@warlock.js/ai-anthropic 4.2.11 → 4.3.0

package/CHANGELOG.md

@@ -6,6 +6,16 @@ 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 wiring against the core `@warlock.js/ai` contract:
+  - **Usage accounting** — `usage.cacheWriteTokens` is now populated from Anthropic's `cache_creation_input_tokens` (the cache-write surcharge), alongside the existing `cachedTokens` (from `cache_read_input_tokens`). In streaming, both are seeded from `message_start` and then refreshed from the cumulative `message_delta` counts. `reasoningTokens` is intentionally left unset because Anthropic bills extended-thinking tokens inside `output_tokens`.
+  - **Extended thinking** — `ModelCallOptions.reasoning` now maps to Anthropic's `thinking: { type: "enabled", budget_tokens }`. `reasoning.maxTokens` sets the budget directly; `reasoning.effort` (`low`/`medium`/`high`) maps to a tiered budget (`1024`/`4096`/`12000`); any budget is floored at the `1024` minimum. `temperature` is dropped when thinking is enabled (Anthropic rejects the combination). New `reasoning?: boolean` model-config override.
+  - **System-prompt prompt caching** — a per-call `ModelCallOptions.cacheControl.breakpoints >= 1` now emits the hoisted system prompt as a `TextBlockParam` carrying `cache_control: { type: "ephemeral" }`, independent of the existing tools-caching `promptCaching` flag.
+  - **Capabilities** — `reasoning` (default `true`, overridable), `promptCaching` (`true`), and `pdf` (`true`) are now advertised on `ModelCapabilities`; `audio` stays absent (Anthropic has no audio-input block).
+
 ## 4.2.0
 
 ### Added

package/cjs/index.cjs

@@ -421,6 +421,24 @@ const LOG_MODULE = "ai.anthropic";
 */
 const DEFAULT_MAX_TOKENS = 4096;
 /**
+* Anthropic rejects an extended-thinking budget below 1024 tokens, so
+* any resolved budget is floored at this minimum before it reaches the
+* wire.
+*/
+const MIN_THINKING_BUDGET = 1024;
+/**
+* Map the neutral `ReasoningEffort` level to an Anthropic
+* `thinking.budget_tokens` value. Anthropic budgets reasoning by token
+* count (unlike OpenAI's opaque `reasoning_effort` enum), so the three
+* neutral levels translate to representative token budgets when the
+* caller doesn't pass an explicit `reasoning.maxTokens`.
+*/
+const EFFORT_THINKING_BUDGET = {
+	low: 1024,
+	medium: 4096,
+	high: 12e3
+};
+/**
 * Anthropic-backed implementation of `ModelContract`.
 *
 * **Role.** The provider-facing bridge between the vendor-neutral
@@ -468,7 +486,10 @@ var AnthropicModel = 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 ?? true,
+			promptCaching: true,
+			pdf: true
 		};
 	}
 	/**
@@ -541,8 +562,10 @@ var AnthropicModel = class {
 			for await (const event of stream) {
 				if (event.type === "message_start") {
 					usage.input = event.message.usage.input_tokens ?? 0;
-					const cached = event.message.usage.cache_read_input_tokens;
-					if (cached !== null && cached !== void 0 && cached > 0) usage.cachedTokens = cached;
+					const cacheRead = event.message.usage.cache_read_input_tokens;
+					if (cacheRead !== null && cacheRead !== void 0 && cacheRead > 0) usage.cachedTokens = cacheRead;
+					const cacheWrite = event.message.usage.cache_creation_input_tokens;
+					if (cacheWrite !== null && cacheWrite !== void 0 && cacheWrite > 0) usage.cacheWriteTokens = cacheWrite;
 					continue;
 				}
 				if (event.type === "content_block_start") {
@@ -581,6 +604,10 @@ var AnthropicModel = class {
 				if (event.type === "message_delta") {
 					rawStopReason = event.delta.stop_reason ?? rawStopReason;
 					usage.output = event.usage.output_tokens ?? usage.output;
+					const cacheRead = event.usage.cache_read_input_tokens;
+					if (cacheRead !== null && cacheRead !== void 0 && cacheRead > 0) usage.cachedTokens = cacheRead;
+					const cacheWrite = event.usage.cache_creation_input_tokens;
+					if (cacheWrite !== null && cacheWrite !== void 0 && cacheWrite > 0) usage.cacheWriteTokens = cacheWrite;
 				}
 			}
 		} catch (thrown) {
@@ -604,22 +631,69 @@ var AnthropicModel = class {
 	* overload resolves to the right return type). Hoists the system
 	* prompt out of `messages`, resolves `max_tokens` (required by
 	* Anthropic) with the documented default, and conditionally attaches
-	* temperature, tools, and native structured output.
+	* temperature, tools, native structured output, extended thinking
+	* (`reasoning`), and a system-prompt cache breakpoint (`cacheControl`).
+	* Temperature is dropped when thinking is enabled, since Anthropic
+	* rejects the two together.
 	*/
 	buildParams(messages, options) {
 		const { system, messages: anthropicMessages } = toAnthropicMessages(messages);
+		const thinking = this.buildThinking(options?.reasoning);
 		const temperature = options?.temperature ?? this.config.temperature;
 		return {
 			model: this.name,
 			max_tokens: options?.maxTokens ?? this.config.maxTokens ?? DEFAULT_MAX_TOKENS,
 			messages: anthropicMessages,
-			...system ? { system } : {},
-			...temperature !== void 0 ? { temperature } : {},
+			...this.buildSystem(system, options?.cacheControl),
+			...temperature !== void 0 && !thinking.thinking ? { temperature } : {},
 			...this.buildTools(options?.tools),
-			...this.buildStructuredOutput(options?.responseSchema)
+			...this.buildStructuredOutput(options?.responseSchema),
+			...thinking
 		};
 	}
 	/**
+	* Spread-friendly `system` fragment. Returns an empty object when no
+	* system prompt was hoisted out of the messages.
+	*
+	* When a per-call `cacheControl.breakpoints` hint is present (≥ 1),
+	* the system prompt is emitted as a single `TextBlockParam` carrying
+	* `cache_control: ephemeral` — the system prompt is the longest stable
+	* prefix on a turn, so one breakpoint there lets multi-turn agents
+	* read it back at the ~0.1x cache-read rate. Without the hint the
+	* system prompt stays a plain string (left uncached) so a one-shot
+	* call never pays the ~1.25x cache-write surcharge.
+	*/
+	buildSystem(system, cacheControl) {
+		if (!system) return {};
+		const breakpoints = cacheControl?.breakpoints ?? 0;
+		if (this.capabilities.promptCaching && breakpoints > 0) return { system: [{
+			type: "text",
+			text: system,
+			cache_control: { type: "ephemeral" }
+		}] };
+		return { system };
+	}
+	/**
+	* Translate the neutral `reasoning` option into Anthropic's
+	* `thinking` request field. Emitted only when the model declares the
+	* `reasoning` capability AND a reasoning option is supplied; otherwise
+	* returns an empty object so the caller can unconditionally spread it.
+	*
+	* Budget resolution: an explicit `reasoning.maxTokens` wins; otherwise
+	* the neutral `effort` level maps to a tiered token budget. Anthropic
+	* requires `budget_tokens` ≥ 1024, so the budget is floored at that
+	* minimum.
+	*/
+	buildThinking(reasoning) {
+		if (!this.capabilities.reasoning || !reasoning) return {};
+		if (reasoning.maxTokens === void 0 && reasoning.effort === void 0) return {};
+		const budget = reasoning.maxTokens ?? EFFORT_THINKING_BUDGET[reasoning.effort ?? "medium"];
+		return { thinking: {
+			type: "enabled",
+			budget_tokens: Math.max(MIN_THINKING_BUDGET, budget)
+		} };
+	}
+	/**
 	* Spread-friendly tools fragment. Returns an empty object when no
 	* tools were supplied so the caller can unconditionally spread it.
 	*
@@ -690,17 +764,25 @@ var AnthropicModel = class {
 	* Normalize Anthropic's `usage` block into the neutral `Usage` shape.
 	* Anthropic reports `input_tokens` / `output_tokens` separately with
 	* no pre-summed total, so `total` is computed. Cache-read tokens are
-	* surfaced as `cachedTokens` only when non-zero.
+	* surfaced as `cachedTokens` and cache-write tokens as
+	* `cacheWriteTokens`, each only when non-zero.
+	*
+	* Note: Anthropic does not report a separate reasoning-token count —
+	* extended-thinking tokens are billed inside `output_tokens` — so
+	* `Usage.reasoningTokens` is intentionally left unset here. Populating
+	* it would double-count against `output`.
 	*/
 	extractUsage(raw) {
 		const input = raw.input_tokens ?? 0;
 		const output = raw.output_tokens ?? 0;
-		const cached = raw.cache_read_input_tokens;
+		const cacheRead = raw.cache_read_input_tokens;
+		const cacheWrite = raw.cache_creation_input_tokens;
 		return {
 			input,
 			output,
 			total: input + output,
-			...cached !== null && cached !== void 0 && cached > 0 ? { cachedTokens: cached } : {}
+			...cacheRead !== null && cacheRead !== void 0 && cacheRead > 0 ? { cachedTokens: cacheRead } : {},
+			...cacheWrite !== null && cacheWrite !== void 0 && cacheWrite > 0 ? { cacheWriteTokens: cacheWrite } : {}
 		};
 	}
 	/**

package/cjs/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","names":["AIError","ProviderTimeoutError","ProviderAuthError","QuotaExceededError","ProviderRateLimitError","ContextLengthExceededError","InvalidRequestError","ProviderError","APIError","APIConnectionTimeoutError","log","Anthropic"],"sources":["../../../../../../@warlock.js/ai-anthropic/src/known-vision-models.ts","../../../../../../@warlock.js/ai-anthropic/src/utils/map-stop-reason.ts","../../../../../../@warlock.js/ai-anthropic/src/utils/to-anthropic-messages.ts","../../../../../../@warlock.js/ai-anthropic/src/utils/to-anthropic-tools.ts","../../../../../../@warlock.js/ai-anthropic/src/utils/wrap-anthropic-error.ts","../../../../../../@warlock.js/ai-anthropic/src/model.ts","../../../../../../@warlock.js/ai-anthropic/src/sdk.ts"],"sourcesContent":["/**\n * Model-name prefixes for Claude families that accept image input\n * (vision) on the Messages API.\n *\n * Every Claude 3, Claude 3.5/3.7, and Claude 4 family model is\n * multimodal, so the list covers both the dotted legacy naming\n * (`claude-3-haiku-...`, `claude-3-5-sonnet-...`) and the current\n * `claude-<tier>-4-*` naming (`claude-opus-4-7`, `claude-sonnet-4-6`,\n * `claude-haiku-4-5`). Pre-3 families (`claude-2`, `claude-instant`)\n * are text-only and intentionally absent.\n *\n * Matched as a prefix so dated variants (`claude-opus-4-20250514`) are\n * covered without listing every release tag. Devs can always override\n * per-model via `anthropic.model({ name, vision: true | false })` —\n * explicit config wins over inference in either direction.\n */\nconst VISION_CAPABLE_PREFIXES = [\n  \"claude-3\",\n  \"claude-4\",\n  \"claude-opus-4\",\n  \"claude-sonnet-4\",\n  \"claude-haiku-4\",\n];\n\n/**\n * Infer whether a given Claude model name supports vision based on the\n * known-prefix list. Unknown models default to `false` so that passing\n * an image attachment to an unsupported model surfaces a clear,\n * agent-side capability error instead of an opaque Anthropic 400.\n *\n * @example\n * inferVisionCapability(\"claude-sonnet-4-6\");      // → true\n * inferVisionCapability(\"claude-3-5-sonnet-latest\"); // → true\n * inferVisionCapability(\"claude-2.1\");             // → false\n * inferVisionCapability(\"custom-proxy-llm\");       // → false\n */\nexport function inferVisionCapability(modelName: string): boolean {\n  const normalized = modelName.toLowerCase();\n\n  return VISION_CAPABLE_PREFIXES.some((prefix) => normalized.startsWith(prefix));\n}\n","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 Anthropic's `stop_reason` to the normalized `FinishReason` union.\n *\n * `end_turn` / `stop_sequence` are both natural stops. `max_tokens`\n * maps to `length`. `tool_use` maps to `tool_calls`. Everything else —\n * `refusal` (policy intervention), `pause_turn` (incomplete\n * long-running turn), `null`, or any value a future API version adds —\n * falls through to `\"error\"` so the agent treats the trip as a\n * non-clean terminal rather than silently accepting a partial result.\n *\n * @example\n * mapStopReason(\"end_turn\");  // \"stop\"\n * mapStopReason(\"tool_use\");  // \"tool_calls\"\n * mapStopReason(\"refusal\");   // \"error\"\n * mapStopReason(null);        // \"error\"\n */\nexport function mapStopReason(raw: string | null | undefined): FinishReason {\n  return stopReasonMap[raw ?? \"\"] ?? \"error\";\n}\n","import type { ContentPart, Message } from \"@warlock.js/ai\";\nimport type Anthropic from \"@anthropic-ai/sdk\";\n\n/**\n * Result of splitting a vendor-neutral `Message[]` for the Anthropic\n * Messages API: the system prompt is hoisted to a top-level `system`\n * string (Anthropic has no `\"system\"` role inside `messages`), and the\n * remaining turns are mapped to `MessageParam[]`.\n */\nexport type AnthropicMessages = {\n  system: string | undefined;\n  messages: Anthropic.MessageParam[];\n};\n\nconst ANTHROPIC_IMAGE_MEDIA_TYPES = [\"image/jpeg\", \"image/png\", \"image/gif\", \"image/webp\"] as const;\n\ntype AnthropicImageMediaType = (typeof ANTHROPIC_IMAGE_MEDIA_TYPES)[number];\n\n/**\n * Convert vendor-neutral `Message[]` into Anthropic's request shape.\n *\n * Anthropic differs from the OpenAI Chat protocol in three ways this\n * function absorbs:\n *\n * 1. **No `system` role.** System messages are concatenated (newline-\n *    separated) and returned separately as the top-level `system`\n *    parameter.\n * 2. **Tool results are `user` turns.** A neutral `tool` message becomes\n *    a `user` message whose content is a single `tool_result` block\n *    keyed by `tool_use_id`.\n * 3. **Tool calls are `tool_use` content blocks.** An assistant message\n *    carrying `toolCalls` becomes an `assistant` message whose content\n *    is an optional leading `text` block followed by one `tool_use`\n *    block per call.\n *\n * Consecutive same-role turns are left as-is — the Messages API merges\n * them server-side.\n *\n * @example\n * const { system, messages } = toAnthropicMessages([\n *   { role: \"system\", content: \"Be concise.\" },\n *   { role: \"user\", content: \"Hi\" },\n * ]);\n * // system === \"Be concise.\"\n * // messages === [{ role: \"user\", content: \"Hi\" }]\n */\nexport function toAnthropicMessages(messages: Message[]): AnthropicMessages {\n  const systemParts: string[] = [];\n  const mapped: Anthropic.MessageParam[] = [];\n\n  for (const message of messages) {\n    if (message.role === \"system\") {\n      systemParts.push(stringifyContent(message.content));\n\n      continue;\n    }\n\n    if (message.role === \"tool\") {\n      mapped.push({\n        role: \"user\",\n        content: [\n          {\n            type: \"tool_result\",\n            tool_use_id: message.toolCallId ?? \"\",\n            content: stringifyContent(message.content),\n          },\n        ],\n      });\n\n      continue;\n    }\n\n    if (message.role === \"assistant\" && message.toolCalls && message.toolCalls.length > 0) {\n      const blocks: Anthropic.ContentBlockParam[] = [];\n      const text = stringifyContent(message.content);\n\n      if (text) {\n        blocks.push({ type: \"text\", text });\n      }\n\n      for (const toolCall of message.toolCalls) {\n        blocks.push({\n          type: \"tool_use\",\n          id: toolCall.id,\n          name: toolCall.name,\n          input: toolCall.input ?? {},\n        });\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(toAnthropicContentBlock),\n      });\n\n      continue;\n    }\n\n    mapped.push({\n      role: message.role === \"assistant\" ? \"assistant\" : \"user\",\n      content: stringifyContent(message.content),\n    });\n  }\n\n  return {\n    system: systemParts.length > 0 ? systemParts.join(\"\\n\\n\") : 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 so the wire\n * format stays valid. Plain 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 single resolved `ContentPart` to an Anthropic content block.\n * Text passes straight through; images become an `image` block with a\n * `url` source (remote) or a `base64` source (inlined bytes). The\n * agent has already resolved every attachment before it reaches here,\n * so this never reads files or fetches URLs.\n */\nfunction toAnthropicContentBlock(part: ContentPart): Anthropic.ContentBlockParam {\n  if (part.type === \"text\") {\n    return { type: \"text\", text: part.text };\n  }\n\n  if (\"url\" in part.source) {\n    return { type: \"image\", source: { type: \"url\", url: part.source.url } };\n  }\n\n  return {\n    type: \"image\",\n    source: {\n      type: \"base64\",\n      media_type: part.source.mediaType as AnthropicImageMediaType,\n      data: part.source.base64,\n    },\n  };\n}\n","import { extractJsonSchema, type ToolConfig } from \"@warlock.js/ai\";\nimport type Anthropic from \"@anthropic-ai/sdk\";\n\n/**\n * Convert vendor-neutral `ToolConfig[]` into Anthropic's `tools` array.\n * Uses the shared `extractJsonSchema` helper; Anthropic requires the\n * input schema to be a JSON-Schema object, so a non-object extraction\n * is coerced into an empty-object schema rather than rejected — the\n * tool still registers and the model simply sees no parameters.\n *\n * @example\n * const tools = toAnthropicTools([weatherTool, calculatorTool]);\n * await client.messages.create({ model, max_tokens, messages, tools });\n */\nexport function toAnthropicTools(\n  tools: ToolConfig<unknown, unknown>[] | undefined,\n): Anthropic.Tool[] | undefined {\n  if (!tools || tools.length === 0) {\n    return undefined;\n  }\n\n  return tools.map((tool) => ({\n    name: tool.name,\n    description: tool.description,\n    input_schema: toInputSchema(tool.input),\n  }));\n}\n\n/**\n * Coerce the extracted JSON Schema into Anthropic's `Tool.InputSchema`\n * shape (root must be `{ type: \"object\" }`). Anything that isn't an\n * object schema degrades to a parameterless object so registration\n * never fails on a malformed extractor result.\n */\nfunction toInputSchema(input: ToolConfig<unknown, unknown>[\"input\"]): Anthropic.Tool.InputSchema {\n  const schema = extractJsonSchema(input);\n\n  if (schema && schema.type === \"object\") {\n    return schema as Anthropic.Tool.InputSchema;\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\";\nimport { APIConnectionTimeoutError, APIError } from \"@anthropic-ai/sdk\";\n\n/**\n * Raw-error fields the wrapper reads off an Anthropic SDK error.\n *\n * `APIError` exposes `status`, `type` (the `error.type` from the\n * response body, e.g. `\"rate_limit_error\"`), `message`, `headers`, and\n * `requestID`. We duck-type because wrapped retries, proxied errors,\n * and custom subclasses sometimes lose the `instanceof` relationship.\n */\ntype AnthropicErrorShape = {\n  status?: number;\n  type?: string | null;\n  message?: string;\n  headers?: HeaderBag | undefined;\n  name?: string;\n  requestId?: string;\n};\n\n/** Either a fetch `Headers` instance or a plain record (duck-typed tests). */\ntype HeaderBag = Headers | Record<string, string>;\n\n/**\n * Wrap any thrown value caught inside the Anthropic adapter into the\n * appropriate `@warlock.js/ai` `AIError` subclass.\n *\n * **Dispatch strategy.** Anthropic has no per-error machine `code`; the\n * stable identifier is `error.type` on the response body (surfaced as\n * `APIError.type`). Dispatch prefers `type`, falls back to `status`\n * when the body was stripped (common with proxies). Name-based\n * detection catches transport-layer timeouts that never produced an\n * HTTP response. The `invalid_request_error` branch additionally\n * sniffs the message for Anthropic's \"prompt is too long\" phrasing,\n * which is the only signal that a 400 was a context-length overflow.\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.messages.create(...);\n * } catch (thrown) {\n *   throw wrapAnthropicError(thrown);\n * }\n */\nexport function wrapAnthropicError(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(thrown, shape)) {\n    return new ProviderTimeoutError(message, { cause: thrown, context });\n  }\n\n  if (shape.type === \"authentication_error\" || shape.type === \"permission_error\") {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.status === 401 || shape.status === 403) {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.type === \"billing_error\") {\n    return new QuotaExceededError(message, { cause: thrown, context });\n  }\n\n  if (shape.type === \"rate_limit_error\" || shape.status === 429) {\n    return new ProviderRateLimitError(message, {\n      cause: thrown,\n      context,\n      retryAfter: parseRetryAfter(shape.headers),\n    });\n  }\n\n  if (shape.type === \"invalid_request_error\" || isClientStatus(shape.status)) {\n    if (/prompt is too long/i.test(message)) {\n      return new ContextLengthExceededError(message, { cause: thrown, context });\n    }\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 APIError`\n * — proxies and re-wrappers strip the prototype chain. Duck-typing on\n * the visible fields is resilient to both.\n */\nfunction toShape(thrown: unknown): AnthropicErrorShape {\n  if (thrown instanceof APIError) {\n    return {\n      status: typeof thrown.status === \"number\" ? thrown.status : undefined,\n      type: thrown.type,\n      message: thrown.message,\n      headers: thrown.headers,\n      name: thrown.name,\n      requestId: thrown.requestID ?? undefined,\n    };\n  }\n\n  if (typeof thrown === \"object\" && thrown !== null) {\n    const raw = thrown as Record<string, unknown>;\n\n    return {\n      status: typeof raw.status === \"number\" ? raw.status : undefined,\n      type: typeof raw.type === \"string\" ? raw.type : undefined,\n      message: typeof raw.message === \"string\" ? raw.message : undefined,\n      headers: isHeaderBag(raw.headers) ? raw.headers : undefined,\n      name: typeof raw.name === \"string\" ? raw.name : undefined,\n      requestId: readRequestId(raw),\n    };\n  }\n\n  return {};\n}\n\n/**\n * Decide whether the thrown value represents a timeout. The Anthropic\n * SDK throws `APIConnectionTimeoutError` for transport-level timeouts;\n * Node surfaces `ETIMEDOUT` / `ECONNABORTED` on the socket layer.\n * Either signal counts.\n */\nfunction isTimeout(thrown: unknown, shape: AnthropicErrorShape): boolean {\n  if (thrown instanceof APIConnectionTimeoutError) {\n    return true;\n  }\n\n  if (shape.name === \"APIConnectionTimeoutError\") {\n    return true;\n  }\n\n  if (typeof thrown === \"object\" && thrown !== null) {\n    const code = (thrown as Record<string, unknown>).code;\n\n    if (code === \"ETIMEDOUT\" || code === \"ECONNABORTED\") {\n      return true;\n    }\n  }\n\n  return false;\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` so consumers\n * have everything the provider surfaced without each subclass having\n * to redeclare them. Never includes `cause` — that lives on\n * `error.cause`.\n */\nfunction buildContext(shape: AnthropicErrorShape): Record<string, unknown> {\n  const context: Record<string, unknown> = {};\n\n  if (shape.status !== undefined) {\n    context.status = shape.status;\n  }\n\n  if (shape.type) {\n    context.type = shape.type;\n  }\n\n  if (shape.requestId) {\n    context.requestId = shape.requestId;\n  }\n\n  return context;\n}\n\n/** Anthropic exposes the request id as `requestID`; some proxies use `request_id`. */\nfunction readRequestId(raw: Record<string, unknown>): string | undefined {\n  if (typeof raw.requestID === \"string\") {\n    return raw.requestID;\n  }\n\n  if (typeof raw.request_id === \"string\") {\n    return raw.request_id;\n  }\n\n  return undefined;\n}\n\nfunction isHeaderBag(value: unknown): value is HeaderBag {\n  return typeof value === \"object\" && value !== null;\n}\n\n/** Read a header value from either a `Headers` instance or a plain record. */\nfunction readHeader(headers: HeaderBag, name: string): string | undefined {\n  if (typeof (headers as Headers).get === \"function\") {\n    return (headers as Headers).get(name) ?? undefined;\n  }\n\n  const record = headers as Record<string, string>;\n\n  return record[name] ?? record[name.toLowerCase()];\n}\n\n/**\n * Parse the `Retry-After` response header (seconds per HTTP spec) into\n * milliseconds so consumers can feed it straight to `setTimeout`.\n * Returns `undefined` when missing or unparseable.\n */\nfunction parseRetryAfter(headers: HeaderBag | undefined): number | undefined {\n  if (!headers) {\n    return undefined;\n  }\n\n  const raw = readHeader(headers, \"retry-after\") ?? readHeader(headers, \"Retry-After\");\n\n  if (!raw) {\n    return undefined;\n  }\n\n  const seconds = Number(raw);\n\n  if (!Number.isFinite(seconds) || seconds < 0) {\n    return undefined;\n  }\n\n  return Math.round(seconds * 1000);\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 type Anthropic from \"@anthropic-ai/sdk\";\nimport type { AnthropicModelConfig } from \"./config.type\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapStopReason, toAnthropicMessages, toAnthropicTools, wrapAnthropicError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.anthropic\";\n\n/**\n * Anthropic requires `max_tokens` on every request (unlike OpenAI,\n * where it is optional). When neither the per-call option nor the\n * model config supplies one, fall back to a generous default so a\n * caller who never thought about token caps still gets a complete\n * answer instead of a 400.\n */\nconst DEFAULT_MAX_TOKENS = 4096;\n\n/**\n * Anthropic-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the official `@anthropic-ai/sdk`\n * Messages API. Agents, workflows, and supervisors never talk to\n * Anthropic directly — they hold a `ModelContract`, and this class is\n * what makes that contract concrete for Claude models.\n *\n * **Responsibility.**\n * - Owns: a long-lived `Anthropic` client + frozen `ModelConfig`\n *   (name, temperature, maxTokens) used as defaults for every call.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Anthropic wire shapes (system hoisting, `tool_use` / `tool_result`\n *   blocks) on the way out, and translating Anthropic's content-block\n *   response (text, tool calls, stop reason, usage) back into the\n *   neutral shapes on the way in.\n * - Does NOT own: dispatching tools, deciding whether to loop, tracking\n *   conversation history, or retrying on failure — those are agent\n *   concerns. The model is a stateless (per-call) protocol adapter.\n *\n * Because it holds a live client and shared defaults, it is modeled as\n * a class (see §4.2 of code-style.md — \"long-lived state across\n * calls\").\n *\n * @example\n * import Anthropic from \"@anthropic-ai/sdk\";\n * const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });\n * const model = new AnthropicModel(client, { name: \"claude-sonnet-4-6\" });\n *\n * const myAgent = agent({\n *   model,\n *   systemPrompt: \"You are a helpful assistant.\",\n *   tools: [searchTool],\n * });\n *\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class AnthropicModel 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: Anthropic;\n  private readonly config: AnthropicModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: Anthropic,\n    config: AnthropicModelConfig,\n    provider: string = \"anthropic\",\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. Sends the full message list to the Messages\n   * endpoint, waits for the terminal response, and reshapes it into a\n   * vendor-neutral `ModelResponse`. Per-call `options` override the\n   * instance's `ModelConfig` defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting call to messages.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response: Anthropic.Message;\n\n    try {\n      response = await this.client.messages.create(\n        { ...this.buildParams(messages, options), stream: false },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = mapStopReason(response.stop_reason);\n    const usage = this.extractUsage(response.usage);\n    const toolCalls = this.extractToolCalls(response.content);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"call to messages.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: this.extractText(response.content),\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion. Yields neutral `ModelStreamChunk`s\n   * — `delta` for text tokens, `tool-call` once a `tool_use` block's\n   * arguments have fully accumulated, and a terminal `done` carrying the\n   * final finish reason + usage totals. Callers consume it with\n   * `for await`.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting streaming call to messages.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let stream: Awaited<ReturnType<Anthropic[\"messages\"][\"create\"]>>;\n\n    try {\n      stream = await this.client.messages.create(\n        { ...this.buildParams(messages, options), stream: true },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawStopReason: string | null = null;\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 stream as AsyncIterable<Anthropic.RawMessageStreamEvent>) {\n        if (event.type === \"message_start\") {\n          usage.input = event.message.usage.input_tokens ?? 0;\n\n          const cached = event.message.usage.cache_read_input_tokens;\n\n          if (cached !== null && cached !== undefined && cached > 0) {\n            usage.cachedTokens = cached;\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_start\") {\n          const block = event.content_block;\n\n          if (block.type === \"tool_use\") {\n            toolBlocks.set(event.index, { id: block.id, name: block.name, json: \"\" });\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_delta\") {\n          if (event.delta.type === \"text_delta\") {\n            yield { type: \"delta\", content: event.delta.text };\n          } else if (event.delta.type === \"input_json_delta\") {\n            const accumulator = toolBlocks.get(event.index);\n\n            if (accumulator) {\n              accumulator.json += event.delta.partial_json;\n            }\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_stop\") {\n          const accumulator = toolBlocks.get(event.index);\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.index);\n          }\n\n          continue;\n        }\n\n        if (event.type === \"message_delta\") {\n          rawStopReason = event.delta.stop_reason ?? rawStopReason;\n          usage.output = event.usage.output_tokens ?? usage.output;\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    usage.total = usage.input + usage.output;\n\n    const finishReason = mapStopReason(rawStopReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Streaming call to messages.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the Anthropic request body shared by `complete()` and\n   * `stream()` (each adds its own `stream` literal so the SDK's create\n   * overload resolves to the right return type). Hoists the system\n   * prompt out of `messages`, resolves `max_tokens` (required by\n   * Anthropic) with the documented default, and conditionally attaches\n   * temperature, tools, and native structured output.\n   */\n  private buildParams(\n    messages: Message[],\n    options: ModelCallOptions | undefined,\n  ): Omit<Anthropic.MessageCreateParamsNonStreaming, \"stream\"> {\n    const { system, messages: anthropicMessages } = toAnthropicMessages(messages);\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      model: this.name,\n      max_tokens: options?.maxTokens ?? this.config.maxTokens ?? DEFAULT_MAX_TOKENS,\n      messages: anthropicMessages,\n      ...(system ? { system } : {}),\n      ...(temperature !== undefined ? { temperature } : {}),\n      ...this.buildTools(options?.tools),\n      ...this.buildStructuredOutput(options?.responseSchema),\n    };\n  }\n\n  /**\n   * Spread-friendly tools fragment. Returns an empty object when no\n   * tools were supplied so the caller can unconditionally spread it.\n   *\n   * When `config.promptCaching` is on, marks the LAST tool with\n   * `cache_control: ephemeral` — Anthropic caches the whole prefix up\n   * to a breakpoint, and tools sit before `system` in that prefix, so\n   * one breakpoint on the final tool caches every tool definition.\n   * Reads bill at ~0.1x after the first write. The system prompt is\n   * deliberately left uncached: it carries per-turn placeholders, so a\n   * breakpoint there would pay the ~1.25x write surcharge every turn\n   * with no reads.\n   */\n  private buildTools(tools: ModelCallOptions[\"tools\"]): { tools?: Anthropic.Tool[] } {\n    const mapped = toAnthropicTools(tools);\n\n    if (!mapped) {\n      return {};\n    }\n\n    if (this.config.promptCaching && mapped.length > 0) {\n      const last = mapped.length - 1;\n      mapped[last] = { ...mapped[last], cache_control: { type: \"ephemeral\" } };\n    }\n\n    return { tools: mapped };\n  }\n\n  /**\n   * Translate the neutral `responseSchema` option into Anthropic's\n   * native `output_config.format` (JSON-schema structured outputs).\n   *\n   * Only emitted when the model declares the `structuredOutput`\n   * capability AND the schema is a proper root-object JSON Schema —\n   * Anthropic rejects non-object roots. When the capability is off\n   * (config override) or the schema is non-object, returns an empty\n   * object: the agent has already injected a soft schema hint into the\n   * system prompt as the fallback, and client-side `validate()` still\n   * enforces shape.\n   */\n  private buildStructuredOutput(responseSchema: Record<string, unknown> | undefined): {\n    output_config?: Anthropic.OutputConfig;\n  } {\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      output_config: {\n        format: { type: \"json_schema\", schema: responseSchema },\n      },\n    };\n  }\n\n  /**\n   * Concatenate every `text` content block into the single neutral\n   * `content` string. `tool_use` and other block types are ignored\n   * here — tool calls are surfaced separately via `extractToolCalls`.\n   */\n  private extractText(content: Anthropic.ContentBlock[]): string {\n    return content\n      .filter((block): block is Anthropic.TextBlock => block.type === \"text\")\n      .map((block) => block.text)\n      .join(\"\");\n  }\n\n  /**\n   * Reshape Anthropic's `tool_use` content blocks into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when the model\n   * requested no tools so callers can branch on presence.\n   */\n  private extractToolCalls(\n    content: Anthropic.ContentBlock[],\n  ): ModelToolCallRequest[] | undefined {\n    const toolUses = content.filter(\n      (block): block is Anthropic.ToolUseBlock => block.type === \"tool_use\",\n    );\n\n    if (toolUses.length === 0) {\n      return undefined;\n    }\n\n    return toolUses.map((block) => ({\n      id: block.id,\n      name: block.name,\n      input: (block.input ?? {}) as Record<string, unknown>,\n    }));\n  }\n\n  /**\n   * Normalize Anthropic's `usage` block into the neutral `Usage` shape.\n   * Anthropic reports `input_tokens` / `output_tokens` separately with\n   * no pre-summed total, so `total` is computed. Cache-read tokens are\n   * surfaced as `cachedTokens` only when non-zero.\n   */\n  private extractUsage(raw: Anthropic.Usage): Usage {\n    const input = raw.input_tokens ?? 0;\n    const output = raw.output_tokens ?? 0;\n    const cached = raw.cache_read_input_tokens;\n\n    return {\n      input,\n      output,\n      total: input + output,\n      ...(cached !== null && cached !== undefined && cached > 0 ? { cachedTokens: cached } : {}),\n    };\n  }\n\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy and\n   * emit the standard error log line before it propagates. Shared by\n   * every catch site so the log shape stays identical.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapAnthropicError(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 Anthropic from \"@anthropic-ai/sdk\";\nimport type { ModelContract, ModelPricing, SDKAdapterContract } from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type { AnthropicModelConfig, AnthropicSDKConfig } from \"./config.type\";\nimport { AnthropicModel } from \"./model\";\n\n/**\n * Anthropic-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for Claude models via the official\n * `@anthropic-ai/sdk`. A single `AnthropicSDK` instance holds one live\n * `Anthropic` client, shared by every `ModelContract` it produces via\n * `model()`. Users construct one SDK per account and reuse it across\n * all agents, workflows, and supervisors that target Anthropic.\n *\n * **Responsibility.**\n * - Owns: a long-lived `Anthropic` client (authentication, base URL)\n *   and its lifetime scope. Factory for `AnthropicModel` instances —\n *   each model call gets a reference to the same client.\n * - Does NOT own: anything per-call (tool execution, message history,\n *   streaming loop) — those live in `AnthropicModel` and the agent\n *   runtime. Does NOT implement `embedder()`: Anthropic ships no\n *   first-party embeddings API (the contract marks it optional).\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"): the `Anthropic` client is heavy to construct\n * and designed to be reused; keeping it on `this` makes that reuse\n * explicit and aligns with the `new Anthropic(...)` upstream\n * convention.\n *\n * @example\n * const anthropic = new AnthropicSDK({ apiKey: process.env.ANTHROPIC_API_KEY! });\n * const model = anthropic.model({ name: \"claude-sonnet-4-6\", temperature: 0.7 });\n * const tokens = await anthropic.count(\"Hello world\");\n *\n * @example\n * // Compose into an `ai.anthropic` namespace for ergonomic agent wiring\n * const ai = { agent, tool, systemPrompt, anthropic: new AnthropicSDK({ apiKey }) };\n * const myAgent = ai.agent({ model: ai.anthropic.model({ name: \"claude-haiku-4-5\" }) });\n */\nexport class AnthropicSDK implements SDKAdapterContract {\n  private readonly client: Anthropic;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: AnthropicSDKConfig) {\n    this.client = new Anthropic({\n      apiKey: config.apiKey,\n      baseURL: config.baseURL,\n    });\n    this.provider = config.provider ?? \"anthropic\";\n    this.pricing = config.pricing;\n  }\n\n  /**\n   * Build an `AnthropicModel` bound to this SDK's client. Each call\n   * returns a fresh model instance, but all instances share the\n   * underlying `Anthropic` client — connection pools, rate limits, and\n   * authentication stay unified across every model produced here. The\n   * SDK's `provider` label is forwarded so every model self-identifies\n   * as coming from the same upstream.\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: AnthropicModelConfig): ModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: AnthropicModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new AnthropicModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate for a given text. Uses the\n   * character-heuristic (`approximateTokenCount`) from the core package\n   * — good enough for budgeting and quota guards, not for billing.\n   * Anthropic does expose a `messages.countTokens` endpoint, but that\n   * is a network round-trip; `count()` is intentionally offline and\n   * synchronous-cost. The optional model id is reserved for future\n   * per-model tokenizer dispatch; currently ignored.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgBA,MAAM,0BAA0B;CAC9B;CACA;CACA;CACA;CACA;AACF;;;;;;;;;;;;;AAcA,SAAgB,sBAAsB,WAA4B;CAChE,MAAM,aAAa,UAAU,YAAY;CAEzC,OAAO,wBAAwB,MAAM,WAAW,WAAW,WAAW,MAAM,CAAC;AAC/E;;;;ACtCA,MAAM,gBAA8C;CAClD,UAAU;CACV,eAAe;CACf,YAAY;CACZ,UAAU;AACZ;;;;;;;;;;;;;;;;;AAkBA,SAAgB,cAAc,KAA8C;CAC1E,OAAO,cAAc,OAAO,OAAO;AACrC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACmBA,SAAgB,oBAAoB,UAAwC;CAC1E,MAAM,cAAwB,CAAC;CAC/B,MAAM,SAAmC,CAAC;CAE1C,KAAK,MAAM,WAAW,UAAU;EAC9B,IAAI,QAAQ,SAAS,UAAU;GAC7B,YAAY,KAAK,iBAAiB,QAAQ,OAAO,CAAC;GAElD;EACF;EAEA,IAAI,QAAQ,SAAS,QAAQ;GAC3B,OAAO,KAAK;IACV,MAAM;IACN,SAAS,CACP;KACE,MAAM;KACN,aAAa,QAAQ,cAAc;KACnC,SAAS,iBAAiB,QAAQ,OAAO;IAC3C,CACF;GACF,CAAC;GAED;EACF;EAEA,IAAI,QAAQ,SAAS,eAAe,QAAQ,aAAa,QAAQ,UAAU,SAAS,GAAG;GACrF,MAAM,SAAwC,CAAC;GAC/C,MAAM,OAAO,iBAAiB,QAAQ,OAAO;GAE7C,IAAI,MACF,OAAO,KAAK;IAAE,MAAM;IAAQ;GAAK,CAAC;GAGpC,KAAK,MAAM,YAAY,QAAQ,WAC7B,OAAO,KAAK;IACV,MAAM;IACN,IAAI,SAAS;IACb,MAAM,SAAS;IACf,OAAO,SAAS,SAAS,CAAC;GAC5B,CAAC;GAGH,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,uBAAuB;GACtD,CAAC;GAED;EACF;EAEA,OAAO,KAAK;GACV,MAAM,QAAQ,SAAS,cAAc,cAAc;GACnD,SAAS,iBAAiB,QAAQ,OAAO;EAC3C,CAAC;CACH;CAEA,OAAO;EACL,QAAQ,YAAY,SAAS,IAAI,YAAY,KAAK,MAAM,IAAI;EAC5D,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;;;;;;;;AASA,SAAS,wBAAwB,MAAgD;CAC/E,IAAI,KAAK,SAAS,QAChB,OAAO;EAAE,MAAM;EAAQ,MAAM,KAAK;CAAK;CAGzC,IAAI,SAAS,KAAK,QAChB,OAAO;EAAE,MAAM;EAAS,QAAQ;GAAE,MAAM;GAAO,KAAK,KAAK,OAAO;EAAI;CAAE;CAGxE,OAAO;EACL,MAAM;EACN,QAAQ;GACN,MAAM;GACN,YAAY,KAAK,OAAO;GACxB,MAAM,KAAK,OAAO;EACpB;CACF;AACF;;;;;;;;;;;;;;;AC7IA,SAAgB,iBACd,OAC8B;CAC9B,IAAI,CAAC,SAAS,MAAM,WAAW,GAC7B;CAGF,OAAO,MAAM,KAAK,UAAU;EAC1B,MAAM,KAAK;EACX,aAAa,KAAK;EAClB,cAAc,cAAc,KAAK,KAAK;CACxC,EAAE;AACJ;;;;;;;AAQA,SAAS,cAAc,OAA0E;CAC/F,MAAM,+CAA2B,KAAK;CAEtC,IAAI,UAAU,OAAO,SAAS,UAC5B,OAAO;CAGT,OAAO,EAAE,MAAM,SAAS;AAC1B;;;;;;;;;;;;;;;;;;;;;;;;;;;ACaA,SAAgB,mBAAmB,QAA0B;CAC3D,IAAI,kBAAkBA,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,QAAQ,KAAK,GACzB,OAAO,IAAIC,oCAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,SAAS,0BAA0B,MAAM,SAAS,oBAC1D,OAAO,IAAIC,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,WAAW,OAAO,MAAM,WAAW,KAC3C,OAAO,IAAIA,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,iBACjB,OAAO,IAAIC,kCAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,SAAS,sBAAsB,MAAM,WAAW,KACxD,OAAO,IAAIC,sCAAuB,SAAS;EACzC,OAAO;EACP;EACA,YAAY,gBAAgB,MAAM,OAAO;CAC3C,CAAC;CAGH,IAAI,MAAM,SAAS,2BAA2B,eAAe,MAAM,MAAM,GAAG;EAC1E,IAAI,sBAAsB,KAAK,OAAO,GACpC,OAAO,IAAIC,0CAA2B,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;EAG3E,OAAO,IAAIC,mCAAoB,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;CACpE;CAEA,OAAO,IAAIC,6BAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;AAOA,SAAS,QAAQ,QAAsC;CACrD,IAAI,kBAAkBC,4BACpB,OAAO;EACL,QAAQ,OAAO,OAAO,WAAW,WAAW,OAAO,SAAS;EAC5D,MAAM,OAAO;EACb,SAAS,OAAO;EAChB,SAAS,OAAO;EAChB,MAAM,OAAO;EACb,WAAW,OAAO,aAAa;CACjC;CAGF,IAAI,OAAO,WAAW,YAAY,WAAW,MAAM;EACjD,MAAM,MAAM;EAEZ,OAAO;GACL,QAAQ,OAAO,IAAI,WAAW,WAAW,IAAI,SAAS;GACtD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,SAAS,OAAO,IAAI,YAAY,WAAW,IAAI,UAAU;GACzD,SAAS,YAAY,IAAI,OAAO,IAAI,IAAI,UAAU;GAClD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,WAAW,cAAc,GAAG;EAC9B;CACF;CAEA,OAAO,CAAC;AACV;;;;;;;AAQA,SAAS,UAAU,QAAiB,OAAqC;CACvE,IAAI,kBAAkBC,6CACpB,OAAO;CAGT,IAAI,MAAM,SAAS,6BACjB,OAAO;CAGT,IAAI,OAAO,WAAW,YAAY,WAAW,MAAM;EACjD,MAAM,OAAQ,OAAmC;EAEjD,IAAI,SAAS,eAAe,SAAS,gBACnC,OAAO;CAEX;CAEA,OAAO;AACT;;AAGA,SAAS,eAAe,QAAqC;CAC3D,OAAO,OAAO,WAAW,YAAY,UAAU,OAAO,SAAS;AACjE;;;;;;;AAQA,SAAS,aAAa,OAAqD;CACzE,MAAM,UAAmC,CAAC;CAE1C,IAAI,MAAM,WAAW,QACnB,QAAQ,SAAS,MAAM;CAGzB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,IAAI,MAAM,WACR,QAAQ,YAAY,MAAM;CAG5B,OAAO;AACT;;AAGA,SAAS,cAAc,KAAkD;CACvE,IAAI,OAAO,IAAI,cAAc,UAC3B,OAAO,IAAI;CAGb,IAAI,OAAO,IAAI,eAAe,UAC5B,OAAO,IAAI;AAIf;AAEA,SAAS,YAAY,OAAoC;CACvD,OAAO,OAAO,UAAU,YAAY,UAAU;AAChD;;AAGA,SAAS,WAAW,SAAoB,MAAkC;CACxE,IAAI,OAAQ,QAAoB,QAAQ,YACtC,OAAQ,QAAoB,IAAI,IAAI,KAAK;CAG3C,MAAM,SAAS;CAEf,OAAO,OAAO,SAAS,OAAO,KAAK,YAAY;AACjD;;;;;;AAOA,SAAS,gBAAgB,SAAoD;CAC3E,IAAI,CAAC,SACH;CAGF,MAAM,MAAM,WAAW,SAAS,aAAa,KAAK,WAAW,SAAS,aAAa;CAEnF,IAAI,CAAC,KACH;CAGF,MAAM,UAAU,OAAO,GAAG;CAE1B,IAAI,CAAC,OAAO,SAAS,OAAO,KAAK,UAAU,GACzC;CAGF,OAAO,KAAK,MAAM,UAAU,GAAI;AAClC;;;;AC5NA,MAAM,aAAa;;;;;;;;AASnB,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwC3B,IAAa,iBAAb,MAAqD;CAUnD,AAAO,YACL,QACA,QACA,WAAmB,aACnB;gBANgCC;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,oCAAoC;GAC3E,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,SAAS,OACpC;IAAE,GAAG,KAAK,YAAY,UAAU,OAAO;IAAG,QAAQ;GAAM,GACxD,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,SAAS,WAAW;EACvD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAC9C,MAAM,YAAY,KAAK,iBAAiB,SAAS,OAAO;EAExD,KAAK,OAAO,MAAM,YAAY,YAAY,qCAAqC;GAC7E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,KAAK,YAAY,SAAS,OAAO;GAC1C;GACA;GACA;EACF;CACF;;;;;;;;CASA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,8CAA8C;GACrF,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,SAAS,MAAM,KAAK,OAAO,SAAS,OAClC;IAAE,GAAG,KAAK,YAAY,UAAU,OAAO;IAAG,QAAQ;GAAK,GACvD,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI,gBAA+B;EACnC,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,6BAAa,IAAI,IAAwD;EAE/E,IAAI;GACF,WAAW,MAAM,SAAS,QAA0D;IAClF,IAAI,MAAM,SAAS,iBAAiB;KAClC,MAAM,QAAQ,MAAM,QAAQ,MAAM,gBAAgB;KAElD,MAAM,SAAS,MAAM,QAAQ,MAAM;KAEnC,IAAI,WAAW,QAAQ,WAAW,UAAa,SAAS,GACtD,MAAM,eAAe;KAGvB;IACF;IAEA,IAAI,MAAM,SAAS,uBAAuB;KACxC,MAAM,QAAQ,MAAM;KAEpB,IAAI,MAAM,SAAS,YACjB,WAAW,IAAI,MAAM,OAAO;MAAE,IAAI,MAAM;MAAI,MAAM,MAAM;MAAM,MAAM;KAAG,CAAC;KAG1E;IACF;IAEA,IAAI,MAAM,SAAS,uBAAuB;KACxC,IAAI,MAAM,MAAM,SAAS,cACvB,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM,MAAM;KAAK;UAC5C,IAAI,MAAM,MAAM,SAAS,oBAAoB;MAClD,MAAM,cAAc,WAAW,IAAI,MAAM,KAAK;MAE9C,IAAI,aACF,YAAY,QAAQ,MAAM,MAAM;KAEpC;KAEA;IACF;IAEA,IAAI,MAAM,SAAS,sBAAsB;KACvC,MAAM,cAAc,WAAW,IAAI,MAAM,KAAK;KAE9C,IAAI,aAAa;MACf,MAAM;OACJ,MAAM;OACN,IAAI,YAAY;OAChB,MAAM,YAAY;OAClB,yCAA8C,YAAY,MAAM,CAAC,CAAC;MACpE;MAEA,WAAW,OAAO,MAAM,KAAK;KAC/B;KAEA;IACF;IAEA,IAAI,MAAM,SAAS,iBAAiB;KAClC,gBAAgB,MAAM,MAAM,eAAe;KAC3C,MAAM,SAAS,MAAM,MAAM,iBAAiB,MAAM;IACpD;GACF;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,QAAQ,MAAM,QAAQ,MAAM;EAElC,MAAM,eAAe,cAAc,aAAa;EAEhD,KAAK,OAAO,MAAM,YAAY,YAAY,+CAA+C;GACvF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;;;CAUA,AAAQ,YACN,UACA,SAC2D;EAC3D,MAAM,EAAE,QAAQ,UAAU,sBAAsB,oBAAoB,QAAQ;EAC5E,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,OAAO,KAAK;GACZ,YAAY,SAAS,aAAa,KAAK,OAAO,aAAa;GAC3D,UAAU;GACV,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;GAC3B,GAAI,gBAAgB,SAAY,EAAE,YAAY,IAAI,CAAC;GACnD,GAAG,KAAK,WAAW,SAAS,KAAK;GACjC,GAAG,KAAK,sBAAsB,SAAS,cAAc;EACvD;CACF;;;;;;;;;;;;;;CAeA,AAAQ,WAAW,OAAgE;EACjF,MAAM,SAAS,iBAAiB,KAAK;EAErC,IAAI,CAAC,QACH,OAAO,CAAC;EAGV,IAAI,KAAK,OAAO,iBAAiB,OAAO,SAAS,GAAG;GAClD,MAAM,OAAO,OAAO,SAAS;GAC7B,OAAO,QAAQ;IAAE,GAAG,OAAO;IAAO,eAAe,EAAE,MAAM,YAAY;GAAE;EACzE;EAEA,OAAO,EAAE,OAAO,OAAO;CACzB;;;;;;;;;;;;;CAcA,AAAQ,sBAAsB,gBAE5B;EACA,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO,EACL,eAAe,EACb,QAAQ;GAAE,MAAM;GAAe,QAAQ;EAAe,EACxD,EACF;CACF;;;;;;CAOA,AAAQ,YAAY,SAA2C;EAC7D,OAAO,QACJ,QAAQ,UAAwC,MAAM,SAAS,MAAM,CAAC,CACtE,KAAK,UAAU,MAAM,IAAI,CAAC,CAC1B,KAAK,EAAE;CACZ;;;;;;CAOA,AAAQ,iBACN,SACoC;EACpC,MAAM,WAAW,QAAQ,QACtB,UAA2C,MAAM,SAAS,UAC7D;EAEA,IAAI,SAAS,WAAW,GACtB;EAGF,OAAO,SAAS,KAAK,WAAW;GAC9B,IAAI,MAAM;GACV,MAAM,MAAM;GACZ,OAAQ,MAAM,SAAS,CAAC;EAC1B,EAAE;CACJ;;;;;;;CAQA,AAAQ,aAAa,KAA6B;EAChD,MAAM,QAAQ,IAAI,gBAAgB;EAClC,MAAM,SAAS,IAAI,iBAAiB;EACpC,MAAM,SAAS,IAAI;EAEnB,OAAO;GACL;GACA;GACA,OAAO,QAAQ;GACf,GAAI,WAAW,QAAQ,WAAW,UAAa,SAAS,IAAI,EAAE,cAAc,OAAO,IAAI,CAAC;EAC1F;CACF;;;;;;CAOA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,mBAAmB,MAAM;EAEzC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;GACtD,MAAM,QAAQ;GACd,SAAS,QAAQ;EACnB,CAAC;EAED,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACrWA,IAAa,eAAb,MAAwD;CAKtD,AAAO,YAAY,QAA4B;EAC7C,KAAK,SAAS,IAAIC,0BAAU;GAC1B,QAAQ,OAAO;GACf,SAAS,OAAO;EAClB,CAAC;EACD,KAAK,WAAW,OAAO,YAAY;EACnC,KAAK,UAAU,OAAO;CACxB;;;;;;;;;;;;;CAcA,AAAO,MAAM,QAA6C;EACxD,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,eAAe,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACtE;;;;;;;;;;CAWA,MAAa,MAAM,MAAc,QAAkC;EACjE,iDAA6B,IAAI;CACnC;AACF"}
+{"version":3,"file":"index.cjs","names":["AIError","ProviderTimeoutError","ProviderAuthError","QuotaExceededError","ProviderRateLimitError","ContextLengthExceededError","InvalidRequestError","ProviderError","APIError","APIConnectionTimeoutError","log","Anthropic"],"sources":["../../../../../../@warlock.js/ai-anthropic/src/known-vision-models.ts","../../../../../../@warlock.js/ai-anthropic/src/utils/map-stop-reason.ts","../../../../../../@warlock.js/ai-anthropic/src/utils/to-anthropic-messages.ts","../../../../../../@warlock.js/ai-anthropic/src/utils/to-anthropic-tools.ts","../../../../../../@warlock.js/ai-anthropic/src/utils/wrap-anthropic-error.ts","../../../../../../@warlock.js/ai-anthropic/src/model.ts","../../../../../../@warlock.js/ai-anthropic/src/sdk.ts"],"sourcesContent":["/**\n * Model-name prefixes for Claude families that accept image input\n * (vision) on the Messages API.\n *\n * Every Claude 3, Claude 3.5/3.7, and Claude 4 family model is\n * multimodal, so the list covers both the dotted legacy naming\n * (`claude-3-haiku-...`, `claude-3-5-sonnet-...`) and the current\n * `claude-<tier>-4-*` naming (`claude-opus-4-7`, `claude-sonnet-4-6`,\n * `claude-haiku-4-5`). Pre-3 families (`claude-2`, `claude-instant`)\n * are text-only and intentionally absent.\n *\n * Matched as a prefix so dated variants (`claude-opus-4-20250514`) are\n * covered without listing every release tag. Devs can always override\n * per-model via `anthropic.model({ name, vision: true | false })` —\n * explicit config wins over inference in either direction.\n */\nconst VISION_CAPABLE_PREFIXES = [\n  \"claude-3\",\n  \"claude-4\",\n  \"claude-opus-4\",\n  \"claude-sonnet-4\",\n  \"claude-haiku-4\",\n];\n\n/**\n * Infer whether a given Claude model name supports vision based on the\n * known-prefix list. Unknown models default to `false` so that passing\n * an image attachment to an unsupported model surfaces a clear,\n * agent-side capability error instead of an opaque Anthropic 400.\n *\n * @example\n * inferVisionCapability(\"claude-sonnet-4-6\");      // → true\n * inferVisionCapability(\"claude-3-5-sonnet-latest\"); // → true\n * inferVisionCapability(\"claude-2.1\");             // → false\n * inferVisionCapability(\"custom-proxy-llm\");       // → false\n */\nexport function inferVisionCapability(modelName: string): boolean {\n  const normalized = modelName.toLowerCase();\n\n  return VISION_CAPABLE_PREFIXES.some((prefix) => normalized.startsWith(prefix));\n}\n","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 Anthropic's `stop_reason` to the normalized `FinishReason` union.\n *\n * `end_turn` / `stop_sequence` are both natural stops. `max_tokens`\n * maps to `length`. `tool_use` maps to `tool_calls`. Everything else —\n * `refusal` (policy intervention), `pause_turn` (incomplete\n * long-running turn), `null`, or any value a future API version adds —\n * falls through to `\"error\"` so the agent treats the trip as a\n * non-clean terminal rather than silently accepting a partial result.\n *\n * @example\n * mapStopReason(\"end_turn\");  // \"stop\"\n * mapStopReason(\"tool_use\");  // \"tool_calls\"\n * mapStopReason(\"refusal\");   // \"error\"\n * mapStopReason(null);        // \"error\"\n */\nexport function mapStopReason(raw: string | null | undefined): FinishReason {\n  return stopReasonMap[raw ?? \"\"] ?? \"error\";\n}\n","import type { ContentPart, Message } from \"@warlock.js/ai\";\nimport type Anthropic from \"@anthropic-ai/sdk\";\n\n/**\n * Result of splitting a vendor-neutral `Message[]` for the Anthropic\n * Messages API: the system prompt is hoisted to a top-level `system`\n * string (Anthropic has no `\"system\"` role inside `messages`), and the\n * remaining turns are mapped to `MessageParam[]`.\n */\nexport type AnthropicMessages = {\n  system: string | undefined;\n  messages: Anthropic.MessageParam[];\n};\n\nconst ANTHROPIC_IMAGE_MEDIA_TYPES = [\"image/jpeg\", \"image/png\", \"image/gif\", \"image/webp\"] as const;\n\ntype AnthropicImageMediaType = (typeof ANTHROPIC_IMAGE_MEDIA_TYPES)[number];\n\n/**\n * Convert vendor-neutral `Message[]` into Anthropic's request shape.\n *\n * Anthropic differs from the OpenAI Chat protocol in three ways this\n * function absorbs:\n *\n * 1. **No `system` role.** System messages are concatenated (newline-\n *    separated) and returned separately as the top-level `system`\n *    parameter.\n * 2. **Tool results are `user` turns.** A neutral `tool` message becomes\n *    a `user` message whose content is a single `tool_result` block\n *    keyed by `tool_use_id`.\n * 3. **Tool calls are `tool_use` content blocks.** An assistant message\n *    carrying `toolCalls` becomes an `assistant` message whose content\n *    is an optional leading `text` block followed by one `tool_use`\n *    block per call.\n *\n * Consecutive same-role turns are left as-is — the Messages API merges\n * them server-side.\n *\n * @example\n * const { system, messages } = toAnthropicMessages([\n *   { role: \"system\", content: \"Be concise.\" },\n *   { role: \"user\", content: \"Hi\" },\n * ]);\n * // system === \"Be concise.\"\n * // messages === [{ role: \"user\", content: \"Hi\" }]\n */\nexport function toAnthropicMessages(messages: Message[]): AnthropicMessages {\n  const systemParts: string[] = [];\n  const mapped: Anthropic.MessageParam[] = [];\n\n  for (const message of messages) {\n    if (message.role === \"system\") {\n      systemParts.push(stringifyContent(message.content));\n\n      continue;\n    }\n\n    if (message.role === \"tool\") {\n      mapped.push({\n        role: \"user\",\n        content: [\n          {\n            type: \"tool_result\",\n            tool_use_id: message.toolCallId ?? \"\",\n            content: stringifyContent(message.content),\n          },\n        ],\n      });\n\n      continue;\n    }\n\n    if (message.role === \"assistant\" && message.toolCalls && message.toolCalls.length > 0) {\n      const blocks: Anthropic.ContentBlockParam[] = [];\n      const text = stringifyContent(message.content);\n\n      if (text) {\n        blocks.push({ type: \"text\", text });\n      }\n\n      for (const toolCall of message.toolCalls) {\n        blocks.push({\n          type: \"tool_use\",\n          id: toolCall.id,\n          name: toolCall.name,\n          input: toolCall.input ?? {},\n        });\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(toAnthropicContentBlock),\n      });\n\n      continue;\n    }\n\n    mapped.push({\n      role: message.role === \"assistant\" ? \"assistant\" : \"user\",\n      content: stringifyContent(message.content),\n    });\n  }\n\n  return {\n    system: systemParts.length > 0 ? systemParts.join(\"\\n\\n\") : 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 so the wire\n * format stays valid. Plain 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 single resolved `ContentPart` to an Anthropic content block.\n * Text passes straight through; images become an `image` block with a\n * `url` source (remote) or a `base64` source (inlined bytes). The\n * agent has already resolved every attachment before it reaches here,\n * so this never reads files or fetches URLs.\n */\nfunction toAnthropicContentBlock(part: ContentPart): Anthropic.ContentBlockParam {\n  if (part.type === \"text\") {\n    return { type: \"text\", text: part.text };\n  }\n\n  if (\"url\" in part.source) {\n    return { type: \"image\", source: { type: \"url\", url: part.source.url } };\n  }\n\n  return {\n    type: \"image\",\n    source: {\n      type: \"base64\",\n      media_type: part.source.mediaType as AnthropicImageMediaType,\n      data: part.source.base64,\n    },\n  };\n}\n","import { extractJsonSchema, type ToolConfig } from \"@warlock.js/ai\";\nimport type Anthropic from \"@anthropic-ai/sdk\";\n\n/**\n * Convert vendor-neutral `ToolConfig[]` into Anthropic's `tools` array.\n * Uses the shared `extractJsonSchema` helper; Anthropic requires the\n * input schema to be a JSON-Schema object, so a non-object extraction\n * is coerced into an empty-object schema rather than rejected — the\n * tool still registers and the model simply sees no parameters.\n *\n * @example\n * const tools = toAnthropicTools([weatherTool, calculatorTool]);\n * await client.messages.create({ model, max_tokens, messages, tools });\n */\nexport function toAnthropicTools(\n  tools: ToolConfig<unknown, unknown>[] | undefined,\n): Anthropic.Tool[] | undefined {\n  if (!tools || tools.length === 0) {\n    return undefined;\n  }\n\n  return tools.map((tool) => ({\n    name: tool.name,\n    description: tool.description,\n    input_schema: toInputSchema(tool.input),\n  }));\n}\n\n/**\n * Coerce the extracted JSON Schema into Anthropic's `Tool.InputSchema`\n * shape (root must be `{ type: \"object\" }`). Anything that isn't an\n * object schema degrades to a parameterless object so registration\n * never fails on a malformed extractor result.\n */\nfunction toInputSchema(input: ToolConfig<unknown, unknown>[\"input\"]): Anthropic.Tool.InputSchema {\n  const schema = extractJsonSchema(input);\n\n  if (schema && schema.type === \"object\") {\n    return schema as Anthropic.Tool.InputSchema;\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\";\nimport { APIConnectionTimeoutError, APIError } from \"@anthropic-ai/sdk\";\n\n/**\n * Raw-error fields the wrapper reads off an Anthropic SDK error.\n *\n * `APIError` exposes `status`, `type` (the `error.type` from the\n * response body, e.g. `\"rate_limit_error\"`), `message`, `headers`, and\n * `requestID`. We duck-type because wrapped retries, proxied errors,\n * and custom subclasses sometimes lose the `instanceof` relationship.\n */\ntype AnthropicErrorShape = {\n  status?: number;\n  type?: string | null;\n  message?: string;\n  headers?: HeaderBag | undefined;\n  name?: string;\n  requestId?: string;\n};\n\n/** Either a fetch `Headers` instance or a plain record (duck-typed tests). */\ntype HeaderBag = Headers | Record<string, string>;\n\n/**\n * Wrap any thrown value caught inside the Anthropic adapter into the\n * appropriate `@warlock.js/ai` `AIError` subclass.\n *\n * **Dispatch strategy.** Anthropic has no per-error machine `code`; the\n * stable identifier is `error.type` on the response body (surfaced as\n * `APIError.type`). Dispatch prefers `type`, falls back to `status`\n * when the body was stripped (common with proxies). Name-based\n * detection catches transport-layer timeouts that never produced an\n * HTTP response. The `invalid_request_error` branch additionally\n * sniffs the message for Anthropic's \"prompt is too long\" phrasing,\n * which is the only signal that a 400 was a context-length overflow.\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.messages.create(...);\n * } catch (thrown) {\n *   throw wrapAnthropicError(thrown);\n * }\n */\nexport function wrapAnthropicError(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(thrown, shape)) {\n    return new ProviderTimeoutError(message, { cause: thrown, context });\n  }\n\n  if (shape.type === \"authentication_error\" || shape.type === \"permission_error\") {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.status === 401 || shape.status === 403) {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.type === \"billing_error\") {\n    return new QuotaExceededError(message, { cause: thrown, context });\n  }\n\n  if (shape.type === \"rate_limit_error\" || shape.status === 429) {\n    return new ProviderRateLimitError(message, {\n      cause: thrown,\n      context,\n      retryAfter: parseRetryAfter(shape.headers),\n    });\n  }\n\n  if (shape.type === \"invalid_request_error\" || isClientStatus(shape.status)) {\n    if (/prompt is too long/i.test(message)) {\n      return new ContextLengthExceededError(message, { cause: thrown, context });\n    }\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 APIError`\n * — proxies and re-wrappers strip the prototype chain. Duck-typing on\n * the visible fields is resilient to both.\n */\nfunction toShape(thrown: unknown): AnthropicErrorShape {\n  if (thrown instanceof APIError) {\n    return {\n      status: typeof thrown.status === \"number\" ? thrown.status : undefined,\n      type: thrown.type,\n      message: thrown.message,\n      headers: thrown.headers,\n      name: thrown.name,\n      requestId: thrown.requestID ?? undefined,\n    };\n  }\n\n  if (typeof thrown === \"object\" && thrown !== null) {\n    const raw = thrown as Record<string, unknown>;\n\n    return {\n      status: typeof raw.status === \"number\" ? raw.status : undefined,\n      type: typeof raw.type === \"string\" ? raw.type : undefined,\n      message: typeof raw.message === \"string\" ? raw.message : undefined,\n      headers: isHeaderBag(raw.headers) ? raw.headers : undefined,\n      name: typeof raw.name === \"string\" ? raw.name : undefined,\n      requestId: readRequestId(raw),\n    };\n  }\n\n  return {};\n}\n\n/**\n * Decide whether the thrown value represents a timeout. The Anthropic\n * SDK throws `APIConnectionTimeoutError` for transport-level timeouts;\n * Node surfaces `ETIMEDOUT` / `ECONNABORTED` on the socket layer.\n * Either signal counts.\n */\nfunction isTimeout(thrown: unknown, shape: AnthropicErrorShape): boolean {\n  if (thrown instanceof APIConnectionTimeoutError) {\n    return true;\n  }\n\n  if (shape.name === \"APIConnectionTimeoutError\") {\n    return true;\n  }\n\n  if (typeof thrown === \"object\" && thrown !== null) {\n    const code = (thrown as Record<string, unknown>).code;\n\n    if (code === \"ETIMEDOUT\" || code === \"ECONNABORTED\") {\n      return true;\n    }\n  }\n\n  return false;\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` so consumers\n * have everything the provider surfaced without each subclass having\n * to redeclare them. Never includes `cause` — that lives on\n * `error.cause`.\n */\nfunction buildContext(shape: AnthropicErrorShape): Record<string, unknown> {\n  const context: Record<string, unknown> = {};\n\n  if (shape.status !== undefined) {\n    context.status = shape.status;\n  }\n\n  if (shape.type) {\n    context.type = shape.type;\n  }\n\n  if (shape.requestId) {\n    context.requestId = shape.requestId;\n  }\n\n  return context;\n}\n\n/** Anthropic exposes the request id as `requestID`; some proxies use `request_id`. */\nfunction readRequestId(raw: Record<string, unknown>): string | undefined {\n  if (typeof raw.requestID === \"string\") {\n    return raw.requestID;\n  }\n\n  if (typeof raw.request_id === \"string\") {\n    return raw.request_id;\n  }\n\n  return undefined;\n}\n\nfunction isHeaderBag(value: unknown): value is HeaderBag {\n  return typeof value === \"object\" && value !== null;\n}\n\n/** Read a header value from either a `Headers` instance or a plain record. */\nfunction readHeader(headers: HeaderBag, name: string): string | undefined {\n  if (typeof (headers as Headers).get === \"function\") {\n    return (headers as Headers).get(name) ?? undefined;\n  }\n\n  const record = headers as Record<string, string>;\n\n  return record[name] ?? record[name.toLowerCase()];\n}\n\n/**\n * Parse the `Retry-After` response header (seconds per HTTP spec) into\n * milliseconds so consumers can feed it straight to `setTimeout`.\n * Returns `undefined` when missing or unparseable.\n */\nfunction parseRetryAfter(headers: HeaderBag | undefined): number | undefined {\n  if (!headers) {\n    return undefined;\n  }\n\n  const raw = readHeader(headers, \"retry-after\") ?? readHeader(headers, \"Retry-After\");\n\n  if (!raw) {\n    return undefined;\n  }\n\n  const seconds = Number(raw);\n\n  if (!Number.isFinite(seconds) || seconds < 0) {\n    return undefined;\n  }\n\n  return Math.round(seconds * 1000);\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 ReasoningEffort,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type Anthropic from \"@anthropic-ai/sdk\";\nimport type { AnthropicModelConfig } from \"./config.type\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapStopReason, toAnthropicMessages, toAnthropicTools, wrapAnthropicError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.anthropic\";\n\n/**\n * Anthropic requires `max_tokens` on every request (unlike OpenAI,\n * where it is optional). When neither the per-call option nor the\n * model config supplies one, fall back to a generous default so a\n * caller who never thought about token caps still gets a complete\n * answer instead of a 400.\n */\nconst DEFAULT_MAX_TOKENS = 4096;\n\n/**\n * Anthropic rejects an extended-thinking budget below 1024 tokens, so\n * any resolved budget is floored at this minimum before it reaches the\n * wire.\n */\nconst MIN_THINKING_BUDGET = 1024;\n\n/**\n * Map the neutral `ReasoningEffort` level to an Anthropic\n * `thinking.budget_tokens` value. Anthropic budgets reasoning by token\n * count (unlike OpenAI's opaque `reasoning_effort` enum), so the three\n * neutral levels translate to representative token budgets when the\n * caller doesn't pass an explicit `reasoning.maxTokens`.\n */\nconst EFFORT_THINKING_BUDGET: Record<ReasoningEffort, number> = {\n  low: 1024,\n  medium: 4096,\n  high: 12000,\n};\n\n/**\n * Anthropic-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the official `@anthropic-ai/sdk`\n * Messages API. Agents, workflows, and supervisors never talk to\n * Anthropic directly — they hold a `ModelContract`, and this class is\n * what makes that contract concrete for Claude models.\n *\n * **Responsibility.**\n * - Owns: a long-lived `Anthropic` client + frozen `ModelConfig`\n *   (name, temperature, maxTokens) used as defaults for every call.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Anthropic wire shapes (system hoisting, `tool_use` / `tool_result`\n *   blocks) on the way out, and translating Anthropic's content-block\n *   response (text, tool calls, stop reason, usage) back into the\n *   neutral shapes on the way in.\n * - Does NOT own: dispatching tools, deciding whether to loop, tracking\n *   conversation history, or retrying on failure — those are agent\n *   concerns. The model is a stateless (per-call) protocol adapter.\n *\n * Because it holds a live client and shared defaults, it is modeled as\n * a class (see §4.2 of code-style.md — \"long-lived state across\n * calls\").\n *\n * @example\n * import Anthropic from \"@anthropic-ai/sdk\";\n * const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });\n * const model = new AnthropicModel(client, { name: \"claude-sonnet-4-6\" });\n *\n * const myAgent = agent({\n *   model,\n *   systemPrompt: \"You are a helpful assistant.\",\n *   tools: [searchTool],\n * });\n *\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class AnthropicModel 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: Anthropic;\n  private readonly config: AnthropicModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: Anthropic,\n    config: AnthropicModelConfig,\n    provider: string = \"anthropic\",\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      // Every modern Claude model accepts Anthropic extended thinking\n      // (`thinking: { type: \"enabled\", budget_tokens }`). Advertise the\n      // reasoning channel so the agent forwards `reasoning` options;\n      // explicit config override wins for proxied/legacy targets.\n      reasoning: config.reasoning ?? true,\n      // Anthropic prompt caching is caller-driven via `cache_control`\n      // breakpoints. The adapter both places those breakpoints (tools\n      // when `config.promptCaching`, system when\n      // `options.cacheControl.breakpoints`) and reports the read/write\n      // accounting (`cachedTokens` / `cacheWriteTokens`), so advertise\n      // the capability unconditionally.\n      promptCaching: true,\n      // The Messages API accepts PDF/document content blocks\n      // (`DocumentBlockParam`) on vision-capable Claude models.\n      pdf: true,\n      // Anthropic has no audio input content block, so `audio` stays\n      // absent (treated as false) and the agent rejects audio\n      // attachments upfront rather than dropping them at the wire layer.\n    };\n  }\n\n  /**\n   * Single-shot completion. Sends the full message list to the Messages\n   * endpoint, waits for the terminal response, and reshapes it into a\n   * vendor-neutral `ModelResponse`. Per-call `options` override the\n   * instance's `ModelConfig` defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting call to messages.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response: Anthropic.Message;\n\n    try {\n      response = await this.client.messages.create(\n        { ...this.buildParams(messages, options), stream: false },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = mapStopReason(response.stop_reason);\n    const usage = this.extractUsage(response.usage);\n    const toolCalls = this.extractToolCalls(response.content);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"call to messages.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: this.extractText(response.content),\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion. Yields neutral `ModelStreamChunk`s\n   * — `delta` for text tokens, `tool-call` once a `tool_use` block's\n   * arguments have fully accumulated, and a terminal `done` carrying the\n   * final finish reason + usage totals. Callers consume it with\n   * `for await`.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting streaming call to messages.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let stream: Awaited<ReturnType<Anthropic[\"messages\"][\"create\"]>>;\n\n    try {\n      stream = await this.client.messages.create(\n        { ...this.buildParams(messages, options), stream: true },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawStopReason: string | null = null;\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 stream as AsyncIterable<Anthropic.RawMessageStreamEvent>) {\n        if (event.type === \"message_start\") {\n          usage.input = event.message.usage.input_tokens ?? 0;\n\n          const cacheRead = event.message.usage.cache_read_input_tokens;\n\n          if (cacheRead !== null && cacheRead !== undefined && cacheRead > 0) {\n            usage.cachedTokens = cacheRead;\n          }\n\n          const cacheWrite = event.message.usage.cache_creation_input_tokens;\n\n          if (cacheWrite !== null && cacheWrite !== undefined && cacheWrite > 0) {\n            usage.cacheWriteTokens = cacheWrite;\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_start\") {\n          const block = event.content_block;\n\n          if (block.type === \"tool_use\") {\n            toolBlocks.set(event.index, { id: block.id, name: block.name, json: \"\" });\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_delta\") {\n          if (event.delta.type === \"text_delta\") {\n            yield { type: \"delta\", content: event.delta.text };\n          } else if (event.delta.type === \"input_json_delta\") {\n            const accumulator = toolBlocks.get(event.index);\n\n            if (accumulator) {\n              accumulator.json += event.delta.partial_json;\n            }\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_stop\") {\n          const accumulator = toolBlocks.get(event.index);\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.index);\n          }\n\n          continue;\n        }\n\n        if (event.type === \"message_delta\") {\n          rawStopReason = event.delta.stop_reason ?? rawStopReason;\n          usage.output = event.usage.output_tokens ?? usage.output;\n\n          // `message_delta.usage` carries the cumulative cache counts —\n          // prefer them over the `message_start` snapshot when present\n          // and non-zero so the terminal `done` reflects the final tally.\n          const cacheRead = event.usage.cache_read_input_tokens;\n\n          if (cacheRead !== null && cacheRead !== undefined && cacheRead > 0) {\n            usage.cachedTokens = cacheRead;\n          }\n\n          const cacheWrite = event.usage.cache_creation_input_tokens;\n\n          if (cacheWrite !== null && cacheWrite !== undefined && cacheWrite > 0) {\n            usage.cacheWriteTokens = cacheWrite;\n          }\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    usage.total = usage.input + usage.output;\n\n    const finishReason = mapStopReason(rawStopReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Streaming call to messages.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the Anthropic request body shared by `complete()` and\n   * `stream()` (each adds its own `stream` literal so the SDK's create\n   * overload resolves to the right return type). Hoists the system\n   * prompt out of `messages`, resolves `max_tokens` (required by\n   * Anthropic) with the documented default, and conditionally attaches\n   * temperature, tools, native structured output, extended thinking\n   * (`reasoning`), and a system-prompt cache breakpoint (`cacheControl`).\n   * Temperature is dropped when thinking is enabled, since Anthropic\n   * rejects the two together.\n   */\n  private buildParams(\n    messages: Message[],\n    options: ModelCallOptions | undefined,\n  ): Omit<Anthropic.MessageCreateParamsNonStreaming, \"stream\"> {\n    const { system, messages: anthropicMessages } = toAnthropicMessages(messages);\n    const thinking = this.buildThinking(options?.reasoning);\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      model: this.name,\n      max_tokens: options?.maxTokens ?? this.config.maxTokens ?? DEFAULT_MAX_TOKENS,\n      messages: anthropicMessages,\n      ...this.buildSystem(system, options?.cacheControl),\n      // Extended thinking pins sampling to the default temperature —\n      // Anthropic 400s when `thinking` is enabled alongside any explicit\n      // `temperature`. Drop temperature in that case rather than letting\n      // the request fail.\n      ...(temperature !== undefined && !thinking.thinking ? { temperature } : {}),\n      ...this.buildTools(options?.tools),\n      ...this.buildStructuredOutput(options?.responseSchema),\n      ...thinking,\n    };\n  }\n\n  /**\n   * Spread-friendly `system` fragment. Returns an empty object when no\n   * system prompt was hoisted out of the messages.\n   *\n   * When a per-call `cacheControl.breakpoints` hint is present (≥ 1),\n   * the system prompt is emitted as a single `TextBlockParam` carrying\n   * `cache_control: ephemeral` — the system prompt is the longest stable\n   * prefix on a turn, so one breakpoint there lets multi-turn agents\n   * read it back at the ~0.1x cache-read rate. Without the hint the\n   * system prompt stays a plain string (left uncached) so a one-shot\n   * call never pays the ~1.25x cache-write surcharge.\n   */\n  private buildSystem(\n    system: string | undefined,\n    cacheControl: ModelCallOptions[\"cacheControl\"],\n  ): { system?: string | Anthropic.TextBlockParam[] } {\n    if (!system) {\n      return {};\n    }\n\n    const breakpoints = cacheControl?.breakpoints ?? 0;\n\n    if (this.capabilities.promptCaching && breakpoints > 0) {\n      return {\n        system: [{ type: \"text\", text: system, cache_control: { type: \"ephemeral\" } }],\n      };\n    }\n\n    return { system };\n  }\n\n  /**\n   * Translate the neutral `reasoning` option into Anthropic's\n   * `thinking` request field. Emitted only when the model declares the\n   * `reasoning` capability AND a reasoning option is supplied; otherwise\n   * returns an empty object so the caller can unconditionally spread it.\n   *\n   * Budget resolution: an explicit `reasoning.maxTokens` wins; otherwise\n   * the neutral `effort` level maps to a tiered token budget. Anthropic\n   * requires `budget_tokens` ≥ 1024, so the budget is floored at that\n   * minimum.\n   */\n  private buildThinking(\n    reasoning: ModelCallOptions[\"reasoning\"],\n  ): { thinking?: Anthropic.ThinkingConfigParam } {\n    if (!this.capabilities.reasoning || !reasoning) {\n      return {};\n    }\n\n    if (reasoning.maxTokens === undefined && reasoning.effort === undefined) {\n      return {};\n    }\n\n    const budget = reasoning.maxTokens ?? EFFORT_THINKING_BUDGET[reasoning.effort ?? \"medium\"];\n\n    return {\n      thinking: {\n        type: \"enabled\",\n        budget_tokens: Math.max(MIN_THINKING_BUDGET, budget),\n      },\n    };\n  }\n\n  /**\n   * Spread-friendly tools fragment. Returns an empty object when no\n   * tools were supplied so the caller can unconditionally spread it.\n   *\n   * When `config.promptCaching` is on, marks the LAST tool with\n   * `cache_control: ephemeral` — Anthropic caches the whole prefix up\n   * to a breakpoint, and tools sit before `system` in that prefix, so\n   * one breakpoint on the final tool caches every tool definition.\n   * Reads bill at ~0.1x after the first write. The system prompt is\n   * deliberately left uncached: it carries per-turn placeholders, so a\n   * breakpoint there would pay the ~1.25x write surcharge every turn\n   * with no reads.\n   */\n  private buildTools(tools: ModelCallOptions[\"tools\"]): { tools?: Anthropic.Tool[] } {\n    const mapped = toAnthropicTools(tools);\n\n    if (!mapped) {\n      return {};\n    }\n\n    if (this.config.promptCaching && mapped.length > 0) {\n      const last = mapped.length - 1;\n      mapped[last] = { ...mapped[last], cache_control: { type: \"ephemeral\" } };\n    }\n\n    return { tools: mapped };\n  }\n\n  /**\n   * Translate the neutral `responseSchema` option into Anthropic's\n   * native `output_config.format` (JSON-schema structured outputs).\n   *\n   * Only emitted when the model declares the `structuredOutput`\n   * capability AND the schema is a proper root-object JSON Schema —\n   * Anthropic rejects non-object roots. When the capability is off\n   * (config override) or the schema is non-object, returns an empty\n   * object: the agent has already injected a soft schema hint into the\n   * system prompt as the fallback, and client-side `validate()` still\n   * enforces shape.\n   */\n  private buildStructuredOutput(responseSchema: Record<string, unknown> | undefined): {\n    output_config?: Anthropic.OutputConfig;\n  } {\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      output_config: {\n        format: { type: \"json_schema\", schema: responseSchema },\n      },\n    };\n  }\n\n  /**\n   * Concatenate every `text` content block into the single neutral\n   * `content` string. `tool_use` and other block types are ignored\n   * here — tool calls are surfaced separately via `extractToolCalls`.\n   */\n  private extractText(content: Anthropic.ContentBlock[]): string {\n    return content\n      .filter((block): block is Anthropic.TextBlock => block.type === \"text\")\n      .map((block) => block.text)\n      .join(\"\");\n  }\n\n  /**\n   * Reshape Anthropic's `tool_use` content blocks into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when the model\n   * requested no tools so callers can branch on presence.\n   */\n  private extractToolCalls(\n    content: Anthropic.ContentBlock[],\n  ): ModelToolCallRequest[] | undefined {\n    const toolUses = content.filter(\n      (block): block is Anthropic.ToolUseBlock => block.type === \"tool_use\",\n    );\n\n    if (toolUses.length === 0) {\n      return undefined;\n    }\n\n    return toolUses.map((block) => ({\n      id: block.id,\n      name: block.name,\n      input: (block.input ?? {}) as Record<string, unknown>,\n    }));\n  }\n\n  /**\n   * Normalize Anthropic's `usage` block into the neutral `Usage` shape.\n   * Anthropic reports `input_tokens` / `output_tokens` separately with\n   * no pre-summed total, so `total` is computed. Cache-read tokens are\n   * surfaced as `cachedTokens` and cache-write tokens as\n   * `cacheWriteTokens`, each only when non-zero.\n   *\n   * Note: Anthropic does not report a separate reasoning-token count —\n   * extended-thinking tokens are billed inside `output_tokens` — so\n   * `Usage.reasoningTokens` is intentionally left unset here. Populating\n   * it would double-count against `output`.\n   */\n  private extractUsage(raw: Anthropic.Usage): Usage {\n    const input = raw.input_tokens ?? 0;\n    const output = raw.output_tokens ?? 0;\n    const cacheRead = raw.cache_read_input_tokens;\n    const cacheWrite = raw.cache_creation_input_tokens;\n\n    return {\n      input,\n      output,\n      total: input + output,\n      ...(cacheRead !== null && cacheRead !== undefined && cacheRead > 0\n        ? { cachedTokens: cacheRead }\n        : {}),\n      ...(cacheWrite !== null && cacheWrite !== undefined && cacheWrite > 0\n        ? { cacheWriteTokens: cacheWrite }\n        : {}),\n    };\n  }\n\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy and\n   * emit the standard error log line before it propagates. Shared by\n   * every catch site so the log shape stays identical.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapAnthropicError(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 Anthropic from \"@anthropic-ai/sdk\";\nimport type { ModelContract, ModelPricing, SDKAdapterContract } from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type { AnthropicModelConfig, AnthropicSDKConfig } from \"./config.type\";\nimport { AnthropicModel } from \"./model\";\n\n/**\n * Anthropic-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for Claude models via the official\n * `@anthropic-ai/sdk`. A single `AnthropicSDK` instance holds one live\n * `Anthropic` client, shared by every `ModelContract` it produces via\n * `model()`. Users construct one SDK per account and reuse it across\n * all agents, workflows, and supervisors that target Anthropic.\n *\n * **Responsibility.**\n * - Owns: a long-lived `Anthropic` client (authentication, base URL)\n *   and its lifetime scope. Factory for `AnthropicModel` instances —\n *   each model call gets a reference to the same client.\n * - Does NOT own: anything per-call (tool execution, message history,\n *   streaming loop) — those live in `AnthropicModel` and the agent\n *   runtime. Does NOT implement `embedder()`: Anthropic ships no\n *   first-party embeddings API (the contract marks it optional).\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"): the `Anthropic` client is heavy to construct\n * and designed to be reused; keeping it on `this` makes that reuse\n * explicit and aligns with the `new Anthropic(...)` upstream\n * convention.\n *\n * @example\n * const anthropic = new AnthropicSDK({ apiKey: process.env.ANTHROPIC_API_KEY! });\n * const model = anthropic.model({ name: \"claude-sonnet-4-6\", temperature: 0.7 });\n * const tokens = await anthropic.count(\"Hello world\");\n *\n * @example\n * // Compose into an `ai.anthropic` namespace for ergonomic agent wiring\n * const ai = { agent, tool, systemPrompt, anthropic: new AnthropicSDK({ apiKey }) };\n * const myAgent = ai.agent({ model: ai.anthropic.model({ name: \"claude-haiku-4-5\" }) });\n */\nexport class AnthropicSDK implements SDKAdapterContract {\n  private readonly client: Anthropic;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: AnthropicSDKConfig) {\n    this.client = new Anthropic({\n      apiKey: config.apiKey,\n      baseURL: config.baseURL,\n    });\n    this.provider = config.provider ?? \"anthropic\";\n    this.pricing = config.pricing;\n  }\n\n  /**\n   * Build an `AnthropicModel` bound to this SDK's client. Each call\n   * returns a fresh model instance, but all instances share the\n   * underlying `Anthropic` client — connection pools, rate limits, and\n   * authentication stay unified across every model produced here. The\n   * SDK's `provider` label is forwarded so every model self-identifies\n   * as coming from the same upstream.\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: AnthropicModelConfig): ModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: AnthropicModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new AnthropicModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate for a given text. Uses the\n   * character-heuristic (`approximateTokenCount`) from the core package\n   * — good enough for budgeting and quota guards, not for billing.\n   * Anthropic does expose a `messages.countTokens` endpoint, but that\n   * is a network round-trip; `count()` is intentionally offline and\n   * synchronous-cost. The optional model id is reserved for future\n   * per-model tokenizer dispatch; currently ignored.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgBA,MAAM,0BAA0B;CAC9B;CACA;CACA;CACA;CACA;AACF;;;;;;;;;;;;;AAcA,SAAgB,sBAAsB,WAA4B;CAChE,MAAM,aAAa,UAAU,YAAY;CAEzC,OAAO,wBAAwB,MAAM,WAAW,WAAW,WAAW,MAAM,CAAC;AAC/E;;;;ACtCA,MAAM,gBAA8C;CAClD,UAAU;CACV,eAAe;CACf,YAAY;CACZ,UAAU;AACZ;;;;;;;;;;;;;;;;;AAkBA,SAAgB,cAAc,KAA8C;CAC1E,OAAO,cAAc,OAAO,OAAO;AACrC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACmBA,SAAgB,oBAAoB,UAAwC;CAC1E,MAAM,cAAwB,CAAC;CAC/B,MAAM,SAAmC,CAAC;CAE1C,KAAK,MAAM,WAAW,UAAU;EAC9B,IAAI,QAAQ,SAAS,UAAU;GAC7B,YAAY,KAAK,iBAAiB,QAAQ,OAAO,CAAC;GAElD;EACF;EAEA,IAAI,QAAQ,SAAS,QAAQ;GAC3B,OAAO,KAAK;IACV,MAAM;IACN,SAAS,CACP;KACE,MAAM;KACN,aAAa,QAAQ,cAAc;KACnC,SAAS,iBAAiB,QAAQ,OAAO;IAC3C,CACF;GACF,CAAC;GAED;EACF;EAEA,IAAI,QAAQ,SAAS,eAAe,QAAQ,aAAa,QAAQ,UAAU,SAAS,GAAG;GACrF,MAAM,SAAwC,CAAC;GAC/C,MAAM,OAAO,iBAAiB,QAAQ,OAAO;GAE7C,IAAI,MACF,OAAO,KAAK;IAAE,MAAM;IAAQ;GAAK,CAAC;GAGpC,KAAK,MAAM,YAAY,QAAQ,WAC7B,OAAO,KAAK;IACV,MAAM;IACN,IAAI,SAAS;IACb,MAAM,SAAS;IACf,OAAO,SAAS,SAAS,CAAC;GAC5B,CAAC;GAGH,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,uBAAuB;GACtD,CAAC;GAED;EACF;EAEA,OAAO,KAAK;GACV,MAAM,QAAQ,SAAS,cAAc,cAAc;GACnD,SAAS,iBAAiB,QAAQ,OAAO;EAC3C,CAAC;CACH;CAEA,OAAO;EACL,QAAQ,YAAY,SAAS,IAAI,YAAY,KAAK,MAAM,IAAI;EAC5D,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;;;;;;;;AASA,SAAS,wBAAwB,MAAgD;CAC/E,IAAI,KAAK,SAAS,QAChB,OAAO;EAAE,MAAM;EAAQ,MAAM,KAAK;CAAK;CAGzC,IAAI,SAAS,KAAK,QAChB,OAAO;EAAE,MAAM;EAAS,QAAQ;GAAE,MAAM;GAAO,KAAK,KAAK,OAAO;EAAI;CAAE;CAGxE,OAAO;EACL,MAAM;EACN,QAAQ;GACN,MAAM;GACN,YAAY,KAAK,OAAO;GACxB,MAAM,KAAK,OAAO;EACpB;CACF;AACF;;;;;;;;;;;;;;;AC7IA,SAAgB,iBACd,OAC8B;CAC9B,IAAI,CAAC,SAAS,MAAM,WAAW,GAC7B;CAGF,OAAO,MAAM,KAAK,UAAU;EAC1B,MAAM,KAAK;EACX,aAAa,KAAK;EAClB,cAAc,cAAc,KAAK,KAAK;CACxC,EAAE;AACJ;;;;;;;AAQA,SAAS,cAAc,OAA0E;CAC/F,MAAM,+CAA2B,KAAK;CAEtC,IAAI,UAAU,OAAO,SAAS,UAC5B,OAAO;CAGT,OAAO,EAAE,MAAM,SAAS;AAC1B;;;;;;;;;;;;;;;;;;;;;;;;;;;ACaA,SAAgB,mBAAmB,QAA0B;CAC3D,IAAI,kBAAkBA,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,QAAQ,KAAK,GACzB,OAAO,IAAIC,oCAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,SAAS,0BAA0B,MAAM,SAAS,oBAC1D,OAAO,IAAIC,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,WAAW,OAAO,MAAM,WAAW,KAC3C,OAAO,IAAIA,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,iBACjB,OAAO,IAAIC,kCAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,SAAS,sBAAsB,MAAM,WAAW,KACxD,OAAO,IAAIC,sCAAuB,SAAS;EACzC,OAAO;EACP;EACA,YAAY,gBAAgB,MAAM,OAAO;CAC3C,CAAC;CAGH,IAAI,MAAM,SAAS,2BAA2B,eAAe,MAAM,MAAM,GAAG;EAC1E,IAAI,sBAAsB,KAAK,OAAO,GACpC,OAAO,IAAIC,0CAA2B,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;EAG3E,OAAO,IAAIC,mCAAoB,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;CACpE;CAEA,OAAO,IAAIC,6BAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;AAOA,SAAS,QAAQ,QAAsC;CACrD,IAAI,kBAAkBC,4BACpB,OAAO;EACL,QAAQ,OAAO,OAAO,WAAW,WAAW,OAAO,SAAS;EAC5D,MAAM,OAAO;EACb,SAAS,OAAO;EAChB,SAAS,OAAO;EAChB,MAAM,OAAO;EACb,WAAW,OAAO,aAAa;CACjC;CAGF,IAAI,OAAO,WAAW,YAAY,WAAW,MAAM;EACjD,MAAM,MAAM;EAEZ,OAAO;GACL,QAAQ,OAAO,IAAI,WAAW,WAAW,IAAI,SAAS;GACtD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,SAAS,OAAO,IAAI,YAAY,WAAW,IAAI,UAAU;GACzD,SAAS,YAAY,IAAI,OAAO,IAAI,IAAI,UAAU;GAClD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,WAAW,cAAc,GAAG;EAC9B;CACF;CAEA,OAAO,CAAC;AACV;;;;;;;AAQA,SAAS,UAAU,QAAiB,OAAqC;CACvE,IAAI,kBAAkBC,6CACpB,OAAO;CAGT,IAAI,MAAM,SAAS,6BACjB,OAAO;CAGT,IAAI,OAAO,WAAW,YAAY,WAAW,MAAM;EACjD,MAAM,OAAQ,OAAmC;EAEjD,IAAI,SAAS,eAAe,SAAS,gBACnC,OAAO;CAEX;CAEA,OAAO;AACT;;AAGA,SAAS,eAAe,QAAqC;CAC3D,OAAO,OAAO,WAAW,YAAY,UAAU,OAAO,SAAS;AACjE;;;;;;;AAQA,SAAS,aAAa,OAAqD;CACzE,MAAM,UAAmC,CAAC;CAE1C,IAAI,MAAM,WAAW,QACnB,QAAQ,SAAS,MAAM;CAGzB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,IAAI,MAAM,WACR,QAAQ,YAAY,MAAM;CAG5B,OAAO;AACT;;AAGA,SAAS,cAAc,KAAkD;CACvE,IAAI,OAAO,IAAI,cAAc,UAC3B,OAAO,IAAI;CAGb,IAAI,OAAO,IAAI,eAAe,UAC5B,OAAO,IAAI;AAIf;AAEA,SAAS,YAAY,OAAoC;CACvD,OAAO,OAAO,UAAU,YAAY,UAAU;AAChD;;AAGA,SAAS,WAAW,SAAoB,MAAkC;CACxE,IAAI,OAAQ,QAAoB,QAAQ,YACtC,OAAQ,QAAoB,IAAI,IAAI,KAAK;CAG3C,MAAM,SAAS;CAEf,OAAO,OAAO,SAAS,OAAO,KAAK,YAAY;AACjD;;;;;;AAOA,SAAS,gBAAgB,SAAoD;CAC3E,IAAI,CAAC,SACH;CAGF,MAAM,MAAM,WAAW,SAAS,aAAa,KAAK,WAAW,SAAS,aAAa;CAEnF,IAAI,CAAC,KACH;CAGF,MAAM,UAAU,OAAO,GAAG;CAE1B,IAAI,CAAC,OAAO,SAAS,OAAO,KAAK,UAAU,GACzC;CAGF,OAAO,KAAK,MAAM,UAAU,GAAI;AAClC;;;;AC3NA,MAAM,aAAa;;;;;;;;AASnB,MAAM,qBAAqB;;;;;;AAO3B,MAAM,sBAAsB;;;;;;;;AAS5B,MAAM,yBAA0D;CAC9D,KAAK;CACL,QAAQ;CACR,MAAM;AACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwCA,IAAa,iBAAb,MAAqD;CAUnD,AAAO,YACL,QACA,QACA,WAAmB,aACnB;gBANgCC;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;GAK1D,WAAW,OAAO,aAAa;GAO/B,eAAe;GAGf,KAAK;EAIP;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,OAAO,MAAM,YAAY,WAAW,oCAAoC;GAC3E,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,SAAS,OACpC;IAAE,GAAG,KAAK,YAAY,UAAU,OAAO;IAAG,QAAQ;GAAM,GACxD,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,SAAS,WAAW;EACvD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAC9C,MAAM,YAAY,KAAK,iBAAiB,SAAS,OAAO;EAExD,KAAK,OAAO,MAAM,YAAY,YAAY,qCAAqC;GAC7E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,KAAK,YAAY,SAAS,OAAO;GAC1C;GACA;GACA;EACF;CACF;;;;;;;;CASA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,8CAA8C;GACrF,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,SAAS,MAAM,KAAK,OAAO,SAAS,OAClC;IAAE,GAAG,KAAK,YAAY,UAAU,OAAO;IAAG,QAAQ;GAAK,GACvD,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI,gBAA+B;EACnC,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,6BAAa,IAAI,IAAwD;EAE/E,IAAI;GACF,WAAW,MAAM,SAAS,QAA0D;IAClF,IAAI,MAAM,SAAS,iBAAiB;KAClC,MAAM,QAAQ,MAAM,QAAQ,MAAM,gBAAgB;KAElD,MAAM,YAAY,MAAM,QAAQ,MAAM;KAEtC,IAAI,cAAc,QAAQ,cAAc,UAAa,YAAY,GAC/D,MAAM,eAAe;KAGvB,MAAM,aAAa,MAAM,QAAQ,MAAM;KAEvC,IAAI,eAAe,QAAQ,eAAe,UAAa,aAAa,GAClE,MAAM,mBAAmB;KAG3B;IACF;IAEA,IAAI,MAAM,SAAS,uBAAuB;KACxC,MAAM,QAAQ,MAAM;KAEpB,IAAI,MAAM,SAAS,YACjB,WAAW,IAAI,MAAM,OAAO;MAAE,IAAI,MAAM;MAAI,MAAM,MAAM;MAAM,MAAM;KAAG,CAAC;KAG1E;IACF;IAEA,IAAI,MAAM,SAAS,uBAAuB;KACxC,IAAI,MAAM,MAAM,SAAS,cACvB,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM,MAAM;KAAK;UAC5C,IAAI,MAAM,MAAM,SAAS,oBAAoB;MAClD,MAAM,cAAc,WAAW,IAAI,MAAM,KAAK;MAE9C,IAAI,aACF,YAAY,QAAQ,MAAM,MAAM;KAEpC;KAEA;IACF;IAEA,IAAI,MAAM,SAAS,sBAAsB;KACvC,MAAM,cAAc,WAAW,IAAI,MAAM,KAAK;KAE9C,IAAI,aAAa;MACf,MAAM;OACJ,MAAM;OACN,IAAI,YAAY;OAChB,MAAM,YAAY;OAClB,yCAA8C,YAAY,MAAM,CAAC,CAAC;MACpE;MAEA,WAAW,OAAO,MAAM,KAAK;KAC/B;KAEA;IACF;IAEA,IAAI,MAAM,SAAS,iBAAiB;KAClC,gBAAgB,MAAM,MAAM,eAAe;KAC3C,MAAM,SAAS,MAAM,MAAM,iBAAiB,MAAM;KAKlD,MAAM,YAAY,MAAM,MAAM;KAE9B,IAAI,cAAc,QAAQ,cAAc,UAAa,YAAY,GAC/D,MAAM,eAAe;KAGvB,MAAM,aAAa,MAAM,MAAM;KAE/B,IAAI,eAAe,QAAQ,eAAe,UAAa,aAAa,GAClE,MAAM,mBAAmB;IAE7B;GACF;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,QAAQ,MAAM,QAAQ,MAAM;EAElC,MAAM,eAAe,cAAc,aAAa;EAEhD,KAAK,OAAO,MAAM,YAAY,YAAY,+CAA+C;GACvF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;;;;;;CAaA,AAAQ,YACN,UACA,SAC2D;EAC3D,MAAM,EAAE,QAAQ,UAAU,sBAAsB,oBAAoB,QAAQ;EAC5E,MAAM,WAAW,KAAK,cAAc,SAAS,SAAS;EACtD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,OAAO,KAAK;GACZ,YAAY,SAAS,aAAa,KAAK,OAAO,aAAa;GAC3D,UAAU;GACV,GAAG,KAAK,YAAY,QAAQ,SAAS,YAAY;GAKjD,GAAI,gBAAgB,UAAa,CAAC,SAAS,WAAW,EAAE,YAAY,IAAI,CAAC;GACzE,GAAG,KAAK,WAAW,SAAS,KAAK;GACjC,GAAG,KAAK,sBAAsB,SAAS,cAAc;GACrD,GAAG;EACL;CACF;;;;;;;;;;;;;CAcA,AAAQ,YACN,QACA,cACkD;EAClD,IAAI,CAAC,QACH,OAAO,CAAC;EAGV,MAAM,cAAc,cAAc,eAAe;EAEjD,IAAI,KAAK,aAAa,iBAAiB,cAAc,GACnD,OAAO,EACL,QAAQ,CAAC;GAAE,MAAM;GAAQ,MAAM;GAAQ,eAAe,EAAE,MAAM,YAAY;EAAE,CAAC,EAC/E;EAGF,OAAO,EAAE,OAAO;CAClB;;;;;;;;;;;;CAaA,AAAQ,cACN,WAC8C;EAC9C,IAAI,CAAC,KAAK,aAAa,aAAa,CAAC,WACnC,OAAO,CAAC;EAGV,IAAI,UAAU,cAAc,UAAa,UAAU,WAAW,QAC5D,OAAO,CAAC;EAGV,MAAM,SAAS,UAAU,aAAa,uBAAuB,UAAU,UAAU;EAEjF,OAAO,EACL,UAAU;GACR,MAAM;GACN,eAAe,KAAK,IAAI,qBAAqB,MAAM;EACrD,EACF;CACF;;;;;;;;;;;;;;CAeA,AAAQ,WAAW,OAAgE;EACjF,MAAM,SAAS,iBAAiB,KAAK;EAErC,IAAI,CAAC,QACH,OAAO,CAAC;EAGV,IAAI,KAAK,OAAO,iBAAiB,OAAO,SAAS,GAAG;GAClD,MAAM,OAAO,OAAO,SAAS;GAC7B,OAAO,QAAQ;IAAE,GAAG,OAAO;IAAO,eAAe,EAAE,MAAM,YAAY;GAAE;EACzE;EAEA,OAAO,EAAE,OAAO,OAAO;CACzB;;;;;;;;;;;;;CAcA,AAAQ,sBAAsB,gBAE5B;EACA,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO,EACL,eAAe,EACb,QAAQ;GAAE,MAAM;GAAe,QAAQ;EAAe,EACxD,EACF;CACF;;;;;;CAOA,AAAQ,YAAY,SAA2C;EAC7D,OAAO,QACJ,QAAQ,UAAwC,MAAM,SAAS,MAAM,CAAC,CACtE,KAAK,UAAU,MAAM,IAAI,CAAC,CAC1B,KAAK,EAAE;CACZ;;;;;;CAOA,AAAQ,iBACN,SACoC;EACpC,MAAM,WAAW,QAAQ,QACtB,UAA2C,MAAM,SAAS,UAC7D;EAEA,IAAI,SAAS,WAAW,GACtB;EAGF,OAAO,SAAS,KAAK,WAAW;GAC9B,IAAI,MAAM;GACV,MAAM,MAAM;GACZ,OAAQ,MAAM,SAAS,CAAC;EAC1B,EAAE;CACJ;;;;;;;;;;;;;CAcA,AAAQ,aAAa,KAA6B;EAChD,MAAM,QAAQ,IAAI,gBAAgB;EAClC,MAAM,SAAS,IAAI,iBAAiB;EACpC,MAAM,YAAY,IAAI;EACtB,MAAM,aAAa,IAAI;EAEvB,OAAO;GACL;GACA;GACA,OAAO,QAAQ;GACf,GAAI,cAAc,QAAQ,cAAc,UAAa,YAAY,IAC7D,EAAE,cAAc,UAAU,IAC1B,CAAC;GACL,GAAI,eAAe,QAAQ,eAAe,UAAa,aAAa,IAChE,EAAE,kBAAkB,WAAW,IAC/B,CAAC;EACP;CACF;;;;;;CAOA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,mBAAmB,MAAM;EAEzC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;GACtD,MAAM,QAAQ;GACd,SAAS,QAAQ;EACnB,CAAC;EAED,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACrfA,IAAa,eAAb,MAAwD;CAKtD,AAAO,YAAY,QAA4B;EAC7C,KAAK,SAAS,IAAIC,0BAAU;GAC1B,QAAQ,OAAO;GACf,SAAS,OAAO;EAClB,CAAC;EACD,KAAK,WAAW,OAAO,YAAY;EACnC,KAAK,UAAU,OAAO;CACxB;;;;;;;;;;;;;CAcA,AAAO,MAAM,QAA6C;EACxD,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,eAAe,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACtE;;;;;;;;;;CAWA,MAAa,MAAM,MAAc,QAAkC;EACjE,iDAA6B,IAAI;CACnC;AACF"}

package/esm/config.type.d.mts

@@ -79,8 +79,23 @@ type AnthropicModelConfig = ModelConfig & {
    * no surcharge. The system prompt is intentionally NOT cached here:
    * it carries per-turn placeholders, so a breakpoint there would pay
    * the write surcharge every turn with no reads.
+   *
+   * Independent of this flag, a per-call
+   * `ModelCallOptions.cacheControl.breakpoints` still places a cache
+   * breakpoint on the system prompt — that hint is opt-in per request.
    */
   promptCaching?: boolean;
+  /**
+   * Override the inferred `reasoning` capability. When omitted, the
+   * adapter advertises reasoning as supported (every modern Claude model
+   * accepts Anthropic extended thinking via `thinking`), so a per-call
+   * `ModelCallOptions.reasoning` is forwarded as
+   * `thinking: { type: "enabled", budget_tokens }`. Set to `false` for
+   * proxied deployments or older targets that reject the `thinking`
+   * param — the adapter then ignores reasoning options rather than
+   * sending an unsupported field.
+   */
+  reasoning?: boolean;
 };
 //#endregion
 export { AnthropicModelConfig, AnthropicSDKConfig };

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

@@ -1 +1 @@
-{"version":3,"file":"config.type.d.mts","names":[],"sources":["../../../../../../@warlock.js/ai-anthropic/src/config.type.ts"],"mappings":";;;;;;AAgCA;;;;;;;;;;;;;;AAOuC;AAWvC;;;;;;;;;AAgCe;;;KAlDH,kBAAA,GAAqB,aAAA;EAC/B,QAAA;;;;;;EAMA,OAAA,GAAU,MAAA,SAAe,YAAA;AAAA;;;;;;;;;KAWf,oBAAA,GAAuB,WAAW;;;;;;;;EAQ5C,MAAA;;;;;;;;;EASA,gBAAA;;;;;;;;;;;;;;;EAeA,aAAA;AAAA"}
+{"version":3,"file":"config.type.d.mts","names":[],"sources":["../../../../../../@warlock.js/ai-anthropic/src/config.type.ts"],"mappings":";;;;;;AAgCA;;;;;;;;;;;;;;AAOuC;AAWvC;;;;;;;;;;AA+CW;;KAjEC,kBAAA,GAAqB,aAAA;EAC/B,QAAA;;;;;;EAMA,OAAA,GAAU,MAAA,SAAe,YAAA;AAAA;;;;;;;;;KAWf,oBAAA,GAAuB,WAAW;;;;;;;;EAQ5C,MAAA;;;;;;;;;EASA,gBAAA;;;;;;;;;;;;;;;;;;;EAmBA,aAAA;;;;;;;;;;;EAWA,SAAA;AAAA"}

package/esm/model.mjs

@@ -18,6 +18,24 @@ const LOG_MODULE = "ai.anthropic";
 */
 const DEFAULT_MAX_TOKENS = 4096;
 /**
+* Anthropic rejects an extended-thinking budget below 1024 tokens, so
+* any resolved budget is floored at this minimum before it reaches the
+* wire.
+*/
+const MIN_THINKING_BUDGET = 1024;
+/**
+* Map the neutral `ReasoningEffort` level to an Anthropic
+* `thinking.budget_tokens` value. Anthropic budgets reasoning by token
+* count (unlike OpenAI's opaque `reasoning_effort` enum), so the three
+* neutral levels translate to representative token budgets when the
+* caller doesn't pass an explicit `reasoning.maxTokens`.
+*/
+const EFFORT_THINKING_BUDGET = {
+	low: 1024,
+	medium: 4096,
+	high: 12e3
+};
+/**
 * Anthropic-backed implementation of `ModelContract`.
 *
 * **Role.** The provider-facing bridge between the vendor-neutral
@@ -65,7 +83,10 @@ var AnthropicModel = 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 ?? true,
+			promptCaching: true,
+			pdf: true
 		};
 	}
 	/**
@@ -138,8 +159,10 @@ var AnthropicModel = class {
 			for await (const event of stream) {
 				if (event.type === "message_start") {
 					usage.input = event.message.usage.input_tokens ?? 0;
-					const cached = event.message.usage.cache_read_input_tokens;
-					if (cached !== null && cached !== void 0 && cached > 0) usage.cachedTokens = cached;
+					const cacheRead = event.message.usage.cache_read_input_tokens;
+					if (cacheRead !== null && cacheRead !== void 0 && cacheRead > 0) usage.cachedTokens = cacheRead;
+					const cacheWrite = event.message.usage.cache_creation_input_tokens;
+					if (cacheWrite !== null && cacheWrite !== void 0 && cacheWrite > 0) usage.cacheWriteTokens = cacheWrite;
 					continue;
 				}
 				if (event.type === "content_block_start") {
@@ -178,6 +201,10 @@ var AnthropicModel = class {
 				if (event.type === "message_delta") {
 					rawStopReason = event.delta.stop_reason ?? rawStopReason;
 					usage.output = event.usage.output_tokens ?? usage.output;
+					const cacheRead = event.usage.cache_read_input_tokens;
+					if (cacheRead !== null && cacheRead !== void 0 && cacheRead > 0) usage.cachedTokens = cacheRead;
+					const cacheWrite = event.usage.cache_creation_input_tokens;
+					if (cacheWrite !== null && cacheWrite !== void 0 && cacheWrite > 0) usage.cacheWriteTokens = cacheWrite;
 				}
 			}
 		} catch (thrown) {
@@ -201,22 +228,69 @@ var AnthropicModel = class {
 	* overload resolves to the right return type). Hoists the system
 	* prompt out of `messages`, resolves `max_tokens` (required by
 	* Anthropic) with the documented default, and conditionally attaches
-	* temperature, tools, and native structured output.
+	* temperature, tools, native structured output, extended thinking
+	* (`reasoning`), and a system-prompt cache breakpoint (`cacheControl`).
+	* Temperature is dropped when thinking is enabled, since Anthropic
+	* rejects the two together.
 	*/
 	buildParams(messages, options) {
 		const { system, messages: anthropicMessages } = toAnthropicMessages(messages);
+		const thinking = this.buildThinking(options?.reasoning);
 		const temperature = options?.temperature ?? this.config.temperature;
 		return {
 			model: this.name,
 			max_tokens: options?.maxTokens ?? this.config.maxTokens ?? DEFAULT_MAX_TOKENS,
 			messages: anthropicMessages,
-			...system ? { system } : {},
-			...temperature !== void 0 ? { temperature } : {},
+			...this.buildSystem(system, options?.cacheControl),
+			...temperature !== void 0 && !thinking.thinking ? { temperature } : {},
 			...this.buildTools(options?.tools),
-			...this.buildStructuredOutput(options?.responseSchema)
+			...this.buildStructuredOutput(options?.responseSchema),
+			...thinking
 		};
 	}
 	/**
+	* Spread-friendly `system` fragment. Returns an empty object when no
+	* system prompt was hoisted out of the messages.
+	*
+	* When a per-call `cacheControl.breakpoints` hint is present (≥ 1),
+	* the system prompt is emitted as a single `TextBlockParam` carrying
+	* `cache_control: ephemeral` — the system prompt is the longest stable
+	* prefix on a turn, so one breakpoint there lets multi-turn agents
+	* read it back at the ~0.1x cache-read rate. Without the hint the
+	* system prompt stays a plain string (left uncached) so a one-shot
+	* call never pays the ~1.25x cache-write surcharge.
+	*/
+	buildSystem(system, cacheControl) {
+		if (!system) return {};
+		const breakpoints = cacheControl?.breakpoints ?? 0;
+		if (this.capabilities.promptCaching && breakpoints > 0) return { system: [{
+			type: "text",
+			text: system,
+			cache_control: { type: "ephemeral" }
+		}] };
+		return { system };
+	}
+	/**
+	* Translate the neutral `reasoning` option into Anthropic's
+	* `thinking` request field. Emitted only when the model declares the
+	* `reasoning` capability AND a reasoning option is supplied; otherwise
+	* returns an empty object so the caller can unconditionally spread it.
+	*
+	* Budget resolution: an explicit `reasoning.maxTokens` wins; otherwise
+	* the neutral `effort` level maps to a tiered token budget. Anthropic
+	* requires `budget_tokens` ≥ 1024, so the budget is floored at that
+	* minimum.
+	*/
+	buildThinking(reasoning) {
+		if (!this.capabilities.reasoning || !reasoning) return {};
+		if (reasoning.maxTokens === void 0 && reasoning.effort === void 0) return {};
+		const budget = reasoning.maxTokens ?? EFFORT_THINKING_BUDGET[reasoning.effort ?? "medium"];
+		return { thinking: {
+			type: "enabled",
+			budget_tokens: Math.max(MIN_THINKING_BUDGET, budget)
+		} };
+	}
+	/**
 	* Spread-friendly tools fragment. Returns an empty object when no
 	* tools were supplied so the caller can unconditionally spread it.
 	*
@@ -287,17 +361,25 @@ var AnthropicModel = class {
 	* Normalize Anthropic's `usage` block into the neutral `Usage` shape.
 	* Anthropic reports `input_tokens` / `output_tokens` separately with
 	* no pre-summed total, so `total` is computed. Cache-read tokens are
-	* surfaced as `cachedTokens` only when non-zero.
+	* surfaced as `cachedTokens` and cache-write tokens as
+	* `cacheWriteTokens`, each only when non-zero.
+	*
+	* Note: Anthropic does not report a separate reasoning-token count —
+	* extended-thinking tokens are billed inside `output_tokens` — so
+	* `Usage.reasoningTokens` is intentionally left unset here. Populating
+	* it would double-count against `output`.
 	*/
 	extractUsage(raw) {
 		const input = raw.input_tokens ?? 0;
 		const output = raw.output_tokens ?? 0;
-		const cached = raw.cache_read_input_tokens;
+		const cacheRead = raw.cache_read_input_tokens;
+		const cacheWrite = raw.cache_creation_input_tokens;
 		return {
 			input,
 			output,
 			total: input + output,
-			...cached !== null && cached !== void 0 && cached > 0 ? { cachedTokens: cached } : {}
+			...cacheRead !== null && cacheRead !== void 0 && cacheRead > 0 ? { cachedTokens: cacheRead } : {},
+			...cacheWrite !== null && cacheWrite !== void 0 && cacheWrite > 0 ? { cacheWriteTokens: cacheWrite } : {}
 		};
 	}
 	/**

package/esm/model.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../@warlock.js/ai-anthropic/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 type Anthropic from \"@anthropic-ai/sdk\";\nimport type { AnthropicModelConfig } from \"./config.type\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapStopReason, toAnthropicMessages, toAnthropicTools, wrapAnthropicError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.anthropic\";\n\n/**\n * Anthropic requires `max_tokens` on every request (unlike OpenAI,\n * where it is optional). When neither the per-call option nor the\n * model config supplies one, fall back to a generous default so a\n * caller who never thought about token caps still gets a complete\n * answer instead of a 400.\n */\nconst DEFAULT_MAX_TOKENS = 4096;\n\n/**\n * Anthropic-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the official `@anthropic-ai/sdk`\n * Messages API. Agents, workflows, and supervisors never talk to\n * Anthropic directly — they hold a `ModelContract`, and this class is\n * what makes that contract concrete for Claude models.\n *\n * **Responsibility.**\n * - Owns: a long-lived `Anthropic` client + frozen `ModelConfig`\n *   (name, temperature, maxTokens) used as defaults for every call.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Anthropic wire shapes (system hoisting, `tool_use` / `tool_result`\n *   blocks) on the way out, and translating Anthropic's content-block\n *   response (text, tool calls, stop reason, usage) back into the\n *   neutral shapes on the way in.\n * - Does NOT own: dispatching tools, deciding whether to loop, tracking\n *   conversation history, or retrying on failure — those are agent\n *   concerns. The model is a stateless (per-call) protocol adapter.\n *\n * Because it holds a live client and shared defaults, it is modeled as\n * a class (see §4.2 of code-style.md — \"long-lived state across\n * calls\").\n *\n * @example\n * import Anthropic from \"@anthropic-ai/sdk\";\n * const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });\n * const model = new AnthropicModel(client, { name: \"claude-sonnet-4-6\" });\n *\n * const myAgent = agent({\n *   model,\n *   systemPrompt: \"You are a helpful assistant.\",\n *   tools: [searchTool],\n * });\n *\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class AnthropicModel 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: Anthropic;\n  private readonly config: AnthropicModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: Anthropic,\n    config: AnthropicModelConfig,\n    provider: string = \"anthropic\",\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. Sends the full message list to the Messages\n   * endpoint, waits for the terminal response, and reshapes it into a\n   * vendor-neutral `ModelResponse`. Per-call `options` override the\n   * instance's `ModelConfig` defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting call to messages.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response: Anthropic.Message;\n\n    try {\n      response = await this.client.messages.create(\n        { ...this.buildParams(messages, options), stream: false },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = mapStopReason(response.stop_reason);\n    const usage = this.extractUsage(response.usage);\n    const toolCalls = this.extractToolCalls(response.content);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"call to messages.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: this.extractText(response.content),\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion. Yields neutral `ModelStreamChunk`s\n   * — `delta` for text tokens, `tool-call` once a `tool_use` block's\n   * arguments have fully accumulated, and a terminal `done` carrying the\n   * final finish reason + usage totals. Callers consume it with\n   * `for await`.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting streaming call to messages.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let stream: Awaited<ReturnType<Anthropic[\"messages\"][\"create\"]>>;\n\n    try {\n      stream = await this.client.messages.create(\n        { ...this.buildParams(messages, options), stream: true },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawStopReason: string | null = null;\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 stream as AsyncIterable<Anthropic.RawMessageStreamEvent>) {\n        if (event.type === \"message_start\") {\n          usage.input = event.message.usage.input_tokens ?? 0;\n\n          const cached = event.message.usage.cache_read_input_tokens;\n\n          if (cached !== null && cached !== undefined && cached > 0) {\n            usage.cachedTokens = cached;\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_start\") {\n          const block = event.content_block;\n\n          if (block.type === \"tool_use\") {\n            toolBlocks.set(event.index, { id: block.id, name: block.name, json: \"\" });\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_delta\") {\n          if (event.delta.type === \"text_delta\") {\n            yield { type: \"delta\", content: event.delta.text };\n          } else if (event.delta.type === \"input_json_delta\") {\n            const accumulator = toolBlocks.get(event.index);\n\n            if (accumulator) {\n              accumulator.json += event.delta.partial_json;\n            }\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_stop\") {\n          const accumulator = toolBlocks.get(event.index);\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.index);\n          }\n\n          continue;\n        }\n\n        if (event.type === \"message_delta\") {\n          rawStopReason = event.delta.stop_reason ?? rawStopReason;\n          usage.output = event.usage.output_tokens ?? usage.output;\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    usage.total = usage.input + usage.output;\n\n    const finishReason = mapStopReason(rawStopReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Streaming call to messages.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the Anthropic request body shared by `complete()` and\n   * `stream()` (each adds its own `stream` literal so the SDK's create\n   * overload resolves to the right return type). Hoists the system\n   * prompt out of `messages`, resolves `max_tokens` (required by\n   * Anthropic) with the documented default, and conditionally attaches\n   * temperature, tools, and native structured output.\n   */\n  private buildParams(\n    messages: Message[],\n    options: ModelCallOptions | undefined,\n  ): Omit<Anthropic.MessageCreateParamsNonStreaming, \"stream\"> {\n    const { system, messages: anthropicMessages } = toAnthropicMessages(messages);\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      model: this.name,\n      max_tokens: options?.maxTokens ?? this.config.maxTokens ?? DEFAULT_MAX_TOKENS,\n      messages: anthropicMessages,\n      ...(system ? { system } : {}),\n      ...(temperature !== undefined ? { temperature } : {}),\n      ...this.buildTools(options?.tools),\n      ...this.buildStructuredOutput(options?.responseSchema),\n    };\n  }\n\n  /**\n   * Spread-friendly tools fragment. Returns an empty object when no\n   * tools were supplied so the caller can unconditionally spread it.\n   *\n   * When `config.promptCaching` is on, marks the LAST tool with\n   * `cache_control: ephemeral` — Anthropic caches the whole prefix up\n   * to a breakpoint, and tools sit before `system` in that prefix, so\n   * one breakpoint on the final tool caches every tool definition.\n   * Reads bill at ~0.1x after the first write. The system prompt is\n   * deliberately left uncached: it carries per-turn placeholders, so a\n   * breakpoint there would pay the ~1.25x write surcharge every turn\n   * with no reads.\n   */\n  private buildTools(tools: ModelCallOptions[\"tools\"]): { tools?: Anthropic.Tool[] } {\n    const mapped = toAnthropicTools(tools);\n\n    if (!mapped) {\n      return {};\n    }\n\n    if (this.config.promptCaching && mapped.length > 0) {\n      const last = mapped.length - 1;\n      mapped[last] = { ...mapped[last], cache_control: { type: \"ephemeral\" } };\n    }\n\n    return { tools: mapped };\n  }\n\n  /**\n   * Translate the neutral `responseSchema` option into Anthropic's\n   * native `output_config.format` (JSON-schema structured outputs).\n   *\n   * Only emitted when the model declares the `structuredOutput`\n   * capability AND the schema is a proper root-object JSON Schema —\n   * Anthropic rejects non-object roots. When the capability is off\n   * (config override) or the schema is non-object, returns an empty\n   * object: the agent has already injected a soft schema hint into the\n   * system prompt as the fallback, and client-side `validate()` still\n   * enforces shape.\n   */\n  private buildStructuredOutput(responseSchema: Record<string, unknown> | undefined): {\n    output_config?: Anthropic.OutputConfig;\n  } {\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      output_config: {\n        format: { type: \"json_schema\", schema: responseSchema },\n      },\n    };\n  }\n\n  /**\n   * Concatenate every `text` content block into the single neutral\n   * `content` string. `tool_use` and other block types are ignored\n   * here — tool calls are surfaced separately via `extractToolCalls`.\n   */\n  private extractText(content: Anthropic.ContentBlock[]): string {\n    return content\n      .filter((block): block is Anthropic.TextBlock => block.type === \"text\")\n      .map((block) => block.text)\n      .join(\"\");\n  }\n\n  /**\n   * Reshape Anthropic's `tool_use` content blocks into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when the model\n   * requested no tools so callers can branch on presence.\n   */\n  private extractToolCalls(\n    content: Anthropic.ContentBlock[],\n  ): ModelToolCallRequest[] | undefined {\n    const toolUses = content.filter(\n      (block): block is Anthropic.ToolUseBlock => block.type === \"tool_use\",\n    );\n\n    if (toolUses.length === 0) {\n      return undefined;\n    }\n\n    return toolUses.map((block) => ({\n      id: block.id,\n      name: block.name,\n      input: (block.input ?? {}) as Record<string, unknown>,\n    }));\n  }\n\n  /**\n   * Normalize Anthropic's `usage` block into the neutral `Usage` shape.\n   * Anthropic reports `input_tokens` / `output_tokens` separately with\n   * no pre-summed total, so `total` is computed. Cache-read tokens are\n   * surfaced as `cachedTokens` only when non-zero.\n   */\n  private extractUsage(raw: Anthropic.Usage): Usage {\n    const input = raw.input_tokens ?? 0;\n    const output = raw.output_tokens ?? 0;\n    const cached = raw.cache_read_input_tokens;\n\n    return {\n      input,\n      output,\n      total: input + output,\n      ...(cached !== null && cached !== undefined && cached > 0 ? { cachedTokens: cached } : {}),\n    };\n  }\n\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy and\n   * emit the standard error log line before it propagates. Shared by\n   * every catch site so the log shape stays identical.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapAnthropicError(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":";;;;;;;;;;AAkBA,MAAM,aAAa;;;;;;;;AASnB,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwC3B,IAAa,iBAAb,MAAqD;CAUnD,AAAO,YACL,QACA,QACA,WAAmB,aACnB;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,oCAAoC;GAC3E,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,SAAS,OACpC;IAAE,GAAG,KAAK,YAAY,UAAU,OAAO;IAAG,QAAQ;GAAM,GACxD,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,SAAS,WAAW;EACvD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAC9C,MAAM,YAAY,KAAK,iBAAiB,SAAS,OAAO;EAExD,KAAK,OAAO,MAAM,YAAY,YAAY,qCAAqC;GAC7E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,KAAK,YAAY,SAAS,OAAO;GAC1C;GACA;GACA;EACF;CACF;;;;;;;;CASA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,8CAA8C;GACrF,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,SAAS,MAAM,KAAK,OAAO,SAAS,OAClC;IAAE,GAAG,KAAK,YAAY,UAAU,OAAO;IAAG,QAAQ;GAAK,GACvD,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI,gBAA+B;EACnC,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,6BAAa,IAAI,IAAwD;EAE/E,IAAI;GACF,WAAW,MAAM,SAAS,QAA0D;IAClF,IAAI,MAAM,SAAS,iBAAiB;KAClC,MAAM,QAAQ,MAAM,QAAQ,MAAM,gBAAgB;KAElD,MAAM,SAAS,MAAM,QAAQ,MAAM;KAEnC,IAAI,WAAW,QAAQ,WAAW,UAAa,SAAS,GACtD,MAAM,eAAe;KAGvB;IACF;IAEA,IAAI,MAAM,SAAS,uBAAuB;KACxC,MAAM,QAAQ,MAAM;KAEpB,IAAI,MAAM,SAAS,YACjB,WAAW,IAAI,MAAM,OAAO;MAAE,IAAI,MAAM;MAAI,MAAM,MAAM;MAAM,MAAM;KAAG,CAAC;KAG1E;IACF;IAEA,IAAI,MAAM,SAAS,uBAAuB;KACxC,IAAI,MAAM,MAAM,SAAS,cACvB,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM,MAAM;KAAK;UAC5C,IAAI,MAAM,MAAM,SAAS,oBAAoB;MAClD,MAAM,cAAc,WAAW,IAAI,MAAM,KAAK;MAE9C,IAAI,aACF,YAAY,QAAQ,MAAM,MAAM;KAEpC;KAEA;IACF;IAEA,IAAI,MAAM,SAAS,sBAAsB;KACvC,MAAM,cAAc,WAAW,IAAI,MAAM,KAAK;KAE9C,IAAI,aAAa;MACf,MAAM;OACJ,MAAM;OACN,IAAI,YAAY;OAChB,MAAM,YAAY;OAClB,OAAO,cAAuC,YAAY,MAAM,CAAC,CAAC;MACpE;MAEA,WAAW,OAAO,MAAM,KAAK;KAC/B;KAEA;IACF;IAEA,IAAI,MAAM,SAAS,iBAAiB;KAClC,gBAAgB,MAAM,MAAM,eAAe;KAC3C,MAAM,SAAS,MAAM,MAAM,iBAAiB,MAAM;IACpD;GACF;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,QAAQ,MAAM,QAAQ,MAAM;EAElC,MAAM,eAAe,cAAc,aAAa;EAEhD,KAAK,OAAO,MAAM,YAAY,YAAY,+CAA+C;GACvF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;;;CAUA,AAAQ,YACN,UACA,SAC2D;EAC3D,MAAM,EAAE,QAAQ,UAAU,sBAAsB,oBAAoB,QAAQ;EAC5E,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,OAAO,KAAK;GACZ,YAAY,SAAS,aAAa,KAAK,OAAO,aAAa;GAC3D,UAAU;GACV,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;GAC3B,GAAI,gBAAgB,SAAY,EAAE,YAAY,IAAI,CAAC;GACnD,GAAG,KAAK,WAAW,SAAS,KAAK;GACjC,GAAG,KAAK,sBAAsB,SAAS,cAAc;EACvD;CACF;;;;;;;;;;;;;;CAeA,AAAQ,WAAW,OAAgE;EACjF,MAAM,SAAS,iBAAiB,KAAK;EAErC,IAAI,CAAC,QACH,OAAO,CAAC;EAGV,IAAI,KAAK,OAAO,iBAAiB,OAAO,SAAS,GAAG;GAClD,MAAM,OAAO,OAAO,SAAS;GAC7B,OAAO,QAAQ;IAAE,GAAG,OAAO;IAAO,eAAe,EAAE,MAAM,YAAY;GAAE;EACzE;EAEA,OAAO,EAAE,OAAO,OAAO;CACzB;;;;;;;;;;;;;CAcA,AAAQ,sBAAsB,gBAE5B;EACA,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO,EACL,eAAe,EACb,QAAQ;GAAE,MAAM;GAAe,QAAQ;EAAe,EACxD,EACF;CACF;;;;;;CAOA,AAAQ,YAAY,SAA2C;EAC7D,OAAO,QACJ,QAAQ,UAAwC,MAAM,SAAS,MAAM,CAAC,CACtE,KAAK,UAAU,MAAM,IAAI,CAAC,CAC1B,KAAK,EAAE;CACZ;;;;;;CAOA,AAAQ,iBACN,SACoC;EACpC,MAAM,WAAW,QAAQ,QACtB,UAA2C,MAAM,SAAS,UAC7D;EAEA,IAAI,SAAS,WAAW,GACtB;EAGF,OAAO,SAAS,KAAK,WAAW;GAC9B,IAAI,MAAM;GACV,MAAM,MAAM;GACZ,OAAQ,MAAM,SAAS,CAAC;EAC1B,EAAE;CACJ;;;;;;;CAQA,AAAQ,aAAa,KAA6B;EAChD,MAAM,QAAQ,IAAI,gBAAgB;EAClC,MAAM,SAAS,IAAI,iBAAiB;EACpC,MAAM,SAAS,IAAI;EAEnB,OAAO;GACL;GACA;GACA,OAAO,QAAQ;GACf,GAAI,WAAW,QAAQ,WAAW,UAAa,SAAS,IAAI,EAAE,cAAc,OAAO,IAAI,CAAC;EAC1F;CACF;;;;;;CAOA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,mBAAmB,MAAM;EAEzC,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-anthropic/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 ReasoningEffort,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type Anthropic from \"@anthropic-ai/sdk\";\nimport type { AnthropicModelConfig } from \"./config.type\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapStopReason, toAnthropicMessages, toAnthropicTools, wrapAnthropicError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.anthropic\";\n\n/**\n * Anthropic requires `max_tokens` on every request (unlike OpenAI,\n * where it is optional). When neither the per-call option nor the\n * model config supplies one, fall back to a generous default so a\n * caller who never thought about token caps still gets a complete\n * answer instead of a 400.\n */\nconst DEFAULT_MAX_TOKENS = 4096;\n\n/**\n * Anthropic rejects an extended-thinking budget below 1024 tokens, so\n * any resolved budget is floored at this minimum before it reaches the\n * wire.\n */\nconst MIN_THINKING_BUDGET = 1024;\n\n/**\n * Map the neutral `ReasoningEffort` level to an Anthropic\n * `thinking.budget_tokens` value. Anthropic budgets reasoning by token\n * count (unlike OpenAI's opaque `reasoning_effort` enum), so the three\n * neutral levels translate to representative token budgets when the\n * caller doesn't pass an explicit `reasoning.maxTokens`.\n */\nconst EFFORT_THINKING_BUDGET: Record<ReasoningEffort, number> = {\n  low: 1024,\n  medium: 4096,\n  high: 12000,\n};\n\n/**\n * Anthropic-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the official `@anthropic-ai/sdk`\n * Messages API. Agents, workflows, and supervisors never talk to\n * Anthropic directly — they hold a `ModelContract`, and this class is\n * what makes that contract concrete for Claude models.\n *\n * **Responsibility.**\n * - Owns: a long-lived `Anthropic` client + frozen `ModelConfig`\n *   (name, temperature, maxTokens) used as defaults for every call.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Anthropic wire shapes (system hoisting, `tool_use` / `tool_result`\n *   blocks) on the way out, and translating Anthropic's content-block\n *   response (text, tool calls, stop reason, usage) back into the\n *   neutral shapes on the way in.\n * - Does NOT own: dispatching tools, deciding whether to loop, tracking\n *   conversation history, or retrying on failure — those are agent\n *   concerns. The model is a stateless (per-call) protocol adapter.\n *\n * Because it holds a live client and shared defaults, it is modeled as\n * a class (see §4.2 of code-style.md — \"long-lived state across\n * calls\").\n *\n * @example\n * import Anthropic from \"@anthropic-ai/sdk\";\n * const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });\n * const model = new AnthropicModel(client, { name: \"claude-sonnet-4-6\" });\n *\n * const myAgent = agent({\n *   model,\n *   systemPrompt: \"You are a helpful assistant.\",\n *   tools: [searchTool],\n * });\n *\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class AnthropicModel 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: Anthropic;\n  private readonly config: AnthropicModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: Anthropic,\n    config: AnthropicModelConfig,\n    provider: string = \"anthropic\",\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      // Every modern Claude model accepts Anthropic extended thinking\n      // (`thinking: { type: \"enabled\", budget_tokens }`). Advertise the\n      // reasoning channel so the agent forwards `reasoning` options;\n      // explicit config override wins for proxied/legacy targets.\n      reasoning: config.reasoning ?? true,\n      // Anthropic prompt caching is caller-driven via `cache_control`\n      // breakpoints. The adapter both places those breakpoints (tools\n      // when `config.promptCaching`, system when\n      // `options.cacheControl.breakpoints`) and reports the read/write\n      // accounting (`cachedTokens` / `cacheWriteTokens`), so advertise\n      // the capability unconditionally.\n      promptCaching: true,\n      // The Messages API accepts PDF/document content blocks\n      // (`DocumentBlockParam`) on vision-capable Claude models.\n      pdf: true,\n      // Anthropic has no audio input content block, so `audio` stays\n      // absent (treated as false) and the agent rejects audio\n      // attachments upfront rather than dropping them at the wire layer.\n    };\n  }\n\n  /**\n   * Single-shot completion. Sends the full message list to the Messages\n   * endpoint, waits for the terminal response, and reshapes it into a\n   * vendor-neutral `ModelResponse`. Per-call `options` override the\n   * instance's `ModelConfig` defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting call to messages.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response: Anthropic.Message;\n\n    try {\n      response = await this.client.messages.create(\n        { ...this.buildParams(messages, options), stream: false },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = mapStopReason(response.stop_reason);\n    const usage = this.extractUsage(response.usage);\n    const toolCalls = this.extractToolCalls(response.content);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"call to messages.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: this.extractText(response.content),\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion. Yields neutral `ModelStreamChunk`s\n   * — `delta` for text tokens, `tool-call` once a `tool_use` block's\n   * arguments have fully accumulated, and a terminal `done` carrying the\n   * final finish reason + usage totals. Callers consume it with\n   * `for await`.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting streaming call to messages.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let stream: Awaited<ReturnType<Anthropic[\"messages\"][\"create\"]>>;\n\n    try {\n      stream = await this.client.messages.create(\n        { ...this.buildParams(messages, options), stream: true },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawStopReason: string | null = null;\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 stream as AsyncIterable<Anthropic.RawMessageStreamEvent>) {\n        if (event.type === \"message_start\") {\n          usage.input = event.message.usage.input_tokens ?? 0;\n\n          const cacheRead = event.message.usage.cache_read_input_tokens;\n\n          if (cacheRead !== null && cacheRead !== undefined && cacheRead > 0) {\n            usage.cachedTokens = cacheRead;\n          }\n\n          const cacheWrite = event.message.usage.cache_creation_input_tokens;\n\n          if (cacheWrite !== null && cacheWrite !== undefined && cacheWrite > 0) {\n            usage.cacheWriteTokens = cacheWrite;\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_start\") {\n          const block = event.content_block;\n\n          if (block.type === \"tool_use\") {\n            toolBlocks.set(event.index, { id: block.id, name: block.name, json: \"\" });\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_delta\") {\n          if (event.delta.type === \"text_delta\") {\n            yield { type: \"delta\", content: event.delta.text };\n          } else if (event.delta.type === \"input_json_delta\") {\n            const accumulator = toolBlocks.get(event.index);\n\n            if (accumulator) {\n              accumulator.json += event.delta.partial_json;\n            }\n          }\n\n          continue;\n        }\n\n        if (event.type === \"content_block_stop\") {\n          const accumulator = toolBlocks.get(event.index);\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.index);\n          }\n\n          continue;\n        }\n\n        if (event.type === \"message_delta\") {\n          rawStopReason = event.delta.stop_reason ?? rawStopReason;\n          usage.output = event.usage.output_tokens ?? usage.output;\n\n          // `message_delta.usage` carries the cumulative cache counts —\n          // prefer them over the `message_start` snapshot when present\n          // and non-zero so the terminal `done` reflects the final tally.\n          const cacheRead = event.usage.cache_read_input_tokens;\n\n          if (cacheRead !== null && cacheRead !== undefined && cacheRead > 0) {\n            usage.cachedTokens = cacheRead;\n          }\n\n          const cacheWrite = event.usage.cache_creation_input_tokens;\n\n          if (cacheWrite !== null && cacheWrite !== undefined && cacheWrite > 0) {\n            usage.cacheWriteTokens = cacheWrite;\n          }\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    usage.total = usage.input + usage.output;\n\n    const finishReason = mapStopReason(rawStopReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Streaming call to messages.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the Anthropic request body shared by `complete()` and\n   * `stream()` (each adds its own `stream` literal so the SDK's create\n   * overload resolves to the right return type). Hoists the system\n   * prompt out of `messages`, resolves `max_tokens` (required by\n   * Anthropic) with the documented default, and conditionally attaches\n   * temperature, tools, native structured output, extended thinking\n   * (`reasoning`), and a system-prompt cache breakpoint (`cacheControl`).\n   * Temperature is dropped when thinking is enabled, since Anthropic\n   * rejects the two together.\n   */\n  private buildParams(\n    messages: Message[],\n    options: ModelCallOptions | undefined,\n  ): Omit<Anthropic.MessageCreateParamsNonStreaming, \"stream\"> {\n    const { system, messages: anthropicMessages } = toAnthropicMessages(messages);\n    const thinking = this.buildThinking(options?.reasoning);\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      model: this.name,\n      max_tokens: options?.maxTokens ?? this.config.maxTokens ?? DEFAULT_MAX_TOKENS,\n      messages: anthropicMessages,\n      ...this.buildSystem(system, options?.cacheControl),\n      // Extended thinking pins sampling to the default temperature —\n      // Anthropic 400s when `thinking` is enabled alongside any explicit\n      // `temperature`. Drop temperature in that case rather than letting\n      // the request fail.\n      ...(temperature !== undefined && !thinking.thinking ? { temperature } : {}),\n      ...this.buildTools(options?.tools),\n      ...this.buildStructuredOutput(options?.responseSchema),\n      ...thinking,\n    };\n  }\n\n  /**\n   * Spread-friendly `system` fragment. Returns an empty object when no\n   * system prompt was hoisted out of the messages.\n   *\n   * When a per-call `cacheControl.breakpoints` hint is present (≥ 1),\n   * the system prompt is emitted as a single `TextBlockParam` carrying\n   * `cache_control: ephemeral` — the system prompt is the longest stable\n   * prefix on a turn, so one breakpoint there lets multi-turn agents\n   * read it back at the ~0.1x cache-read rate. Without the hint the\n   * system prompt stays a plain string (left uncached) so a one-shot\n   * call never pays the ~1.25x cache-write surcharge.\n   */\n  private buildSystem(\n    system: string | undefined,\n    cacheControl: ModelCallOptions[\"cacheControl\"],\n  ): { system?: string | Anthropic.TextBlockParam[] } {\n    if (!system) {\n      return {};\n    }\n\n    const breakpoints = cacheControl?.breakpoints ?? 0;\n\n    if (this.capabilities.promptCaching && breakpoints > 0) {\n      return {\n        system: [{ type: \"text\", text: system, cache_control: { type: \"ephemeral\" } }],\n      };\n    }\n\n    return { system };\n  }\n\n  /**\n   * Translate the neutral `reasoning` option into Anthropic's\n   * `thinking` request field. Emitted only when the model declares the\n   * `reasoning` capability AND a reasoning option is supplied; otherwise\n   * returns an empty object so the caller can unconditionally spread it.\n   *\n   * Budget resolution: an explicit `reasoning.maxTokens` wins; otherwise\n   * the neutral `effort` level maps to a tiered token budget. Anthropic\n   * requires `budget_tokens` ≥ 1024, so the budget is floored at that\n   * minimum.\n   */\n  private buildThinking(\n    reasoning: ModelCallOptions[\"reasoning\"],\n  ): { thinking?: Anthropic.ThinkingConfigParam } {\n    if (!this.capabilities.reasoning || !reasoning) {\n      return {};\n    }\n\n    if (reasoning.maxTokens === undefined && reasoning.effort === undefined) {\n      return {};\n    }\n\n    const budget = reasoning.maxTokens ?? EFFORT_THINKING_BUDGET[reasoning.effort ?? \"medium\"];\n\n    return {\n      thinking: {\n        type: \"enabled\",\n        budget_tokens: Math.max(MIN_THINKING_BUDGET, budget),\n      },\n    };\n  }\n\n  /**\n   * Spread-friendly tools fragment. Returns an empty object when no\n   * tools were supplied so the caller can unconditionally spread it.\n   *\n   * When `config.promptCaching` is on, marks the LAST tool with\n   * `cache_control: ephemeral` — Anthropic caches the whole prefix up\n   * to a breakpoint, and tools sit before `system` in that prefix, so\n   * one breakpoint on the final tool caches every tool definition.\n   * Reads bill at ~0.1x after the first write. The system prompt is\n   * deliberately left uncached: it carries per-turn placeholders, so a\n   * breakpoint there would pay the ~1.25x write surcharge every turn\n   * with no reads.\n   */\n  private buildTools(tools: ModelCallOptions[\"tools\"]): { tools?: Anthropic.Tool[] } {\n    const mapped = toAnthropicTools(tools);\n\n    if (!mapped) {\n      return {};\n    }\n\n    if (this.config.promptCaching && mapped.length > 0) {\n      const last = mapped.length - 1;\n      mapped[last] = { ...mapped[last], cache_control: { type: \"ephemeral\" } };\n    }\n\n    return { tools: mapped };\n  }\n\n  /**\n   * Translate the neutral `responseSchema` option into Anthropic's\n   * native `output_config.format` (JSON-schema structured outputs).\n   *\n   * Only emitted when the model declares the `structuredOutput`\n   * capability AND the schema is a proper root-object JSON Schema —\n   * Anthropic rejects non-object roots. When the capability is off\n   * (config override) or the schema is non-object, returns an empty\n   * object: the agent has already injected a soft schema hint into the\n   * system prompt as the fallback, and client-side `validate()` still\n   * enforces shape.\n   */\n  private buildStructuredOutput(responseSchema: Record<string, unknown> | undefined): {\n    output_config?: Anthropic.OutputConfig;\n  } {\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      output_config: {\n        format: { type: \"json_schema\", schema: responseSchema },\n      },\n    };\n  }\n\n  /**\n   * Concatenate every `text` content block into the single neutral\n   * `content` string. `tool_use` and other block types are ignored\n   * here — tool calls are surfaced separately via `extractToolCalls`.\n   */\n  private extractText(content: Anthropic.ContentBlock[]): string {\n    return content\n      .filter((block): block is Anthropic.TextBlock => block.type === \"text\")\n      .map((block) => block.text)\n      .join(\"\");\n  }\n\n  /**\n   * Reshape Anthropic's `tool_use` content blocks into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when the model\n   * requested no tools so callers can branch on presence.\n   */\n  private extractToolCalls(\n    content: Anthropic.ContentBlock[],\n  ): ModelToolCallRequest[] | undefined {\n    const toolUses = content.filter(\n      (block): block is Anthropic.ToolUseBlock => block.type === \"tool_use\",\n    );\n\n    if (toolUses.length === 0) {\n      return undefined;\n    }\n\n    return toolUses.map((block) => ({\n      id: block.id,\n      name: block.name,\n      input: (block.input ?? {}) as Record<string, unknown>,\n    }));\n  }\n\n  /**\n   * Normalize Anthropic's `usage` block into the neutral `Usage` shape.\n   * Anthropic reports `input_tokens` / `output_tokens` separately with\n   * no pre-summed total, so `total` is computed. Cache-read tokens are\n   * surfaced as `cachedTokens` and cache-write tokens as\n   * `cacheWriteTokens`, each only when non-zero.\n   *\n   * Note: Anthropic does not report a separate reasoning-token count —\n   * extended-thinking tokens are billed inside `output_tokens` — so\n   * `Usage.reasoningTokens` is intentionally left unset here. Populating\n   * it would double-count against `output`.\n   */\n  private extractUsage(raw: Anthropic.Usage): Usage {\n    const input = raw.input_tokens ?? 0;\n    const output = raw.output_tokens ?? 0;\n    const cacheRead = raw.cache_read_input_tokens;\n    const cacheWrite = raw.cache_creation_input_tokens;\n\n    return {\n      input,\n      output,\n      total: input + output,\n      ...(cacheRead !== null && cacheRead !== undefined && cacheRead > 0\n        ? { cachedTokens: cacheRead }\n        : {}),\n      ...(cacheWrite !== null && cacheWrite !== undefined && cacheWrite > 0\n        ? { cacheWriteTokens: cacheWrite }\n        : {}),\n    };\n  }\n\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy and\n   * emit the standard error log line before it propagates. Shared by\n   * every catch site so the log shape stays identical.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapAnthropicError(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":";;;;;;;;;;AAmBA,MAAM,aAAa;;;;;;;;AASnB,MAAM,qBAAqB;;;;;;AAO3B,MAAM,sBAAsB;;;;;;;;AAS5B,MAAM,yBAA0D;CAC9D,KAAK;CACL,QAAQ;CACR,MAAM;AACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwCA,IAAa,iBAAb,MAAqD;CAUnD,AAAO,YACL,QACA,QACA,WAAmB,aACnB;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;GAK1D,WAAW,OAAO,aAAa;GAO/B,eAAe;GAGf,KAAK;EAIP;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,OAAO,MAAM,YAAY,WAAW,oCAAoC;GAC3E,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,SAAS,OACpC;IAAE,GAAG,KAAK,YAAY,UAAU,OAAO;IAAG,QAAQ;GAAM,GACxD,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,SAAS,WAAW;EACvD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAC9C,MAAM,YAAY,KAAK,iBAAiB,SAAS,OAAO;EAExD,KAAK,OAAO,MAAM,YAAY,YAAY,qCAAqC;GAC7E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,KAAK,YAAY,SAAS,OAAO;GAC1C;GACA;GACA;EACF;CACF;;;;;;;;CASA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,8CAA8C;GACrF,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,SAAS,MAAM,KAAK,OAAO,SAAS,OAClC;IAAE,GAAG,KAAK,YAAY,UAAU,OAAO;IAAG,QAAQ;GAAK,GACvD,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI,gBAA+B;EACnC,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,6BAAa,IAAI,IAAwD;EAE/E,IAAI;GACF,WAAW,MAAM,SAAS,QAA0D;IAClF,IAAI,MAAM,SAAS,iBAAiB;KAClC,MAAM,QAAQ,MAAM,QAAQ,MAAM,gBAAgB;KAElD,MAAM,YAAY,MAAM,QAAQ,MAAM;KAEtC,IAAI,cAAc,QAAQ,cAAc,UAAa,YAAY,GAC/D,MAAM,eAAe;KAGvB,MAAM,aAAa,MAAM,QAAQ,MAAM;KAEvC,IAAI,eAAe,QAAQ,eAAe,UAAa,aAAa,GAClE,MAAM,mBAAmB;KAG3B;IACF;IAEA,IAAI,MAAM,SAAS,uBAAuB;KACxC,MAAM,QAAQ,MAAM;KAEpB,IAAI,MAAM,SAAS,YACjB,WAAW,IAAI,MAAM,OAAO;MAAE,IAAI,MAAM;MAAI,MAAM,MAAM;MAAM,MAAM;KAAG,CAAC;KAG1E;IACF;IAEA,IAAI,MAAM,SAAS,uBAAuB;KACxC,IAAI,MAAM,MAAM,SAAS,cACvB,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM,MAAM;KAAK;UAC5C,IAAI,MAAM,MAAM,SAAS,oBAAoB;MAClD,MAAM,cAAc,WAAW,IAAI,MAAM,KAAK;MAE9C,IAAI,aACF,YAAY,QAAQ,MAAM,MAAM;KAEpC;KAEA;IACF;IAEA,IAAI,MAAM,SAAS,sBAAsB;KACvC,MAAM,cAAc,WAAW,IAAI,MAAM,KAAK;KAE9C,IAAI,aAAa;MACf,MAAM;OACJ,MAAM;OACN,IAAI,YAAY;OAChB,MAAM,YAAY;OAClB,OAAO,cAAuC,YAAY,MAAM,CAAC,CAAC;MACpE;MAEA,WAAW,OAAO,MAAM,KAAK;KAC/B;KAEA;IACF;IAEA,IAAI,MAAM,SAAS,iBAAiB;KAClC,gBAAgB,MAAM,MAAM,eAAe;KAC3C,MAAM,SAAS,MAAM,MAAM,iBAAiB,MAAM;KAKlD,MAAM,YAAY,MAAM,MAAM;KAE9B,IAAI,cAAc,QAAQ,cAAc,UAAa,YAAY,GAC/D,MAAM,eAAe;KAGvB,MAAM,aAAa,MAAM,MAAM;KAE/B,IAAI,eAAe,QAAQ,eAAe,UAAa,aAAa,GAClE,MAAM,mBAAmB;IAE7B;GACF;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,QAAQ,MAAM,QAAQ,MAAM;EAElC,MAAM,eAAe,cAAc,aAAa;EAEhD,KAAK,OAAO,MAAM,YAAY,YAAY,+CAA+C;GACvF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;;;;;;CAaA,AAAQ,YACN,UACA,SAC2D;EAC3D,MAAM,EAAE,QAAQ,UAAU,sBAAsB,oBAAoB,QAAQ;EAC5E,MAAM,WAAW,KAAK,cAAc,SAAS,SAAS;EACtD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,OAAO,KAAK;GACZ,YAAY,SAAS,aAAa,KAAK,OAAO,aAAa;GAC3D,UAAU;GACV,GAAG,KAAK,YAAY,QAAQ,SAAS,YAAY;GAKjD,GAAI,gBAAgB,UAAa,CAAC,SAAS,WAAW,EAAE,YAAY,IAAI,CAAC;GACzE,GAAG,KAAK,WAAW,SAAS,KAAK;GACjC,GAAG,KAAK,sBAAsB,SAAS,cAAc;GACrD,GAAG;EACL;CACF;;;;;;;;;;;;;CAcA,AAAQ,YACN,QACA,cACkD;EAClD,IAAI,CAAC,QACH,OAAO,CAAC;EAGV,MAAM,cAAc,cAAc,eAAe;EAEjD,IAAI,KAAK,aAAa,iBAAiB,cAAc,GACnD,OAAO,EACL,QAAQ,CAAC;GAAE,MAAM;GAAQ,MAAM;GAAQ,eAAe,EAAE,MAAM,YAAY;EAAE,CAAC,EAC/E;EAGF,OAAO,EAAE,OAAO;CAClB;;;;;;;;;;;;CAaA,AAAQ,cACN,WAC8C;EAC9C,IAAI,CAAC,KAAK,aAAa,aAAa,CAAC,WACnC,OAAO,CAAC;EAGV,IAAI,UAAU,cAAc,UAAa,UAAU,WAAW,QAC5D,OAAO,CAAC;EAGV,MAAM,SAAS,UAAU,aAAa,uBAAuB,UAAU,UAAU;EAEjF,OAAO,EACL,UAAU;GACR,MAAM;GACN,eAAe,KAAK,IAAI,qBAAqB,MAAM;EACrD,EACF;CACF;;;;;;;;;;;;;;CAeA,AAAQ,WAAW,OAAgE;EACjF,MAAM,SAAS,iBAAiB,KAAK;EAErC,IAAI,CAAC,QACH,OAAO,CAAC;EAGV,IAAI,KAAK,OAAO,iBAAiB,OAAO,SAAS,GAAG;GAClD,MAAM,OAAO,OAAO,SAAS;GAC7B,OAAO,QAAQ;IAAE,GAAG,OAAO;IAAO,eAAe,EAAE,MAAM,YAAY;GAAE;EACzE;EAEA,OAAO,EAAE,OAAO,OAAO;CACzB;;;;;;;;;;;;;CAcA,AAAQ,sBAAsB,gBAE5B;EACA,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO,EACL,eAAe,EACb,QAAQ;GAAE,MAAM;GAAe,QAAQ;EAAe,EACxD,EACF;CACF;;;;;;CAOA,AAAQ,YAAY,SAA2C;EAC7D,OAAO,QACJ,QAAQ,UAAwC,MAAM,SAAS,MAAM,CAAC,CACtE,KAAK,UAAU,MAAM,IAAI,CAAC,CAC1B,KAAK,EAAE;CACZ;;;;;;CAOA,AAAQ,iBACN,SACoC;EACpC,MAAM,WAAW,QAAQ,QACtB,UAA2C,MAAM,SAAS,UAC7D;EAEA,IAAI,SAAS,WAAW,GACtB;EAGF,OAAO,SAAS,KAAK,WAAW;GAC9B,IAAI,MAAM;GACV,MAAM,MAAM;GACZ,OAAQ,MAAM,SAAS,CAAC;EAC1B,EAAE;CACJ;;;;;;;;;;;;;CAcA,AAAQ,aAAa,KAA6B;EAChD,MAAM,QAAQ,IAAI,gBAAgB;EAClC,MAAM,SAAS,IAAI,iBAAiB;EACpC,MAAM,YAAY,IAAI;EACtB,MAAM,aAAa,IAAI;EAEvB,OAAO;GACL;GACA;GACA,OAAO,QAAQ;GACf,GAAI,cAAc,QAAQ,cAAc,UAAa,YAAY,IAC7D,EAAE,cAAc,UAAU,IAC1B,CAAC;GACL,GAAI,eAAe,QAAQ,eAAe,UAAa,aAAa,IAChE,EAAE,kBAAkB,WAAW,IAC/B,CAAC;EACP;CACF;;;;;;CAOA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,mBAAmB,MAAM;EAEzC,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-anthropic
-description: 'Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), no first-party embeddings. Triggers: `AnthropicSDK`, `anthropic.model`, `anthropic.count`, `maxTokens`, `claude-sonnet-4-6`, `claude-haiku-4-5`, `claude-opus-4-7`; "wire claude into warlock agent", "configure anthropic provider", "use claude sonnet", "anthropic gateway baseURL"; typical import `import { AnthropicSDK } from "@warlock.js/ai-anthropic"`. Skip: embeddings — `@warlock.js/ai-openai/setup-openai/SKILL.md`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `@anthropic-ai/sdk`; Vercel `@ai-sdk/anthropic`.'
+description: 'Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, reasoning?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → thinking budget_tokens, prompt caching via promptCaching + options.cacheControl, cost-truth usage (cachedTokens/cacheWriteTokens), no first-party embeddings. Triggers: `AnthropicSDK`, `anthropic.model`, `anthropic.count`, `maxTokens`, `reasoning`, `thinking`, `budget_tokens`, `cacheControl`, `promptCaching`, `cachedTokens`, `cacheWriteTokens`, `claude-sonnet-4-6`, `claude-haiku-4-5`, `claude-opus-4-7`; "wire claude into warlock agent", "configure anthropic provider", "use claude sonnet", "anthropic gateway baseURL", "claude extended thinking", "anthropic prompt caching cost"; typical import `import { AnthropicSDK } from "@warlock.js/ai-anthropic"`. Skip: embeddings — `@warlock.js/ai-openai/setup-openai/SKILL.md`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `@anthropic-ai/sdk`; Vercel `@ai-sdk/anthropic`.'
 ---
 
 # `@warlock.js/ai-anthropic`
@@ -46,8 +46,12 @@ anthropic.model({ name: "claude-opus-4-7", maxTokens: 8192 })   // raise the cap
 | --- | --- |
 | `structuredOutput` | `true` (via Anthropic's native `output_config.format`) |
 | `vision` | Inferred from model name. `true` for Claude 3 / 3.5 / 3.7 / 4 family; `false` for pre-3 and unknown. |
+| `reasoning` | `true` — every modern Claude model accepts extended thinking. Override with `reasoning: false` for legacy/proxied targets that reject the `thinking` param. |
+| `promptCaching` | `true` — the adapter places `cache_control` breakpoints and reports both cache reads (`Usage.cachedTokens`) and writes (`Usage.cacheWriteTokens`). |
+| `pdf` | `true` — the Messages API accepts PDF/document content blocks on vision-capable models. |
+| `audio` | absent (`false`) — Anthropic has no audio-input block, so the agent rejects audio attachments upfront. |
 
-Explicit config always wins.
+Explicit config always wins (`structuredOutput`, `vision`, `reasoning`).
 
 ## Pricing & cost
 
@@ -67,7 +71,19 @@ const anthropic = new AnthropicSDK({
 anthropic.model({ name: "claude-sonnet-4-6", pricing: { input: 3, output: 15 } });
 ```
 
-Resolution at `model()` time: per-model `pricing` > SDK registry entry for that name > `undefined`. `ModelPricing` is `{ input, output, cachedInput?, cachedOutput? }`. Anthropic meters cache **reads** — when present they surface as `usage.cachedTokens` and bill at the `cachedInput` rate (falling back to `input` when unset).
+Resolution at `model()` time: per-model `pricing` > SDK registry entry for that name > `undefined`. `ModelPricing` is `{ input, output, cachedInput?, cachedOutput? }`.
+
+## Tokens & usage accounting
+
+Every `ModelResponse.usage` (and the streaming terminal `done`) is normalized to the neutral `Usage` shape. Anthropic reports `input_tokens` / `output_tokens` separately with no pre-summed total, so `total` is computed (`input + output`). The two cache channels are surfaced only when non-zero:
+
+| Neutral field | Anthropic source | Meaning / billing |
+| --- | --- | --- |
+| `usage.cachedTokens` | `cache_read_input_tokens` | Subset of input served from the prompt cache. Bills at `cachedInput` (falls back to `input` when unset). |
+| `usage.cacheWriteTokens` | `cache_creation_input_tokens` | Input tokens **written** to the cache on this call (the ~1.25x write surcharge). Bills at `cachedOutput` when set. |
+| `usage.reasoningTokens` | — *(not reported separately)* | Anthropic bills extended-thinking tokens **inside** `output_tokens`, so this stays unset — populating it would double-count against `output`. |
+
+In streaming, `cachedTokens` / `cacheWriteTokens` are seeded from `message_start` and then **overwritten by the cumulative counts on `message_delta`** when those are present and non-zero, so the terminal `done` carries the final tally.
 
 ## `max_tokens` is required
 
@@ -88,6 +104,33 @@ Anthropic has no `"system"` role inside `messages`. The adapter hoists every neu
 
 When the agent passes `responseSchema` and the model is `structuredOutput`-capable, an **object-root** schema is forwarded as `output_config: { format: { type: "json_schema", schema } }` (Anthropic native). Non-object schemas or `structuredOutput: false` omit it; the agent's soft system-prompt hint + client-side `validate()` still enforce shape.
 
+## Extended thinking (reasoning)
+
+When the model is `reasoning`-capable (default `true`) and the agent passes `options.reasoning`, the adapter forwards Anthropic extended thinking as `thinking: { type: "enabled", budget_tokens }`:
+
+- `reasoning.maxTokens` → `budget_tokens` verbatim.
+- `reasoning.effort` (`"low" | "medium" | "high"`) → a tiered budget (`1024` / `4096` / `12000`) when no explicit `maxTokens` is given.
+- Any resolved budget is floored at Anthropic's `1024`-token minimum.
+- `reasoning` with neither `effort` nor `maxTokens` emits nothing.
+
+Because Anthropic 400s when `thinking` is combined with an explicit `temperature`, the adapter **drops `temperature`** for that request (thinking pins sampling to the default). Set `reasoning: false` on the model config to suppress the param entirely for proxied/legacy targets.
+
+```ts
+await agent.execute("Prove it.", { reasoning: { effort: "high" } });   // budget_tokens: 12000
+await model.complete(messages, { reasoning: { maxTokens: 8000 } });     // budget_tokens: 8000
+```
+
+## Prompt caching (`cache_control`)
+
+Two independent breakpoint sites, both reported back via `usage.cachedTokens` / `usage.cacheWriteTokens`:
+
+- **Tools** — set `model({ promptCaching: true })` to mark the *last* tool definition with `cache_control: ephemeral`. One breakpoint caches the whole tool prefix; off by default since a write costs ~1.25x and only pays off across multiple trips.
+- **System prompt** — a per-call `options.cacheControl.breakpoints >= 1` emits the system prompt as a `TextBlockParam` carrying `cache_control: ephemeral` (the longest stable prefix on a turn). Without the hint the system prompt stays a plain string (uncached), so a one-shot call never pays the write surcharge. No system prompt → no block.
+
+```ts
+await agent.execute("…", { cacheControl: { breakpoints: 1 } });  // cache the system prefix
+```
+
 ## Multipart messages (vision)
 
 `ContentPart[]` user content maps to Anthropic content blocks:
@@ -102,7 +145,7 @@ When the agent passes `responseSchema` and the model is `structuredOutput`-capab
 
 - `content_block_delta` text → `{ type: "delta", content }`
 - `tool_use` blocks accumulate `input_json_delta` fragments and emit one consolidated `{ type: "tool-call", ... }` at `content_block_stop` (no partial-arg `{}` artifact — input is parsed once, whole)
-- terminal `{ type: "done", finishReason, usage }` — `usage.input` from `message_start`, `usage.output` from `message_delta`, `total` computed
+- terminal `{ type: "done", finishReason, usage }` — `usage.input` from `message_start`, `usage.output` from `message_delta`, `total` computed; `cachedTokens` / `cacheWriteTokens` from `message_start` then the cumulative `message_delta` counts (see [Tokens & usage accounting](#tokens--usage-accounting))
 
 ## Finish-reason mapping
 

package/llms.txt

@@ -6,4 +6,4 @@
 
 ## Skills
 
-- [setup-anthropic](@warlock.js/ai-anthropic/setup-anthropic/SKILL.md): Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), no first-party embeddings. Triggers: `AnthropicSDK`, `anthropic.model`, `anthropic.count`, `maxTokens`, `claude-sonnet-4-6`, `claude-haiku-4-5`, `claude-opus-4-7`; "wire claude into warlock agent", "configure anthropic provider", "use claude sonnet", "anthropic gateway baseURL"; typical import `import { AnthropicSDK } from "@warlock.js/ai-anthropic"`. Skip: embeddings — `@warlock.js/ai-openai/setup-openai/SKILL.md`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `@anthropic-ai/sdk`; Vercel `@ai-sdk/anthropic`.
+- [setup-anthropic](@warlock.js/ai-anthropic/setup-anthropic/SKILL.md): Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, reasoning?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → thinking budget_tokens, prompt caching via promptCaching + options.cacheControl, cost-truth usage (cachedTokens/cacheWriteTokens), no first-party embeddings. Triggers: `AnthropicSDK`, `anthropic.model`, `anthropic.count`, `maxTokens`, `reasoning`, `thinking`, `budget_tokens`, `cacheControl`, `promptCaching`, `cachedTokens`, `cacheWriteTokens`, `claude-sonnet-4-6`, `claude-haiku-4-5`, `claude-opus-4-7`; "wire claude into warlock agent", "configure anthropic provider", "use claude sonnet", "anthropic gateway baseURL", "claude extended thinking", "anthropic prompt caching cost"; typical import `import { AnthropicSDK } from "@warlock.js/ai-anthropic"`. Skip: embeddings — `@warlock.js/ai-openai/setup-openai/SKILL.md`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `@anthropic-ai/sdk`; Vercel `@ai-sdk/anthropic`.

package/package.json

@@ -15,12 +15,12 @@
   },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.96.0",
-    "@warlock.js/logger": "4.2.11"
+    "@warlock.js/logger": "4.3.0"
   },
   "peerDependencies": {
-    "@warlock.js/ai": "4.2.11"
+    "@warlock.js/ai": "4.3.0"
   },
-  "version": "4.2.11",
+  "version": "4.3.0",
   "main": "./cjs/index.cjs",
   "module": "./esm/index.mjs",
   "types": "./esm/index.d.mts",

package/skills/README.md

@@ -6,4 +6,4 @@ Per-task skills. All cross-references use the form `@warlock.js/<pkg>/<skill>/SK
 
 ### [`setup-anthropic/`](./setup-anthropic/SKILL.md)
 
-Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), no first-party embeddings. Load when wiring a Claude-backed model into a @warlock.js agent.
+Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, reasoning?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → thinking budget_tokens, prompt caching via promptCaching + options.cacheControl with cost-truth usage (cachedTokens/cacheWriteTokens), no first-party embeddings. Load when wiring a Claude-backed model into a @warlock.js agent.

package/skills/setup-anthropic/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: setup-anthropic
-description: 'Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), no first-party embeddings. Triggers: `AnthropicSDK`, `anthropic.model`, `anthropic.count`, `maxTokens`, `claude-sonnet-4-6`, `claude-haiku-4-5`, `claude-opus-4-7`; "wire claude into warlock agent", "configure anthropic provider", "use claude sonnet", "anthropic gateway baseURL"; typical import `import { AnthropicSDK } from "@warlock.js/ai-anthropic"`. Skip: embeddings — `@warlock.js/ai-openai/setup-openai/SKILL.md`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `@anthropic-ai/sdk`; Vercel `@ai-sdk/anthropic`.'
+description: 'Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, reasoning?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → thinking budget_tokens, prompt caching via promptCaching + options.cacheControl, cost-truth usage (cachedTokens/cacheWriteTokens), no first-party embeddings. Triggers: `AnthropicSDK`, `anthropic.model`, `anthropic.count`, `maxTokens`, `reasoning`, `thinking`, `budget_tokens`, `cacheControl`, `promptCaching`, `cachedTokens`, `cacheWriteTokens`, `claude-sonnet-4-6`, `claude-haiku-4-5`, `claude-opus-4-7`; "wire claude into warlock agent", "configure anthropic provider", "use claude sonnet", "anthropic gateway baseURL", "claude extended thinking", "anthropic prompt caching cost"; typical import `import { AnthropicSDK } from "@warlock.js/ai-anthropic"`. Skip: embeddings — `@warlock.js/ai-openai/setup-openai/SKILL.md`; sibling adapters `@warlock.js/ai-openai`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `@anthropic-ai/sdk`; Vercel `@ai-sdk/anthropic`.'
 ---
 
 # `@warlock.js/ai-anthropic`
@@ -38,8 +38,12 @@ anthropic.model({ name: "claude-opus-4-7", maxTokens: 8192 })   // raise the cap
 | --- | --- |
 | `structuredOutput` | `true` (via Anthropic's native `output_config.format`) |
 | `vision` | Inferred from model name. `true` for Claude 3 / 3.5 / 3.7 / 4 family; `false` for pre-3 and unknown. |
+| `reasoning` | `true` — every modern Claude model accepts extended thinking. Override with `reasoning: false` for legacy/proxied targets that reject the `thinking` param. |
+| `promptCaching` | `true` — the adapter places `cache_control` breakpoints and reports both cache reads (`Usage.cachedTokens`) and writes (`Usage.cacheWriteTokens`). |
+| `pdf` | `true` — the Messages API accepts PDF/document content blocks on vision-capable models. |
+| `audio` | absent (`false`) — Anthropic has no audio-input block, so the agent rejects audio attachments upfront. |
 
-Explicit config always wins.
+Explicit config always wins (`structuredOutput`, `vision`, `reasoning`).
 
 ## Pricing & cost
 
@@ -59,7 +63,19 @@ const anthropic = new AnthropicSDK({
 anthropic.model({ name: "claude-sonnet-4-6", pricing: { input: 3, output: 15 } });
 ```
 
-Resolution at `model()` time: per-model `pricing` > SDK registry entry for that name > `undefined`. `ModelPricing` is `{ input, output, cachedInput?, cachedOutput? }`. Anthropic meters cache **reads** — when present they surface as `usage.cachedTokens` and bill at the `cachedInput` rate (falling back to `input` when unset).
+Resolution at `model()` time: per-model `pricing` > SDK registry entry for that name > `undefined`. `ModelPricing` is `{ input, output, cachedInput?, cachedOutput? }`.
+
+## Tokens & usage accounting
+
+Every `ModelResponse.usage` (and the streaming terminal `done`) is normalized to the neutral `Usage` shape. Anthropic reports `input_tokens` / `output_tokens` separately with no pre-summed total, so `total` is computed (`input + output`). The two cache channels are surfaced only when non-zero:
+
+| Neutral field | Anthropic source | Meaning / billing |
+| --- | --- | --- |
+| `usage.cachedTokens` | `cache_read_input_tokens` | Subset of input served from the prompt cache. Bills at `cachedInput` (falls back to `input` when unset). |
+| `usage.cacheWriteTokens` | `cache_creation_input_tokens` | Input tokens **written** to the cache on this call (the ~1.25x write surcharge). Bills at `cachedOutput` when set. |
+| `usage.reasoningTokens` | — *(not reported separately)* | Anthropic bills extended-thinking tokens **inside** `output_tokens`, so this stays unset — populating it would double-count against `output`. |
+
+In streaming, `cachedTokens` / `cacheWriteTokens` are seeded from `message_start` and then **overwritten by the cumulative counts on `message_delta`** when those are present and non-zero, so the terminal `done` carries the final tally.
 
 ## `max_tokens` is required
 
@@ -80,6 +96,33 @@ Anthropic has no `"system"` role inside `messages`. The adapter hoists every neu
 
 When the agent passes `responseSchema` and the model is `structuredOutput`-capable, an **object-root** schema is forwarded as `output_config: { format: { type: "json_schema", schema } }` (Anthropic native). Non-object schemas or `structuredOutput: false` omit it; the agent's soft system-prompt hint + client-side `validate()` still enforce shape.
 
+## Extended thinking (reasoning)
+
+When the model is `reasoning`-capable (default `true`) and the agent passes `options.reasoning`, the adapter forwards Anthropic extended thinking as `thinking: { type: "enabled", budget_tokens }`:
+
+- `reasoning.maxTokens` → `budget_tokens` verbatim.
+- `reasoning.effort` (`"low" | "medium" | "high"`) → a tiered budget (`1024` / `4096` / `12000`) when no explicit `maxTokens` is given.
+- Any resolved budget is floored at Anthropic's `1024`-token minimum.
+- `reasoning` with neither `effort` nor `maxTokens` emits nothing.
+
+Because Anthropic 400s when `thinking` is combined with an explicit `temperature`, the adapter **drops `temperature`** for that request (thinking pins sampling to the default). Set `reasoning: false` on the model config to suppress the param entirely for proxied/legacy targets.
+
+```ts
+await agent.execute("Prove it.", { reasoning: { effort: "high" } });   // budget_tokens: 12000
+await model.complete(messages, { reasoning: { maxTokens: 8000 } });     // budget_tokens: 8000
+```
+
+## Prompt caching (`cache_control`)
+
+Two independent breakpoint sites, both reported back via `usage.cachedTokens` / `usage.cacheWriteTokens`:
+
+- **Tools** — set `model({ promptCaching: true })` to mark the *last* tool definition with `cache_control: ephemeral`. One breakpoint caches the whole tool prefix; off by default since a write costs ~1.25x and only pays off across multiple trips.
+- **System prompt** — a per-call `options.cacheControl.breakpoints >= 1` emits the system prompt as a `TextBlockParam` carrying `cache_control: ephemeral` (the longest stable prefix on a turn). Without the hint the system prompt stays a plain string (uncached), so a one-shot call never pays the write surcharge. No system prompt → no block.
+
+```ts
+await agent.execute("…", { cacheControl: { breakpoints: 1 } });  // cache the system prefix
+```
+
 ## Multipart messages (vision)
 
 `ContentPart[]` user content maps to Anthropic content blocks:
@@ -94,7 +137,7 @@ When the agent passes `responseSchema` and the model is `structuredOutput`-capab
 
 - `content_block_delta` text → `{ type: "delta", content }`
 - `tool_use` blocks accumulate `input_json_delta` fragments and emit one consolidated `{ type: "tool-call", ... }` at `content_block_stop` (no partial-arg `{}` artifact — input is parsed once, whole)
-- terminal `{ type: "done", finishReason, usage }` — `usage.input` from `message_start`, `usage.output` from `message_delta`, `total` computed
+- terminal `{ type: "done", finishReason, usage }` — `usage.input` from `message_start`, `usage.output` from `message_delta`, `total` computed; `cachedTokens` / `cacheWriteTokens` from `message_start` then the cumulative `message_delta` counts (see [Tokens & usage accounting](#tokens--usage-accounting))
 
 ## Finish-reason mapping