\n   */\n  public embedder(config: BedrockEmbedderConfig): EmbedderContract {\n    return new BedrockEmbedder(this.client, config, this.provider);\n  }\n}\n"],"mappings":";;;;;;AAEA,MAAM,gBAA8C;CAClD,UAAU;CACV,eAAe;CACf,YAAY;CACZ,UAAU;AACZ;;;;;;;;;;;;;;;;;;AAmBA,SAAgB,cAAc,KAA8C;CAC1E,OAAO,cAAc,OAAO,OAAO;AACrC;;;;ACTA,MAAM,uBAAoD;CACxD,cAAc;CACd,aAAa;CACb,aAAa;CACb,cAAc;AAChB;;;;;;;;;;;;;;;;;;;;;;AAuBA,SAAgB,kBAAkB,UAAsC;CACtE,MAAM,SAA+B,CAAC;CACtC,MAAM,SAA2B,CAAC;CAElC,KAAK,MAAM,WAAW,UAAU;EAC9B,IAAI,QAAQ,SAAS,UAAU;GAC7B,OAAO,KAAK,EAAE,MAAM,iBAAiB,QAAQ,OAAO,EAAE,CAAC;GAEvD;EACF;EAEA,IAAI,QAAQ,SAAS,QAAQ;GAC3B,OAAO,KAAK;IACV,MAAM;IACN,SAAS,CACP,EACE,YAAY;KACV,WAAW,QAAQ,cAAc;KACjC,SAAS,CAAC,EAAE,MAAM,iBAAiB,QAAQ,OAAO,EAAE,CAAC;IACvD,EACF,CACF;GACF,CAAC;GAED;EACF;EAEA,IAAI,QAAQ,SAAS,eAAe,QAAQ,aAAa,QAAQ,UAAU,SAAS,GAAG;GACrF,MAAM,SAAyB,CAAC;GAChC,MAAM,OAAO,iBAAiB,QAAQ,OAAO;GAE7C,IAAI,MACF,OAAO,KAAK,EAAE,KAAK,CAAC;GAGtB,KAAK,MAAM,YAAY,QAAQ,WAC7B,OAAO,KAAK,EACV,SAAS;IACP,WAAW,SAAS;IACpB,MAAM,SAAS;IACf,OAAO,SAAS,SAAS,CAAC;GAC5B,EACF,CAAiB;GAGnB,OAAO,KAAK;IAAE,MAAM;IAAa,SAAS;GAAO,CAAC;GAElD;EACF;EAEA,IAAI,QAAQ,SAAS,UAAU,MAAM,QAAQ,QAAQ,OAAO,GAAG;GAC7D,OAAO,KAAK;IACV,MAAM;IACN,SAAS,QAAQ,QAAQ,IAAI,qBAAqB;GACpD,CAAC;GAED;EACF;EAEA,OAAO,KAAK;GACV,MAAM,QAAQ,SAAS,cAAc,cAAc;GACnD,SAAS,CAAC,EAAE,MAAM,iBAAiB,QAAQ,OAAO,EAAE,CAAC;EACvD,CAAC;CACH;CAEA,OAAO;EACL,QAAQ,OAAO,SAAS,IAAI,SAAS;EACrC,UAAU;CACZ;AACF;;;;;;AAOA,SAAS,iBAAiB,SAAyC;CACjE,IAAI,OAAO,YAAY,UACrB,OAAO;CAGT,OAAO,QACJ,QAAQ,SAAiD,KAAK,SAAS,MAAM,CAAC,CAC9E,KAAK,SAAS,KAAK,IAAI,CAAC,CACxB,KAAK,EAAE;AACZ;;;;;;;;;AAUA,SAAS,sBAAsB,MAAiC;CAC9D,IAAI,KAAK,SAAS,QAChB,OAAO,EAAE,MAAM,KAAK,KAAK;CAG3B,IAAI,SAAS,KAAK,QAChB,MAAM,IAAIA,mCACR,gGACF;CAGF,MAAM,SAAS,qBAAqB,KAAK,OAAO;CAEhD,IAAI,CAAC,QACH,MAAM,IAAIA,mCACR,8CAA8C,KAAK,OAAO,UAAU,8DACtE;CAGF,OAAO,EACL,OAAO;EACL;EACA,QAAQ,EAAE,OAAO,OAAO,KAAK,KAAK,OAAO,QAAQ,QAAQ,EAAE;CAC7D,EACF;AACF;;;;;;;;;;;;;;;;;;;ACrJA,SAAgB,oBACd,OAC+B;CAC/B,IAAI,CAAC,SAAS,MAAM,WAAW,GAC7B;CAGF,OAAO,EACL,OAAO,MAAM,KACV,UAAgB,EACf,UAAU;EACR,MAAM,KAAK;EACX,aAAa,KAAK;EAClB,aAAa,EAAE,MAAM,aAAa,KAAK,KAAK,EAAE;CAChD,EACF,EACF,EACF;AACF;;;;;;;AAQA,SAAS,aAAa,OAAuE;CAC3F,MAAM,+CAA2B,KAAK;CAEtC,IAAI,UAAU,OAAO,SAAS,UAC5B,OAAO;CAGT,OAAO,EAAE,MAAM,SAAS;AAC1B;;;;AC1BA,MAAM,gBAAgB,IAAI,IAAI;CAC5B;CACA;CACA;CACA;AACF,CAAC;;;;;;;;;;;;;;;;;;;;;;;AAwBD,SAAgB,iBAAiB,QAA0B;CACzD,IAAI,kBAAkBC,wBACpB,OAAO;CAGT,MAAM,QAAQ,QAAQ,MAAM;CAC5B,MAAM,UAAU,aAAa,KAAK;CAClC,MAAM,UAAU,MAAM,YAAY,kBAAkB,QAAQ,OAAO,UAAU,OAAO,MAAM;CAE1F,IAAI,UAAU,KAAK,GACjB,OAAO,IAAIC,oCAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,SAAS,2BAA2B,MAAM,mBAAmB,KACrE,OAAO,IAAIC,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,mBAAmB,KAC3B,OAAO,IAAIA,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,iCACjB,OAAO,IAAIC,kCAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,SAAS,yBAAyB,MAAM,mBAAmB,KACnE,OAAO,IAAIC,sCAAuB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGvE,IAAI,MAAM,SAAS,uBAAuB;EACxC,IAAI,+DAA+D,KAAK,OAAO,GAC7E,OAAO,IAAIC,0CAA2B,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;EAG3E,OAAO,IAAIC,mCAAoB,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;CACpE;CAEA,IACE,MAAM,SAAS,+BACf,MAAM,SAAS,uBACf,eAAe,MAAM,cAAc,GAEnC,OAAO,IAAIA,mCAAoB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGpE,OAAO,IAAIC,6BAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;AAOA,SAAS,QAAQ,QAAoC;CACnD,IAAI,OAAO,WAAW,YAAY,WAAW,MAC3C,OAAO,CAAC;CAGV,MAAM,MAAM;CACZ,MAAM,WAAW,IAAI;CAErB,OAAO;EACL,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;EAChD,SAAS,OAAO,IAAI,YAAY,WAAW,IAAI,UAAU;EACzD,gBACE,YAAY,OAAO,SAAS,mBAAmB,WAC3C,SAAS,iBACT,OAAO,IAAI,WAAW,WACnB,IAAI,SACL;EACR,WAAW,YAAY,OAAO,SAAS,cAAc,WAAW,SAAS,YAAY;EACrF,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;CAClD;AACF;;;;;;AAOA,SAAS,UAAU,OAAmC;CACpD,IAAI,MAAM,QAAQ,cAAc,IAAI,MAAM,IAAI,GAC5C,OAAO;CAGT,OAAO,MAAM,SAAS,eAAe,MAAM,SAAS;AACtD;;AAGA,SAAS,eAAe,QAAqC;CAC3D,OAAO,OAAO,WAAW,YAAY,UAAU,OAAO,SAAS;AACjE;;;;;;AAOA,SAAS,aAAa,OAAmD;CACvE,MAAM,UAAmC,CAAC;CAE1C,IAAI,MAAM,mBAAmB,QAC3B,QAAQ,SAAS,MAAM;CAGzB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,IAAI,MAAM,WACR,QAAQ,YAAY,MAAM;CAG5B,OAAO;AACT;;;;AC9JA,MAAMC,eAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCnB,IAAa,kBAAb,MAAyD;CASvD,AAAO,YACL,QACA,QACA,WAAmB,WACnB;gBANgCC;EAOhC,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,uBAAuB,OAAO;EACnC,KAAK,aAAa,OAAO,cAAc;CACzC;CAEA,MAAa,MAAM,OAAyC;EAC1D,MAAM,EAAE,QAAQ,WAAW,MAAM,KAAK,OAAO,KAAK;EAElD,OAAO;GACL;GACA,YAAY,KAAK;GACjB,OAAO;IAAE,cAAc;IAAQ,aAAa;GAAO;EACrD;CACF;CAEA,MAAa,UAAU,QAAiD;EACtE,MAAM,UAAsB,CAAC;EAC7B,IAAI,SAAS;EAEb,KAAK,MAAM,SAAS,QAAQ;GAC1B,MAAM,SAAS,MAAM,KAAK,OAAO,KAAK;GAEtC,QAAQ,KAAK,OAAO,MAAM;GAC1B,UAAU,OAAO;EACnB;EAEA,MAAM,QAAwB;GAAE,cAAc;GAAQ,aAAa;EAAO;EAE1E,OAAO;GAAE;GAAS,YAAY,KAAK;GAAY;EAAM;CACvD;;;;;;CAOA,MAAc,OAAO,OAA8D;EACjF,KAAK,OAAO,MAAMD,cAAY,oBAAoB,0BAA0B,EAC1E,OAAO,KAAK,KACd,CAAC;EAED,MAAM,OAAO,KAAK,UAAU;GAC1B,WAAW;GACX,GAAI,KAAK,yBAAyB,SAC9B,EAAE,YAAY,KAAK,qBAAqB,IACxC,CAAC;EACP,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,MAAM,MAAM,KAAK,OAAO,KACtB,IAAIE,mDAAmB;IACrB,SAAS,KAAK;IACd,aAAa;IACb,QAAQ;IACR,MAAM,IAAI,YAAY,CAAC,CAAC,OAAO,IAAI;GACrC,CAAC,CACH;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,iBAAiB,MAAM;GAEvC,KAAK,OAAO,MAAMF,cAAY,kBAAkB,QAAQ,SAAS;IAC/D,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,UAAU,KAAK,MAAM,IAAI,YAAY,CAAC,CAAC,OAAO,IAAI,IAAI,CAAC;EAE7D,IAAI,KAAK,eAAe,GACtB,KAAK,aAAa,QAAQ,UAAU;EAGtC,KAAK,OAAO,MAAMA,cAAY,qBAAqB,mCAAmC;GACpF,YAAY,QAAQ,UAAU;GAC9B,QAAQ,QAAQ;EAClB,CAAC;EAED,OAAO;GAAE,QAAQ,QAAQ;GAAW,QAAQ,QAAQ;EAAoB;CAC1E;AACF;;;;;;;;;;;;;;;;;;;;ACnIA,MAAM,4BAA4B;CAChC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;;;;;;;;;;;;;;AAeA,SAAgB,sBAAsB,SAA0B;CAC9D,MAAM,aAAa,QAAQ,YAAY;CAEvC,OAAO,0BAA0B,MAAM,aAAa,WAAW,SAAS,QAAQ,CAAC;AACnF;;;;ACrBA,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCnB,IAAa,eAAb,MAAmD;CAUjD,AAAO,YACL,QACA,QACA,WAAmB,WACnB;gBANgCG;EAOhC,KAAK,SAAS;EACd,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,KAAK,eAAe;GAClB,kBAAkB,OAAO,oBAAoB;GAC7C,QAAQ,OAAO,UAAU,sBAAsB,OAAO,IAAI;EAC5D;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,OAAO,MAAM,YAAY,WAAW,0BAA0B;GACjE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAC3B,IAAIC,gDAAgB,KAAK,aAAa,UAAU,OAAO,CAAC,GACxD,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,MACtD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,SAAS,SAAS,QAAQ,SAAS,WAAW,CAAC;EACrD,MAAM,eAAe,cAAc,SAAS,UAAU;EACtD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAC9C,MAAM,YAAY,KAAK,iBAAiB,MAAM;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,2BAA2B;GAAE;GAAc;EAAM,CAAC;EAE5F,OAAO;GACL,SAAS,KAAK,YAAY,MAAM;GAChC;GACA;GACA;EACF;CACF;;;;;;;CAQA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,gCAAgC;GACvE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAC3B,IAAIC,sDAAsB,KAAK,aAAa,UAAU,OAAO,CAAC,GAC9D,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,MACtD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI;EACJ,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,6BAAa,IAAI,IAAwD;EAE/E,IAAI;GACF,WAAW,MAAM,SAAS,SAAS,UAAU,CAAC,GAAG;IAC/C,IAAI,MAAM,mBAAmB,OAAO,SAAS;KAC3C,MAAM,QAAQ,MAAM,kBAAkB,MAAM;KAE5C,WAAW,IAAI,MAAM,kBAAkB,qBAAqB,GAAG;MAC7D,IAAI,MAAM,aAAa;MACvB,MAAM,MAAM,QAAQ;MACpB,MAAM;KACR,CAAC;KAED;IACF;IAEA,IAAI,MAAM,mBAAmB,OAAO;KAClC,MAAM,QAAQ,MAAM,kBAAkB;KAEtC,IAAI,MAAM,MACR,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM;KAAK;UACtC,IAAI,MAAM,SAAS;MACxB,MAAM,cAAc,WAAW,IAAI,MAAM,kBAAkB,qBAAqB,CAAC;MAEjF,IAAI,aACF,YAAY,QAAQ,MAAM,QAAQ,SAAS;KAE/C;KAEA;IACF;IAEA,IAAI,MAAM,kBAAkB;KAC1B,MAAM,cAAc,WAAW,IAAI,MAAM,iBAAiB,qBAAqB,CAAC;KAEhF,IAAI,aAAa;MACf,MAAM;OACJ,MAAM;OACN,IAAI,YAAY;OAChB,MAAM,YAAY;OAClB,yCAA8C,YAAY,MAAM,CAAC,CAAC;MACpE;MAEA,WAAW,OAAO,MAAM,iBAAiB,qBAAqB,CAAC;KACjE;KAEA;IACF;IAEA,IAAI,MAAM,aACR,gBAAgB,MAAM,YAAY;IAGpC,IAAI,MAAM,UAAU,OAAO;KACzB,MAAM,MAAM,MAAM,SAAS;KAE3B,MAAM,QAAQ,IAAI,eAAe;KACjC,MAAM,SAAS,IAAI,gBAAgB;KACnC,MAAM,QAAQ,IAAI,eAAe,MAAM,QAAQ,MAAM;KAErD,IAAI,IAAI,wBAAwB,IAAI,uBAAuB,GACzD,MAAM,eAAe,IAAI;IAE7B;GACF;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,aAAa;EAEhD,KAAK,OAAO,MAAM,YAAY,YAAY,iCAAiC;GACzE;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;CAQA,AAAQ,aACN,UACA,SACiB;EACjB,MAAM,EAAE,QAAQ,UAAU,oBAAoB,kBAAkB,QAAQ;EACxE,MAAM,YAAY,SAAS,aAAa,KAAK,OAAO;EACpD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,SAAS,KAAK;GACd,UAAU;GACV,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;GAC3B,iBAAiB;IACf,GAAI,cAAc,SAAY,EAAE,UAAU,IAAI,CAAC;IAC/C,GAAI,gBAAgB,SAAY,EAAE,YAAY,IAAI,CAAC;GACrD;GACA,GAAG,KAAK,gBAAgB,SAAS,KAAK;GACtC,GAAG,KAAK,kBAAkB,SAAS,cAAc;EACnD;CACF;;;;;CAMA,AAAQ,gBAAgB,OAAuE;EAC7F,MAAM,aAAa,oBAAoB,KAAK;EAE5C,OAAO,aAAa,EAAE,WAAW,IAAI,CAAC;CACxC;;;;;;;;;;CAWA,AAAQ,kBACN,gBACuC;EACvC,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO,EACL,cAAc,EACZ,YAAY;GACV,MAAM;GACN,WAAW,EACT,YAAY;IAAE,MAAM;IAAY,QAAQ,KAAK,UAAU,cAAc;GAAE,EACzE;EACF,EACF,EACF;CACF;;;;;;CAOA,AAAQ,YAAY,QAAgC;EAClD,OAAO,OACJ,KAAK,UAAW,UAAU,SAAS,OAAO,MAAM,SAAS,WAAW,MAAM,OAAO,EAAG,CAAC,CACrF,KAAK,EAAE;CACZ;;;;;;CAOA,AAAQ,iBAAiB,QAA4D;EACnF,MAAM,YAAoC,CAAC;EAE3C,KAAK,MAAM,SAAS,QAClB,IAAI,aAAa,SAAS,MAAM,SAC9B,UAAU,KAAK;GACb,IAAI,MAAM,QAAQ,aAAa;GAC/B,MAAM,MAAM,QAAQ,QAAQ;GAC5B,OAAQ,MAAM,QAAQ,SAAS,CAAC;EAClC,CAAC;EAIL,OAAO,UAAU,SAAS,IAAI,YAAY;CAC5C;;;;;;CAOA,AAAQ,aAAa,KAAoC;EACvD,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,QAAQ,IAAI,eAAe;EACjC,MAAM,SAAS,IAAI,gBAAgB;EACnC,MAAM,SAAS,IAAI;EAEnB,OAAO;GACL;GACA;GACA,OAAO,IAAI,eAAe,QAAQ;GAClC,GAAI,UAAU,SAAS,IAAI,EAAE,cAAc,OAAO,IAAI,CAAC;EACzD;CACF;;;;;;CAOA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,iBAAiB,MAAM;EAEvC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;GACtD,MAAM,QAAQ;GACd,SAAS,QAAQ;EACnB,CAAC;EAED,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC7UA,IAAa,aAAb,MAAsD;CAKpD,AAAO,YAAY,QAA0B;EAC3C,MAAM,EAAE,UAAU,SAAS,GAAG,iBAAiB;EAE/C,KAAK,SAAS,IAAIC,qDAAqB,YAAY;EACnD,KAAK,WAAW,YAAY;EAC5B,KAAK,UAAU;CACjB;;;;;;;;;;;CAYA,AAAO,MAAM,QAA2C;EACtD,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,aAAa,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACpE;;;;;;;CAQA,MAAa,MAAM,MAAc,QAAkC;EACjE,iDAA6B,IAAI;CACnC;;;;;;;;;CAUA,AAAO,SAAS,QAAiD;EAC/D,OAAO,IAAI,gBAAgB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CAC/D;AACF"}
+{"version":3,"file":"index.cjs","names":["InvalidRequestError","AIError","ProviderTimeoutError","ProviderAuthError","QuotaExceededError","ProviderRateLimitError","ContextLengthExceededError","InvalidRequestError","ProviderError","LOG_MODULE","log","InvokeModelCommand","log","ConverseCommand","ConverseStreamCommand","BedrockRuntimeClient"],"sources":["../../../../../../@warlock.js/ai-bedrock/src/utils/map-stop-reason.ts","../../../../../../@warlock.js/ai-bedrock/src/utils/to-bedrock-messages.ts","../../../../../../@warlock.js/ai-bedrock/src/utils/to-bedrock-tools.ts","../../../../../../@warlock.js/ai-bedrock/src/utils/wrap-bedrock-error.ts","../../../../../../@warlock.js/ai-bedrock/src/embedder.ts","../../../../../../@warlock.js/ai-bedrock/src/known-capabilities.ts","../../../../../../@warlock.js/ai-bedrock/src/known-vision-models.ts","../../../../../../@warlock.js/ai-bedrock/src/model.ts","../../../../../../@warlock.js/ai-bedrock/src/sdk.ts"],"sourcesContent":["import type { FinishReason } from \"@warlock.js/ai\";\n\nconst stopReasonMap: Record<string, FinishReason> = {\n  end_turn: \"stop\",\n  stop_sequence: \"stop\",\n  max_tokens: \"length\",\n  tool_use: \"tool_calls\",\n};\n\n/**\n * Map Bedrock Converse's `stopReason` to the normalized `FinishReason`\n * union.\n *\n * `end_turn` / `stop_sequence` are natural stops. `max_tokens` maps to\n * `length`. `tool_use` maps to `tool_calls`. Everything else —\n * `content_filtered`, `guardrail_intervened`, `malformed_tool_use`,\n * `malformed_model_output`, `model_context_window_exceeded`, `null`,\n * or any future value — falls through to `\"error\"`: none produced a\n * clean terminal answer, so the agent must not treat them as success.\n *\n * @example\n * mapStopReason(\"end_turn\");              // \"stop\"\n * mapStopReason(\"tool_use\");              // \"tool_calls\"\n * mapStopReason(\"guardrail_intervened\");  // \"error\"\n * mapStopReason(undefined);               // \"error\"\n */\nexport function mapStopReason(raw: string | null | undefined): FinishReason {\n  return stopReasonMap[raw ?? \"\"] ?? \"error\";\n}\n","import { InvalidRequestError, type ContentPart, type Message } from \"@warlock.js/ai\";\nimport type {\n  ContentBlock,\n  ImageFormat,\n  Message as BedrockMessage,\n  SystemContentBlock,\n} from \"@aws-sdk/client-bedrock-runtime\";\n\n/**\n * Result of splitting a vendor-neutral `Message[]` for the Bedrock\n * Converse API: system prompts hoist to a separate `SystemContentBlock[]`\n * (Converse has no `\"system\"` role inside `messages`), and the\n * remaining turns map to Bedrock `Message[]`.\n */\nexport type BedrockMessages = {\n  system: SystemContentBlock[] | undefined;\n  messages: BedrockMessage[];\n};\n\nconst MEDIA_TYPE_TO_FORMAT: Record<string, ImageFormat> = {\n  \"image/jpeg\": \"jpeg\",\n  \"image/png\": \"png\",\n  \"image/gif\": \"gif\",\n  \"image/webp\": \"webp\",\n};\n\n/**\n * Convert vendor-neutral `Message[]` into Bedrock Converse's request\n * shape.\n *\n * Converse differs from the OpenAI Chat protocol in three ways this\n * function absorbs:\n *\n * 1. **No `system` role.** System messages become a separate\n *    `SystemContentBlock[]` (one `{ text }` block each).\n * 2. **Tool results are `user` turns.** A neutral `tool` message\n *    becomes a `user` message carrying a single `toolResult` block.\n * 3. **Tool calls are `toolUse` content blocks.** An assistant message\n *    with `toolCalls` becomes an `assistant` message: an optional\n *    leading `text` block followed by one `toolUse` block per call.\n *\n * @example\n * const { system, messages } = toBedrockMessages([\n *   { role: \"system\", content: \"Be concise.\" },\n *   { role: \"user\", content: \"Hi\" },\n * ]);\n */\nexport function toBedrockMessages(messages: Message[]): BedrockMessages {\n  const system: SystemContentBlock[] = [];\n  const mapped: BedrockMessage[] = [];\n\n  for (const message of messages) {\n    if (message.role === \"system\") {\n      system.push({ text: stringifyContent(message.content) });\n\n      continue;\n    }\n\n    if (message.role === \"tool\") {\n      mapped.push({\n        role: \"user\",\n        content: [\n          {\n            toolResult: {\n              toolUseId: message.toolCallId ?? \"\",\n              content: [{ text: stringifyContent(message.content) }],\n            },\n          },\n        ],\n      });\n\n      continue;\n    }\n\n    if (message.role === \"assistant\" && message.toolCalls && message.toolCalls.length > 0) {\n      const blocks: ContentBlock[] = [];\n      const text = stringifyContent(message.content);\n\n      if (text) {\n        blocks.push({ text });\n      }\n\n      for (const toolCall of message.toolCalls) {\n        blocks.push({\n          toolUse: {\n            toolUseId: toolCall.id,\n            name: toolCall.name,\n            input: toolCall.input ?? {},\n          },\n        } as ContentBlock);\n      }\n\n      mapped.push({ role: \"assistant\", content: blocks });\n\n      continue;\n    }\n\n    if (message.role === \"user\" && Array.isArray(message.content)) {\n      mapped.push({\n        role: \"user\",\n        content: message.content.map(toBedrockContentBlock),\n      });\n\n      continue;\n    }\n\n    mapped.push({\n      role: message.role === \"assistant\" ? \"assistant\" : \"user\",\n      content: [{ text: stringifyContent(message.content) }],\n    });\n  }\n\n  return {\n    system: system.length > 0 ? system : undefined,\n    messages: mapped,\n  };\n}\n\n/**\n * Multipart content is only meaningful on user messages — for any other\n * role collapse a `ContentPart[]` to its concatenated text. Plain\n * strings pass through unchanged.\n */\nfunction stringifyContent(content: string | ContentPart[]): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n\n  return content\n    .filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n    .map((part) => part.text)\n    .join(\"\");\n}\n\n/**\n * Map a resolved `ContentPart` to a Bedrock `ContentBlock`. Bedrock's\n * `ImageSource` only accepts raw bytes or an S3 location — there is no\n * remote-URL source. A neutral `{ url }` image therefore cannot be\n * sent and surfaces a typed `InvalidRequestError` upfront rather than\n * a downstream Bedrock validation fault. The agent has already\n * resolved attachments, so this never fetches or reads anything.\n */\nfunction toBedrockContentBlock(part: ContentPart): ContentBlock {\n  if (part.type === \"text\") {\n    return { text: part.text };\n  }\n\n  if (\"url\" in part.source) {\n    throw new InvalidRequestError(\n      \"Bedrock Converse does not support remote-URL image sources; supply base64 image bytes instead.\",\n    );\n  }\n\n  const format = MEDIA_TYPE_TO_FORMAT[part.source.mediaType];\n\n  if (!format) {\n    throw new InvalidRequestError(\n      `Unsupported image media type for Bedrock: \"${part.source.mediaType}\" (expected image/jpeg, image/png, image/gif, or image/webp).`,\n    );\n  }\n\n  return {\n    image: {\n      format,\n      source: { bytes: Buffer.from(part.source.base64, \"base64\") },\n    },\n  };\n}\n","import { extractJsonSchema, type ToolConfig } from \"@warlock.js/ai\";\nimport type { Tool, ToolConfiguration, ToolInputSchema } from \"@aws-sdk/client-bedrock-runtime\";\n\n/**\n * Convert vendor-neutral `ToolConfig[]` into Bedrock Converse's\n * `ToolConfiguration`. Each tool becomes a `toolSpec` with a JSON\n * `inputSchema`. Bedrock requires the schema root to be an object —\n * a non-object extraction degrades to a parameterless object schema\n * so registration never fails.\n *\n * Returns `undefined` when there are no tools so the caller can omit\n * `toolConfig` from the request entirely (Bedrock rejects an empty\n * `tools` array).\n *\n * @example\n * const toolConfig = toBedrockToolConfig([weatherTool]);\n * await client.send(new ConverseCommand({ modelId, messages, toolConfig }));\n */\nexport function toBedrockToolConfig(\n  tools: ToolConfig<unknown, unknown>[] | undefined,\n): ToolConfiguration | undefined {\n  if (!tools || tools.length === 0) {\n    return undefined;\n  }\n\n  return {\n    tools: tools.map(\n      (tool): Tool => ({\n        toolSpec: {\n          name: tool.name,\n          description: tool.description,\n          inputSchema: { json: toJsonSchema(tool.input) } as ToolInputSchema,\n        },\n      }),\n    ),\n  };\n}\n\n/**\n * Resolve a tool's input schema to a JSON-Schema object. Bedrock's\n * `ToolInputSchema.json` requires an object root; anything else (or a\n * failed extraction) degrades to a parameterless object so the tool\n * still registers.\n */\nfunction toJsonSchema(input: ToolConfig<unknown, unknown>[\"input\"]): Record<string, unknown> {\n  const schema = extractJsonSchema(input);\n\n  if (schema && schema.type === \"object\") {\n    return schema;\n  }\n\n  return { type: \"object\" };\n}\n","import {\n  AIError,\n  ContextLengthExceededError,\n  InvalidRequestError,\n  ProviderAuthError,\n  ProviderError,\n  ProviderRateLimitError,\n  ProviderTimeoutError,\n  QuotaExceededError,\n} from \"@warlock.js/ai\";\n\n/**\n * Raw-error fields the wrapper reads off an AWS SDK exception. Every\n * Bedrock error is a Smithy `__BaseException` with a stable `name`\n * (`\"ThrottlingException\"`, `\"ValidationException\"`, …) and a\n * `$metadata` carrying `httpStatusCode` + `requestId`. We duck-type\n * because retries and proxies sometimes flatten the prototype chain.\n */\ntype BedrockErrorShape = {\n  name?: string;\n  message?: string;\n  httpStatusCode?: number;\n  requestId?: string;\n  code?: string;\n};\n\nconst TIMEOUT_NAMES = new Set([\n  \"ModelTimeoutException\",\n  \"TimeoutError\",\n  \"RequestTimeout\",\n  \"RequestTimeoutException\",\n]);\n\n/**\n * Wrap any thrown value caught inside the Bedrock adapter into the\n * appropriate `@warlock.js/ai` `AIError` subclass.\n *\n * **Dispatch strategy.** AWS errors carry no provider machine `code`;\n * the stable identifier is the Smithy exception `name`. Dispatch keys\n * on `name`, falls back to `$metadata.httpStatusCode` when the name is\n * missing (flattened/proxied errors). `ValidationException` is split:\n * the \"input is too long / exceeds context window\" phrasing maps to\n * `ContextLengthExceededError`, everything else to\n * `InvalidRequestError`.\n *\n * `AIError` instances pass through unchanged so `catch/throw wrap(e)`\n * pipelines never double-wrap.\n *\n * @example\n * try {\n *   return await this.client.send(new ConverseCommand(...));\n * } catch (thrown) {\n *   throw wrapBedrockError(thrown);\n * }\n */\nexport function wrapBedrockError(thrown: unknown): AIError {\n  if (thrown instanceof AIError) {\n    return thrown;\n  }\n\n  const shape = toShape(thrown);\n  const context = buildContext(shape);\n  const message = shape.message ?? (thrown instanceof Error ? thrown.message : String(thrown));\n\n  if (isTimeout(shape)) {\n    return new ProviderTimeoutError(message, { cause: thrown, context });\n  }\n\n  if (shape.name === \"AccessDeniedException\" || shape.httpStatusCode === 403) {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.httpStatusCode === 401) {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.name === \"ServiceQuotaExceededException\") {\n    return new QuotaExceededError(message, { cause: thrown, context });\n  }\n\n  if (shape.name === \"ThrottlingException\" || shape.httpStatusCode === 429) {\n    return new ProviderRateLimitError(message, { cause: thrown, context });\n  }\n\n  if (shape.name === \"ValidationException\") {\n    if (/too long|context window|maximum context|exceeds the maximum/i.test(message)) {\n      return new ContextLengthExceededError(message, { cause: thrown, context });\n    }\n\n    return new InvalidRequestError(message, { cause: thrown, context });\n  }\n\n  if (\n    shape.name === \"ResourceNotFoundException\" ||\n    shape.name === \"ConflictException\" ||\n    isClientStatus(shape.httpStatusCode)\n  ) {\n    return new InvalidRequestError(message, { cause: thrown, context });\n  }\n\n  return new ProviderError(message, { cause: thrown, context });\n}\n\n/**\n * Read the raw error shape without depending on `instanceof`. AWS\n * exceptions expose `$metadata`; plain/proxied errors may carry\n * `status` / `code` instead.\n */\nfunction toShape(thrown: unknown): BedrockErrorShape {\n  if (typeof thrown !== \"object\" || thrown === null) {\n    return {};\n  }\n\n  const raw = thrown as Record<string, unknown>;\n  const metadata = raw.$metadata as { httpStatusCode?: number; requestId?: string } | undefined;\n\n  return {\n    name: typeof raw.name === \"string\" ? raw.name : undefined,\n    message: typeof raw.message === \"string\" ? raw.message : undefined,\n    httpStatusCode:\n      metadata && typeof metadata.httpStatusCode === \"number\"\n        ? metadata.httpStatusCode\n        : typeof raw.status === \"number\"\n          ? (raw.status as number)\n          : undefined,\n    requestId: metadata && typeof metadata.requestId === \"string\" ? metadata.requestId : undefined,\n    code: typeof raw.code === \"string\" ? raw.code : undefined,\n  };\n}\n\n/**\n * Decide whether the error is a timeout. Bedrock surfaces\n * `ModelTimeoutException`; the AWS transport layer surfaces\n * `TimeoutError` / `ETIMEDOUT` / `ECONNABORTED`.\n */\nfunction isTimeout(shape: BedrockErrorShape): boolean {\n  if (shape.name && TIMEOUT_NAMES.has(shape.name)) {\n    return true;\n  }\n\n  return shape.code === \"ETIMEDOUT\" || shape.code === \"ECONNABORTED\";\n}\n\n/** True for HTTP 4xx — a client-side request problem, not a server fault. */\nfunction isClientStatus(status: number | undefined): boolean {\n  return typeof status === \"number\" && status >= 400 && status < 500;\n}\n\n/**\n * Attach the raw diagnostic fields to `error.context`. The Smithy\n * exception `name` is the closest thing Bedrock has to a stable error\n * code, so it lands on `context.code`.\n */\nfunction buildContext(shape: BedrockErrorShape): Record<string, unknown> {\n  const context: Record<string, unknown> = {};\n\n  if (shape.httpStatusCode !== undefined) {\n    context.status = shape.httpStatusCode;\n  }\n\n  if (shape.name) {\n    context.code = shape.name;\n  }\n\n  if (shape.requestId) {\n    context.requestId = shape.requestId;\n  }\n\n  return context;\n}\n","import {\n  type EmbeddingBatchResult,\n  type EmbeddingResult,\n  type EmbeddingUsage,\n  type EmbedderContract,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport { InvokeModelCommand, type BedrockRuntimeClient } from \"@aws-sdk/client-bedrock-runtime\";\nimport type { BedrockEmbedderConfig } from \"./config.type\";\nimport { wrapBedrockError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.bedrock\";\n\n/** Shape of the Amazon Titan Text Embeddings response body. */\ntype TitanEmbeddingResponse = {\n  embedding: number[];\n  inputTextTokenCount: number;\n};\n\n/**\n * Bedrock-backed implementation of `EmbedderContract`, targeting the\n * Amazon Titan Text Embeddings family\n * (`amazon.titan-embed-text-v2:0` / v1) via `InvokeModel`.\n *\n * **Role.** Converts text into floating-point vectors. Standalone\n * primitive — unrelated to Converse / tools / the agent loop.\n *\n * **Single-input only upstream.** Titan's `InvokeModel` body accepts\n * one `inputText` per call — there is no batch endpoint. `embedMany`\n * therefore issues one request per input sequentially and aggregates\n * token usage. This is a deliberate, documented trade-off: a real\n * batch API does not exist for Titan on Bedrock, so the alternative\n * (failing `embedMany`) would be worse. Cohere embeddings on Bedrock\n * *do* batch but use an incompatible body shape — out of scope; use\n * the OpenAI adapter or a future Cohere adapter when batch throughput\n * matters.\n *\n * **Dimensions.** When no `dimensions` override is given,\n * `this.dimensions` starts at `0` and is populated from the first\n * response's vector length, then cached. Passing `dimensions` forwards\n * Titan v2's truncation hint (256 / 512 / 1024) and sets the initial\n * value immediately.\n *\n * @example\n * const embedder = new BedrockEmbedder(client, { name: \"amazon.titan-embed-text-v2:0\" });\n * const { vector } = await embedder.embed(\"Hello world\");\n * const { vectors } = await embedder.embedMany([\"doc 1\", \"doc 2\"]);\n */\nexport class BedrockEmbedder implements EmbedderContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public dimensions: number;\n\n  private readonly client: BedrockRuntimeClient;\n  private readonly configuredDimensions: number | undefined;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: BedrockRuntimeClient,\n    config: BedrockEmbedderConfig,\n    provider: string = \"bedrock\",\n  ) {\n    this.client = client;\n    this.name = config.name;\n    this.provider = provider;\n    this.configuredDimensions = config.dimensions;\n    this.dimensions = config.dimensions ?? 0;\n  }\n\n  public async embed(input: string): Promise<EmbeddingResult> {\n    const { vector, tokens } = await this.invoke(input);\n\n    return {\n      vector,\n      dimensions: this.dimensions,\n      usage: { promptTokens: tokens, totalTokens: tokens },\n    };\n  }\n\n  public async embedMany(inputs: string[]): Promise<EmbeddingBatchResult> {\n    const vectors: number[][] = [];\n    let tokens = 0;\n\n    for (const input of inputs) {\n      const result = await this.invoke(input);\n\n      vectors.push(result.vector);\n      tokens += result.tokens;\n    }\n\n    const usage: EmbeddingUsage = { promptTokens: tokens, totalTokens: tokens };\n\n    return { vectors, dimensions: this.dimensions, usage };\n  }\n\n  /**\n   * Issue a single Titan `InvokeModel` embedding request: encode the\n   * JSON body, send, wrap provider errors, decode the response, and\n   * cache `dimensions` on the first successful call.\n   */\n  private async invoke(input: string): Promise<{ vector: number[]; tokens: number }> {\n    this.logger.debug(LOG_MODULE, \"embedder.request\", \"InvokeModel embeddings\", {\n      model: this.name,\n    });\n\n    const body = JSON.stringify({\n      inputText: input,\n      ...(this.configuredDimensions !== undefined\n        ? { dimensions: this.configuredDimensions }\n        : {}),\n    });\n\n    let raw;\n\n    try {\n      raw = await this.client.send(\n        new InvokeModelCommand({\n          modelId: this.name,\n          contentType: \"application/json\",\n          accept: \"application/json\",\n          body: new TextEncoder().encode(body),\n        }),\n      );\n    } catch (thrown) {\n      const wrapped = wrapBedrockError(thrown);\n\n      this.logger.error(LOG_MODULE, \"embedder.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    const decoded = JSON.parse(new TextDecoder().decode(raw.body)) as TitanEmbeddingResponse;\n\n    if (this.dimensions === 0) {\n      this.dimensions = decoded.embedding.length;\n    }\n\n    this.logger.debug(LOG_MODULE, \"embedder.response\", \"InvokeModel embeddings returned\", {\n      dimensions: decoded.embedding.length,\n      tokens: decoded.inputTextTokenCount,\n    });\n\n    return { vector: decoded.embedding, tokens: decoded.inputTextTokenCount };\n  }\n}\n","/**\n * Cost-truth capability inference for Bedrock Converse model ids.\n *\n * Bedrock model ids are provider-prefixed and version-suffixed\n * (`anthropic.claude-3-7-sonnet-20250219-v1:0`, `us.amazon.nova-pro-v1:0`),\n * so — exactly like `known-vision-models.ts` — a lowercase substring scan\n * is the only robust check across cross-region inference-profile prefixes\n * (`us.`, `eu.`, `apac.`) and date/version tags.\n *\n * Each predicate answers a single `ModelCapabilities` flag the agent reads\n * to decide whether to forward a cost-truth option (`reasoning`,\n * `cacheControl`) or up-front-reject an attachment (`pdf`). Unknown ids\n * default to `false` so an unsupported request fails fast with a clear\n * capability error instead of an opaque Bedrock `ValidationException`.\n * Every inference is overridable per-model via `bedrock.model({ name, … })`.\n */\n\n/**\n * Families that expose Anthropic-style extended thinking on Bedrock\n * Converse via `additionalModelRequestFields.thinking`. Only Claude 3.7\n * and the Claude 4 line (Sonnet / Opus / Haiku) support a configurable\n * thinking budget; earlier Claude, Nova, Llama, Mistral and Cohere do\n * not, so they are intentionally absent.\n */\nconst REASONING_CAPABLE_SUBSTRINGS = [\n  \"claude-3-7\",\n  \"claude-sonnet-4\",\n  \"claude-opus-4\",\n  \"claude-haiku-4\",\n];\n\n/**\n * Families that honor Converse `cachePoint` prompt-cache breakpoints.\n * Anthropic Claude 3.5+ / 3.7 / 4 and the Amazon Nova line support\n * cache points; text-only legacy families do not.\n */\nconst PROMPT_CACHING_CAPABLE_SUBSTRINGS = [\n  \"claude-3-5\",\n  \"claude-3-7\",\n  \"claude-sonnet-4\",\n  \"claude-opus-4\",\n  \"claude-haiku-4\",\n  \"nova-lite\",\n  \"nova-pro\",\n  \"nova-premier\",\n  \"nova-micro\",\n];\n\n/**\n * Families that accept Converse `document` content blocks (PDF / docx /\n * txt input). The multimodal Claude 3+ and Nova families support\n * document blocks; the substring set mirrors the vision-capable list\n * minus the image-only Llama entries (Llama on Bedrock takes images but\n * not document blocks via Converse).\n */\nconst PDF_CAPABLE_SUBSTRINGS = [\n  \"claude-3\",\n  \"claude-sonnet-4\",\n  \"claude-opus-4\",\n  \"claude-haiku-4\",\n  \"nova-lite\",\n  \"nova-pro\",\n  \"nova-premier\",\n];\n\nfunction matchesAny(modelId: string, fragments: string[]): boolean {\n  const normalized = modelId.toLowerCase();\n\n  return fragments.some((fragment) => normalized.includes(fragment));\n}\n\n/**\n * Infer whether a Bedrock model id exposes extended-thinking / reasoning\n * (Claude 3.7 + Claude 4). When true the adapter forwards\n * `ModelCallOptions.reasoning` as Converse\n * `additionalModelRequestFields.thinking`.\n *\n * @example\n * inferReasoningCapability(\"anthropic.claude-3-7-sonnet-20250219-v1:0\"); // → true\n * inferReasoningCapability(\"us.amazon.nova-pro-v1:0\");                    // → false\n */\nexport function inferReasoningCapability(modelId: string): boolean {\n  return matchesAny(modelId, REASONING_CAPABLE_SUBSTRINGS);\n}\n\n/**\n * Infer whether a Bedrock model id honors Converse `cachePoint`\n * breakpoints (Claude 3.5+ / Nova). When true the adapter both maps\n * `ModelCallOptions.cacheControl` write breakpoints to cache points and\n * reports `Usage.cachedTokens` / `Usage.cacheWriteTokens`.\n *\n * @example\n * inferPromptCachingCapability(\"us.amazon.nova-pro-v1:0\");                       // → true\n * inferPromptCachingCapability(\"meta.llama3-1-8b-instruct-v1:0\");                // → false\n */\nexport function inferPromptCachingCapability(modelId: string): boolean {\n  return matchesAny(modelId, PROMPT_CACHING_CAPABLE_SUBSTRINGS);\n}\n\n/**\n * Infer whether a Bedrock model id accepts Converse `document` content\n * blocks (PDF / document input — Claude 3+ / Nova). When false the agent\n * rejects a PDF attachment up front instead of dropping it at the wire.\n *\n * @example\n * inferPdfCapability(\"anthropic.claude-3-5-sonnet-20240620-v1:0\"); // → true\n * inferPdfCapability(\"meta.llama3-2-90b-instruct-v1:0\");           // → false\n */\nexport function inferPdfCapability(modelId: string): boolean {\n  return matchesAny(modelId, PDF_CAPABLE_SUBSTRINGS);\n}\n","/**\n * Substrings that identify Bedrock model ids whose family accepts image\n * input on the Converse API.\n *\n * Bedrock model ids are provider-prefixed and version-suffixed\n * (`anthropic.claude-3-5-sonnet-20240620-v1:0`, `us.amazon.nova-pro-v1:0`,\n * `meta.llama3-2-90b-instruct-v1:0`), so a substring match is the only\n * robust check across the cross-region inference-profile prefixes\n * (`us.`, `eu.`, `apac.`) and date/version tags.\n *\n * Multimodal families covered: Anthropic Claude 3 / 3.5 / 3.7 / 4,\n * Amazon Nova Lite/Pro/Premier, Meta Llama 3.2 (11B/90B) and Llama 4.\n * Text-only families (Llama 3/3.1, Titan Text, Mistral 7B, Cohere\n * Command) are intentionally absent. Override per-model via\n * `bedrock.model({ name, vision: true | false })`.\n */\nconst VISION_CAPABLE_SUBSTRINGS = [\n  \"claude-3\",\n  \"claude-sonnet-4\",\n  \"claude-opus-4\",\n  \"claude-haiku-4\",\n  \"nova-lite\",\n  \"nova-pro\",\n  \"nova-premier\",\n  \"llama3-2-11b\",\n  \"llama3-2-90b\",\n  \"llama4\",\n];\n\n/**\n * Infer whether a Bedrock model id supports vision based on the known\n * multimodal-family substrings. Unknown ids default to `false` so that\n * passing an image attachment to an unsupported model surfaces a clear,\n * agent-side capability error instead of an opaque Bedrock validation\n * fault.\n *\n * @example\n * inferVisionCapability(\"anthropic.claude-3-5-sonnet-20240620-v1:0\"); // → true\n * inferVisionCapability(\"us.amazon.nova-pro-v1:0\");                    // → true\n * inferVisionCapability(\"meta.llama3-1-8b-instruct-v1:0\");             // → false\n * inferVisionCapability(\"amazon.titan-text-express-v1\");               // → false\n */\nexport function inferVisionCapability(modelId: string): boolean {\n  const normalized = modelId.toLowerCase();\n\n  return VISION_CAPABLE_SUBSTRINGS.some((fragment) => normalized.includes(fragment));\n}\n","import {\n  safeJsonParse,\n  type Message,\n  type ModelCallOptions,\n  type ModelCapabilities,\n  type ModelContract,\n  type ModelPricing,\n  type ModelResponse,\n  type ModelStreamChunk,\n  type ModelToolCallRequest,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport {\n  ConverseCommand,\n  ConverseStreamCommand,\n  type BedrockRuntimeClient,\n  type ContentBlock,\n  type ConverseRequest,\n  type TokenUsage,\n} from \"@aws-sdk/client-bedrock-runtime\";\nimport type { BedrockModelConfig } from \"./config.type\";\nimport {\n  inferPdfCapability,\n  inferPromptCachingCapability,\n  inferReasoningCapability,\n} from \"./known-capabilities\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapStopReason, toBedrockMessages, toBedrockToolConfig, wrapBedrockError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.bedrock\";\n\n/**\n * Conventional extended-thinking token budgets for the neutral\n * `reasoning.effort` levels, used when the caller asks for an effort\n * tier without naming an explicit `reasoning.maxTokens` budget. Mirrors\n * the low / medium / high spread other reasoning adapters expose so the\n * vendor-neutral option behaves consistently across providers.\n */\nconst EFFORT_THINKING_BUDGET: Record<string, number | undefined> = {\n  low: 1024,\n  medium: 4096,\n  high: 16384,\n};\n\n/**\n * Bedrock-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and AWS Bedrock's Converse /\n * ConverseStream API. Converse is the model-agnostic surface — one\n * wire mapping covers every Bedrock-hosted family (Anthropic Claude,\n * Amazon Nova, Meta Llama, Mistral, Cohere) instead of per-family\n * `InvokeModel` body shapes.\n *\n * **Responsibility.**\n * - Owns: a long-lived `BedrockRuntimeClient` + frozen `ModelConfig`\n *   (modelId, temperature, maxTokens) used as per-call defaults.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Converse shapes (system hoisting, `toolUse` / `toolResult` blocks,\n *   image bytes) on the way out, and Converse's content-block response\n *   (text, tool calls, stop reason, token usage) back into the neutral\n *   shapes on the way in.\n * - Does NOT own: dispatching tools, looping, history, retries — those\n *   are agent concerns. The model is a per-call protocol adapter.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across calls\"): the AWS client is heavy to construct and reused for\n * the SDK's lifetime.\n *\n * @example\n * import { BedrockRuntimeClient } from \"@aws-sdk/client-bedrock-runtime\";\n * const client = new BedrockRuntimeClient({ region: \"us-east-1\" });\n * const model = new BedrockModel(client, {\n *   name: \"anthropic.claude-sonnet-4-5-20250929-v1:0\",\n * });\n *\n * const myAgent = agent({ model, tools: [searchTool] });\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class BedrockModel implements ModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly capabilities: ModelCapabilities;\n  public readonly pricing?: ModelPricing;\n\n  private readonly client: BedrockRuntimeClient;\n  private readonly config: BedrockModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: BedrockRuntimeClient,\n    config: BedrockModelConfig,\n    provider: string = \"bedrock\",\n  ) {\n    this.client = client;\n    this.config = config;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    this.capabilities = {\n      structuredOutput: config.structuredOutput ?? true,\n      vision: config.vision ?? inferVisionCapability(config.name),\n      reasoning: config.reasoning ?? inferReasoningCapability(config.name),\n      promptCaching: config.promptCaching ?? inferPromptCachingCapability(config.name),\n      pdf: config.pdf ?? inferPdfCapability(config.name),\n      audio: config.audio ?? false,\n    };\n  }\n\n  /**\n   * Single-shot completion via the Converse API. Sends the full\n   * message list, waits for the terminal response, and reshapes it\n   * into a vendor-neutral `ModelResponse`. Per-call `options` override\n   * the instance defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting Converse call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response;\n\n    try {\n      response = await this.client.send(\n        new ConverseCommand(this.buildRequest(messages, options)),\n        options?.signal ? { abortSignal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const blocks = response.output?.message?.content ?? [];\n    const finishReason = mapStopReason(response.stopReason);\n    const usage = this.extractUsage(response.usage);\n    const toolCalls = this.extractToolCalls(blocks);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Converse call succeeded\", { finishReason, usage });\n\n    return {\n      content: this.extractText(blocks),\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion via ConverseStream. Yields neutral\n   * `ModelStreamChunk`s — `delta` for text, `tool-call` once a\n   * `toolUse` block's accumulated input JSON is complete, and a\n   * terminal `done` with the final finish reason + usage totals.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting ConverseStream call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response;\n\n    try {\n      response = await this.client.send(\n        new ConverseStreamCommand(this.buildRequest(messages, options)),\n        options?.signal ? { abortSignal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawStopReason: string | undefined;\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n    const toolBlocks = new Map<number, { id: string; name: string; json: string }>();\n\n    try {\n      for await (const event of response.stream ?? []) {\n        if (event.contentBlockStart?.start?.toolUse) {\n          const start = event.contentBlockStart.start.toolUse;\n\n          toolBlocks.set(event.contentBlockStart.contentBlockIndex ?? 0, {\n            id: start.toolUseId ?? \"\",\n            name: start.name ?? \"\",\n            json: \"\",\n          });\n\n          continue;\n        }\n\n        if (event.contentBlockDelta?.delta) {\n          const delta = event.contentBlockDelta.delta;\n\n          if (delta.text) {\n            yield { type: \"delta\", content: delta.text };\n          } else if (delta.toolUse) {\n            const accumulator = toolBlocks.get(event.contentBlockDelta.contentBlockIndex ?? 0);\n\n            if (accumulator) {\n              accumulator.json += delta.toolUse.input ?? \"\";\n            }\n          }\n\n          continue;\n        }\n\n        if (event.contentBlockStop) {\n          const accumulator = toolBlocks.get(event.contentBlockStop.contentBlockIndex ?? 0);\n\n          if (accumulator) {\n            yield {\n              type: \"tool-call\",\n              id: accumulator.id,\n              name: accumulator.name,\n              input: safeJsonParse<Record<string, unknown>>(accumulator.json, {}),\n            };\n\n            toolBlocks.delete(event.contentBlockStop.contentBlockIndex ?? 0);\n          }\n\n          continue;\n        }\n\n        if (event.messageStop) {\n          rawStopReason = event.messageStop.stopReason;\n        }\n\n        if (event.metadata?.usage) {\n          const raw = event.metadata.usage;\n\n          usage.input = raw.inputTokens ?? 0;\n          usage.output = raw.outputTokens ?? 0;\n          usage.total = raw.totalTokens ?? usage.input + usage.output;\n\n          if (raw.cacheReadInputTokens && raw.cacheReadInputTokens > 0) {\n            usage.cachedTokens = raw.cacheReadInputTokens;\n          }\n\n          if (raw.cacheWriteInputTokens && raw.cacheWriteInputTokens > 0) {\n            usage.cacheWriteTokens = raw.cacheWriteInputTokens;\n          }\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = mapStopReason(rawStopReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"ConverseStream call succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the Converse request shared by `complete()` and\n   * `stream()` (both command shapes take the same input). Hoists the\n   * system prompt, maps inference params, and conditionally attaches\n   * tools and native structured output.\n   */\n  private buildRequest(\n    messages: Message[],\n    options: ModelCallOptions | undefined,\n  ): ConverseRequest {\n    const { system, messages: bedrockMessages } = toBedrockMessages(messages);\n    const maxTokens = options?.maxTokens ?? this.config.maxTokens;\n    const temperature = options?.temperature ?? this.config.temperature;\n    const cachedMessages = this.applyCacheBreakpoints(bedrockMessages, options?.cacheControl);\n\n    return {\n      modelId: this.name,\n      messages: cachedMessages,\n      ...(system ? { system } : {}),\n      inferenceConfig: {\n        ...(maxTokens !== undefined ? { maxTokens } : {}),\n        ...(temperature !== undefined ? { temperature } : {}),\n      },\n      ...this.buildToolConfig(options?.tools),\n      ...this.buildOutputConfig(options?.responseSchema),\n      ...this.buildReasoningConfig(options?.reasoning),\n    };\n  }\n\n  /**\n   * Append a Converse `cachePoint` block to the LAST message when the\n   * caller supplies a `cacheControl` write breakpoint and the model is\n   * `promptCaching`-capable. A cache point tells Bedrock to cache the\n   * whole prefix up to that block, so subsequent calls reusing the same\n   * prefix bill the cached portion at the discounted read rate\n   * (surfaced as `Usage.cachedTokens`). No-ops gracefully when caching\n   * is unsupported, no breakpoint was requested, or there are no\n   * messages to mark — Bedrock then prices the call normally.\n   *\n   * Bedrock only honors `CachePointType.DEFAULT`; the neutral\n   * `breakpoints` count is a presence hint (one trailing breakpoint is\n   * the only placement Converse supports without manual block surgery),\n   * so any positive value marks the trailing message.\n   */\n  private applyCacheBreakpoints(\n    messages: ConverseRequest[\"messages\"],\n    cacheControl: ModelCallOptions[\"cacheControl\"],\n  ): ConverseRequest[\"messages\"] {\n    const breakpoints = cacheControl?.breakpoints ?? 0;\n\n    if (!this.capabilities.promptCaching || breakpoints <= 0 || !messages || messages.length === 0) {\n      return messages;\n    }\n\n    const last = messages.length - 1;\n    const lastMessage = messages[last];\n\n    return [\n      ...messages.slice(0, last),\n      {\n        ...lastMessage,\n        content: [...(lastMessage.content ?? []), { cachePoint: { type: \"default\" } }],\n      },\n    ];\n  }\n\n  /**\n   * Translate the neutral `reasoning` option into Claude-on-Bedrock's\n   * extended-thinking control, carried in Converse's escape hatch\n   * `additionalModelRequestFields.thinking`. Emitted only when the model\n   * is `reasoning`-capable and a budget can be resolved — `maxTokens`\n   * (explicit thinking budget) wins, otherwise `effort` maps to a\n   * conventional token budget so callers can opt in without picking a\n   * number. Returns an empty object (no-op) for non-reasoning models or\n   * when no reasoning option was supplied, so unsupported params never\n   * reach the wire.\n   */\n  private buildReasoningConfig(\n    reasoning: ModelCallOptions[\"reasoning\"],\n  ): Pick<ConverseRequest, \"additionalModelRequestFields\"> {\n    if (!this.capabilities.reasoning || !reasoning) {\n      return {};\n    }\n\n    const budgetTokens = reasoning.maxTokens ?? EFFORT_THINKING_BUDGET[reasoning.effort ?? \"\"];\n\n    if (budgetTokens === undefined) {\n      return {};\n    }\n\n    return {\n      additionalModelRequestFields: {\n        thinking: { type: \"enabled\", budget_tokens: budgetTokens },\n      },\n    };\n  }\n\n  /**\n   * Spread-friendly tool fragment. Returns an empty object when no\n   * tools were supplied (Bedrock rejects an empty `tools` array).\n   */\n  private buildToolConfig(tools: ModelCallOptions[\"tools\"]): Pick<ConverseRequest, \"toolConfig\"> {\n    const toolConfig = toBedrockToolConfig(tools);\n\n    return toolConfig ? { toolConfig } : {};\n  }\n\n  /**\n   * Translate the neutral `responseSchema` into Converse's native\n   * `outputConfig.textFormat` (JSON-schema structured output). Bedrock\n   * requires the schema as a stringified JSON document and only\n   * accepts an object root. Emitted only when the model is\n   * `structuredOutput`-capable and the schema is an object — otherwise\n   * the agent's soft system-prompt hint + client-side `validate()`\n   * carry shape (same degradation philosophy as the OpenAI adapter).\n   */\n  private buildOutputConfig(\n    responseSchema: Record<string, unknown> | undefined,\n  ): Pick<ConverseRequest, \"outputConfig\"> {\n    if (!responseSchema || !this.capabilities.structuredOutput) {\n      return {};\n    }\n\n    if (responseSchema.type !== \"object\" || typeof responseSchema.properties !== \"object\") {\n      return {};\n    }\n\n    return {\n      outputConfig: {\n        textFormat: {\n          type: \"json_schema\",\n          structure: {\n            jsonSchema: { name: \"response\", schema: JSON.stringify(responseSchema) },\n          },\n        },\n      },\n    };\n  }\n\n  /**\n   * Concatenate every `text` content block into the single neutral\n   * `content` string. `toolUse` and other block types are surfaced\n   * separately via `extractToolCalls`.\n   */\n  private extractText(blocks: ContentBlock[]): string {\n    return blocks\n      .map((block) => (\"text\" in block && typeof block.text === \"string\" ? block.text : \"\"))\n      .join(\"\");\n  }\n\n  /**\n   * Reshape Converse `toolUse` content blocks into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when no tools were\n   * requested so callers can branch on presence.\n   */\n  private extractToolCalls(blocks: ContentBlock[]): ModelToolCallRequest[] | undefined {\n    const toolCalls: ModelToolCallRequest[] = [];\n\n    for (const block of blocks) {\n      if (\"toolUse\" in block && block.toolUse) {\n        toolCalls.push({\n          id: block.toolUse.toolUseId ?? \"\",\n          name: block.toolUse.name ?? \"\",\n          input: (block.toolUse.input ?? {}) as Record<string, unknown>,\n        });\n      }\n    }\n\n    return toolCalls.length > 0 ? toolCalls : undefined;\n  }\n\n  /**\n   * Normalize Converse's `TokenUsage` into the neutral `Usage` shape.\n   * Bedrock supplies a pre-summed `totalTokens`; cache-read and\n   * cache-write tokens are surfaced as `cachedTokens` /\n   * `cacheWriteTokens` only when non-zero so callers can price the\n   * discounted read rate and the one-time write cost separately.\n   * Bedrock's Converse `TokenUsage` carries no reasoning-token channel,\n   * so `Usage.reasoningTokens` is intentionally left unset here.\n   */\n  private extractUsage(raw: TokenUsage | undefined): Usage {\n    if (!raw) {\n      return { input: 0, output: 0, total: 0 };\n    }\n\n    const input = raw.inputTokens ?? 0;\n    const output = raw.outputTokens ?? 0;\n    const cached = raw.cacheReadInputTokens;\n    const cacheWrite = raw.cacheWriteInputTokens;\n\n    return {\n      input,\n      output,\n      total: raw.totalTokens ?? input + output,\n      ...(cached && cached > 0 ? { cachedTokens: cached } : {}),\n      ...(cacheWrite && cacheWrite > 0 ? { cacheWriteTokens: cacheWrite } : {}),\n    };\n  }\n\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy\n   * and emit the standard error log line before it propagates. Shared\n   * by every catch site so the log shape stays identical.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapBedrockError(thrown);\n\n    this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n      code: wrapped.code,\n      context: wrapped.context,\n    });\n\n    return wrapped;\n  }\n}\n","import { BedrockRuntimeClient } from \"@aws-sdk/client-bedrock-runtime\";\nimport type {\n  EmbedderContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n} from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type {\n  BedrockEmbedderConfig,\n  BedrockModelConfig,\n  BedrockSDKConfig,\n} from \"./config.type\";\nimport { BedrockEmbedder } from \"./embedder\";\nimport { BedrockModel } from \"./model\";\n\n/**\n * AWS Bedrock-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for any Bedrock-hosted model via\n * the Converse API. A single `BedrockSDK` holds one live\n * `BedrockRuntimeClient`, shared by every `ModelContract` and\n * `EmbedderContract` it produces. Construct one SDK per AWS\n * account/region and reuse it everywhere.\n *\n * **Responsibility.**\n * - Owns: a long-lived `BedrockRuntimeClient` (region, credential\n *   chain) and its lifetime. Factory for `BedrockModel` /\n *   `BedrockEmbedder` instances sharing that client.\n * - Does NOT own: anything per-call — those live in `BedrockModel` /\n *   `BedrockEmbedder` and the agent runtime.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"): the AWS client is heavy to construct and\n * designed for reuse; keeping it on `this` aligns with the\n * `new BedrockRuntimeClient(...)` upstream convention.\n *\n * @example\n * const bedrock = new BedrockSDK({ region: \"us-east-1\" });\n * const model = bedrock.model({ name: \"anthropic.claude-sonnet-4-5-20250929-v1:0\" });\n * const embedder = bedrock.embedder({ name: \"amazon.titan-embed-text-v2:0\" });\n */\nexport class BedrockSDK implements SDKAdapterContract {\n  private readonly client: BedrockRuntimeClient;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: BedrockSDKConfig) {\n    const { provider, pricing, ...clientConfig } = config;\n\n    this.client = new BedrockRuntimeClient(clientConfig);\n    this.provider = provider ?? \"bedrock\";\n    this.pricing = pricing;\n  }\n\n  /**\n   * Build a `BedrockModel` bound to this SDK's client. Each call\n   * returns a fresh instance; all instances share the underlying AWS\n   * client so connection pools, credential refresh, and retry config\n   * stay unified. The SDK's `provider` label is forwarded.\n   *\n   * Pricing resolution: per-model `config.pricing` wins; otherwise the\n   * SDK-level registry entry keyed by `config.name`; otherwise\n   * `undefined` (no cost computed).\n   */\n  public model(config: BedrockModelConfig): ModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: BedrockModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new BedrockModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate. Uses the character-heuristic\n   * (`approximateTokenCount`) from the core package — Bedrock has no\n   * offline tokenizer and the per-model tokenizers differ; good enough\n   * for budgeting and quota guards, not for billing.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n\n  /**\n   * Build a `BedrockEmbedder` (Amazon Titan Text Embeddings) bound to\n   * this SDK's client.\n   *\n   * @example\n   * const embedder = bedrock.embedder({ name: \"amazon.titan-embed-text-v2:0\" });\n   * const { vector } = await embedder.embed(\"Hello world\");\n   */\n  public embedder(config: BedrockEmbedderConfig): EmbedderContract {\n    return new BedrockEmbedder(this.client, config, this.provider);\n  }\n}\n"],"mappings":";;;;;;AAEA,MAAM,gBAA8C;CAClD,UAAU;CACV,eAAe;CACf,YAAY;CACZ,UAAU;AACZ;;;;;;;;;;;;;;;;;;AAmBA,SAAgB,cAAc,KAA8C;CAC1E,OAAO,cAAc,OAAO,OAAO;AACrC;;;;ACTA,MAAM,uBAAoD;CACxD,cAAc;CACd,aAAa;CACb,aAAa;CACb,cAAc;AAChB;;;;;;;;;;;;;;;;;;;;;;AAuBA,SAAgB,kBAAkB,UAAsC;CACtE,MAAM,SAA+B,CAAC;CACtC,MAAM,SAA2B,CAAC;CAElC,KAAK,MAAM,WAAW,UAAU;EAC9B,IAAI,QAAQ,SAAS,UAAU;GAC7B,OAAO,KAAK,EAAE,MAAM,iBAAiB,QAAQ,OAAO,EAAE,CAAC;GAEvD;EACF;EAEA,IAAI,QAAQ,SAAS,QAAQ;GAC3B,OAAO,KAAK;IACV,MAAM;IACN,SAAS,CACP,EACE,YAAY;KACV,WAAW,QAAQ,cAAc;KACjC,SAAS,CAAC,EAAE,MAAM,iBAAiB,QAAQ,OAAO,EAAE,CAAC;IACvD,EACF,CACF;GACF,CAAC;GAED;EACF;EAEA,IAAI,QAAQ,SAAS,eAAe,QAAQ,aAAa,QAAQ,UAAU,SAAS,GAAG;GACrF,MAAM,SAAyB,CAAC;GAChC,MAAM,OAAO,iBAAiB,QAAQ,OAAO;GAE7C,IAAI,MACF,OAAO,KAAK,EAAE,KAAK,CAAC;GAGtB,KAAK,MAAM,YAAY,QAAQ,WAC7B,OAAO,KAAK,EACV,SAAS;IACP,WAAW,SAAS;IACpB,MAAM,SAAS;IACf,OAAO,SAAS,SAAS,CAAC;GAC5B,EACF,CAAiB;GAGnB,OAAO,KAAK;IAAE,MAAM;IAAa,SAAS;GAAO,CAAC;GAElD;EACF;EAEA,IAAI,QAAQ,SAAS,UAAU,MAAM,QAAQ,QAAQ,OAAO,GAAG;GAC7D,OAAO,KAAK;IACV,MAAM;IACN,SAAS,QAAQ,QAAQ,IAAI,qBAAqB;GACpD,CAAC;GAED;EACF;EAEA,OAAO,KAAK;GACV,MAAM,QAAQ,SAAS,cAAc,cAAc;GACnD,SAAS,CAAC,EAAE,MAAM,iBAAiB,QAAQ,OAAO,EAAE,CAAC;EACvD,CAAC;CACH;CAEA,OAAO;EACL,QAAQ,OAAO,SAAS,IAAI,SAAS;EACrC,UAAU;CACZ;AACF;;;;;;AAOA,SAAS,iBAAiB,SAAyC;CACjE,IAAI,OAAO,YAAY,UACrB,OAAO;CAGT,OAAO,QACJ,QAAQ,SAAiD,KAAK,SAAS,MAAM,CAAC,CAC9E,KAAK,SAAS,KAAK,IAAI,CAAC,CACxB,KAAK,EAAE;AACZ;;;;;;;;;AAUA,SAAS,sBAAsB,MAAiC;CAC9D,IAAI,KAAK,SAAS,QAChB,OAAO,EAAE,MAAM,KAAK,KAAK;CAG3B,IAAI,SAAS,KAAK,QAChB,MAAM,IAAIA,mCACR,gGACF;CAGF,MAAM,SAAS,qBAAqB,KAAK,OAAO;CAEhD,IAAI,CAAC,QACH,MAAM,IAAIA,mCACR,8CAA8C,KAAK,OAAO,UAAU,8DACtE;CAGF,OAAO,EACL,OAAO;EACL;EACA,QAAQ,EAAE,OAAO,OAAO,KAAK,KAAK,OAAO,QAAQ,QAAQ,EAAE;CAC7D,EACF;AACF;;;;;;;;;;;;;;;;;;;ACrJA,SAAgB,oBACd,OAC+B;CAC/B,IAAI,CAAC,SAAS,MAAM,WAAW,GAC7B;CAGF,OAAO,EACL,OAAO,MAAM,KACV,UAAgB,EACf,UAAU;EACR,MAAM,KAAK;EACX,aAAa,KAAK;EAClB,aAAa,EAAE,MAAM,aAAa,KAAK,KAAK,EAAE;CAChD,EACF,EACF,EACF;AACF;;;;;;;AAQA,SAAS,aAAa,OAAuE;CAC3F,MAAM,+CAA2B,KAAK;CAEtC,IAAI,UAAU,OAAO,SAAS,UAC5B,OAAO;CAGT,OAAO,EAAE,MAAM,SAAS;AAC1B;;;;AC1BA,MAAM,gBAAgB,IAAI,IAAI;CAC5B;CACA;CACA;CACA;AACF,CAAC;;;;;;;;;;;;;;;;;;;;;;;AAwBD,SAAgB,iBAAiB,QAA0B;CACzD,IAAI,kBAAkBC,wBACpB,OAAO;CAGT,MAAM,QAAQ,QAAQ,MAAM;CAC5B,MAAM,UAAU,aAAa,KAAK;CAClC,MAAM,UAAU,MAAM,YAAY,kBAAkB,QAAQ,OAAO,UAAU,OAAO,MAAM;CAE1F,IAAI,UAAU,KAAK,GACjB,OAAO,IAAIC,oCAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,SAAS,2BAA2B,MAAM,mBAAmB,KACrE,OAAO,IAAIC,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,mBAAmB,KAC3B,OAAO,IAAIA,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,iCACjB,OAAO,IAAIC,kCAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,SAAS,yBAAyB,MAAM,mBAAmB,KACnE,OAAO,IAAIC,sCAAuB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGvE,IAAI,MAAM,SAAS,uBAAuB;EACxC,IAAI,+DAA+D,KAAK,OAAO,GAC7E,OAAO,IAAIC,0CAA2B,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;EAG3E,OAAO,IAAIC,mCAAoB,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;CACpE;CAEA,IACE,MAAM,SAAS,+BACf,MAAM,SAAS,uBACf,eAAe,MAAM,cAAc,GAEnC,OAAO,IAAIA,mCAAoB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGpE,OAAO,IAAIC,6BAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;AAOA,SAAS,QAAQ,QAAoC;CACnD,IAAI,OAAO,WAAW,YAAY,WAAW,MAC3C,OAAO,CAAC;CAGV,MAAM,MAAM;CACZ,MAAM,WAAW,IAAI;CAErB,OAAO;EACL,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;EAChD,SAAS,OAAO,IAAI,YAAY,WAAW,IAAI,UAAU;EACzD,gBACE,YAAY,OAAO,SAAS,mBAAmB,WAC3C,SAAS,iBACT,OAAO,IAAI,WAAW,WACnB,IAAI,SACL;EACR,WAAW,YAAY,OAAO,SAAS,cAAc,WAAW,SAAS,YAAY;EACrF,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;CAClD;AACF;;;;;;AAOA,SAAS,UAAU,OAAmC;CACpD,IAAI,MAAM,QAAQ,cAAc,IAAI,MAAM,IAAI,GAC5C,OAAO;CAGT,OAAO,MAAM,SAAS,eAAe,MAAM,SAAS;AACtD;;AAGA,SAAS,eAAe,QAAqC;CAC3D,OAAO,OAAO,WAAW,YAAY,UAAU,OAAO,SAAS;AACjE;;;;;;AAOA,SAAS,aAAa,OAAmD;CACvE,MAAM,UAAmC,CAAC;CAE1C,IAAI,MAAM,mBAAmB,QAC3B,QAAQ,SAAS,MAAM;CAGzB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,IAAI,MAAM,WACR,QAAQ,YAAY,MAAM;CAG5B,OAAO;AACT;;;;AC9JA,MAAMC,eAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCnB,IAAa,kBAAb,MAAyD;CASvD,AAAO,YACL,QACA,QACA,WAAmB,WACnB;gBANgCC;EAOhC,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,uBAAuB,OAAO;EACnC,KAAK,aAAa,OAAO,cAAc;CACzC;CAEA,MAAa,MAAM,OAAyC;EAC1D,MAAM,EAAE,QAAQ,WAAW,MAAM,KAAK,OAAO,KAAK;EAElD,OAAO;GACL;GACA,YAAY,KAAK;GACjB,OAAO;IAAE,cAAc;IAAQ,aAAa;GAAO;EACrD;CACF;CAEA,MAAa,UAAU,QAAiD;EACtE,MAAM,UAAsB,CAAC;EAC7B,IAAI,SAAS;EAEb,KAAK,MAAM,SAAS,QAAQ;GAC1B,MAAM,SAAS,MAAM,KAAK,OAAO,KAAK;GAEtC,QAAQ,KAAK,OAAO,MAAM;GAC1B,UAAU,OAAO;EACnB;EAEA,MAAM,QAAwB;GAAE,cAAc;GAAQ,aAAa;EAAO;EAE1E,OAAO;GAAE;GAAS,YAAY,KAAK;GAAY;EAAM;CACvD;;;;;;CAOA,MAAc,OAAO,OAA8D;EACjF,KAAK,OAAO,MAAMD,cAAY,oBAAoB,0BAA0B,EAC1E,OAAO,KAAK,KACd,CAAC;EAED,MAAM,OAAO,KAAK,UAAU;GAC1B,WAAW;GACX,GAAI,KAAK,yBAAyB,SAC9B,EAAE,YAAY,KAAK,qBAAqB,IACxC,CAAC;EACP,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,MAAM,MAAM,KAAK,OAAO,KACtB,IAAIE,mDAAmB;IACrB,SAAS,KAAK;IACd,aAAa;IACb,QAAQ;IACR,MAAM,IAAI,YAAY,CAAC,CAAC,OAAO,IAAI;GACrC,CAAC,CACH;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,iBAAiB,MAAM;GAEvC,KAAK,OAAO,MAAMF,cAAY,kBAAkB,QAAQ,SAAS;IAC/D,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,UAAU,KAAK,MAAM,IAAI,YAAY,CAAC,CAAC,OAAO,IAAI,IAAI,CAAC;EAE7D,IAAI,KAAK,eAAe,GACtB,KAAK,aAAa,QAAQ,UAAU;EAGtC,KAAK,OAAO,MAAMA,cAAY,qBAAqB,mCAAmC;GACpF,YAAY,QAAQ,UAAU;GAC9B,QAAQ,QAAQ;EAClB,CAAC;EAED,OAAO;GAAE,QAAQ,QAAQ;GAAW,QAAQ,QAAQ;EAAoB;CAC1E;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;AC3HA,MAAM,+BAA+B;CACnC;CACA;CACA;CACA;AACF;;;;;;AAOA,MAAM,oCAAoC;CACxC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;;;;;;;;AASA,MAAM,yBAAyB;CAC7B;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,SAAS,WAAW,SAAiB,WAA8B;CACjE,MAAM,aAAa,QAAQ,YAAY;CAEvC,OAAO,UAAU,MAAM,aAAa,WAAW,SAAS,QAAQ,CAAC;AACnE;;;;;;;;;;;AAYA,SAAgB,yBAAyB,SAA0B;CACjE,OAAO,WAAW,SAAS,4BAA4B;AACzD;;;;;;;;;;;AAYA,SAAgB,6BAA6B,SAA0B;CACrE,OAAO,WAAW,SAAS,iCAAiC;AAC9D;;;;;;;;;;AAWA,SAAgB,mBAAmB,SAA0B;CAC3D,OAAO,WAAW,SAAS,sBAAsB;AACnD;;;;;;;;;;;;;;;;;;;;AC9FA,MAAM,4BAA4B;CAChC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;;;;;;;;;;;;;;AAeA,SAAgB,sBAAsB,SAA0B;CAC9D,MAAM,aAAa,QAAQ,YAAY;CAEvC,OAAO,0BAA0B,MAAM,aAAa,WAAW,SAAS,QAAQ,CAAC;AACnF;;;;AChBA,MAAM,aAAa;;;;;;;;AASnB,MAAM,yBAA6D;CACjE,KAAK;CACL,QAAQ;CACR,MAAM;AACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,IAAa,eAAb,MAAmD;CAUjD,AAAO,YACL,QACA,QACA,WAAmB,WACnB;gBANgCG;EAOhC,KAAK,SAAS;EACd,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,KAAK,eAAe;GAClB,kBAAkB,OAAO,oBAAoB;GAC7C,QAAQ,OAAO,UAAU,sBAAsB,OAAO,IAAI;GAC1D,WAAW,OAAO,aAAa,yBAAyB,OAAO,IAAI;GACnE,eAAe,OAAO,iBAAiB,6BAA6B,OAAO,IAAI;GAC/E,KAAK,OAAO,OAAO,mBAAmB,OAAO,IAAI;GACjD,OAAO,OAAO,SAAS;EACzB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,OAAO,MAAM,YAAY,WAAW,0BAA0B;GACjE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAC3B,IAAIC,gDAAgB,KAAK,aAAa,UAAU,OAAO,CAAC,GACxD,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,MACtD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,SAAS,SAAS,QAAQ,SAAS,WAAW,CAAC;EACrD,MAAM,eAAe,cAAc,SAAS,UAAU;EACtD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAC9C,MAAM,YAAY,KAAK,iBAAiB,MAAM;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,2BAA2B;GAAE;GAAc;EAAM,CAAC;EAE5F,OAAO;GACL,SAAS,KAAK,YAAY,MAAM;GAChC;GACA;GACA;EACF;CACF;;;;;;;CAQA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,gCAAgC;GACvE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAC3B,IAAIC,sDAAsB,KAAK,aAAa,UAAU,OAAO,CAAC,GAC9D,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,MACtD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI;EACJ,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,6BAAa,IAAI,IAAwD;EAE/E,IAAI;GACF,WAAW,MAAM,SAAS,SAAS,UAAU,CAAC,GAAG;IAC/C,IAAI,MAAM,mBAAmB,OAAO,SAAS;KAC3C,MAAM,QAAQ,MAAM,kBAAkB,MAAM;KAE5C,WAAW,IAAI,MAAM,kBAAkB,qBAAqB,GAAG;MAC7D,IAAI,MAAM,aAAa;MACvB,MAAM,MAAM,QAAQ;MACpB,MAAM;KACR,CAAC;KAED;IACF;IAEA,IAAI,MAAM,mBAAmB,OAAO;KAClC,MAAM,QAAQ,MAAM,kBAAkB;KAEtC,IAAI,MAAM,MACR,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM;KAAK;UACtC,IAAI,MAAM,SAAS;MACxB,MAAM,cAAc,WAAW,IAAI,MAAM,kBAAkB,qBAAqB,CAAC;MAEjF,IAAI,aACF,YAAY,QAAQ,MAAM,QAAQ,SAAS;KAE/C;KAEA;IACF;IAEA,IAAI,MAAM,kBAAkB;KAC1B,MAAM,cAAc,WAAW,IAAI,MAAM,iBAAiB,qBAAqB,CAAC;KAEhF,IAAI,aAAa;MACf,MAAM;OACJ,MAAM;OACN,IAAI,YAAY;OAChB,MAAM,YAAY;OAClB,yCAA8C,YAAY,MAAM,CAAC,CAAC;MACpE;MAEA,WAAW,OAAO,MAAM,iBAAiB,qBAAqB,CAAC;KACjE;KAEA;IACF;IAEA,IAAI,MAAM,aACR,gBAAgB,MAAM,YAAY;IAGpC,IAAI,MAAM,UAAU,OAAO;KACzB,MAAM,MAAM,MAAM,SAAS;KAE3B,MAAM,QAAQ,IAAI,eAAe;KACjC,MAAM,SAAS,IAAI,gBAAgB;KACnC,MAAM,QAAQ,IAAI,eAAe,MAAM,QAAQ,MAAM;KAErD,IAAI,IAAI,wBAAwB,IAAI,uBAAuB,GACzD,MAAM,eAAe,IAAI;KAG3B,IAAI,IAAI,yBAAyB,IAAI,wBAAwB,GAC3D,MAAM,mBAAmB,IAAI;IAEjC;GACF;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,aAAa;EAEhD,KAAK,OAAO,MAAM,YAAY,YAAY,iCAAiC;GACzE;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;CAQA,AAAQ,aACN,UACA,SACiB;EACjB,MAAM,EAAE,QAAQ,UAAU,oBAAoB,kBAAkB,QAAQ;EACxE,MAAM,YAAY,SAAS,aAAa,KAAK,OAAO;EACpD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EACxD,MAAM,iBAAiB,KAAK,sBAAsB,iBAAiB,SAAS,YAAY;EAExF,OAAO;GACL,SAAS,KAAK;GACd,UAAU;GACV,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;GAC3B,iBAAiB;IACf,GAAI,cAAc,SAAY,EAAE,UAAU,IAAI,CAAC;IAC/C,GAAI,gBAAgB,SAAY,EAAE,YAAY,IAAI,CAAC;GACrD;GACA,GAAG,KAAK,gBAAgB,SAAS,KAAK;GACtC,GAAG,KAAK,kBAAkB,SAAS,cAAc;GACjD,GAAG,KAAK,qBAAqB,SAAS,SAAS;EACjD;CACF;;;;;;;;;;;;;;;;CAiBA,AAAQ,sBACN,UACA,cAC6B;EAC7B,MAAM,cAAc,cAAc,eAAe;EAEjD,IAAI,CAAC,KAAK,aAAa,iBAAiB,eAAe,KAAK,CAAC,YAAY,SAAS,WAAW,GAC3F,OAAO;EAGT,MAAM,OAAO,SAAS,SAAS;EAC/B,MAAM,cAAc,SAAS;EAE7B,OAAO,CACL,GAAG,SAAS,MAAM,GAAG,IAAI,GACzB;GACE,GAAG;GACH,SAAS,CAAC,GAAI,YAAY,WAAW,CAAC,GAAI,EAAE,YAAY,EAAE,MAAM,UAAU,EAAE,CAAC;EAC/E,CACF;CACF;;;;;;;;;;;;CAaA,AAAQ,qBACN,WACuD;EACvD,IAAI,CAAC,KAAK,aAAa,aAAa,CAAC,WACnC,OAAO,CAAC;EAGV,MAAM,eAAe,UAAU,aAAa,uBAAuB,UAAU,UAAU;EAEvF,IAAI,iBAAiB,QACnB,OAAO,CAAC;EAGV,OAAO,EACL,8BAA8B,EAC5B,UAAU;GAAE,MAAM;GAAW,eAAe;EAAa,EAC3D,EACF;CACF;;;;;CAMA,AAAQ,gBAAgB,OAAuE;EAC7F,MAAM,aAAa,oBAAoB,KAAK;EAE5C,OAAO,aAAa,EAAE,WAAW,IAAI,CAAC;CACxC;;;;;;;;;;CAWA,AAAQ,kBACN,gBACuC;EACvC,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO,EACL,cAAc,EACZ,YAAY;GACV,MAAM;GACN,WAAW,EACT,YAAY;IAAE,MAAM;IAAY,QAAQ,KAAK,UAAU,cAAc;GAAE,EACzE;EACF,EACF,EACF;CACF;;;;;;CAOA,AAAQ,YAAY,QAAgC;EAClD,OAAO,OACJ,KAAK,UAAW,UAAU,SAAS,OAAO,MAAM,SAAS,WAAW,MAAM,OAAO,EAAG,CAAC,CACrF,KAAK,EAAE;CACZ;;;;;;CAOA,AAAQ,iBAAiB,QAA4D;EACnF,MAAM,YAAoC,CAAC;EAE3C,KAAK,MAAM,SAAS,QAClB,IAAI,aAAa,SAAS,MAAM,SAC9B,UAAU,KAAK;GACb,IAAI,MAAM,QAAQ,aAAa;GAC/B,MAAM,MAAM,QAAQ,QAAQ;GAC5B,OAAQ,MAAM,QAAQ,SAAS,CAAC;EAClC,CAAC;EAIL,OAAO,UAAU,SAAS,IAAI,YAAY;CAC5C;;;;;;;;;;CAWA,AAAQ,aAAa,KAAoC;EACvD,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,QAAQ,IAAI,eAAe;EACjC,MAAM,SAAS,IAAI,gBAAgB;EACnC,MAAM,SAAS,IAAI;EACnB,MAAM,aAAa,IAAI;EAEvB,OAAO;GACL;GACA;GACA,OAAO,IAAI,eAAe,QAAQ;GAClC,GAAI,UAAU,SAAS,IAAI,EAAE,cAAc,OAAO,IAAI,CAAC;GACvD,GAAI,cAAc,aAAa,IAAI,EAAE,kBAAkB,WAAW,IAAI,CAAC;EACzE;CACF;;;;;;CAOA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,iBAAiB,MAAM;EAEvC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;GACtD,MAAM,QAAQ;GACd,SAAS,QAAQ;EACnB,CAAC;EAED,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACnbA,IAAa,aAAb,MAAsD;CAKpD,AAAO,YAAY,QAA0B;EAC3C,MAAM,EAAE,UAAU,SAAS,GAAG,iBAAiB;EAE/C,KAAK,SAAS,IAAIC,qDAAqB,YAAY;EACnD,KAAK,WAAW,YAAY;EAC5B,KAAK,UAAU;CACjB;;;;;;;;;;;CAYA,AAAO,MAAM,QAA2C;EACtD,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,aAAa,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACpE;;;;;;;CAQA,MAAa,MAAM,MAAc,QAAkC;EACjE,iDAA6B,IAAI;CACnC;;;;;;;;;CAUA,AAAO,SAAS,QAAiD;EAC/D,OAAO,IAAI,gBAAgB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CAC/D;AACF"}

package/esm/config.type.d.mts

@@ -72,6 +72,36 @@ type BedrockModelConfig = ModelConfig & {
    * hint into the system prompt instead.
    */
   structuredOutput?: boolean;
+  /**
+   * Override the auto-inferred reasoning / extended-thinking capability.
+   * When omitted, the adapter infers `true` for Claude 3.7 + Claude 4
+   * families (see `known-capabilities.ts`). When capable, the adapter
+   * forwards `ModelCallOptions.reasoning` as Converse
+   * `additionalModelRequestFields.thinking`.
+   */
+  reasoning?: boolean;
+  /**
+   * Override the auto-inferred prompt-caching capability. When omitted,
+   * the adapter infers `true` for Claude 3.5+/3.7/4 and Nova families.
+   * When capable, the adapter maps `ModelCallOptions.cacheControl` write
+   * breakpoints to Converse `cachePoint` blocks and reports
+   * `Usage.cachedTokens` / `Usage.cacheWriteTokens` from the response.
+   */
+  promptCaching?: boolean;
+  /**
+   * Override the auto-inferred PDF / document-input capability. When
+   * omitted, the adapter infers `true` for Claude 3+ / Nova families
+   * that accept Converse `document` content blocks. When false the
+   * agent rejects a PDF attachment up front.
+   */
+  pdf?: boolean;
+  /**
+   * Override the audio-input capability. Defaults to `false` — Bedrock
+   * Converse does not accept audio content blocks for the families this
+   * adapter targets. Set `true` only for a model id you have confirmed
+   * supports audio input on Converse.
+   */
+  audio?: boolean;
 };
 /**
  * Per-embedder configuration for `BedrockSDK.embedder()`. `name` is the

package/esm/config.type.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"config.type.d.mts","names":[],"sources":["../../../../../../@warlock.js/ai-bedrock/src/config.type.ts"],"mappings":";;;;;;AAoCA;;;;;;;;;;;;;;AAQuC;AAavC;;;;;;;;AAgBkB;AAclB;;;;AAAkD;;;KAnDtC,gBAAA,GAAmB,0BAAA;EAC7B,QAAA;;;;;;;EAOA,OAAA,GAAU,MAAA,SAAe,YAAA;AAAA;;;;;;;;;;;KAaf,kBAAA,GAAqB,WAAW;;;;;;;EAO1C,MAAA;;;;;;;;;EASA,gBAAA;AAAA;;;;;;;;;;;;KAcU,qBAAA,GAAwB,cAAc"}
+{"version":3,"file":"config.type.d.mts","names":[],"sources":["../../../../../../@warlock.js/ai-bedrock/src/config.type.ts"],"mappings":";;;;;;AAoCA;;;;;;;;;;;;;;AAQuC;AAavC;;;;;;;;;;;;AA8CO;AAcP;;;KAjFY,gBAAA,GAAmB,0BAAA;EAC7B,QAAA;;;;;;;EAOA,OAAA,GAAU,MAAA,SAAe,YAAA;AAAA;;;;;;;;;;;KAaf,kBAAA,GAAqB,WAAW;;;;;;;EAO1C,MAAA;;;;;;;;;EASA,gBAAA;;;;;;;;EAQA,SAAA;;;;;;;;EAQA,aAAA;;;;;;;EAOA,GAAA;;;;;;;EAOA,KAAA;AAAA;;;;;;;;;;;;KAcU,qBAAA,GAAwB,cAAc"}

package/esm/known-capabilities.mjs

@@ -0,0 +1,108 @@
+//#region ../@warlock.js/ai-bedrock/src/known-capabilities.ts
+/**
+* Cost-truth capability inference for Bedrock Converse model ids.
+*
+* Bedrock model ids are provider-prefixed and version-suffixed
+* (`anthropic.claude-3-7-sonnet-20250219-v1:0`, `us.amazon.nova-pro-v1:0`),
+* so — exactly like `known-vision-models.ts` — a lowercase substring scan
+* is the only robust check across cross-region inference-profile prefixes
+* (`us.`, `eu.`, `apac.`) and date/version tags.
+*
+* Each predicate answers a single `ModelCapabilities` flag the agent reads
+* to decide whether to forward a cost-truth option (`reasoning`,
+* `cacheControl`) or up-front-reject an attachment (`pdf`). Unknown ids
+* default to `false` so an unsupported request fails fast with a clear
+* capability error instead of an opaque Bedrock `ValidationException`.
+* Every inference is overridable per-model via `bedrock.model({ name, … })`.
+*/
+/**
+* Families that expose Anthropic-style extended thinking on Bedrock
+* Converse via `additionalModelRequestFields.thinking`. Only Claude 3.7
+* and the Claude 4 line (Sonnet / Opus / Haiku) support a configurable
+* thinking budget; earlier Claude, Nova, Llama, Mistral and Cohere do
+* not, so they are intentionally absent.
+*/
+const REASONING_CAPABLE_SUBSTRINGS = [
+	"claude-3-7",
+	"claude-sonnet-4",
+	"claude-opus-4",
+	"claude-haiku-4"
+];
+/**
+* Families that honor Converse `cachePoint` prompt-cache breakpoints.
+* Anthropic Claude 3.5+ / 3.7 / 4 and the Amazon Nova line support
+* cache points; text-only legacy families do not.
+*/
+const PROMPT_CACHING_CAPABLE_SUBSTRINGS = [
+	"claude-3-5",
+	"claude-3-7",
+	"claude-sonnet-4",
+	"claude-opus-4",
+	"claude-haiku-4",
+	"nova-lite",
+	"nova-pro",
+	"nova-premier",
+	"nova-micro"
+];
+/**
+* Families that accept Converse `document` content blocks (PDF / docx /
+* txt input). The multimodal Claude 3+ and Nova families support
+* document blocks; the substring set mirrors the vision-capable list
+* minus the image-only Llama entries (Llama on Bedrock takes images but
+* not document blocks via Converse).
+*/
+const PDF_CAPABLE_SUBSTRINGS = [
+	"claude-3",
+	"claude-sonnet-4",
+	"claude-opus-4",
+	"claude-haiku-4",
+	"nova-lite",
+	"nova-pro",
+	"nova-premier"
+];
+function matchesAny(modelId, fragments) {
+	const normalized = modelId.toLowerCase();
+	return fragments.some((fragment) => normalized.includes(fragment));
+}
+/**
+* Infer whether a Bedrock model id exposes extended-thinking / reasoning
+* (Claude 3.7 + Claude 4). When true the adapter forwards
+* `ModelCallOptions.reasoning` as Converse
+* `additionalModelRequestFields.thinking`.
+*
+* @example
+* inferReasoningCapability("anthropic.claude-3-7-sonnet-20250219-v1:0"); // → true
+* inferReasoningCapability("us.amazon.nova-pro-v1:0");                    // → false
+*/
+function inferReasoningCapability(modelId) {
+	return matchesAny(modelId, REASONING_CAPABLE_SUBSTRINGS);
+}
+/**
+* Infer whether a Bedrock model id honors Converse `cachePoint`
+* breakpoints (Claude 3.5+ / Nova). When true the adapter both maps
+* `ModelCallOptions.cacheControl` write breakpoints to cache points and
+* reports `Usage.cachedTokens` / `Usage.cacheWriteTokens`.
+*
+* @example
+* inferPromptCachingCapability("us.amazon.nova-pro-v1:0");                       // → true
+* inferPromptCachingCapability("meta.llama3-1-8b-instruct-v1:0");                // → false
+*/
+function inferPromptCachingCapability(modelId) {
+	return matchesAny(modelId, PROMPT_CACHING_CAPABLE_SUBSTRINGS);
+}
+/**
+* Infer whether a Bedrock model id accepts Converse `document` content
+* blocks (PDF / document input — Claude 3+ / Nova). When false the agent
+* rejects a PDF attachment up front instead of dropping it at the wire.
+*
+* @example
+* inferPdfCapability("anthropic.claude-3-5-sonnet-20240620-v1:0"); // → true
+* inferPdfCapability("meta.llama3-2-90b-instruct-v1:0");           // → false
+*/
+function inferPdfCapability(modelId) {
+	return matchesAny(modelId, PDF_CAPABLE_SUBSTRINGS);
+}
+
+//#endregion
+export { inferPdfCapability, inferPromptCachingCapability, inferReasoningCapability };
+//# sourceMappingURL=known-capabilities.mjs.map

package/esm/known-capabilities.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"known-capabilities.mjs","names":[],"sources":["../../../../../../@warlock.js/ai-bedrock/src/known-capabilities.ts"],"sourcesContent":["/**\n * Cost-truth capability inference for Bedrock Converse model ids.\n *\n * Bedrock model ids are provider-prefixed and version-suffixed\n * (`anthropic.claude-3-7-sonnet-20250219-v1:0`, `us.amazon.nova-pro-v1:0`),\n * so — exactly like `known-vision-models.ts` — a lowercase substring scan\n * is the only robust check across cross-region inference-profile prefixes\n * (`us.`, `eu.`, `apac.`) and date/version tags.\n *\n * Each predicate answers a single `ModelCapabilities` flag the agent reads\n * to decide whether to forward a cost-truth option (`reasoning`,\n * `cacheControl`) or up-front-reject an attachment (`pdf`). Unknown ids\n * default to `false` so an unsupported request fails fast with a clear\n * capability error instead of an opaque Bedrock `ValidationException`.\n * Every inference is overridable per-model via `bedrock.model({ name, … })`.\n */\n\n/**\n * Families that expose Anthropic-style extended thinking on Bedrock\n * Converse via `additionalModelRequestFields.thinking`. Only Claude 3.7\n * and the Claude 4 line (Sonnet / Opus / Haiku) support a configurable\n * thinking budget; earlier Claude, Nova, Llama, Mistral and Cohere do\n * not, so they are intentionally absent.\n */\nconst REASONING_CAPABLE_SUBSTRINGS = [\n  \"claude-3-7\",\n  \"claude-sonnet-4\",\n  \"claude-opus-4\",\n  \"claude-haiku-4\",\n];\n\n/**\n * Families that honor Converse `cachePoint` prompt-cache breakpoints.\n * Anthropic Claude 3.5+ / 3.7 / 4 and the Amazon Nova line support\n * cache points; text-only legacy families do not.\n */\nconst PROMPT_CACHING_CAPABLE_SUBSTRINGS = [\n  \"claude-3-5\",\n  \"claude-3-7\",\n  \"claude-sonnet-4\",\n  \"claude-opus-4\",\n  \"claude-haiku-4\",\n  \"nova-lite\",\n  \"nova-pro\",\n  \"nova-premier\",\n  \"nova-micro\",\n];\n\n/**\n * Families that accept Converse `document` content blocks (PDF / docx /\n * txt input). The multimodal Claude 3+ and Nova families support\n * document blocks; the substring set mirrors the vision-capable list\n * minus the image-only Llama entries (Llama on Bedrock takes images but\n * not document blocks via Converse).\n */\nconst PDF_CAPABLE_SUBSTRINGS = [\n  \"claude-3\",\n  \"claude-sonnet-4\",\n  \"claude-opus-4\",\n  \"claude-haiku-4\",\n  \"nova-lite\",\n  \"nova-pro\",\n  \"nova-premier\",\n];\n\nfunction matchesAny(modelId: string, fragments: string[]): boolean {\n  const normalized = modelId.toLowerCase();\n\n  return fragments.some((fragment) => normalized.includes(fragment));\n}\n\n/**\n * Infer whether a Bedrock model id exposes extended-thinking / reasoning\n * (Claude 3.7 + Claude 4). When true the adapter forwards\n * `ModelCallOptions.reasoning` as Converse\n * `additionalModelRequestFields.thinking`.\n *\n * @example\n * inferReasoningCapability(\"anthropic.claude-3-7-sonnet-20250219-v1:0\"); // → true\n * inferReasoningCapability(\"us.amazon.nova-pro-v1:0\");                    // → false\n */\nexport function inferReasoningCapability(modelId: string): boolean {\n  return matchesAny(modelId, REASONING_CAPABLE_SUBSTRINGS);\n}\n\n/**\n * Infer whether a Bedrock model id honors Converse `cachePoint`\n * breakpoints (Claude 3.5+ / Nova). When true the adapter both maps\n * `ModelCallOptions.cacheControl` write breakpoints to cache points and\n * reports `Usage.cachedTokens` / `Usage.cacheWriteTokens`.\n *\n * @example\n * inferPromptCachingCapability(\"us.amazon.nova-pro-v1:0\");                       // → true\n * inferPromptCachingCapability(\"meta.llama3-1-8b-instruct-v1:0\");                // → false\n */\nexport function inferPromptCachingCapability(modelId: string): boolean {\n  return matchesAny(modelId, PROMPT_CACHING_CAPABLE_SUBSTRINGS);\n}\n\n/**\n * Infer whether a Bedrock model id accepts Converse `document` content\n * blocks (PDF / document input — Claude 3+ / Nova). When false the agent\n * rejects a PDF attachment up front instead of dropping it at the wire.\n *\n * @example\n * inferPdfCapability(\"anthropic.claude-3-5-sonnet-20240620-v1:0\"); // → true\n * inferPdfCapability(\"meta.llama3-2-90b-instruct-v1:0\");           // → false\n */\nexport function inferPdfCapability(modelId: string): boolean {\n  return matchesAny(modelId, PDF_CAPABLE_SUBSTRINGS);\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;AAwBA,MAAM,+BAA+B;CACnC;CACA;CACA;CACA;AACF;;;;;;AAOA,MAAM,oCAAoC;CACxC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;;;;;;;;AASA,MAAM,yBAAyB;CAC7B;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,SAAS,WAAW,SAAiB,WAA8B;CACjE,MAAM,aAAa,QAAQ,YAAY;CAEvC,OAAO,UAAU,MAAM,aAAa,WAAW,SAAS,QAAQ,CAAC;AACnE;;;;;;;;;;;AAYA,SAAgB,yBAAyB,SAA0B;CACjE,OAAO,WAAW,SAAS,4BAA4B;AACzD;;;;;;;;;;;AAYA,SAAgB,6BAA6B,SAA0B;CACrE,OAAO,WAAW,SAAS,iCAAiC;AAC9D;;;;;;;;;;AAWA,SAAgB,mBAAmB,SAA0B;CAC3D,OAAO,WAAW,SAAS,sBAAsB;AACnD"}

package/esm/model.mjs

@@ -3,6 +3,7 @@ import { toBedrockMessages } from "./utils/to-bedrock-messages.mjs";
 import { toBedrockToolConfig } from "./utils/to-bedrock-tools.mjs";
 import { wrapBedrockError } from "./utils/wrap-bedrock-error.mjs";
 import "./utils/index.mjs";
+import { inferPdfCapability, inferPromptCachingCapability, inferReasoningCapability } from "./known-capabilities.mjs";
 import { inferVisionCapability } from "./known-vision-models.mjs";
 import { ConverseCommand, ConverseStreamCommand } from "@aws-sdk/client-bedrock-runtime";
 import { safeJsonParse } from "@warlock.js/ai";
@@ -11,6 +12,18 @@ import { log } from "@warlock.js/logger";
 //#region ../@warlock.js/ai-bedrock/src/model.ts
 const LOG_MODULE = "ai.bedrock";
 /**
+* Conventional extended-thinking token budgets for the neutral
+* `reasoning.effort` levels, used when the caller asks for an effort
+* tier without naming an explicit `reasoning.maxTokens` budget. Mirrors
+* the low / medium / high spread other reasoning adapters expose so the
+* vendor-neutral option behaves consistently across providers.
+*/
+const EFFORT_THINKING_BUDGET = {
+	low: 1024,
+	medium: 4096,
+	high: 16384
+};
+/**
 * Bedrock-backed implementation of `ModelContract`.
 *
 * **Role.** The provider-facing bridge between the vendor-neutral
@@ -55,7 +68,11 @@ var BedrockModel = class {
 		this.pricing = config.pricing;
 		this.capabilities = {
 			structuredOutput: config.structuredOutput ?? true,
-			vision: config.vision ?? inferVisionCapability(config.name)
+			vision: config.vision ?? inferVisionCapability(config.name),
+			reasoning: config.reasoning ?? inferReasoningCapability(config.name),
+			promptCaching: config.promptCaching ?? inferPromptCachingCapability(config.name),
+			pdf: config.pdf ?? inferPdfCapability(config.name),
+			audio: config.audio ?? false
 		};
 	}
 	/**
@@ -161,6 +178,7 @@ var BedrockModel = class {
 					usage.output = raw.outputTokens ?? 0;
 					usage.total = raw.totalTokens ?? usage.input + usage.output;
 					if (raw.cacheReadInputTokens && raw.cacheReadInputTokens > 0) usage.cachedTokens = raw.cacheReadInputTokens;
+					if (raw.cacheWriteInputTokens && raw.cacheWriteInputTokens > 0) usage.cacheWriteTokens = raw.cacheWriteInputTokens;
 				}
 			}
 		} catch (thrown) {
@@ -187,19 +205,66 @@ var BedrockModel = class {
 		const { system, messages: bedrockMessages } = toBedrockMessages(messages);
 		const maxTokens = options?.maxTokens ?? this.config.maxTokens;
 		const temperature = options?.temperature ?? this.config.temperature;
+		const cachedMessages = this.applyCacheBreakpoints(bedrockMessages, options?.cacheControl);
 		return {
 			modelId: this.name,
-			messages: bedrockMessages,
+			messages: cachedMessages,
 			...system ? { system } : {},
 			inferenceConfig: {
 				...maxTokens !== void 0 ? { maxTokens } : {},
 				...temperature !== void 0 ? { temperature } : {}
 			},
 			...this.buildToolConfig(options?.tools),
-			...this.buildOutputConfig(options?.responseSchema)
+			...this.buildOutputConfig(options?.responseSchema),
+			...this.buildReasoningConfig(options?.reasoning)
 		};
 	}
 	/**
+	* Append a Converse `cachePoint` block to the LAST message when the
+	* caller supplies a `cacheControl` write breakpoint and the model is
+	* `promptCaching`-capable. A cache point tells Bedrock to cache the
+	* whole prefix up to that block, so subsequent calls reusing the same
+	* prefix bill the cached portion at the discounted read rate
+	* (surfaced as `Usage.cachedTokens`). No-ops gracefully when caching
+	* is unsupported, no breakpoint was requested, or there are no
+	* messages to mark — Bedrock then prices the call normally.
+	*
+	* Bedrock only honors `CachePointType.DEFAULT`; the neutral
+	* `breakpoints` count is a presence hint (one trailing breakpoint is
+	* the only placement Converse supports without manual block surgery),
+	* so any positive value marks the trailing message.
+	*/
+	applyCacheBreakpoints(messages, cacheControl) {
+		const breakpoints = cacheControl?.breakpoints ?? 0;
+		if (!this.capabilities.promptCaching || breakpoints <= 0 || !messages || messages.length === 0) return messages;
+		const last = messages.length - 1;
+		const lastMessage = messages[last];
+		return [...messages.slice(0, last), {
+			...lastMessage,
+			content: [...lastMessage.content ?? [], { cachePoint: { type: "default" } }]
+		}];
+	}
+	/**
+	* Translate the neutral `reasoning` option into Claude-on-Bedrock's
+	* extended-thinking control, carried in Converse's escape hatch
+	* `additionalModelRequestFields.thinking`. Emitted only when the model
+	* is `reasoning`-capable and a budget can be resolved — `maxTokens`
+	* (explicit thinking budget) wins, otherwise `effort` maps to a
+	* conventional token budget so callers can opt in without picking a
+	* number. Returns an empty object (no-op) for non-reasoning models or
+	* when no reasoning option was supplied, so unsupported params never
+	* reach the wire.
+	*/
+	buildReasoningConfig(reasoning) {
+		if (!this.capabilities.reasoning || !reasoning) return {};
+		const budgetTokens = reasoning.maxTokens ?? EFFORT_THINKING_BUDGET[reasoning.effort ?? ""];
+		if (budgetTokens === void 0) return {};
+		return { additionalModelRequestFields: { thinking: {
+			type: "enabled",
+			budget_tokens: budgetTokens
+		} } };
+	}
+	/**
 	* Spread-friendly tool fragment. Returns an empty object when no
 	* tools were supplied (Bedrock rejects an empty `tools` array).
 	*/
@@ -251,8 +316,12 @@ var BedrockModel = class {
 	}
 	/**
 	* Normalize Converse's `TokenUsage` into the neutral `Usage` shape.
-	* Bedrock supplies a pre-summed `totalTokens`; cache-read tokens are
-	* surfaced as `cachedTokens` only when non-zero.
+	* Bedrock supplies a pre-summed `totalTokens`; cache-read and
+	* cache-write tokens are surfaced as `cachedTokens` /
+	* `cacheWriteTokens` only when non-zero so callers can price the
+	* discounted read rate and the one-time write cost separately.
+	* Bedrock's Converse `TokenUsage` carries no reasoning-token channel,
+	* so `Usage.reasoningTokens` is intentionally left unset here.
 	*/
 	extractUsage(raw) {
 		if (!raw) return {
@@ -263,11 +332,13 @@ var BedrockModel = class {
 		const input = raw.inputTokens ?? 0;
 		const output = raw.outputTokens ?? 0;
 		const cached = raw.cacheReadInputTokens;
+		const cacheWrite = raw.cacheWriteInputTokens;
 		return {
 			input,
 			output,
 			total: raw.totalTokens ?? input + output,
-			...cached && cached > 0 ? { cachedTokens: cached } : {}
+			...cached && cached > 0 ? { cachedTokens: cached } : {},
+			...cacheWrite && cacheWrite > 0 ? { cacheWriteTokens: cacheWrite } : {}
 		};
 	}
 	/**

package/esm/model.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../@warlock.js/ai-bedrock/src/model.ts"],"sourcesContent":["import {\n  safeJsonParse,\n  type Message,\n  type ModelCallOptions,\n  type ModelCapabilities,\n  type ModelContract,\n  type ModelPricing,\n  type ModelResponse,\n  type ModelStreamChunk,\n  type ModelToolCallRequest,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport {\n  ConverseCommand,\n  ConverseStreamCommand,\n  type BedrockRuntimeClient,\n  type ContentBlock,\n  type ConverseRequest,\n  type TokenUsage,\n} from \"@aws-sdk/client-bedrock-runtime\";\nimport type { BedrockModelConfig } from \"./config.type\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapStopReason, toBedrockMessages, toBedrockToolConfig, wrapBedrockError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.bedrock\";\n\n/**\n * Bedrock-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and AWS Bedrock's Converse /\n * ConverseStream API. Converse is the model-agnostic surface — one\n * wire mapping covers every Bedrock-hosted family (Anthropic Claude,\n * Amazon Nova, Meta Llama, Mistral, Cohere) instead of per-family\n * `InvokeModel` body shapes.\n *\n * **Responsibility.**\n * - Owns: a long-lived `BedrockRuntimeClient` + frozen `ModelConfig`\n *   (modelId, temperature, maxTokens) used as per-call defaults.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Converse shapes (system hoisting, `toolUse` / `toolResult` blocks,\n *   image bytes) on the way out, and Converse's content-block response\n *   (text, tool calls, stop reason, token usage) back into the neutral\n *   shapes on the way in.\n * - Does NOT own: dispatching tools, looping, history, retries — those\n *   are agent concerns. The model is a per-call protocol adapter.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across calls\"): the AWS client is heavy to construct and reused for\n * the SDK's lifetime.\n *\n * @example\n * import { BedrockRuntimeClient } from \"@aws-sdk/client-bedrock-runtime\";\n * const client = new BedrockRuntimeClient({ region: \"us-east-1\" });\n * const model = new BedrockModel(client, {\n *   name: \"anthropic.claude-sonnet-4-5-20250929-v1:0\",\n * });\n *\n * const myAgent = agent({ model, tools: [searchTool] });\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class BedrockModel implements ModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly capabilities: ModelCapabilities;\n  public readonly pricing?: ModelPricing;\n\n  private readonly client: BedrockRuntimeClient;\n  private readonly config: BedrockModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: BedrockRuntimeClient,\n    config: BedrockModelConfig,\n    provider: string = \"bedrock\",\n  ) {\n    this.client = client;\n    this.config = config;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    this.capabilities = {\n      structuredOutput: config.structuredOutput ?? true,\n      vision: config.vision ?? inferVisionCapability(config.name),\n    };\n  }\n\n  /**\n   * Single-shot completion via the Converse API. Sends the full\n   * message list, waits for the terminal response, and reshapes it\n   * into a vendor-neutral `ModelResponse`. Per-call `options` override\n   * the instance defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting Converse call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response;\n\n    try {\n      response = await this.client.send(\n        new ConverseCommand(this.buildRequest(messages, options)),\n        options?.signal ? { abortSignal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const blocks = response.output?.message?.content ?? [];\n    const finishReason = mapStopReason(response.stopReason);\n    const usage = this.extractUsage(response.usage);\n    const toolCalls = this.extractToolCalls(blocks);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Converse call succeeded\", { finishReason, usage });\n\n    return {\n      content: this.extractText(blocks),\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion via ConverseStream. Yields neutral\n   * `ModelStreamChunk`s — `delta` for text, `tool-call` once a\n   * `toolUse` block's accumulated input JSON is complete, and a\n   * terminal `done` with the final finish reason + usage totals.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting ConverseStream call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response;\n\n    try {\n      response = await this.client.send(\n        new ConverseStreamCommand(this.buildRequest(messages, options)),\n        options?.signal ? { abortSignal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawStopReason: string | undefined;\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n    const toolBlocks = new Map<number, { id: string; name: string; json: string }>();\n\n    try {\n      for await (const event of response.stream ?? []) {\n        if (event.contentBlockStart?.start?.toolUse) {\n          const start = event.contentBlockStart.start.toolUse;\n\n          toolBlocks.set(event.contentBlockStart.contentBlockIndex ?? 0, {\n            id: start.toolUseId ?? \"\",\n            name: start.name ?? \"\",\n            json: \"\",\n          });\n\n          continue;\n        }\n\n        if (event.contentBlockDelta?.delta) {\n          const delta = event.contentBlockDelta.delta;\n\n          if (delta.text) {\n            yield { type: \"delta\", content: delta.text };\n          } else if (delta.toolUse) {\n            const accumulator = toolBlocks.get(event.contentBlockDelta.contentBlockIndex ?? 0);\n\n            if (accumulator) {\n              accumulator.json += delta.toolUse.input ?? \"\";\n            }\n          }\n\n          continue;\n        }\n\n        if (event.contentBlockStop) {\n          const accumulator = toolBlocks.get(event.contentBlockStop.contentBlockIndex ?? 0);\n\n          if (accumulator) {\n            yield {\n              type: \"tool-call\",\n              id: accumulator.id,\n              name: accumulator.name,\n              input: safeJsonParse<Record<string, unknown>>(accumulator.json, {}),\n            };\n\n            toolBlocks.delete(event.contentBlockStop.contentBlockIndex ?? 0);\n          }\n\n          continue;\n        }\n\n        if (event.messageStop) {\n          rawStopReason = event.messageStop.stopReason;\n        }\n\n        if (event.metadata?.usage) {\n          const raw = event.metadata.usage;\n\n          usage.input = raw.inputTokens ?? 0;\n          usage.output = raw.outputTokens ?? 0;\n          usage.total = raw.totalTokens ?? usage.input + usage.output;\n\n          if (raw.cacheReadInputTokens && raw.cacheReadInputTokens > 0) {\n            usage.cachedTokens = raw.cacheReadInputTokens;\n          }\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = mapStopReason(rawStopReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"ConverseStream call succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the Converse request shared by `complete()` and\n   * `stream()` (both command shapes take the same input). Hoists the\n   * system prompt, maps inference params, and conditionally attaches\n   * tools and native structured output.\n   */\n  private buildRequest(\n    messages: Message[],\n    options: ModelCallOptions | undefined,\n  ): ConverseRequest {\n    const { system, messages: bedrockMessages } = toBedrockMessages(messages);\n    const maxTokens = options?.maxTokens ?? this.config.maxTokens;\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      modelId: this.name,\n      messages: bedrockMessages,\n      ...(system ? { system } : {}),\n      inferenceConfig: {\n        ...(maxTokens !== undefined ? { maxTokens } : {}),\n        ...(temperature !== undefined ? { temperature } : {}),\n      },\n      ...this.buildToolConfig(options?.tools),\n      ...this.buildOutputConfig(options?.responseSchema),\n    };\n  }\n\n  /**\n   * Spread-friendly tool fragment. Returns an empty object when no\n   * tools were supplied (Bedrock rejects an empty `tools` array).\n   */\n  private buildToolConfig(tools: ModelCallOptions[\"tools\"]): Pick<ConverseRequest, \"toolConfig\"> {\n    const toolConfig = toBedrockToolConfig(tools);\n\n    return toolConfig ? { toolConfig } : {};\n  }\n\n  /**\n   * Translate the neutral `responseSchema` into Converse's native\n   * `outputConfig.textFormat` (JSON-schema structured output). Bedrock\n   * requires the schema as a stringified JSON document and only\n   * accepts an object root. Emitted only when the model is\n   * `structuredOutput`-capable and the schema is an object — otherwise\n   * the agent's soft system-prompt hint + client-side `validate()`\n   * carry shape (same degradation philosophy as the OpenAI adapter).\n   */\n  private buildOutputConfig(\n    responseSchema: Record<string, unknown> | undefined,\n  ): Pick<ConverseRequest, \"outputConfig\"> {\n    if (!responseSchema || !this.capabilities.structuredOutput) {\n      return {};\n    }\n\n    if (responseSchema.type !== \"object\" || typeof responseSchema.properties !== \"object\") {\n      return {};\n    }\n\n    return {\n      outputConfig: {\n        textFormat: {\n          type: \"json_schema\",\n          structure: {\n            jsonSchema: { name: \"response\", schema: JSON.stringify(responseSchema) },\n          },\n        },\n      },\n    };\n  }\n\n  /**\n   * Concatenate every `text` content block into the single neutral\n   * `content` string. `toolUse` and other block types are surfaced\n   * separately via `extractToolCalls`.\n   */\n  private extractText(blocks: ContentBlock[]): string {\n    return blocks\n      .map((block) => (\"text\" in block && typeof block.text === \"string\" ? block.text : \"\"))\n      .join(\"\");\n  }\n\n  /**\n   * Reshape Converse `toolUse` content blocks into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when no tools were\n   * requested so callers can branch on presence.\n   */\n  private extractToolCalls(blocks: ContentBlock[]): ModelToolCallRequest[] | undefined {\n    const toolCalls: ModelToolCallRequest[] = [];\n\n    for (const block of blocks) {\n      if (\"toolUse\" in block && block.toolUse) {\n        toolCalls.push({\n          id: block.toolUse.toolUseId ?? \"\",\n          name: block.toolUse.name ?? \"\",\n          input: (block.toolUse.input ?? {}) as Record<string, unknown>,\n        });\n      }\n    }\n\n    return toolCalls.length > 0 ? toolCalls : undefined;\n  }\n\n  /**\n   * Normalize Converse's `TokenUsage` into the neutral `Usage` shape.\n   * Bedrock supplies a pre-summed `totalTokens`; cache-read tokens are\n   * surfaced as `cachedTokens` only when non-zero.\n   */\n  private extractUsage(raw: TokenUsage | undefined): Usage {\n    if (!raw) {\n      return { input: 0, output: 0, total: 0 };\n    }\n\n    const input = raw.inputTokens ?? 0;\n    const output = raw.outputTokens ?? 0;\n    const cached = raw.cacheReadInputTokens;\n\n    return {\n      input,\n      output,\n      total: raw.totalTokens ?? input + output,\n      ...(cached && cached > 0 ? { cachedTokens: cached } : {}),\n    };\n  }\n\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy\n   * and emit the standard error log line before it propagates. Shared\n   * by every catch site so the log shape stays identical.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapBedrockError(thrown);\n\n    this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n      code: wrapped.code,\n      context: wrapped.context,\n    });\n\n    return wrapped;\n  }\n}\n"],"mappings":";;;;;;;;;;;AAyBA,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCnB,IAAa,eAAb,MAAmD;CAUjD,AAAO,YACL,QACA,QACA,WAAmB,WACnB;gBANgC;EAOhC,KAAK,SAAS;EACd,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,KAAK,eAAe;GAClB,kBAAkB,OAAO,oBAAoB;GAC7C,QAAQ,OAAO,UAAU,sBAAsB,OAAO,IAAI;EAC5D;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,OAAO,MAAM,YAAY,WAAW,0BAA0B;GACjE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAC3B,IAAI,gBAAgB,KAAK,aAAa,UAAU,OAAO,CAAC,GACxD,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,MACtD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,SAAS,SAAS,QAAQ,SAAS,WAAW,CAAC;EACrD,MAAM,eAAe,cAAc,SAAS,UAAU;EACtD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAC9C,MAAM,YAAY,KAAK,iBAAiB,MAAM;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,2BAA2B;GAAE;GAAc;EAAM,CAAC;EAE5F,OAAO;GACL,SAAS,KAAK,YAAY,MAAM;GAChC;GACA;GACA;EACF;CACF;;;;;;;CAQA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,gCAAgC;GACvE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAC3B,IAAI,sBAAsB,KAAK,aAAa,UAAU,OAAO,CAAC,GAC9D,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,MACtD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI;EACJ,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,6BAAa,IAAI,IAAwD;EAE/E,IAAI;GACF,WAAW,MAAM,SAAS,SAAS,UAAU,CAAC,GAAG;IAC/C,IAAI,MAAM,mBAAmB,OAAO,SAAS;KAC3C,MAAM,QAAQ,MAAM,kBAAkB,MAAM;KAE5C,WAAW,IAAI,MAAM,kBAAkB,qBAAqB,GAAG;MAC7D,IAAI,MAAM,aAAa;MACvB,MAAM,MAAM,QAAQ;MACpB,MAAM;KACR,CAAC;KAED;IACF;IAEA,IAAI,MAAM,mBAAmB,OAAO;KAClC,MAAM,QAAQ,MAAM,kBAAkB;KAEtC,IAAI,MAAM,MACR,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM;KAAK;UACtC,IAAI,MAAM,SAAS;MACxB,MAAM,cAAc,WAAW,IAAI,MAAM,kBAAkB,qBAAqB,CAAC;MAEjF,IAAI,aACF,YAAY,QAAQ,MAAM,QAAQ,SAAS;KAE/C;KAEA;IACF;IAEA,IAAI,MAAM,kBAAkB;KAC1B,MAAM,cAAc,WAAW,IAAI,MAAM,iBAAiB,qBAAqB,CAAC;KAEhF,IAAI,aAAa;MACf,MAAM;OACJ,MAAM;OACN,IAAI,YAAY;OAChB,MAAM,YAAY;OAClB,OAAO,cAAuC,YAAY,MAAM,CAAC,CAAC;MACpE;MAEA,WAAW,OAAO,MAAM,iBAAiB,qBAAqB,CAAC;KACjE;KAEA;IACF;IAEA,IAAI,MAAM,aACR,gBAAgB,MAAM,YAAY;IAGpC,IAAI,MAAM,UAAU,OAAO;KACzB,MAAM,MAAM,MAAM,SAAS;KAE3B,MAAM,QAAQ,IAAI,eAAe;KACjC,MAAM,SAAS,IAAI,gBAAgB;KACnC,MAAM,QAAQ,IAAI,eAAe,MAAM,QAAQ,MAAM;KAErD,IAAI,IAAI,wBAAwB,IAAI,uBAAuB,GACzD,MAAM,eAAe,IAAI;IAE7B;GACF;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,aAAa;EAEhD,KAAK,OAAO,MAAM,YAAY,YAAY,iCAAiC;GACzE;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;CAQA,AAAQ,aACN,UACA,SACiB;EACjB,MAAM,EAAE,QAAQ,UAAU,oBAAoB,kBAAkB,QAAQ;EACxE,MAAM,YAAY,SAAS,aAAa,KAAK,OAAO;EACpD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,SAAS,KAAK;GACd,UAAU;GACV,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;GAC3B,iBAAiB;IACf,GAAI,cAAc,SAAY,EAAE,UAAU,IAAI,CAAC;IAC/C,GAAI,gBAAgB,SAAY,EAAE,YAAY,IAAI,CAAC;GACrD;GACA,GAAG,KAAK,gBAAgB,SAAS,KAAK;GACtC,GAAG,KAAK,kBAAkB,SAAS,cAAc;EACnD;CACF;;;;;CAMA,AAAQ,gBAAgB,OAAuE;EAC7F,MAAM,aAAa,oBAAoB,KAAK;EAE5C,OAAO,aAAa,EAAE,WAAW,IAAI,CAAC;CACxC;;;;;;;;;;CAWA,AAAQ,kBACN,gBACuC;EACvC,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO,EACL,cAAc,EACZ,YAAY;GACV,MAAM;GACN,WAAW,EACT,YAAY;IAAE,MAAM;IAAY,QAAQ,KAAK,UAAU,cAAc;GAAE,EACzE;EACF,EACF,EACF;CACF;;;;;;CAOA,AAAQ,YAAY,QAAgC;EAClD,OAAO,OACJ,KAAK,UAAW,UAAU,SAAS,OAAO,MAAM,SAAS,WAAW,MAAM,OAAO,EAAG,CAAC,CACrF,KAAK,EAAE;CACZ;;;;;;CAOA,AAAQ,iBAAiB,QAA4D;EACnF,MAAM,YAAoC,CAAC;EAE3C,KAAK,MAAM,SAAS,QAClB,IAAI,aAAa,SAAS,MAAM,SAC9B,UAAU,KAAK;GACb,IAAI,MAAM,QAAQ,aAAa;GAC/B,MAAM,MAAM,QAAQ,QAAQ;GAC5B,OAAQ,MAAM,QAAQ,SAAS,CAAC;EAClC,CAAC;EAIL,OAAO,UAAU,SAAS,IAAI,YAAY;CAC5C;;;;;;CAOA,AAAQ,aAAa,KAAoC;EACvD,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,QAAQ,IAAI,eAAe;EACjC,MAAM,SAAS,IAAI,gBAAgB;EACnC,MAAM,SAAS,IAAI;EAEnB,OAAO;GACL;GACA;GACA,OAAO,IAAI,eAAe,QAAQ;GAClC,GAAI,UAAU,SAAS,IAAI,EAAE,cAAc,OAAO,IAAI,CAAC;EACzD;CACF;;;;;;CAOA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,iBAAiB,MAAM;EAEvC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;GACtD,MAAM,QAAQ;GACd,SAAS,QAAQ;EACnB,CAAC;EAED,OAAO;CACT;AACF"}
+{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../@warlock.js/ai-bedrock/src/model.ts"],"sourcesContent":["import {\n  safeJsonParse,\n  type Message,\n  type ModelCallOptions,\n  type ModelCapabilities,\n  type ModelContract,\n  type ModelPricing,\n  type ModelResponse,\n  type ModelStreamChunk,\n  type ModelToolCallRequest,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport {\n  ConverseCommand,\n  ConverseStreamCommand,\n  type BedrockRuntimeClient,\n  type ContentBlock,\n  type ConverseRequest,\n  type TokenUsage,\n} from \"@aws-sdk/client-bedrock-runtime\";\nimport type { BedrockModelConfig } from \"./config.type\";\nimport {\n  inferPdfCapability,\n  inferPromptCachingCapability,\n  inferReasoningCapability,\n} from \"./known-capabilities\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapStopReason, toBedrockMessages, toBedrockToolConfig, wrapBedrockError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.bedrock\";\n\n/**\n * Conventional extended-thinking token budgets for the neutral\n * `reasoning.effort` levels, used when the caller asks for an effort\n * tier without naming an explicit `reasoning.maxTokens` budget. Mirrors\n * the low / medium / high spread other reasoning adapters expose so the\n * vendor-neutral option behaves consistently across providers.\n */\nconst EFFORT_THINKING_BUDGET: Record<string, number | undefined> = {\n  low: 1024,\n  medium: 4096,\n  high: 16384,\n};\n\n/**\n * Bedrock-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and AWS Bedrock's Converse /\n * ConverseStream API. Converse is the model-agnostic surface — one\n * wire mapping covers every Bedrock-hosted family (Anthropic Claude,\n * Amazon Nova, Meta Llama, Mistral, Cohere) instead of per-family\n * `InvokeModel` body shapes.\n *\n * **Responsibility.**\n * - Owns: a long-lived `BedrockRuntimeClient` + frozen `ModelConfig`\n *   (modelId, temperature, maxTokens) used as per-call defaults.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Converse shapes (system hoisting, `toolUse` / `toolResult` blocks,\n *   image bytes) on the way out, and Converse's content-block response\n *   (text, tool calls, stop reason, token usage) back into the neutral\n *   shapes on the way in.\n * - Does NOT own: dispatching tools, looping, history, retries — those\n *   are agent concerns. The model is a per-call protocol adapter.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across calls\"): the AWS client is heavy to construct and reused for\n * the SDK's lifetime.\n *\n * @example\n * import { BedrockRuntimeClient } from \"@aws-sdk/client-bedrock-runtime\";\n * const client = new BedrockRuntimeClient({ region: \"us-east-1\" });\n * const model = new BedrockModel(client, {\n *   name: \"anthropic.claude-sonnet-4-5-20250929-v1:0\",\n * });\n *\n * const myAgent = agent({ model, tools: [searchTool] });\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class BedrockModel implements ModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly capabilities: ModelCapabilities;\n  public readonly pricing?: ModelPricing;\n\n  private readonly client: BedrockRuntimeClient;\n  private readonly config: BedrockModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: BedrockRuntimeClient,\n    config: BedrockModelConfig,\n    provider: string = \"bedrock\",\n  ) {\n    this.client = client;\n    this.config = config;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    this.capabilities = {\n      structuredOutput: config.structuredOutput ?? true,\n      vision: config.vision ?? inferVisionCapability(config.name),\n      reasoning: config.reasoning ?? inferReasoningCapability(config.name),\n      promptCaching: config.promptCaching ?? inferPromptCachingCapability(config.name),\n      pdf: config.pdf ?? inferPdfCapability(config.name),\n      audio: config.audio ?? false,\n    };\n  }\n\n  /**\n   * Single-shot completion via the Converse API. Sends the full\n   * message list, waits for the terminal response, and reshapes it\n   * into a vendor-neutral `ModelResponse`. Per-call `options` override\n   * the instance defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting Converse call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response;\n\n    try {\n      response = await this.client.send(\n        new ConverseCommand(this.buildRequest(messages, options)),\n        options?.signal ? { abortSignal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const blocks = response.output?.message?.content ?? [];\n    const finishReason = mapStopReason(response.stopReason);\n    const usage = this.extractUsage(response.usage);\n    const toolCalls = this.extractToolCalls(blocks);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Converse call succeeded\", { finishReason, usage });\n\n    return {\n      content: this.extractText(blocks),\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion via ConverseStream. Yields neutral\n   * `ModelStreamChunk`s — `delta` for text, `tool-call` once a\n   * `toolUse` block's accumulated input JSON is complete, and a\n   * terminal `done` with the final finish reason + usage totals.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting ConverseStream call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response;\n\n    try {\n      response = await this.client.send(\n        new ConverseStreamCommand(this.buildRequest(messages, options)),\n        options?.signal ? { abortSignal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawStopReason: string | undefined;\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n    const toolBlocks = new Map<number, { id: string; name: string; json: string }>();\n\n    try {\n      for await (const event of response.stream ?? []) {\n        if (event.contentBlockStart?.start?.toolUse) {\n          const start = event.contentBlockStart.start.toolUse;\n\n          toolBlocks.set(event.contentBlockStart.contentBlockIndex ?? 0, {\n            id: start.toolUseId ?? \"\",\n            name: start.name ?? \"\",\n            json: \"\",\n          });\n\n          continue;\n        }\n\n        if (event.contentBlockDelta?.delta) {\n          const delta = event.contentBlockDelta.delta;\n\n          if (delta.text) {\n            yield { type: \"delta\", content: delta.text };\n          } else if (delta.toolUse) {\n            const accumulator = toolBlocks.get(event.contentBlockDelta.contentBlockIndex ?? 0);\n\n            if (accumulator) {\n              accumulator.json += delta.toolUse.input ?? \"\";\n            }\n          }\n\n          continue;\n        }\n\n        if (event.contentBlockStop) {\n          const accumulator = toolBlocks.get(event.contentBlockStop.contentBlockIndex ?? 0);\n\n          if (accumulator) {\n            yield {\n              type: \"tool-call\",\n              id: accumulator.id,\n              name: accumulator.name,\n              input: safeJsonParse<Record<string, unknown>>(accumulator.json, {}),\n            };\n\n            toolBlocks.delete(event.contentBlockStop.contentBlockIndex ?? 0);\n          }\n\n          continue;\n        }\n\n        if (event.messageStop) {\n          rawStopReason = event.messageStop.stopReason;\n        }\n\n        if (event.metadata?.usage) {\n          const raw = event.metadata.usage;\n\n          usage.input = raw.inputTokens ?? 0;\n          usage.output = raw.outputTokens ?? 0;\n          usage.total = raw.totalTokens ?? usage.input + usage.output;\n\n          if (raw.cacheReadInputTokens && raw.cacheReadInputTokens > 0) {\n            usage.cachedTokens = raw.cacheReadInputTokens;\n          }\n\n          if (raw.cacheWriteInputTokens && raw.cacheWriteInputTokens > 0) {\n            usage.cacheWriteTokens = raw.cacheWriteInputTokens;\n          }\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = mapStopReason(rawStopReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"ConverseStream call succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the Converse request shared by `complete()` and\n   * `stream()` (both command shapes take the same input). Hoists the\n   * system prompt, maps inference params, and conditionally attaches\n   * tools and native structured output.\n   */\n  private buildRequest(\n    messages: Message[],\n    options: ModelCallOptions | undefined,\n  ): ConverseRequest {\n    const { system, messages: bedrockMessages } = toBedrockMessages(messages);\n    const maxTokens = options?.maxTokens ?? this.config.maxTokens;\n    const temperature = options?.temperature ?? this.config.temperature;\n    const cachedMessages = this.applyCacheBreakpoints(bedrockMessages, options?.cacheControl);\n\n    return {\n      modelId: this.name,\n      messages: cachedMessages,\n      ...(system ? { system } : {}),\n      inferenceConfig: {\n        ...(maxTokens !== undefined ? { maxTokens } : {}),\n        ...(temperature !== undefined ? { temperature } : {}),\n      },\n      ...this.buildToolConfig(options?.tools),\n      ...this.buildOutputConfig(options?.responseSchema),\n      ...this.buildReasoningConfig(options?.reasoning),\n    };\n  }\n\n  /**\n   * Append a Converse `cachePoint` block to the LAST message when the\n   * caller supplies a `cacheControl` write breakpoint and the model is\n   * `promptCaching`-capable. A cache point tells Bedrock to cache the\n   * whole prefix up to that block, so subsequent calls reusing the same\n   * prefix bill the cached portion at the discounted read rate\n   * (surfaced as `Usage.cachedTokens`). No-ops gracefully when caching\n   * is unsupported, no breakpoint was requested, or there are no\n   * messages to mark — Bedrock then prices the call normally.\n   *\n   * Bedrock only honors `CachePointType.DEFAULT`; the neutral\n   * `breakpoints` count is a presence hint (one trailing breakpoint is\n   * the only placement Converse supports without manual block surgery),\n   * so any positive value marks the trailing message.\n   */\n  private applyCacheBreakpoints(\n    messages: ConverseRequest[\"messages\"],\n    cacheControl: ModelCallOptions[\"cacheControl\"],\n  ): ConverseRequest[\"messages\"] {\n    const breakpoints = cacheControl?.breakpoints ?? 0;\n\n    if (!this.capabilities.promptCaching || breakpoints <= 0 || !messages || messages.length === 0) {\n      return messages;\n    }\n\n    const last = messages.length - 1;\n    const lastMessage = messages[last];\n\n    return [\n      ...messages.slice(0, last),\n      {\n        ...lastMessage,\n        content: [...(lastMessage.content ?? []), { cachePoint: { type: \"default\" } }],\n      },\n    ];\n  }\n\n  /**\n   * Translate the neutral `reasoning` option into Claude-on-Bedrock's\n   * extended-thinking control, carried in Converse's escape hatch\n   * `additionalModelRequestFields.thinking`. Emitted only when the model\n   * is `reasoning`-capable and a budget can be resolved — `maxTokens`\n   * (explicit thinking budget) wins, otherwise `effort` maps to a\n   * conventional token budget so callers can opt in without picking a\n   * number. Returns an empty object (no-op) for non-reasoning models or\n   * when no reasoning option was supplied, so unsupported params never\n   * reach the wire.\n   */\n  private buildReasoningConfig(\n    reasoning: ModelCallOptions[\"reasoning\"],\n  ): Pick<ConverseRequest, \"additionalModelRequestFields\"> {\n    if (!this.capabilities.reasoning || !reasoning) {\n      return {};\n    }\n\n    const budgetTokens = reasoning.maxTokens ?? EFFORT_THINKING_BUDGET[reasoning.effort ?? \"\"];\n\n    if (budgetTokens === undefined) {\n      return {};\n    }\n\n    return {\n      additionalModelRequestFields: {\n        thinking: { type: \"enabled\", budget_tokens: budgetTokens },\n      },\n    };\n  }\n\n  /**\n   * Spread-friendly tool fragment. Returns an empty object when no\n   * tools were supplied (Bedrock rejects an empty `tools` array).\n   */\n  private buildToolConfig(tools: ModelCallOptions[\"tools\"]): Pick<ConverseRequest, \"toolConfig\"> {\n    const toolConfig = toBedrockToolConfig(tools);\n\n    return toolConfig ? { toolConfig } : {};\n  }\n\n  /**\n   * Translate the neutral `responseSchema` into Converse's native\n   * `outputConfig.textFormat` (JSON-schema structured output). Bedrock\n   * requires the schema as a stringified JSON document and only\n   * accepts an object root. Emitted only when the model is\n   * `structuredOutput`-capable and the schema is an object — otherwise\n   * the agent's soft system-prompt hint + client-side `validate()`\n   * carry shape (same degradation philosophy as the OpenAI adapter).\n   */\n  private buildOutputConfig(\n    responseSchema: Record<string, unknown> | undefined,\n  ): Pick<ConverseRequest, \"outputConfig\"> {\n    if (!responseSchema || !this.capabilities.structuredOutput) {\n      return {};\n    }\n\n    if (responseSchema.type !== \"object\" || typeof responseSchema.properties !== \"object\") {\n      return {};\n    }\n\n    return {\n      outputConfig: {\n        textFormat: {\n          type: \"json_schema\",\n          structure: {\n            jsonSchema: { name: \"response\", schema: JSON.stringify(responseSchema) },\n          },\n        },\n      },\n    };\n  }\n\n  /**\n   * Concatenate every `text` content block into the single neutral\n   * `content` string. `toolUse` and other block types are surfaced\n   * separately via `extractToolCalls`.\n   */\n  private extractText(blocks: ContentBlock[]): string {\n    return blocks\n      .map((block) => (\"text\" in block && typeof block.text === \"string\" ? block.text : \"\"))\n      .join(\"\");\n  }\n\n  /**\n   * Reshape Converse `toolUse` content blocks into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when no tools were\n   * requested so callers can branch on presence.\n   */\n  private extractToolCalls(blocks: ContentBlock[]): ModelToolCallRequest[] | undefined {\n    const toolCalls: ModelToolCallRequest[] = [];\n\n    for (const block of blocks) {\n      if (\"toolUse\" in block && block.toolUse) {\n        toolCalls.push({\n          id: block.toolUse.toolUseId ?? \"\",\n          name: block.toolUse.name ?? \"\",\n          input: (block.toolUse.input ?? {}) as Record<string, unknown>,\n        });\n      }\n    }\n\n    return toolCalls.length > 0 ? toolCalls : undefined;\n  }\n\n  /**\n   * Normalize Converse's `TokenUsage` into the neutral `Usage` shape.\n   * Bedrock supplies a pre-summed `totalTokens`; cache-read and\n   * cache-write tokens are surfaced as `cachedTokens` /\n   * `cacheWriteTokens` only when non-zero so callers can price the\n   * discounted read rate and the one-time write cost separately.\n   * Bedrock's Converse `TokenUsage` carries no reasoning-token channel,\n   * so `Usage.reasoningTokens` is intentionally left unset here.\n   */\n  private extractUsage(raw: TokenUsage | undefined): Usage {\n    if (!raw) {\n      return { input: 0, output: 0, total: 0 };\n    }\n\n    const input = raw.inputTokens ?? 0;\n    const output = raw.outputTokens ?? 0;\n    const cached = raw.cacheReadInputTokens;\n    const cacheWrite = raw.cacheWriteInputTokens;\n\n    return {\n      input,\n      output,\n      total: raw.totalTokens ?? input + output,\n      ...(cached && cached > 0 ? { cachedTokens: cached } : {}),\n      ...(cacheWrite && cacheWrite > 0 ? { cacheWriteTokens: cacheWrite } : {}),\n    };\n  }\n\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy\n   * and emit the standard error log line before it propagates. Shared\n   * by every catch site so the log shape stays identical.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapBedrockError(thrown);\n\n    this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n      code: wrapped.code,\n      context: wrapped.context,\n    });\n\n    return wrapped;\n  }\n}\n"],"mappings":";;;;;;;;;;;;AA8BA,MAAM,aAAa;;;;;;;;AASnB,MAAM,yBAA6D;CACjE,KAAK;CACL,QAAQ;CACR,MAAM;AACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,IAAa,eAAb,MAAmD;CAUjD,AAAO,YACL,QACA,QACA,WAAmB,WACnB;gBANgC;EAOhC,KAAK,SAAS;EACd,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,KAAK,eAAe;GAClB,kBAAkB,OAAO,oBAAoB;GAC7C,QAAQ,OAAO,UAAU,sBAAsB,OAAO,IAAI;GAC1D,WAAW,OAAO,aAAa,yBAAyB,OAAO,IAAI;GACnE,eAAe,OAAO,iBAAiB,6BAA6B,OAAO,IAAI;GAC/E,KAAK,OAAO,OAAO,mBAAmB,OAAO,IAAI;GACjD,OAAO,OAAO,SAAS;EACzB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,OAAO,MAAM,YAAY,WAAW,0BAA0B;GACjE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAC3B,IAAI,gBAAgB,KAAK,aAAa,UAAU,OAAO,CAAC,GACxD,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,MACtD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,SAAS,SAAS,QAAQ,SAAS,WAAW,CAAC;EACrD,MAAM,eAAe,cAAc,SAAS,UAAU;EACtD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAC9C,MAAM,YAAY,KAAK,iBAAiB,MAAM;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,2BAA2B;GAAE;GAAc;EAAM,CAAC;EAE5F,OAAO;GACL,SAAS,KAAK,YAAY,MAAM;GAChC;GACA;GACA;EACF;CACF;;;;;;;CAQA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,gCAAgC;GACvE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAC3B,IAAI,sBAAsB,KAAK,aAAa,UAAU,OAAO,CAAC,GAC9D,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,MACtD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI;EACJ,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,6BAAa,IAAI,IAAwD;EAE/E,IAAI;GACF,WAAW,MAAM,SAAS,SAAS,UAAU,CAAC,GAAG;IAC/C,IAAI,MAAM,mBAAmB,OAAO,SAAS;KAC3C,MAAM,QAAQ,MAAM,kBAAkB,MAAM;KAE5C,WAAW,IAAI,MAAM,kBAAkB,qBAAqB,GAAG;MAC7D,IAAI,MAAM,aAAa;MACvB,MAAM,MAAM,QAAQ;MACpB,MAAM;KACR,CAAC;KAED;IACF;IAEA,IAAI,MAAM,mBAAmB,OAAO;KAClC,MAAM,QAAQ,MAAM,kBAAkB;KAEtC,IAAI,MAAM,MACR,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM;KAAK;UACtC,IAAI,MAAM,SAAS;MACxB,MAAM,cAAc,WAAW,IAAI,MAAM,kBAAkB,qBAAqB,CAAC;MAEjF,IAAI,aACF,YAAY,QAAQ,MAAM,QAAQ,SAAS;KAE/C;KAEA;IACF;IAEA,IAAI,MAAM,kBAAkB;KAC1B,MAAM,cAAc,WAAW,IAAI,MAAM,iBAAiB,qBAAqB,CAAC;KAEhF,IAAI,aAAa;MACf,MAAM;OACJ,MAAM;OACN,IAAI,YAAY;OAChB,MAAM,YAAY;OAClB,OAAO,cAAuC,YAAY,MAAM,CAAC,CAAC;MACpE;MAEA,WAAW,OAAO,MAAM,iBAAiB,qBAAqB,CAAC;KACjE;KAEA;IACF;IAEA,IAAI,MAAM,aACR,gBAAgB,MAAM,YAAY;IAGpC,IAAI,MAAM,UAAU,OAAO;KACzB,MAAM,MAAM,MAAM,SAAS;KAE3B,MAAM,QAAQ,IAAI,eAAe;KACjC,MAAM,SAAS,IAAI,gBAAgB;KACnC,MAAM,QAAQ,IAAI,eAAe,MAAM,QAAQ,MAAM;KAErD,IAAI,IAAI,wBAAwB,IAAI,uBAAuB,GACzD,MAAM,eAAe,IAAI;KAG3B,IAAI,IAAI,yBAAyB,IAAI,wBAAwB,GAC3D,MAAM,mBAAmB,IAAI;IAEjC;GACF;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,aAAa;EAEhD,KAAK,OAAO,MAAM,YAAY,YAAY,iCAAiC;GACzE;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;CAQA,AAAQ,aACN,UACA,SACiB;EACjB,MAAM,EAAE,QAAQ,UAAU,oBAAoB,kBAAkB,QAAQ;EACxE,MAAM,YAAY,SAAS,aAAa,KAAK,OAAO;EACpD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EACxD,MAAM,iBAAiB,KAAK,sBAAsB,iBAAiB,SAAS,YAAY;EAExF,OAAO;GACL,SAAS,KAAK;GACd,UAAU;GACV,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;GAC3B,iBAAiB;IACf,GAAI,cAAc,SAAY,EAAE,UAAU,IAAI,CAAC;IAC/C,GAAI,gBAAgB,SAAY,EAAE,YAAY,IAAI,CAAC;GACrD;GACA,GAAG,KAAK,gBAAgB,SAAS,KAAK;GACtC,GAAG,KAAK,kBAAkB,SAAS,cAAc;GACjD,GAAG,KAAK,qBAAqB,SAAS,SAAS;EACjD;CACF;;;;;;;;;;;;;;;;CAiBA,AAAQ,sBACN,UACA,cAC6B;EAC7B,MAAM,cAAc,cAAc,eAAe;EAEjD,IAAI,CAAC,KAAK,aAAa,iBAAiB,eAAe,KAAK,CAAC,YAAY,SAAS,WAAW,GAC3F,OAAO;EAGT,MAAM,OAAO,SAAS,SAAS;EAC/B,MAAM,cAAc,SAAS;EAE7B,OAAO,CACL,GAAG,SAAS,MAAM,GAAG,IAAI,GACzB;GACE,GAAG;GACH,SAAS,CAAC,GAAI,YAAY,WAAW,CAAC,GAAI,EAAE,YAAY,EAAE,MAAM,UAAU,EAAE,CAAC;EAC/E,CACF;CACF;;;;;;;;;;;;CAaA,AAAQ,qBACN,WACuD;EACvD,IAAI,CAAC,KAAK,aAAa,aAAa,CAAC,WACnC,OAAO,CAAC;EAGV,MAAM,eAAe,UAAU,aAAa,uBAAuB,UAAU,UAAU;EAEvF,IAAI,iBAAiB,QACnB,OAAO,CAAC;EAGV,OAAO,EACL,8BAA8B,EAC5B,UAAU;GAAE,MAAM;GAAW,eAAe;EAAa,EAC3D,EACF;CACF;;;;;CAMA,AAAQ,gBAAgB,OAAuE;EAC7F,MAAM,aAAa,oBAAoB,KAAK;EAE5C,OAAO,aAAa,EAAE,WAAW,IAAI,CAAC;CACxC;;;;;;;;;;CAWA,AAAQ,kBACN,gBACuC;EACvC,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO,EACL,cAAc,EACZ,YAAY;GACV,MAAM;GACN,WAAW,EACT,YAAY;IAAE,MAAM;IAAY,QAAQ,KAAK,UAAU,cAAc;GAAE,EACzE;EACF,EACF,EACF;CACF;;;;;;CAOA,AAAQ,YAAY,QAAgC;EAClD,OAAO,OACJ,KAAK,UAAW,UAAU,SAAS,OAAO,MAAM,SAAS,WAAW,MAAM,OAAO,EAAG,CAAC,CACrF,KAAK,EAAE;CACZ;;;;;;CAOA,AAAQ,iBAAiB,QAA4D;EACnF,MAAM,YAAoC,CAAC;EAE3C,KAAK,MAAM,SAAS,QAClB,IAAI,aAAa,SAAS,MAAM,SAC9B,UAAU,KAAK;GACb,IAAI,MAAM,QAAQ,aAAa;GAC/B,MAAM,MAAM,QAAQ,QAAQ;GAC5B,OAAQ,MAAM,QAAQ,SAAS,CAAC;EAClC,CAAC;EAIL,OAAO,UAAU,SAAS,IAAI,YAAY;CAC5C;;;;;;;;;;CAWA,AAAQ,aAAa,KAAoC;EACvD,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,QAAQ,IAAI,eAAe;EACjC,MAAM,SAAS,IAAI,gBAAgB;EACnC,MAAM,SAAS,IAAI;EACnB,MAAM,aAAa,IAAI;EAEvB,OAAO;GACL;GACA;GACA,OAAO,IAAI,eAAe,QAAQ;GAClC,GAAI,UAAU,SAAS,IAAI,EAAE,cAAc,OAAO,IAAI,CAAC;GACvD,GAAI,cAAc,aAAa,IAAI,EAAE,kBAAkB,WAAW,IAAI,CAAC;EACzE;CACF;;;;;;CAOA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,iBAAiB,MAAM;EAEvC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;GACtD,MAAM,QAAQ;GACd,SAAS,QAAQ;EACnB,CAAC;EAED,OAAO;CACT;AACF"}

package/llms-full.txt

@@ -8,7 +8,7 @@
 
 ---
 name: setup-bedrock
-description: 'Wire @warlock.js/ai-bedrock — new BedrockSDK({region, credentials?, provider?}) for AWS Bedrock Converse API + Titan embeddings. AWS credential chain (no apiKey). Triggers: `BedrockSDK`, `bedrock.model`, `bedrock.embedder`, `bedrock.count`, `BedrockRuntimeClient`; "how do I use AWS Bedrock", "wire Claude/Nova/Llama on Bedrock", "Titan embeddings", "Bedrock Converse API"; typical import `import { BedrockSDK } from "@warlock.js/ai-bedrock"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; provider choice — `@warlock.js/ai/pick-ai-provider/SKILL.md`; embeddings concepts — `@warlock.js/ai/embed-text/SKILL.md`; competing libs `@aws-sdk/client-bedrock-runtime`, `@ai-sdk/amazon-bedrock`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-anthropic`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`.'
+description: 'Wire @warlock.js/ai-bedrock — new BedrockSDK({region, credentials?, provider?}) for AWS Bedrock Converse API + Titan embeddings. AWS credential chain (no apiKey). .model({name, vision?, reasoning?, promptCaching?, pdf?, audio?}) with cost-truth capabilities inferred per family, extended thinking via options.reasoning → Converse thinking budget_tokens, prompt caching via options.cacheControl → cachePoint, usage cachedTokens/cacheWriteTokens. Triggers: `BedrockSDK`, `bedrock.model`, `bedrock.embedder`, `bedrock.count`, `BedrockRuntimeClient`, `reasoning`, `thinking`, `budget_tokens`, `cachePoint`, `cacheControl`, `promptCaching`, `cachedTokens`, `cacheWriteTokens`; "how do I use AWS Bedrock", "wire Claude/Nova/Llama on Bedrock", "Titan embeddings", "Bedrock Converse API", "Claude 3.7/4 extended thinking on Bedrock", "Bedrock prompt caching cost"; typical import `import { BedrockSDK } from "@warlock.js/ai-bedrock"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; provider choice — `@warlock.js/ai/pick-ai-provider/SKILL.md`; embeddings concepts — `@warlock.js/ai/embed-text/SKILL.md`; competing libs `@aws-sdk/client-bedrock-runtime`, `@ai-sdk/amazon-bedrock`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-anthropic`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`.'
 ---
 
 # `@warlock.js/ai-bedrock`
@@ -51,8 +51,36 @@ Uses the model-agnostic **Converse / ConverseStream** API — one wire mapping c
 | --- | --- |
 | `structuredOutput` | `true` (via Converse `outputConfig.textFormat`) |
 | `vision` | Inferred from model id substring. `true` for Claude 3/3.5/3.7/4, Nova Lite/Pro/Premier, Llama 3.2-11B/90B, Llama 4. |
+| `reasoning` | Inferred from model id. `true` for Claude 3.7 + Claude 4 (Sonnet/Opus/Haiku) — they expose a configurable thinking budget on Converse. |
+| `promptCaching` | Inferred from model id. `true` for Claude 3.5+/3.7/4 and Nova Micro/Lite/Pro/Premier — the families that honor Converse `cachePoint`. |
+| `pdf` | Inferred from model id. `true` for Claude 3+ and Nova Lite/Pro/Premier (Converse `document` blocks). |
+| `audio` | `false` by default — Converse does not accept audio blocks for the families this adapter targets. Pass `audio: true` for a confirmed audio model. |
 
-Explicit config always wins.
+Explicit config always wins — every flag is overridable per model, e.g. `bedrock.model({ name, reasoning: false, promptCaching: true, pdf: false, audio: true })`.
+
+## Reasoning / extended thinking
+
+For a `reasoning`-capable model (Claude 3.7 / 4), `ModelCallOptions.reasoning` maps to Converse's escape hatch `additionalModelRequestFields.thinking = { type: "enabled", budget_tokens }`:
+
+```ts
+// Explicit thinking budget (token cap):
+await model.complete(messages, { reasoning: { maxTokens: 2048 } });
+
+// Effort tier → conventional budget (low 1024 / medium 4096 / high 16384):
+await model.complete(messages, { reasoning: { effort: "high" } });
+```
+
+`maxTokens` wins over `effort` when both are set. For non-reasoning models — or when neither `effort` nor `maxTokens` is given — the option is a no-op, so unsupported params never reach the wire. Bedrock's Converse `TokenUsage` reports no reasoning-token channel, so `usage.reasoningTokens` is left unset (reasoning text arrives as `reasoningContent` content blocks, not a token count).
+
+## Prompt caching
+
+For a `promptCaching`-capable model, `ModelCallOptions.cacheControl.breakpoints` (any positive value) appends one Converse `cachePoint: { type: "default" }` block to the LAST message — Bedrock then caches the whole prefix up to that point:
+
+```ts
+await model.complete(messages, { cacheControl: { breakpoints: 1 } });
+```
+
+Earlier turns are untouched; only the trailing message is marked. No-ops gracefully when the model is not caching-capable or no breakpoint was requested. Read-side accounting works without any breakpoint — see the usage section below.
 
 ## System prompt
 
@@ -81,7 +109,7 @@ Object-root `responseSchema` + `structuredOutput`-capable → `outputConfig.text
 
 - `contentBlockDelta.delta.text` → `{ type: "delta", content }`
 - `toolUse`: `contentBlockStart` opens an accumulator, `contentBlockDelta.delta.toolUse.input` fragments append, `contentBlockStop` emits one consolidated `{ type: "tool-call", ... }`
-- terminal `{ type: "done", finishReason, usage }` — usage from `metadata.usage` (a `TokenUsage`)
+- terminal `{ type: "done", finishReason, usage }` — usage from `metadata.usage` (a `TokenUsage`), including `cachedTokens` / `cacheWriteTokens` when reported (see the usage table below)
 
 ## Finish-reason mapping
 
@@ -136,7 +164,22 @@ const { usage } = await ai.agent({ model: bedrock.model({ name: "us.amazon.nova-
 usage.cost;  // per-channel USD breakdown of THIS run, computed from tokens × pricing[model]
 ```
 
-Resolution at `model()` time: per-model `pricing` (`bedrock.model({ name, pricing })`) > SDK registry > `undefined` (no cost computed). When the same model id appears under a cross-region inference profile, list each id you actually invoke (`us.`, `eu.`, `apac.` prefixes are distinct keys). Bedrock meters cache-read tokens via Converse `usage` — those surface as `usage.cachedTokens`, so set `cachedInput` to have them priced at the discounted rate. See [`@warlock.js/ai/pick-ai-provider/SKILL.md`](@warlock.js/ai/pick-ai-provider/SKILL.md).
+Resolution at `model()` time: per-model `pricing` (`bedrock.model({ name, pricing })`) > SDK registry > `undefined` (no cost computed). When the same model id appears under a cross-region inference profile, list each id you actually invoke (`us.`, `eu.`, `apac.` prefixes are distinct keys). See [`@warlock.js/ai/pick-ai-provider/SKILL.md`](@warlock.js/ai/pick-ai-provider/SKILL.md).
+
+### Tokens & usage — cost truth
+
+Both `complete()` and `stream()` normalize Converse's `TokenUsage` into the neutral `Usage` shape:
+
+| Converse `TokenUsage` field | Neutral `Usage` field | Notes |
+| --- | --- | --- |
+| `inputTokens` | `usage.input` | Includes cached read tokens. |
+| `outputTokens` | `usage.output` | |
+| `totalTokens` | `usage.total` | Falls back to `input + output` when absent. |
+| `cacheReadInputTokens` | `usage.cachedTokens` | Only when `> 0`. Cache-hit reads — price at `pricing.cachedInput`. |
+| `cacheWriteInputTokens` | `usage.cacheWriteTokens` | Only when `> 0`. One-time cache write — price at `pricing.cachedOutput`. |
+| — | `usage.reasoningTokens` | **Never set** — Converse exposes no reasoning-token count. |
+
+Set `cachedInput` / `cachedOutput` in `ModelPricing` to have cache reads/writes billed at their distinct rates. Read-side accounting (`cachedTokens` / `cacheWriteTokens`) is populated whenever Bedrock reports it, independent of whether you sent a `cacheControl` breakpoint.
 
 ## When NOT to use this skill
 

package/llms.txt

@@ -6,4 +6,4 @@
 
 ## Skills
 
-- [setup-bedrock](@warlock.js/ai-bedrock/setup-bedrock/SKILL.md): Wire @warlock.js/ai-bedrock — new BedrockSDK({region, credentials?, provider?}) for AWS Bedrock Converse API + Titan embeddings. AWS credential chain (no apiKey). Triggers: `BedrockSDK`, `bedrock.model`, `bedrock.embedder`, `bedrock.count`, `BedrockRuntimeClient`; "how do I use AWS Bedrock", "wire Claude/Nova/Llama on Bedrock", "Titan embeddings", "Bedrock Converse API"; typical import `import { BedrockSDK } from "@warlock.js/ai-bedrock"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; provider choice — `@warlock.js/ai/pick-ai-provider/SKILL.md`; embeddings concepts — `@warlock.js/ai/embed-text/SKILL.md`; competing libs `@aws-sdk/client-bedrock-runtime`, `@ai-sdk/amazon-bedrock`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-anthropic`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`.
+- [setup-bedrock](@warlock.js/ai-bedrock/setup-bedrock/SKILL.md): Wire @warlock.js/ai-bedrock — new BedrockSDK({region, credentials?, provider?}) for AWS Bedrock Converse API + Titan embeddings. AWS credential chain (no apiKey). .model({name, vision?, reasoning?, promptCaching?, pdf?, audio?}) with cost-truth capabilities inferred per family, extended thinking via options.reasoning → Converse thinking budget_tokens, prompt caching via options.cacheControl → cachePoint, usage cachedTokens/cacheWriteTokens. Triggers: `BedrockSDK`, `bedrock.model`, `bedrock.embedder`, `bedrock.count`, `BedrockRuntimeClient`, `reasoning`, `thinking`, `budget_tokens`, `cachePoint`, `cacheControl`, `promptCaching`, `cachedTokens`, `cacheWriteTokens`; "how do I use AWS Bedrock", "wire Claude/Nova/Llama on Bedrock", "Titan embeddings", "Bedrock Converse API", "Claude 3.7/4 extended thinking on Bedrock", "Bedrock prompt caching cost"; typical import `import { BedrockSDK } from "@warlock.js/ai-bedrock"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; provider choice — `@warlock.js/ai/pick-ai-provider/SKILL.md`; embeddings concepts — `@warlock.js/ai/embed-text/SKILL.md`; competing libs `@aws-sdk/client-bedrock-runtime`, `@ai-sdk/amazon-bedrock`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-anthropic`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`.

package/package.json

@@ -15,12 +15,12 @@
   },
   "dependencies": {
     "@aws-sdk/client-bedrock-runtime": "^3.1048.0",
-    "@warlock.js/logger": "4.2.10"
+    "@warlock.js/logger": "4.3.0"
   },
   "peerDependencies": {
-    "@warlock.js/ai": "4.2.10"
+    "@warlock.js/ai": "4.3.0"
   },
-  "version": "4.2.10",
+  "version": "4.3.0",
   "main": "./cjs/index.cjs",
   "module": "./esm/index.mjs",
   "types": "./esm/index.d.mts",

package/skills/setup-bedrock/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: setup-bedrock
-description: 'Wire @warlock.js/ai-bedrock — new BedrockSDK({region, credentials?, provider?}) for AWS Bedrock Converse API + Titan embeddings. AWS credential chain (no apiKey). Triggers: `BedrockSDK`, `bedrock.model`, `bedrock.embedder`, `bedrock.count`, `BedrockRuntimeClient`; "how do I use AWS Bedrock", "wire Claude/Nova/Llama on Bedrock", "Titan embeddings", "Bedrock Converse API"; typical import `import { BedrockSDK } from "@warlock.js/ai-bedrock"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; provider choice — `@warlock.js/ai/pick-ai-provider/SKILL.md`; embeddings concepts — `@warlock.js/ai/embed-text/SKILL.md`; competing libs `@aws-sdk/client-bedrock-runtime`, `@ai-sdk/amazon-bedrock`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-anthropic`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`.'
+description: 'Wire @warlock.js/ai-bedrock — new BedrockSDK({region, credentials?, provider?}) for AWS Bedrock Converse API + Titan embeddings. AWS credential chain (no apiKey). .model({name, vision?, reasoning?, promptCaching?, pdf?, audio?}) with cost-truth capabilities inferred per family, extended thinking via options.reasoning → Converse thinking budget_tokens, prompt caching via options.cacheControl → cachePoint, usage cachedTokens/cacheWriteTokens. Triggers: `BedrockSDK`, `bedrock.model`, `bedrock.embedder`, `bedrock.count`, `BedrockRuntimeClient`, `reasoning`, `thinking`, `budget_tokens`, `cachePoint`, `cacheControl`, `promptCaching`, `cachedTokens`, `cacheWriteTokens`; "how do I use AWS Bedrock", "wire Claude/Nova/Llama on Bedrock", "Titan embeddings", "Bedrock Converse API", "Claude 3.7/4 extended thinking on Bedrock", "Bedrock prompt caching cost"; typical import `import { BedrockSDK } from "@warlock.js/ai-bedrock"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; provider choice — `@warlock.js/ai/pick-ai-provider/SKILL.md`; embeddings concepts — `@warlock.js/ai/embed-text/SKILL.md`; competing libs `@aws-sdk/client-bedrock-runtime`, `@ai-sdk/amazon-bedrock`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-anthropic`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`.'
 ---
 
 # `@warlock.js/ai-bedrock`
@@ -43,8 +43,36 @@ Uses the model-agnostic **Converse / ConverseStream** API — one wire mapping c
 | --- | --- |
 | `structuredOutput` | `true` (via Converse `outputConfig.textFormat`) |
 | `vision` | Inferred from model id substring. `true` for Claude 3/3.5/3.7/4, Nova Lite/Pro/Premier, Llama 3.2-11B/90B, Llama 4. |
+| `reasoning` | Inferred from model id. `true` for Claude 3.7 + Claude 4 (Sonnet/Opus/Haiku) — they expose a configurable thinking budget on Converse. |
+| `promptCaching` | Inferred from model id. `true` for Claude 3.5+/3.7/4 and Nova Micro/Lite/Pro/Premier — the families that honor Converse `cachePoint`. |
+| `pdf` | Inferred from model id. `true` for Claude 3+ and Nova Lite/Pro/Premier (Converse `document` blocks). |
+| `audio` | `false` by default — Converse does not accept audio blocks for the families this adapter targets. Pass `audio: true` for a confirmed audio model. |
 
-Explicit config always wins.
+Explicit config always wins — every flag is overridable per model, e.g. `bedrock.model({ name, reasoning: false, promptCaching: true, pdf: false, audio: true })`.
+
+## Reasoning / extended thinking
+
+For a `reasoning`-capable model (Claude 3.7 / 4), `ModelCallOptions.reasoning` maps to Converse's escape hatch `additionalModelRequestFields.thinking = { type: "enabled", budget_tokens }`:
+
+```ts
+// Explicit thinking budget (token cap):
+await model.complete(messages, { reasoning: { maxTokens: 2048 } });
+
+// Effort tier → conventional budget (low 1024 / medium 4096 / high 16384):
+await model.complete(messages, { reasoning: { effort: "high" } });
+```
+
+`maxTokens` wins over `effort` when both are set. For non-reasoning models — or when neither `effort` nor `maxTokens` is given — the option is a no-op, so unsupported params never reach the wire. Bedrock's Converse `TokenUsage` reports no reasoning-token channel, so `usage.reasoningTokens` is left unset (reasoning text arrives as `reasoningContent` content blocks, not a token count).
+
+## Prompt caching
+
+For a `promptCaching`-capable model, `ModelCallOptions.cacheControl.breakpoints` (any positive value) appends one Converse `cachePoint: { type: "default" }` block to the LAST message — Bedrock then caches the whole prefix up to that point:
+
+```ts
+await model.complete(messages, { cacheControl: { breakpoints: 1 } });
+```
+
+Earlier turns are untouched; only the trailing message is marked. No-ops gracefully when the model is not caching-capable or no breakpoint was requested. Read-side accounting works without any breakpoint — see the usage section below.
 
 ## System prompt
 
@@ -73,7 +101,7 @@ Object-root `responseSchema` + `structuredOutput`-capable → `outputConfig.text
 
 - `contentBlockDelta.delta.text` → `{ type: "delta", content }`
 - `toolUse`: `contentBlockStart` opens an accumulator, `contentBlockDelta.delta.toolUse.input` fragments append, `contentBlockStop` emits one consolidated `{ type: "tool-call", ... }`
-- terminal `{ type: "done", finishReason, usage }` — usage from `metadata.usage` (a `TokenUsage`)
+- terminal `{ type: "done", finishReason, usage }` — usage from `metadata.usage` (a `TokenUsage`), including `cachedTokens` / `cacheWriteTokens` when reported (see the usage table below)
 
 ## Finish-reason mapping
 
@@ -128,7 +156,22 @@ const { usage } = await ai.agent({ model: bedrock.model({ name: "us.amazon.nova-
 usage.cost;  // per-channel USD breakdown of THIS run, computed from tokens × pricing[model]
 ```
 
-Resolution at `model()` time: per-model `pricing` (`bedrock.model({ name, pricing })`) > SDK registry > `undefined` (no cost computed). When the same model id appears under a cross-region inference profile, list each id you actually invoke (`us.`, `eu.`, `apac.` prefixes are distinct keys). Bedrock meters cache-read tokens via Converse `usage` — those surface as `usage.cachedTokens`, so set `cachedInput` to have them priced at the discounted rate. See [`@warlock.js/ai/pick-ai-provider/SKILL.md`](@warlock.js/ai/pick-ai-provider/SKILL.md).
+Resolution at `model()` time: per-model `pricing` (`bedrock.model({ name, pricing })`) > SDK registry > `undefined` (no cost computed). When the same model id appears under a cross-region inference profile, list each id you actually invoke (`us.`, `eu.`, `apac.` prefixes are distinct keys). See [`@warlock.js/ai/pick-ai-provider/SKILL.md`](@warlock.js/ai/pick-ai-provider/SKILL.md).
+
+### Tokens & usage — cost truth
+
+Both `complete()` and `stream()` normalize Converse's `TokenUsage` into the neutral `Usage` shape:
+
+| Converse `TokenUsage` field | Neutral `Usage` field | Notes |
+| --- | --- | --- |
+| `inputTokens` | `usage.input` | Includes cached read tokens. |
+| `outputTokens` | `usage.output` | |
+| `totalTokens` | `usage.total` | Falls back to `input + output` when absent. |
+| `cacheReadInputTokens` | `usage.cachedTokens` | Only when `> 0`. Cache-hit reads — price at `pricing.cachedInput`. |
+| `cacheWriteInputTokens` | `usage.cacheWriteTokens` | Only when `> 0`. One-time cache write — price at `pricing.cachedOutput`. |
+| — | `usage.reasoningTokens` | **Never set** — Converse exposes no reasoning-token count. |
+
+Set `cachedInput` / `cachedOutput` in `ModelPricing` to have cache reads/writes billed at their distinct rates. Read-side accounting (`cachedTokens` / `cacheWriteTokens`) is populated whenever Bedrock reports it, independent of whether you sent a `cacheControl` breakpoint.
 
 ## When NOT to use this skill