@warlock.js/ai-anthropic 5.28.0 → 5.30.0

package/esm/utils/to-anthropic-messages.mjs

@@ -1,4 +1,6 @@
-//#region ../ai-anthropic/src/utils/to-anthropic-messages.ts
+import { fromProviderBlocks } from "./provider-blocks.mjs";
+
+//#region ../../packages/ai-anthropic/src/utils/to-anthropic-messages.ts
 /**
 * Convert vendor-neutral `Message[]` into Anthropic's request shape.
 *
@@ -9,15 +11,24 @@
 *    separated) and returned separately as the top-level `system`
 *    parameter.
 * 2. **Tool results are `user` turns.** A neutral `tool` message becomes
-*    a `user` message whose content is a single `tool_result` block
-*    keyed by `tool_use_id`.
+*    a `tool_result` block keyed by `tool_use_id`. Consecutive tool
+*    messages (the results of one parallel tool-use turn) are grouped
+*    into a single `user` message, as Anthropic expects — splitting them
+*    across messages teaches the model to stop calling tools in parallel.
 * 3. **Tool calls are `tool_use` content blocks.** An assistant message
 *    carrying `toolCalls` becomes an `assistant` message whose content
 *    is an optional leading `text` block followed by one `tool_use`
 *    block per call.
 *
-* Consecutive same-role turns are left as-is — the Messages API merges
-* them server-side.
+* An assistant message that carries Anthropic `blocks` (thinking,
+* redacted thinking, compaction, server-tool blocks — see
+* `ProviderContentBlock`) is replayed verbatim from those blocks, which
+* keeps signed thinking and compaction state valid. Blocks from another
+* provider, or blocks whose `tool_use` ids no longer match `toolCalls`,
+* are ignored and the turn is rebuilt from `content` + `toolCalls`.
+*
+* Other consecutive same-role turns are left as-is — the Messages API
+* merges them server-side.
 *
 * @example
 * const { system, messages } = toAnthropicMessages([
@@ -36,16 +47,29 @@ function toAnthropicMessages(messages) {
 			continue;
 		}
 		if (message.role === "tool") {
-			mapped.push({
+			const toolResult = {
+				type: "tool_result",
+				tool_use_id: message.toolCallId ?? "",
+				content: stringifyContent(message.content)
+			};
+			const previous = mapped[mapped.length - 1];
+			if (previous && previous.role === "user" && isToolResultTurn(previous)) previous.content.push(toolResult);
+			else mapped.push({
 				role: "user",
-				content: [{
-					type: "tool_result",
-					tool_use_id: message.toolCallId ?? "",
-					content: stringifyContent(message.content)
-				}]
+				content: [toolResult]
 			});
 			continue;
 		}
+		if (message.role === "assistant") {
+			const replayed = fromProviderBlocks(message.blocks, message.toolCalls);
+			if (replayed) {
+				mapped.push({
+					role: "assistant",
+					content: replayed
+				});
+				continue;
+			}
+		}
 		if (message.role === "assistant" && message.toolCalls && message.toolCalls.length > 0) {
 			const blocks = extractThinkingBlocks(message.toolCalls);
 			const text = stringifyContent(message.content);
@@ -108,6 +132,14 @@ function toAnthropicThinkingBlock(block) {
 	}];
 	return [];
 }
+/**
+* A `user` turn this function built from tool results — its content is a
+* non-empty array made only of `tool_result` blocks. Later tool results
+* of the same tool-use turn are appended to it.
+*/
+function isToolResultTurn(message) {
+	return Array.isArray(message.content) && message.content.length > 0 && message.content.every((block) => block.type === "tool_result");
+}
 function isRecord(value) {
 	return typeof value === "object" && value !== null;
 }

package/esm/utils/to-anthropic-messages.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"to-anthropic-messages.mjs","names":[],"sources":["../../../../../../../ai-anthropic/src/utils/to-anthropic-messages.ts"],"sourcesContent":["import type { ContentPart, Message, ModelToolCallRequest } 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[] = extractThinkingBlocks(message.toolCalls);\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 * Recover signed extended-thinking blocks stored by this adapter on a tool\n * request. Anthropic requires these blocks to be replayed before the\n * assistant's text and tool uses on the following request.\n */\nfunction extractThinkingBlocks(toolCalls: ModelToolCallRequest[]): Anthropic.ContentBlockParam[] {\n  for (const toolCall of toolCalls) {\n    const anthropic = toolCall.providerMetadata?.anthropic;\n\n    if (!isRecord(anthropic) || !Array.isArray(anthropic.thinkingBlocks)) {\n      continue;\n    }\n\n    return anthropic.thinkingBlocks.flatMap(toAnthropicThinkingBlock);\n  }\n\n  return [];\n}\n\nfunction toAnthropicThinkingBlock(block: unknown): Anthropic.ContentBlockParam[] {\n  if (!isRecord(block) || typeof block.type !== \"string\") {\n    return [];\n  }\n\n  if (\n    block.type === \"thinking\" &&\n    typeof block.thinking === \"string\" &&\n    typeof block.signature === \"string\"\n  ) {\n    return [{ type: \"thinking\", thinking: block.thinking, signature: block.signature }];\n  }\n\n  if (block.type === \"redacted_thinking\" && typeof block.data === \"string\") {\n    return [{ type: \"redacted_thinking\", data: block.data }];\n  }\n\n  return [];\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n  return typeof value === \"object\" && value !== null;\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  // PDF → Anthropic `document` block (A2). The Messages API accepts a\n  // url or base64 `application/pdf` source on PDF-capable Claude models.\n  if (part.type === \"pdf\") {\n    const source =\n      \"url\" in part.source\n        ? { type: \"url\", url: part.source.url }\n        : { type: \"base64\", media_type: \"application/pdf\", data: part.source.base64 };\n    return { type: \"document\", source } as unknown as Anthropic.ContentBlockParam;\n  }\n\n  // Anthropic has no audio content block; the agent's capability gate\n  // normally prevents this, but stay loud rather than mis-map it.\n  if (part.type === \"audio\") {\n    throw new Error(\"Anthropic adapter: audio attachments are not supported\");\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"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8CA,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,sBAAsB,QAAQ,SAAS;GACrF,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,sBAAsB,WAAkE;CAC/F,KAAK,MAAM,YAAY,WAAW;EAChC,MAAM,YAAY,SAAS,kBAAkB;EAE7C,IAAI,CAAC,SAAS,SAAS,KAAK,CAAC,MAAM,QAAQ,UAAU,cAAc,GACjE;EAGF,OAAO,UAAU,eAAe,QAAQ,wBAAwB;CAClE;CAEA,OAAO,CAAC;AACV;AAEA,SAAS,yBAAyB,OAA+C;CAC/E,IAAI,CAAC,SAAS,KAAK,KAAK,OAAO,MAAM,SAAS,UAC5C,OAAO,CAAC;CAGV,IACE,MAAM,SAAS,cACf,OAAO,MAAM,aAAa,YAC1B,OAAO,MAAM,cAAc,UAE3B,OAAO,CAAC;EAAE,MAAM;EAAY,UAAU,MAAM;EAAU,WAAW,MAAM;CAAU,CAAC;CAGpF,IAAI,MAAM,SAAS,uBAAuB,OAAO,MAAM,SAAS,UAC9D,OAAO,CAAC;EAAE,MAAM;EAAqB,MAAM,MAAM;CAAK,CAAC;CAGzD,OAAO,CAAC;AACV;AAEA,SAAS,SAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU;AAChD;;;;;;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;CAKzC,IAAI,KAAK,SAAS,OAKhB,OAAO;EAAE,MAAM;EAAY,QAHzB,SAAS,KAAK,SACV;GAAE,MAAM;GAAO,KAAK,KAAK,OAAO;EAAI,IACpC;GAAE,MAAM;GAAU,YAAY;GAAmB,MAAM,KAAK,OAAO;EAAO;CAC9C;CAKpC,IAAI,KAAK,SAAS,SAChB,MAAM,IAAI,MAAM,wDAAwD;CAG1E,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"}
+{"version":3,"file":"to-anthropic-messages.mjs","names":[],"sources":["../../../../../../../../packages/ai-anthropic/src/utils/to-anthropic-messages.ts"],"sourcesContent":["import type { ContentPart, Message, ModelToolCallRequest } from \"@warlock.js/ai\";\nimport type Anthropic from \"@anthropic-ai/sdk\";\nimport { fromProviderBlocks } from \"./provider-blocks\";\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 `tool_result` block keyed by `tool_use_id`. Consecutive tool\n *    messages (the results of one parallel tool-use turn) are grouped\n *    into a single `user` message, as Anthropic expects — splitting them\n *    across messages teaches the model to stop calling tools in parallel.\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 * An assistant message that carries Anthropic `blocks` (thinking,\n * redacted thinking, compaction, server-tool blocks — see\n * `ProviderContentBlock`) is replayed verbatim from those blocks, which\n * keeps signed thinking and compaction state valid. Blocks from another\n * provider, or blocks whose `tool_use` ids no longer match `toolCalls`,\n * are ignored and the turn is rebuilt from `content` + `toolCalls`.\n *\n * Other consecutive same-role turns are left as-is — the Messages API\n * merges 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      const toolResult: Anthropic.ToolResultBlockParam = {\n        type: \"tool_result\",\n        tool_use_id: message.toolCallId ?? \"\",\n        content: stringifyContent(message.content),\n      };\n      const previous = mapped[mapped.length - 1];\n\n      if (previous && previous.role === \"user\" && isToolResultTurn(previous)) {\n        previous.content.push(toolResult);\n      } else {\n        mapped.push({ role: \"user\", content: [toolResult] });\n      }\n\n      continue;\n    }\n\n    if (message.role === \"assistant\") {\n      const replayed = fromProviderBlocks(message.blocks, message.toolCalls);\n\n      if (replayed) {\n        mapped.push({ role: \"assistant\", content: replayed });\n\n        continue;\n      }\n    }\n\n    if (message.role === \"assistant\" && message.toolCalls && message.toolCalls.length > 0) {\n      const blocks: Anthropic.ContentBlockParam[] = extractThinkingBlocks(message.toolCalls);\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 * Recover signed extended-thinking blocks stored by this adapter on a tool\n * request. Anthropic requires these blocks to be replayed before the\n * assistant's text and tool uses on the following request.\n */\nfunction extractThinkingBlocks(toolCalls: ModelToolCallRequest[]): Anthropic.ContentBlockParam[] {\n  for (const toolCall of toolCalls) {\n    const anthropic = toolCall.providerMetadata?.anthropic;\n\n    if (!isRecord(anthropic) || !Array.isArray(anthropic.thinkingBlocks)) {\n      continue;\n    }\n\n    return anthropic.thinkingBlocks.flatMap(toAnthropicThinkingBlock);\n  }\n\n  return [];\n}\n\nfunction toAnthropicThinkingBlock(block: unknown): Anthropic.ContentBlockParam[] {\n  if (!isRecord(block) || typeof block.type !== \"string\") {\n    return [];\n  }\n\n  if (\n    block.type === \"thinking\" &&\n    typeof block.thinking === \"string\" &&\n    typeof block.signature === \"string\"\n  ) {\n    return [{ type: \"thinking\", thinking: block.thinking, signature: block.signature }];\n  }\n\n  if (block.type === \"redacted_thinking\" && typeof block.data === \"string\") {\n    return [{ type: \"redacted_thinking\", data: block.data }];\n  }\n\n  return [];\n}\n\n/**\n * A `user` turn this function built from tool results — its content is a\n * non-empty array made only of `tool_result` blocks. Later tool results\n * of the same tool-use turn are appended to it.\n */\nfunction isToolResultTurn(\n  message: Anthropic.MessageParam,\n): message is Anthropic.MessageParam & { content: Anthropic.ContentBlockParam[] } {\n  return (\n    Array.isArray(message.content) &&\n    message.content.length > 0 &&\n    message.content.every((block) => block.type === \"tool_result\")\n  );\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n  return typeof value === \"object\" && value !== null;\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  // PDF → Anthropic `document` block (A2). The Messages API accepts a\n  // url or base64 `application/pdf` source on PDF-capable Claude models.\n  if (part.type === \"pdf\") {\n    const source =\n      \"url\" in part.source\n        ? { type: \"url\", url: part.source.url }\n        : { type: \"base64\", media_type: \"application/pdf\", data: part.source.base64 };\n    return { type: \"document\", source } as unknown as Anthropic.ContentBlockParam;\n  }\n\n  // Anthropic has no audio content block; the agent's capability gate\n  // normally prevents this, but stay loud rather than mis-map it.\n  if (part.type === \"audio\") {\n    throw new Error(\"Anthropic adapter: audio attachments are not supported\");\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"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwDA,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,MAAM,aAA6C;IACjD,MAAM;IACN,aAAa,QAAQ,cAAc;IACnC,SAAS,iBAAiB,QAAQ,OAAO;GAC3C;GACA,MAAM,WAAW,OAAO,OAAO,SAAS;GAExC,IAAI,YAAY,SAAS,SAAS,UAAU,iBAAiB,QAAQ,GACnE,SAAS,QAAQ,KAAK,UAAU;QAEhC,OAAO,KAAK;IAAE,MAAM;IAAQ,SAAS,CAAC,UAAU;GAAE,CAAC;GAGrD;EACF;EAEA,IAAI,QAAQ,SAAS,aAAa;GAChC,MAAM,WAAW,mBAAmB,QAAQ,QAAQ,QAAQ,SAAS;GAErE,IAAI,UAAU;IACZ,OAAO,KAAK;KAAE,MAAM;KAAa,SAAS;IAAS,CAAC;IAEpD;GACF;EACF;EAEA,IAAI,QAAQ,SAAS,eAAe,QAAQ,aAAa,QAAQ,UAAU,SAAS,GAAG;GACrF,MAAM,SAAwC,sBAAsB,QAAQ,SAAS;GACrF,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,sBAAsB,WAAkE;CAC/F,KAAK,MAAM,YAAY,WAAW;EAChC,MAAM,YAAY,SAAS,kBAAkB;EAE7C,IAAI,CAAC,SAAS,SAAS,KAAK,CAAC,MAAM,QAAQ,UAAU,cAAc,GACjE;EAGF,OAAO,UAAU,eAAe,QAAQ,wBAAwB;CAClE;CAEA,OAAO,CAAC;AACV;AAEA,SAAS,yBAAyB,OAA+C;CAC/E,IAAI,CAAC,SAAS,KAAK,KAAK,OAAO,MAAM,SAAS,UAC5C,OAAO,CAAC;CAGV,IACE,MAAM,SAAS,cACf,OAAO,MAAM,aAAa,YAC1B,OAAO,MAAM,cAAc,UAE3B,OAAO,CAAC;EAAE,MAAM;EAAY,UAAU,MAAM;EAAU,WAAW,MAAM;CAAU,CAAC;CAGpF,IAAI,MAAM,SAAS,uBAAuB,OAAO,MAAM,SAAS,UAC9D,OAAO,CAAC;EAAE,MAAM;EAAqB,MAAM,MAAM;CAAK,CAAC;CAGzD,OAAO,CAAC;AACV;;;;;;AAOA,SAAS,iBACP,SACgF;CAChF,OACE,MAAM,QAAQ,QAAQ,OAAO,KAC7B,QAAQ,QAAQ,SAAS,KACzB,QAAQ,QAAQ,OAAO,UAAU,MAAM,SAAS,aAAa;AAEjE;AAEA,SAAS,SAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU;AAChD;;;;;;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;CAKzC,IAAI,KAAK,SAAS,OAKhB,OAAO;EAAE,MAAM;EAAY,QAHzB,SAAS,KAAK,SACV;GAAE,MAAM;GAAO,KAAK,KAAK,OAAO;EAAI,IACpC;GAAE,MAAM;GAAU,YAAY;GAAmB,MAAM,KAAK,OAAO;EAAO;CAC9C;CAKpC,IAAI,KAAK,SAAS,SAChB,MAAM,IAAI,MAAM,wDAAwD;CAG1E,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"}

package/esm/utils/to-anthropic-tools.mjs

@@ -1,6 +1,6 @@
 import { extractJsonSchema } from "@warlock.js/ai";
 
-//#region ../ai-anthropic/src/utils/to-anthropic-tools.ts
+//#region ../../packages/ai-anthropic/src/utils/to-anthropic-tools.ts
 /**
 * Convert vendor-neutral `ToolConfig[]` into Anthropic's `tools` array.
 * Uses the shared `extractJsonSchema` helper; Anthropic requires the

package/esm/utils/to-anthropic-tools.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"to-anthropic-tools.mjs","names":[],"sources":["../../../../../../../ai-anthropic/src/utils/to-anthropic-tools.ts"],"sourcesContent":["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"],"mappings":";;;;;;;;;;;;;;AAcA,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,SAAS,kBAAkB,KAAK;CAEtC,IAAI,UAAU,OAAO,SAAS,UAC5B,OAAO;CAGT,OAAO,EAAE,MAAM,SAAS;AAC1B"}
+{"version":3,"file":"to-anthropic-tools.mjs","names":[],"sources":["../../../../../../../../packages/ai-anthropic/src/utils/to-anthropic-tools.ts"],"sourcesContent":["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"],"mappings":";;;;;;;;;;;;;;AAcA,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,SAAS,kBAAkB,KAAK;CAEtC,IAAI,UAAU,OAAO,SAAS,UAC5B,OAAO;CAGT,OAAO,EAAE,MAAM,SAAS;AAC1B"}

package/esm/utils/wrap-anthropic-error.mjs

@@ -1,7 +1,7 @@
 import { APIConnectionTimeoutError, APIError } from "@anthropic-ai/sdk";
 import { AIError, ContextLengthExceededError, InvalidRequestError, ProviderAuthError, ProviderError, ProviderRateLimitError, ProviderTimeoutError, QuotaExceededError } from "@warlock.js/ai";
 
-//#region ../ai-anthropic/src/utils/wrap-anthropic-error.ts
+//#region ../../packages/ai-anthropic/src/utils/wrap-anthropic-error.ts
 /**
 * Wrap any thrown value caught inside the Anthropic adapter into the
 * appropriate `@warlock.js/ai` `AIError` subclass.

package/esm/utils/wrap-anthropic-error.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"wrap-anthropic-error.mjs","names":[],"sources":["../../../../../../../ai-anthropic/src/utils/wrap-anthropic-error.ts"],"sourcesContent":["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"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AAuDA,SAAgB,mBAAmB,QAA0B;CAC3D,IAAI,kBAAkB,SACpB,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,IAAI,qBAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,SAAS,0BAA0B,MAAM,SAAS,oBAC1D,OAAO,IAAI,kBAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,WAAW,OAAO,MAAM,WAAW,KAC3C,OAAO,IAAI,kBAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,iBACjB,OAAO,IAAI,mBAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,SAAS,sBAAsB,MAAM,WAAW,KACxD,OAAO,IAAI,uBAAuB,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,IAAI,2BAA2B,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;EAG3E,OAAO,IAAI,oBAAoB,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;CACpE;CAEA,OAAO,IAAI,cAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;AAOA,SAAS,QAAQ,QAAsC;CACrD,IAAI,kBAAkB,UACpB,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,kBAAkB,2BACpB,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"}
+{"version":3,"file":"wrap-anthropic-error.mjs","names":[],"sources":["../../../../../../../../packages/ai-anthropic/src/utils/wrap-anthropic-error.ts"],"sourcesContent":["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"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AAuDA,SAAgB,mBAAmB,QAA0B;CAC3D,IAAI,kBAAkB,SACpB,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,IAAI,qBAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,SAAS,0BAA0B,MAAM,SAAS,oBAC1D,OAAO,IAAI,kBAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,WAAW,OAAO,MAAM,WAAW,KAC3C,OAAO,IAAI,kBAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,iBACjB,OAAO,IAAI,mBAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,SAAS,sBAAsB,MAAM,WAAW,KACxD,OAAO,IAAI,uBAAuB,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,IAAI,2BAA2B,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;EAG3E,OAAO,IAAI,oBAAoB,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;CACpE;CAEA,OAAO,IAAI,cAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;AAOA,SAAS,QAAQ,QAAsC;CACrD,IAAI,kBAAkB,UACpB,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,kBAAkB,2BACpB,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"}

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?, reasoning?, thinkingMode?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → adaptive thinking + output_config.effort (current/unknown models) or budget_tokens (Haiku 4.5, 4.6 and older; override with thinkingMode), 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`, `thinkingMode`, `adaptive`, `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`.'
+description: 'Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, reasoning?, thinkingMode?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → adaptive thinking + output_config.effort (current/unknown models) or budget_tokens (Haiku 4.5, 4.6 and older; override with thinkingMode), prompt caching via promptCaching + options.cacheControl (per-call ttl), provider blocks (thinking/compaction) replayed verbatim via ModelResponse.blocks / Message.blocks, providerOptions.anthropic (betas, body fields such as context_management / fallbacks, thinkingDisplay), model.countTokens(), distinct finish reasons (pause_turn, refusal, context_exceeded), cost-truth usage (cachedTokens/cacheWriteTokens), no first-party embeddings. Triggers: `AnthropicSDK`, `anthropic.model`, `anthropic.count`, `maxTokens`, `reasoning`, `thinking`, `budget_tokens`, `thinkingMode`, `adaptive`, `cacheControl`, `providerOptions`, `blocks`, `countTokens`, `pause_turn`, `refusal`, `context_management`, `compaction`, `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 — the `setup-openai` topic of the `warlock-js-ai-openai` skill; 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`
@@ -101,7 +101,7 @@ Anthropic has no `"system"` role inside `messages`. The adapter hoists every neu
 
 - Outgoing: vendor-neutral tools → Anthropic `tools` with `input_schema` (non-object schemas degrade to a parameterless object so registration never fails).
 - Assistant tool calls in history → `assistant` message with an optional leading `text` block + one `tool_use` block per call.
-- Tool results (`role: "tool"`) → a `user` turn carrying a single `tool_result` block keyed by `tool_use_id`.
+- Tool results (`role: "tool"`) → `tool_result` blocks keyed by `tool_use_id`. Consecutive tool messages (the results of one parallel tool-use turn) are grouped into **one** `user` turn, as Anthropic expects.
 - Response `tool_use` blocks → neutral `toolCalls`; `stop_reason: "tool_use"` → `finishReason: "tool_calls"`.
 
 ## Structured output
@@ -121,7 +121,7 @@ Adaptive models reject `budget_tokens` with a 400, so the shape matters.
 
 - **Adaptive:** `reasoning.effort` (`"low" | "medium" | "high"`) → `output_config.effort` of the same name. With only `reasoning.maxTokens` set, effort is `"medium"`; `maxTokens` then only sizes the reserved `max_tokens` room, it is not sent as a budget.
 - **Budget:** `reasoning.maxTokens` → `budget_tokens` verbatim; `reasoning.effort` → a tiered budget (`1024` / `4096` / `12000`) when no `maxTokens` is given. Any resolved budget is floored at Anthropic's `1024` minimum.
-- `reasoning` with neither `effort` nor `maxTokens`, or `effort: "none"`, emits nothing.
+- `reasoning` with neither `effort` nor `maxTokens`, or `effort: "none"`, emits nothing — except `effort: "none"` on Claude Sonnet 5.5, which sends `thinking: { type: "between_tools" }` (see below).
 
 ### `thinkingMode` override
 
@@ -152,7 +152,63 @@ await model.complete(messages, { reasoning: { maxTokens: 8000 } });
 // adaptive model: effort "medium"; budget model: budget_tokens 8000
 ```
 
-When a thinking-enabled response also makes tool calls, its signed `thinking` and `redacted_thinking` blocks are retained in each call's `providerMetadata.anthropic.thinkingBlocks`. The adapter replays those opaque blocks automatically on the next assistant turn, preserving Anthropic's required signatures; application code should leave that metadata intact when persisting message history.
+**Claude 5 family (Opus 5.5, Sonnet 5.5, Fable 5 / 5.1):** thinking is always on. These models reject `thinking: { type: "disabled" }` and `budget_tokens` with a 400, so the adapter never sends either: they are adaptive and `reasoning.maxTokens` only sizes `max_tokens`. `effort: "none"` is mapped per model:
+
+| Model                   | `effort: "none"` sends                                                                 | Result                                                                                                                                 |
+| ----------------------- | -------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- |
+| Sonnet 5.5              | `thinking: { type: "between_tools" }`, alone (no `output_config.effort`, no `display`) | Thinking off. `between_tools` takes no other field and needs effort `high` or below; the model default, `high`, applies.               |
+| Opus 5.5, Fable 5 / 5.1 | no `thinking` field                                                                    | No thinking-off form exists, so the model still thinks at its default effort (Opus 5.5: `medium`). Pass `effort: "low"` to think less. |
+
+`thinkingDisplay` is applied only to adaptive thinking, never beside `between_tools`.
+
+`thinking.display` is set with `providerOptions.anthropic.thinkingDisplay` (`"summarized"` for readable summaries; the 5-family default is `"omitted"`, empty thinking text). It is applied whenever the adapter sends `thinking` to an adaptive model, so pair it with a `reasoning` option.
+
+When a thinking-enabled response also makes tool calls, its signed `thinking` and `redacted_thinking` blocks are retained in each call's `providerMetadata.anthropic.thinkingBlocks` and replayed on the next assistant turn. The full turn is also returned as `ModelResponse.blocks` — see [Provider blocks](#provider-blocks-thinking-compaction-server-tools).
+
+## Provider blocks (thinking, compaction, server tools)
+
+When a response contains a block the neutral `content` / `toolCalls` cannot carry (thinking, redacted thinking, `compaction`, server-tool use/results, cited text), the adapter returns the whole ordered turn as `ModelResponse.blocks` (and on the streaming `done` chunk): `{ provider: "anthropic", type, data }` per block. Plain text / tool-use turns have no `blocks`.
+
+The agent copies them onto the assistant `Message.blocks` (tool-call trips, `pause_turn` trips, repair turns) and onto `LLMTrip.blocks`; `AgentResult.blocks` is set only for a single-trip run. On the next request the adapter replays an assistant message **verbatim** from its `blocks` — required for preserved thinking and server-side compaction. It falls back to rebuilding from `content` + `toolCalls` when the blocks come from another provider, a block's `data.type` differs from its `type` or is not an assistant-turn type, or their `tool_use` ids no longer match `toolCalls`.
+
+A replayed thinking block requires every earlier turn unchanged, including the run's tool-call turns and tool results. Never store just `{ content: result.text, blocks }` after a multi-trip run. Keep the run's turns:
+
+```ts
+// agent created with `captureMessages: true`
+const result = await agent.execute(input, { history });
+const finalTrip = result.report.trips.at(-1);
+
+history = [
+  ...(result.report.messages ?? []).filter((message) => message.role !== "system"),
+  { role: "assistant", content: finalTrip?.output ?? "", blocks: finalTrip?.blocks },
+];
+```
+
+## Provider options (`providerOptions.anthropic`)
+
+Anthropic-only request options, per call or in the agent's `modelOptions`. Other adapters ignore the key. Type: `AnthropicCallOptions`.
+
+| Field             | Effect                                                                                                                                         |
+| ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
+| `betas`           | Beta flags. When non-empty the request goes through `client.beta.messages` (needed for beta body fields and beta blocks such as `compaction`). |
+| `body`            | Extra body fields merged **last** (caller wins): `context_management`, `fallbacks`, `inference_geo`, `tool_choice`, … Not validated, except that `model`, `messages`, `system`, `tools`, `mcp_servers` and `stream` are refused (`InvalidRequestError`). |
+| `thinkingDisplay` | `thinking.display` on adaptive models: `"summarized"`, `"omitted"`, `"updates"` (the last needs its beta).                                     |
+
+```ts
+const agent = ai.agent({
+  model: anthropic.model({ name: "claude-opus-5-5" }),
+  modelOptions: {
+    providerOptions: {
+      anthropic: {
+        betas: ["compact-2026-01-12"],
+        body: { context_management: { edits: [{ type: "compact_20260112" }] } },
+      },
+    },
+  },
+});
+```
+
+Cache TTL per call: `options.cacheControl.ttl` (`"5m"` | `"1h"`) overrides the model's `promptCacheTtl` for that call.
 
 ## Prompt caching (`cache_control`)
 
@@ -190,11 +246,14 @@ await agent.execute("…", { cacheControl: { breakpoints: 1 } }); // cache the s
 
 - `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; `cachedTokens` / `cacheWriteTokens` from `message_start` then the cumulative `message_delta` counts (see [Tokens & usage accounting](#tokens--usage-accounting))
+- every content block is also assembled by index (text, thinking + signature, server-tool input, beta `compaction_delta`) and returned as `done.blocks` when the turn has a non-neutral block
+- terminal `{ type: "done", finishReason, usage, blocks? }` — `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
 
-`end_turn` / `stop_sequence` → `stop` · `max_tokens` → `length` · `tool_use` → `tool_calls` · `refusal` / `pause_turn` / unknown / null → `error`.
+`end_turn` / `stop_sequence` / beta `compaction` → `stop` · `max_tokens` → `length` · `model_context_window_exceeded` → `context_exceeded` · `pause_turn` → `pause_turn` · `refusal` → `refusal` · `tool_use` → `tool_calls` · unknown / null → `error`.
+
+The agent resumes `pause_turn` automatically (it replays the paused turn and calls again, within `maxTrips`). `refusal` and `context_exceeded` end the run like `stop`; check `report.trips.at(-1).finishReason` to fall back to another model or compact the history.
 
 ## Errors
 
@@ -211,11 +270,16 @@ Raw SDK errors are wrapped into the typed `@warlock.js/ai` `AIError` hierarchy.
 
 ```ts
 await anthropic.count("some text"); // approximate heuristic, offline
+
+const model = anthropic.model({ name: "claude-opus-5-5" }); // an AnthropicModel
+await model.countTokens(messages, { tools }); // exact, via messages.countTokens (network)
 ```
 
+`countTokens` sends the same system prompt, tools, thinking and `providerOptions.anthropic.betas` a real call would, keeping only the fields the endpoint accepts.
+
 ## No embeddings
 
-Anthropic ships no first-party embeddings API, so `AnthropicSDK` intentionally does **not** implement `embedder()` (the `SDKAdapterContract.embedder` slot is optional). Use [`@warlock.js/ai-openai`](@warlock.js/ai-openai/setup-openai/SKILL.md) (or another embeddings-capable adapter) for vectors.
+Anthropic ships no first-party embeddings API, so `AnthropicSDK` intentionally does **not** implement `embedder()` (the `SDKAdapterContract.embedder` slot is optional). Use `@warlock.js/ai-openai` (or another embeddings-capable adapter; see the `setup-openai` topic of the `warlock-js-ai-openai` skill) for vectors.
 
 ## When NOT to use this skill
 
@@ -224,8 +288,8 @@ Anthropic ships no first-party embeddings API, so `AnthropicSDK` intentionally d
 
 ## See also
 
-- [`@warlock.js/ai/run-ai-agent/SKILL.md`](@warlock.js/ai/run-ai-agent/SKILL.md)
-- [`@warlock.js/ai/pick-ai-provider/SKILL.md`](@warlock.js/ai/pick-ai-provider/SKILL.md)
-- [`@warlock.js/ai/handle-ai-errors/SKILL.md`](@warlock.js/ai/handle-ai-errors/SKILL.md)
+- The `run-ai-agent` topic of the `warlock-js-ai` skill
+- The `pick-ai-provider` topic of the `warlock-js-ai` skill
+- The `handle-ai-errors` topic of the `warlock-js-ai` skill
 
 

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?, reasoning?, thinkingMode?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → adaptive thinking + output_config.effort (current/unknown models) or budget_tokens (Haiku 4.5, 4.6 and older; override with thinkingMode), 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`, `thinkingMode`, `adaptive`, `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`.
+- [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?, thinkingMode?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → adaptive thinking + output_config.effort (current/unknown models) or budget_tokens (Haiku 4.5, 4.6 and older; override with thinkingMode), prompt caching via promptCaching + options.cacheControl (per-call ttl), provider blocks (thinking/compaction) replayed verbatim via ModelResponse.blocks / Message.blocks, providerOptions.anthropic (betas, body fields such as context_management / fallbacks, thinkingDisplay), model.countTokens(), distinct finish reasons (pause_turn, refusal, context_exceeded), cost-truth usage (cachedTokens/cacheWriteTokens), no first-party embeddings. Triggers: `AnthropicSDK`, `anthropic.model`, `anthropic.count`, `maxTokens`, `reasoning`, `thinking`, `budget_tokens`, `thinkingMode`, `adaptive`, `cacheControl`, `providerOptions`, `blocks`, `countTokens`, `pause_turn`, `refusal`, `context_management`, `compaction`, `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 — the `setup-openai` topic of the `warlock-js-ai-openai` skill; 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

@@ -11,16 +11,17 @@
   "license": "MIT",
   "repository": {
     "type": "git",
-    "url": "https://github.com/warlockjs/ai-anthropic"
+    "url": "git+https://github.com/warlockjs/framework.git",
+    "directory": "packages/ai-anthropic"
   },
   "peerDependencies": {
-    "@warlock.js/ai": "5.28.0"
+    "@warlock.js/ai": "5.30.0"
   },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.129.0",
-    "@warlock.js/logger": "5.28.0"
+    "@warlock.js/logger": "5.30.0"
   },
-  "version": "5.28.0",
+  "version": "5.30.0",
   "main": "./cjs/index.cjs",
   "module": "./esm/index.mjs",
   "types": "./esm/index.d.mts",

package/skills/README.md

@@ -1,9 +1,9 @@
 # `@warlock.js/ai-anthropic` — skills index
 
-Per-task skills. All cross-references use the form `@warlock.js/<pkg>/<skill>/SKILL.md`.
+Per-task skills. Cross-references name the topic ("the `<topic>` topic"), and for another package add its skill ("the `<topic>` topic of the `warlock-js-<pkg>` skill").
 
 ## Skills
 
-### [`setup-anthropic/`](./setup-anthropic/SKILL.md)
+### `setup-anthropic`
 
 Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, reasoning?, thinkingMode?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → adaptive thinking + output_config.effort (current/unknown models) or budget_tokens (Haiku 4.5, 4.6 and older; override with thinkingMode), 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?, reasoning?, thinkingMode?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → adaptive thinking + output_config.effort (current/unknown models) or budget_tokens (Haiku 4.5, 4.6 and older; override with thinkingMode), 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`, `thinkingMode`, `adaptive`, `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`.'
+description: 'Wire @warlock.js/ai-anthropic — new AnthropicSDK({apiKey, baseURL?, provider?}) for Claude, .model({name, vision?, structuredOutput?, reasoning?, thinkingMode?, promptCaching?, maxTokens?}). System-prompt hoisting, max_tokens required (default 4096), extended thinking via options.reasoning → adaptive thinking + output_config.effort (current/unknown models) or budget_tokens (Haiku 4.5, 4.6 and older; override with thinkingMode), prompt caching via promptCaching + options.cacheControl (per-call ttl), provider blocks (thinking/compaction) replayed verbatim via ModelResponse.blocks / Message.blocks, providerOptions.anthropic (betas, body fields such as context_management / fallbacks, thinkingDisplay), model.countTokens(), distinct finish reasons (pause_turn, refusal, context_exceeded), cost-truth usage (cachedTokens/cacheWriteTokens), no first-party embeddings. Triggers: `AnthropicSDK`, `anthropic.model`, `anthropic.count`, `maxTokens`, `reasoning`, `thinking`, `budget_tokens`, `thinkingMode`, `adaptive`, `cacheControl`, `providerOptions`, `blocks`, `countTokens`, `pause_turn`, `refusal`, `context_management`, `compaction`, `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 — the `setup-openai` topic of the `warlock-js-ai-openai` skill; 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`
@@ -93,7 +93,7 @@ Anthropic has no `"system"` role inside `messages`. The adapter hoists every neu
 
 - Outgoing: vendor-neutral tools → Anthropic `tools` with `input_schema` (non-object schemas degrade to a parameterless object so registration never fails).
 - Assistant tool calls in history → `assistant` message with an optional leading `text` block + one `tool_use` block per call.
-- Tool results (`role: "tool"`) → a `user` turn carrying a single `tool_result` block keyed by `tool_use_id`.
+- Tool results (`role: "tool"`) → `tool_result` blocks keyed by `tool_use_id`. Consecutive tool messages (the results of one parallel tool-use turn) are grouped into **one** `user` turn, as Anthropic expects.
 - Response `tool_use` blocks → neutral `toolCalls`; `stop_reason: "tool_use"` → `finishReason: "tool_calls"`.
 
 ## Structured output
@@ -113,7 +113,7 @@ Adaptive models reject `budget_tokens` with a 400, so the shape matters.
 
 - **Adaptive:** `reasoning.effort` (`"low" | "medium" | "high"`) → `output_config.effort` of the same name. With only `reasoning.maxTokens` set, effort is `"medium"`; `maxTokens` then only sizes the reserved `max_tokens` room, it is not sent as a budget.
 - **Budget:** `reasoning.maxTokens` → `budget_tokens` verbatim; `reasoning.effort` → a tiered budget (`1024` / `4096` / `12000`) when no `maxTokens` is given. Any resolved budget is floored at Anthropic's `1024` minimum.
-- `reasoning` with neither `effort` nor `maxTokens`, or `effort: "none"`, emits nothing.
+- `reasoning` with neither `effort` nor `maxTokens`, or `effort: "none"`, emits nothing — except `effort: "none"` on Claude Sonnet 5.5, which sends `thinking: { type: "between_tools" }` (see below).
 
 ### `thinkingMode` override
 
@@ -144,7 +144,63 @@ await model.complete(messages, { reasoning: { maxTokens: 8000 } });
 // adaptive model: effort "medium"; budget model: budget_tokens 8000
 ```
 
-When a thinking-enabled response also makes tool calls, its signed `thinking` and `redacted_thinking` blocks are retained in each call's `providerMetadata.anthropic.thinkingBlocks`. The adapter replays those opaque blocks automatically on the next assistant turn, preserving Anthropic's required signatures; application code should leave that metadata intact when persisting message history.
+**Claude 5 family (Opus 5.5, Sonnet 5.5, Fable 5 / 5.1):** thinking is always on. These models reject `thinking: { type: "disabled" }` and `budget_tokens` with a 400, so the adapter never sends either: they are adaptive and `reasoning.maxTokens` only sizes `max_tokens`. `effort: "none"` is mapped per model:
+
+| Model                   | `effort: "none"` sends                                                                 | Result                                                                                                                                 |
+| ----------------------- | -------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- |
+| Sonnet 5.5              | `thinking: { type: "between_tools" }`, alone (no `output_config.effort`, no `display`) | Thinking off. `between_tools` takes no other field and needs effort `high` or below; the model default, `high`, applies.               |
+| Opus 5.5, Fable 5 / 5.1 | no `thinking` field                                                                    | No thinking-off form exists, so the model still thinks at its default effort (Opus 5.5: `medium`). Pass `effort: "low"` to think less. |
+
+`thinkingDisplay` is applied only to adaptive thinking, never beside `between_tools`.
+
+`thinking.display` is set with `providerOptions.anthropic.thinkingDisplay` (`"summarized"` for readable summaries; the 5-family default is `"omitted"`, empty thinking text). It is applied whenever the adapter sends `thinking` to an adaptive model, so pair it with a `reasoning` option.
+
+When a thinking-enabled response also makes tool calls, its signed `thinking` and `redacted_thinking` blocks are retained in each call's `providerMetadata.anthropic.thinkingBlocks` and replayed on the next assistant turn. The full turn is also returned as `ModelResponse.blocks` — see [Provider blocks](#provider-blocks-thinking-compaction-server-tools).
+
+## Provider blocks (thinking, compaction, server tools)
+
+When a response contains a block the neutral `content` / `toolCalls` cannot carry (thinking, redacted thinking, `compaction`, server-tool use/results, cited text), the adapter returns the whole ordered turn as `ModelResponse.blocks` (and on the streaming `done` chunk): `{ provider: "anthropic", type, data }` per block. Plain text / tool-use turns have no `blocks`.
+
+The agent copies them onto the assistant `Message.blocks` (tool-call trips, `pause_turn` trips, repair turns) and onto `LLMTrip.blocks`; `AgentResult.blocks` is set only for a single-trip run. On the next request the adapter replays an assistant message **verbatim** from its `blocks` — required for preserved thinking and server-side compaction. It falls back to rebuilding from `content` + `toolCalls` when the blocks come from another provider, a block's `data.type` differs from its `type` or is not an assistant-turn type, or their `tool_use` ids no longer match `toolCalls`.
+
+A replayed thinking block requires every earlier turn unchanged, including the run's tool-call turns and tool results. Never store just `{ content: result.text, blocks }` after a multi-trip run. Keep the run's turns:
+
+```ts
+// agent created with `captureMessages: true`
+const result = await agent.execute(input, { history });
+const finalTrip = result.report.trips.at(-1);
+
+history = [
+  ...(result.report.messages ?? []).filter((message) => message.role !== "system"),
+  { role: "assistant", content: finalTrip?.output ?? "", blocks: finalTrip?.blocks },
+];
+```
+
+## Provider options (`providerOptions.anthropic`)
+
+Anthropic-only request options, per call or in the agent's `modelOptions`. Other adapters ignore the key. Type: `AnthropicCallOptions`.
+
+| Field             | Effect                                                                                                                                                                                                                                                   |
+| ----------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `betas`           | Beta flags. When non-empty the request goes through `client.beta.messages` (needed for beta body fields and beta blocks such as `compaction`).                                                                                                           |
+| `body`            | Extra body fields merged **last** (caller wins): `context_management`, `fallbacks`, `inference_geo`, `tool_choice`, … Not validated, except that `model`, `messages`, `system`, `tools`, `mcp_servers` and `stream` are refused (`InvalidRequestError`). |
+| `thinkingDisplay` | `thinking.display` on adaptive models: `"summarized"`, `"omitted"`, `"updates"` (the last needs its beta).                                                                                                                                               |
+
+```ts
+const agent = ai.agent({
+  model: anthropic.model({ name: "claude-opus-5-5" }),
+  modelOptions: {
+    providerOptions: {
+      anthropic: {
+        betas: ["compact-2026-01-12"],
+        body: { context_management: { edits: [{ type: "compact_20260112" }] } },
+      },
+    },
+  },
+});
+```
+
+Cache TTL per call: `options.cacheControl.ttl` (`"5m"` | `"1h"`) overrides the model's `promptCacheTtl` for that call.
 
 ## Prompt caching (`cache_control`)
 
@@ -182,11 +238,14 @@ await agent.execute("…", { cacheControl: { breakpoints: 1 } }); // cache the s
 
 - `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; `cachedTokens` / `cacheWriteTokens` from `message_start` then the cumulative `message_delta` counts (see [Tokens & usage accounting](#tokens--usage-accounting))
+- every content block is also assembled by index (text, thinking + signature, server-tool input, beta `compaction_delta`) and returned as `done.blocks` when the turn has a non-neutral block
+- terminal `{ type: "done", finishReason, usage, blocks? }` — `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
 
-`end_turn` / `stop_sequence` → `stop` · `max_tokens` → `length` · `tool_use` → `tool_calls` · `refusal` / `pause_turn` / unknown / null → `error`.
+`end_turn` / `stop_sequence` / beta `compaction` → `stop` · `max_tokens` → `length` · `model_context_window_exceeded` → `context_exceeded` · `pause_turn` → `pause_turn` · `refusal` → `refusal` · `tool_use` → `tool_calls` · unknown / null → `error`.
+
+The agent resumes `pause_turn` automatically (it replays the paused turn and calls again, within `maxTrips`). `refusal` and `context_exceeded` end the run like `stop`; check `report.trips.at(-1).finishReason` to fall back to another model or compact the history.
 
 ## Errors
 
@@ -203,11 +262,16 @@ Raw SDK errors are wrapped into the typed `@warlock.js/ai` `AIError` hierarchy.
 
 ```ts
 await anthropic.count("some text"); // approximate heuristic, offline
+
+const model = anthropic.model({ name: "claude-opus-5-5" }); // an AnthropicModel
+await model.countTokens(messages, { tools }); // exact, via messages.countTokens (network)
 ```
 
+`countTokens` sends the same system prompt, tools, thinking and `providerOptions.anthropic.betas` a real call would, keeping only the fields the endpoint accepts.
+
 ## No embeddings
 
-Anthropic ships no first-party embeddings API, so `AnthropicSDK` intentionally does **not** implement `embedder()` (the `SDKAdapterContract.embedder` slot is optional). Use [`@warlock.js/ai-openai`](@warlock.js/ai-openai/setup-openai/SKILL.md) (or another embeddings-capable adapter) for vectors.
+Anthropic ships no first-party embeddings API, so `AnthropicSDK` intentionally does **not** implement `embedder()` (the `SDKAdapterContract.embedder` slot is optional). Use `@warlock.js/ai-openai` (or another embeddings-capable adapter; see the `setup-openai` topic of the `warlock-js-ai-openai` skill) for vectors.
 
 ## When NOT to use this skill
 
@@ -216,6 +280,6 @@ Anthropic ships no first-party embeddings API, so `AnthropicSDK` intentionally d
 
 ## See also
 
-- [`@warlock.js/ai/run-ai-agent/SKILL.md`](@warlock.js/ai/run-ai-agent/SKILL.md)
-- [`@warlock.js/ai/pick-ai-provider/SKILL.md`](@warlock.js/ai/pick-ai-provider/SKILL.md)
-- [`@warlock.js/ai/handle-ai-errors/SKILL.md`](@warlock.js/ai/handle-ai-errors/SKILL.md)
+- The `run-ai-agent` topic of the `warlock-js-ai` skill
+- The `pick-ai-provider` topic of the `warlock-js-ai` skill
+- The `handle-ai-errors` topic of the `warlock-js-ai` skill