@webskill/sdk 0.9.0 → 0.10.0

package/dist/{dist-CFmkV45C.js → dist-BViUeszk.js}

@@ -103,11 +103,11 @@ const SCHEMA_KEYS = /* @__PURE__ */ new Set([
 	"then",
 	"else"
 ]);
-function isRecord$6(value) {
+function isRecord$7(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function isSelfRef(value) {
-	return isRecord$6(value) && value["$ref"] === SELF_REF;
+	return isRecord$7(value) && value["$ref"] === SELF_REF;
 }
 /**
 * 按 JSON Schema 关键字下降到子 schema。
@@ -117,7 +117,7 @@ function isSelfRef(value) {
 function forEachSubSchema(node, visit) {
 	for (const [key, value] of Object.entries(node)) {
 		if (SCHEMA_MAP_KEYS.has(key)) {
-			if (isRecord$6(value)) for (const name of Object.keys(value)) visit(value[name], [key, name]);
+			if (isRecord$7(value)) for (const name of Object.keys(value)) visit(value[name], [key, name]);
 			continue;
 		}
 		if (SCHEMA_LIST_KEYS.has(key)) {
@@ -136,7 +136,7 @@ function pathKey(path) {
 * `#` 解析到最近的带 `$id` 的祖先；没有则是文档根。
 */
 function collectSelfRefRoots(node, path, resourceRoot, out) {
-	if (!isRecord$6(node)) return;
+	if (!isRecord$7(node)) return;
 	if (isSelfRef(node)) {
 		out.set(pathKey(resourceRoot), resourceRoot);
 		return;
@@ -149,7 +149,7 @@ function collectSelfRefRoots(node, path, resourceRoot, out) {
 function getAt(schema, path) {
 	let current = schema;
 	for (const segment of path) if (Array.isArray(current) && typeof segment === "number") current = current[segment];
-	else if (isRecord$6(current) && typeof segment === "string") current = current[segment];
+	else if (isRecord$7(current) && typeof segment === "string") current = current[segment];
 	else return void 0;
 	return current;
 }
@@ -175,7 +175,7 @@ function setAt(schema, path, replacement) {
 * 正是严格解析器无法解析 `#` 的直接原因。
 */
 function rewriteSelfRefs(node, target, depth) {
-	if (!isRecord$6(node)) return node;
+	if (!isRecord$7(node)) return node;
 	if (depth > 8) throw new WebSkillError("TOOL_SCHEMA_UNAVAILABLE", `Tool schema nesting exceeds 8 levels; cannot produce a portable form`);
 	if (isSelfRef(node)) {
 		const rest = {};
@@ -188,7 +188,7 @@ function rewriteSelfRefs(node, target, depth) {
 	const result = {};
 	for (const [key, value] of Object.entries(node)) {
 		if (key === "$id" || key === "$schema") continue;
-		if (SCHEMA_MAP_KEYS.has(key) && isRecord$6(value)) {
+		if (SCHEMA_MAP_KEYS.has(key) && isRecord$7(value)) {
 			const mapped = {};
 			for (const [name, child] of Object.entries(value)) mapped[name] = rewriteSelfRefs(child, target, depth + 1);
 			result[key] = mapped;
@@ -226,8 +226,8 @@ function toPortableToolSchema(schema) {
 	if (roots.size > 1) throw new WebSkillError("TOOL_SCHEMA_UNAVAILABLE", "Tool schema contains self references in more than one schema resource; cannot produce a portable form");
 	const rootPath = [...roots.values()][0] ?? [];
 	const resource = getAt(schema, rootPath);
-	if (!isRecord$6(resource)) return schema;
-	const existingDefs = isRecord$6(schema[DEFS_KEY]) ? schema[DEFS_KEY] : void 0;
+	if (!isRecord$7(resource)) return schema;
+	const existingDefs = isRecord$7(schema[DEFS_KEY]) ? schema[DEFS_KEY] : void 0;
 	const taken = new Set(Object.keys(existingDefs ?? {}));
 	const nodeName = uniqueName(NODE_NAME, taken);
 	taken.add(nodeName);
@@ -258,6 +258,63 @@ function toPortableToolSchema(schema) {
 		}
 	};
 }
+function isRecord$6(value) {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+/**
+* 与 `interaction/schemaToForm` 的 file 判定保持一致：JSON Schema 里二进制内容的
+* 两种标准写法（`format: binary` / `contentEncoding: base64`）。
+*/
+function isModelUnfillable(node) {
+	return isRecord$6(node) && node["type"] === "string" && (node["format"] === "binary" || node["contentEncoding"] === "base64");
+}
+function stripNode(node) {
+	if (Array.isArray(node)) return node.map(stripNode);
+	if (!isRecord$6(node)) return node;
+	const out = {};
+	for (const [key, value] of Object.entries(node)) {
+		if (SCHEMA_MAP_KEYS.has(key) && isRecord$6(value)) {
+			const mapped = {};
+			for (const [name, child] of Object.entries(value)) {
+				if (key === "properties" && isModelUnfillable(child)) continue;
+				mapped[name] = stripNode(child);
+			}
+			out[key] = mapped;
+			continue;
+		}
+		if (SCHEMA_LIST_KEYS.has(key) && Array.isArray(value)) {
+			out[key] = value.map(stripNode);
+			continue;
+		}
+		if (SCHEMA_KEYS.has(key)) {
+			out[key] = stripNode(value);
+			continue;
+		}
+		out[key] = value;
+	}
+	const required = out["required"];
+	const properties = out["properties"];
+	if (Array.isArray(required) && isRecord$6(properties)) out["required"] = required.filter((name) => typeof name !== "string" || name in properties);
+	return out;
+}
+/**
+* 把模型不可能自己产出的字段（本地文件二进制）从**出站**工具 schema 里剥离。
+*
+* 模型拿不到用户磁盘上的文件，留着这个字段只有两种下场：模型拒绝调用工具
+* （实测 DeepSeek 直接回文本「请上传文件」，file-pick 永远触发不了），或编造一个
+* 占位字符串把技能跑挂。剥掉后模型以缺参形式调用，引擎的 missing-params 链路
+* 会弹 file-pick 让用户用系统文件选择器补这个字段。
+*
+* 这是**传输层适配**，只应在出站请求体上调用；不要下沉到 AgentLoop——
+* 引擎侧的 missing 校验看的必须是技能作者写的完整 inputSchema。
+*
+* 已知边界（0.10.0 复核注记）：只剥 `properties` 里的 inline 二进制字段；
+* 通过 `$ref` 指向 `$defs` 中二进制定义的字段仍会下发（剥掉会留悬空 $ref），
+* 这类技能依旧可能拿不到 file-pick——需要完整修复时再补「剥 $def + 剥指向它的 $ref」。
+*/
+function stripModelUnfillableParams(schema) {
+	return stripNode(schema);
+}
 /** 各家都认的基础关键字。Google `FunctionDeclaration.parameters` 是其中最窄的一家。 */
 const BASE_KEYWORDS = /* @__PURE__ */ new Set([
 	"type",
@@ -499,7 +556,7 @@ const toOpenAiTools = (tools) => tools.map((tool) => ({
 	function: {
 		name: tool.name,
 		...tool.description ? { description: tool.description } : {},
-		parameters: sanitizeForVendor(toPortableToolSchema(tool.inputSchema), OPENAI_SCHEMA_PROFILE, tool.name)
+		parameters: sanitizeForVendor(toPortableToolSchema(stripModelUnfillableParams(tool.inputSchema)), OPENAI_SCHEMA_PROFILE, tool.name)
 	}
 }));
 const errorMessage$2 = (e) => e instanceof Error ? e.message : String(e);
@@ -539,6 +596,7 @@ var OpenAiCompatibleClient = class {
 		const reader = res.body.getReader();
 		const frames = createSseFrameReader();
 		let done = false;
+		let usage;
 		const handleFrame = function* (data) {
 			if (data === "[DONE]") {
 				done = true;
@@ -550,8 +608,17 @@ var OpenAiCompatibleClient = class {
 			} catch (e) {
 				throw new WebSkillError("LLM_REQUEST_FAILED", `Failed to parse LLM stream frame: ${errorMessage$2(e)}`, data);
 			}
+			const usageRaw = chunk.usage;
+			if (usageRaw && (typeof usageRaw["prompt_tokens"] === "number" || typeof usageRaw["completion_tokens"] === "number")) usage = {
+				inputTokens: typeof usageRaw["prompt_tokens"] === "number" ? usageRaw["prompt_tokens"] : 0,
+				outputTokens: typeof usageRaw["completion_tokens"] === "number" ? usageRaw["completion_tokens"] : 0
+			};
 			const delta = chunk.choices?.[0]?.delta;
 			if (!delta) return;
+			if (typeof delta["reasoning_content"] === "string" && delta["reasoning_content"] !== "") yield {
+				type: "thinking-delta",
+				delta: delta["reasoning_content"]
+			};
 			if (typeof delta["content"] === "string" && delta["content"] !== "") yield {
 				type: "text-delta",
 				delta: delta["content"]
@@ -601,7 +668,10 @@ var OpenAiCompatibleClient = class {
 				};
 			})
 		};
-		yield { type: "done" };
+		yield {
+			type: "done",
+			...usage ? { usage } : {}
+		};
 	}
 	async #postChat(input, stream) {
 		const { baseUrl, apiKey, model } = this.#requireConfig();
@@ -615,7 +685,10 @@ var OpenAiCompatibleClient = class {
 			body["tool_choice"] = "auto";
 		}
 		if (input.temperature !== void 0) body["temperature"] = input.temperature;
-		if (stream) body["stream"] = true;
+		if (stream) {
+			body["stream"] = true;
+			body["stream_options"] = { include_usage: true };
+		}
 		let res;
 		try {
 			res = await this.#fetch(`${baseUrl}/chat/completions`, {
@@ -665,9 +738,17 @@ var OpenAiCompatibleClient = class {
 				};
 			});
 			const content = choice["content"];
+			const thinking = choice["reasoning_content"];
+			const usageRaw = data.usage;
+			const usage = usageRaw && (typeof usageRaw["prompt_tokens"] === "number" || typeof usageRaw["completion_tokens"] === "number") ? {
+				inputTokens: typeof usageRaw["prompt_tokens"] === "number" ? usageRaw["prompt_tokens"] : 0,
+				outputTokens: typeof usageRaw["completion_tokens"] === "number" ? usageRaw["completion_tokens"] : 0
+			} : void 0;
 			return {
 				content: typeof content === "string" && content !== "" ? textParts(content) : void 0,
 				toolCalls: toolCalls?.length ? toolCalls : void 0,
+				...typeof thinking === "string" && thinking !== "" ? { thinking } : {},
+				...usage ? { usage } : {},
 				raw: data
 			};
 		} catch (e) {
@@ -756,7 +837,7 @@ function toAnthropicMessages(messages) {
 const toAnthropicTools = (tools) => tools.map((tool) => ({
 	name: tool.name,
 	...tool.description ? { description: tool.description } : {},
-	input_schema: sanitizeForVendor(toPortableToolSchema(tool.inputSchema), ANTHROPIC_SCHEMA_PROFILE, tool.name)
+	input_schema: sanitizeForVendor(toPortableToolSchema(stripModelUnfillableParams(tool.inputSchema)), ANTHROPIC_SCHEMA_PROFILE, tool.name)
 }));
 /** Anthropic Messages API 客户端（零依赖 fetch；Node/浏览器通用） */
 var AnthropicClient = class {
@@ -795,6 +876,8 @@ var AnthropicClient = class {
 		const decoder = new TextDecoder();
 		const reader = res.body.getReader();
 		const frames = createSseFrameReader();
+		let inputTokens;
+		let outputTokens;
 		const handleFrame = function* (data) {
 			let chunk;
 			try {
@@ -804,6 +887,17 @@ var AnthropicClient = class {
 			}
 			const type = chunk["type"];
 			if (type === "error") throw new WebSkillError("LLM_REQUEST_FAILED", `LLM stream error: ${chunk["error"]?.message ?? data}`);
+			if (type === "message_start") {
+				const usage = chunk["message"]?.["usage"];
+				if (typeof usage?.["input_tokens"] === "number") inputTokens = usage["input_tokens"];
+				if (typeof usage?.["output_tokens"] === "number") outputTokens = usage["output_tokens"];
+				return;
+			}
+			if (type === "message_delta") {
+				const usage = chunk["usage"];
+				if (typeof usage?.["output_tokens"] === "number") outputTokens = usage["output_tokens"];
+				return;
+			}
 			if (type === "content_block_start") {
 				const block = chunk["content_block"];
 				if (block?.["type"] === "tool_use") {
@@ -826,6 +920,13 @@ var AnthropicClient = class {
 					};
 					return;
 				}
+				if (delta?.["type"] === "thinking_delta" && typeof delta["thinking"] === "string" && delta["thinking"] !== "") {
+					yield {
+						type: "thinking-delta",
+						delta: delta["thinking"]
+					};
+					return;
+				}
 				if (delta?.["type"] === "input_json_delta" && typeof delta["partial_json"] === "string") {
 					const acc = toolCallsByIndex.get(index) ?? {
 						id: "",
@@ -866,7 +967,14 @@ var AnthropicClient = class {
 				};
 			})
 		};
-		yield { type: "done" };
+		const usage = inputTokens !== void 0 || outputTokens !== void 0 ? {
+			inputTokens: inputTokens ?? 0,
+			outputTokens: outputTokens ?? 0
+		} : void 0;
+		yield {
+			type: "done",
+			...usage ? { usage } : {}
+		};
 	}
 	async #post(input, stream) {
 		validateLlmMessages(input.messages);
@@ -878,7 +986,12 @@ var AnthropicClient = class {
 		};
 		if (system !== void 0) body["system"] = system;
 		if (input.tools?.length) body["tools"] = toAnthropicTools(input.tools);
-		if (input.temperature !== void 0) body["temperature"] = input.temperature;
+		const thinkingBudget = this.#config.thinkingBudgetTokens;
+		if (thinkingBudget !== void 0 && Number.isInteger(thinkingBudget) && thinkingBudget > 0) body["thinking"] = {
+			type: "enabled",
+			budget_tokens: thinkingBudget
+		};
+		else if (input.temperature !== void 0) body["temperature"] = input.temperature;
 		if (stream) body["stream"] = true;
 		let res;
 		try {
@@ -926,14 +1039,22 @@ var AnthropicClient = class {
 			const blocks = data.content;
 			if (!Array.isArray(blocks)) throw new Error("response has no content blocks");
 			const text = blocks.filter((b) => b["type"] === "text").map((b) => String(b["text"] ?? "")).join("");
+			const thinking = blocks.filter((b) => b["type"] === "thinking").map((b) => String(b["thinking"] ?? "")).join("");
 			const toolCalls = blocks.filter((b) => b["type"] === "tool_use").map((b) => ({
 				id: typeof b["id"] === "string" ? b["id"] : "",
 				name: typeof b["name"] === "string" ? b["name"] : "",
 				arguments: typeof b["input"] === "object" && b["input"] !== null ? b["input"] : {}
 			}));
+			const usageRaw = data.usage;
+			const usage = typeof usageRaw?.["input_tokens"] === "number" || typeof usageRaw?.["output_tokens"] === "number" ? {
+				inputTokens: typeof usageRaw["input_tokens"] === "number" ? usageRaw["input_tokens"] : 0,
+				outputTokens: typeof usageRaw["output_tokens"] === "number" ? usageRaw["output_tokens"] : 0
+			} : void 0;
 			return {
 				content: text === "" ? void 0 : textParts(text),
 				toolCalls: toolCalls.length > 0 ? toolCalls : void 0,
+				...thinking === "" ? {} : { thinking },
+				...usage ? { usage } : {},
 				raw: data
 			};
 		} catch (e) {
@@ -942,6 +1063,18 @@ var AnthropicClient = class {
 	}
 };
 const errorMessage = (e) => e instanceof Error ? e.message : String(e);
+/** usageMetadata → LlmTokenUsage（promptTokenCount / candidatesTokenCount；无字段返回 undefined） */
+function parseGenAiUsage(data) {
+	const usage = data?.usageMetadata;
+	if (!usage) return void 0;
+	const input = usage["promptTokenCount"];
+	const output = usage["candidatesTokenCount"];
+	if (typeof input !== "number" && typeof output !== "number") return void 0;
+	return {
+		inputTokens: typeof input === "number" ? input : 0,
+		outputTokens: typeof output === "number" ? output : 0
+	};
+}
 /** parts → GenAI parts（二进制统一走 inlineData） */
 const toGenAiParts = (parts, where) => parts.map((part) => {
 	switch (part.type) {
@@ -1015,7 +1148,7 @@ function toGenAiContents(messages) {
 const toGenAiTools = (tools) => [{ functionDeclarations: tools.map((tool) => ({
 	name: tool.name,
 	...tool.description ? { description: tool.description } : {},
-	parameters: sanitizeForVendor(toPortableToolSchema(tool.inputSchema), GOOGLE_SCHEMA_PROFILE, tool.name)
+	parameters: sanitizeForVendor(toPortableToolSchema(stripModelUnfillableParams(tool.inputSchema)), GOOGLE_SCHEMA_PROFILE, tool.name)
 })) }];
 /** Google GenAI（generateContent / streamGenerateContent）客户端（零依赖 fetch） */
 var GoogleGenAiClient = class {
@@ -1039,9 +1172,11 @@ var GoogleGenAiClient = class {
 		const parts = this.#responseParts(data);
 		const text = parts.filter((p) => typeof p["text"] === "string").map((p) => String(p["text"])).join("");
 		const toolCalls = this.#functionCalls(parts);
+		const usage = parseGenAiUsage(data);
 		return {
 			content: text === "" ? void 0 : textParts(text),
 			toolCalls: toolCalls.length > 0 ? toolCalls : void 0,
+			...usage ? { usage } : {},
 			raw: data
 		};
 	}
@@ -1053,6 +1188,7 @@ var GoogleGenAiClient = class {
 		const decoder = new TextDecoder();
 		const reader = res.body.getReader();
 		const frames = createSseFrameReader();
+		let usage;
 		const handleFrame = function* (data) {
 			let chunk;
 			try {
@@ -1060,6 +1196,7 @@ var GoogleGenAiClient = class {
 			} catch (e) {
 				throw new WebSkillError("LLM_REQUEST_FAILED", `Failed to parse LLM stream frame: ${errorMessage(e)}`, data);
 			}
+			usage = parseGenAiUsage(chunk) ?? usage;
 			const parts = chunk.candidates?.[0]?.content?.parts;
 			for (const part of parts ?? []) {
 				if (typeof part["text"] === "string" && part["text"] !== "") {
@@ -1091,7 +1228,10 @@ var GoogleGenAiClient = class {
 			type: "tool-calls",
 			toolCalls
 		};
-		yield { type: "done" };
+		yield {
+			type: "done",
+			...usage ? { usage } : {}
+		};
 	}
 	async #post(input, stream) {
 		const model = input.model ?? this.#config.model;
@@ -1152,6 +1292,17 @@ var GoogleGenAiClient = class {
 		return out;
 	}
 };
+/** v5 { inputTokens, outputTokens } 与 v4 { promptTokens, completionTokens } 两种形状都接 */
+function fromVercelUsage(raw) {
+	const u = raw ?? {};
+	const input = u["inputTokens"] ?? u["promptTokens"];
+	const output = u["outputTokens"] ?? u["completionTokens"];
+	if (typeof input !== "number" && typeof output !== "number") return void 0;
+	return {
+		inputTokens: typeof input === "number" ? input : 0,
+		outputTokens: typeof output === "number" ? output : 0
+	};
+}
 function toVercelToolSpecs(tools) {
 	const out = {};
 	for (const tool of tools) out[tool.name] = {
@@ -1160,6 +1311,16 @@ function toVercelToolSpecs(tools) {
 	};
 	return out;
 }
+/**
+* 归一化供应商思考内容：v5 的 `reasoningText`（字符串）与 v4 的 `reasoning`
+* （Array<{ text; type }>）都能落到同一条 `LlmResponse.thinking`。
+*/
+function normalizeThinking(value) {
+	if (typeof value === "string") return value !== "" ? value : void 0;
+	if (!Array.isArray(value)) return void 0;
+	const text = value.map((part) => typeof part?.text === "string" ? part.text : "").join("");
+	return text !== "" ? text : void 0;
+}
 /** 接受 generateText 的返回值形状：{ text?, toolCalls?: [{ toolCallId, toolName, input|args }] } */
 function fromVercelResult(result) {
 	const r = result;
@@ -1171,9 +1332,13 @@ function fromVercelResult(result) {
 			arguments: typeof args === "object" && args !== null ? args : {}
 		};
 	});
+	const thinking = normalizeThinking(r.reasoningText ?? r.reasoning);
+	const usage = fromVercelUsage(r.usage);
 	return {
 		content: typeof r.text === "string" && r.text !== "" ? textParts(r.text) : void 0,
 		toolCalls: toolCalls.length ? toolCalls : void 0,
+		...thinking !== void 0 ? { thinking } : {},
+		...usage ? { usage } : {},
 		raw: result
 	};
 }
@@ -1191,6 +1356,13 @@ function fromVercelStreamPart(part) {
 				delta
 			} : void 0;
 		}
+		case "reasoning-delta": {
+			const delta = p["text"] ?? p["delta"];
+			return typeof delta === "string" && delta !== "" ? {
+				type: "thinking-delta",
+				delta
+			} : void 0;
+		}
 		case "tool-call": return {
 			type: "tool-calls",
 			toolCalls: [{
@@ -1199,7 +1371,13 @@ function fromVercelStreamPart(part) {
 				arguments: typeof p["input"] === "object" && p["input"] !== null ? p["input"] : {}
 			}]
 		};
-		case "finish": return { type: "done" };
+		case "finish": {
+			const usage = fromVercelUsage(p["totalUsage"] ?? p["usage"]);
+			return {
+				type: "done",
+				...usage ? { usage } : {}
+			};
+		}
 		default: return;
 	}
 }
@@ -1803,7 +1981,7 @@ const DEFAULT_USER_PROFILE_LIMITS = {
 	recordLimit: 500,
 	valueMaxChars: 2e3,
 	sampleLimit: 80,
-	injectMaxBytes: 2e3,
+	injectMaxBytes: 4e3,
 	entryLimit: 40
 };
 /** @experimental */
@@ -2053,8 +2231,8 @@ function extractSkillCandidate(data) {
 */
 const DEFAULT_LOOP_LIMITS = {
 	maxTurns: 10,
-	totalTimeoutMs: 12e4,
-	toolTimeoutMs: 3e4
+	totalTimeoutMs: 36e5,
+	toolTimeoutMs: 6e5
 };
 const RUN_SNAPSHOT_SCHEMA_VERSION = 3;
 /** @experimental */
@@ -2294,7 +2472,7 @@ var AgentLoop = class {
 			maxTurns: config.maxTurns ?? DEFAULT_LOOP_LIMITS.maxTurns,
 			totalTimeoutMs: config.totalTimeoutMs ?? DEFAULT_LOOP_LIMITS.totalTimeoutMs,
 			toolTimeoutMs: config.toolTimeoutMs ?? DEFAULT_LOOP_LIMITS.toolTimeoutMs,
-			toolResultMaxBytes: config.toolResultMaxBytes ?? 1e5,
+			toolResultMaxBytes: config.toolResultMaxBytes ?? 256e3,
 			paramHistoryLimit: config.paramHistoryLimit ?? 50,
 			toolCallingDisabled: config.toolCallingDisabled ?? false,
 			maxUnknownToolRetries: config.maxUnknownToolRetries ?? 2,
@@ -2357,6 +2535,7 @@ var AgentLoop = class {
 			surfacePatchCount: 0,
 			processedSurfaceActionNonces: /* @__PURE__ */ new Set(),
 			emittedToolEvents: /* @__PURE__ */ new Set(),
+			artifactPaths: /* @__PURE__ */ new Set(),
 			unknownToolCalls: 0,
 			startMs,
 			pausedMs: 0,
@@ -2507,6 +2686,23 @@ var AgentLoop = class {
 					const code = e instanceof WebSkillError && (e.code === "LLM_UNAVAILABLE" || e.code === "LLM_REQUEST_FAILED") ? e.code : "LLM_REQUEST_FAILED";
 					return finish("failed", "llm-error", messageOf(e), code);
 				}
+				if (response.thinking !== void 0 && response.thinking !== "") {
+					state.trace.record("llm.thinking", { data: {
+						turn,
+						text: response.thinking
+					} });
+					const bus = this.#deps.eventBus;
+					if (!llm.stream && bus && bus.listenerCount() > 0) bus.emit({
+						phase: "execute",
+						runId: state.runId,
+						sessionId: state.run.sessionId,
+						ts: state.now(),
+						data: {
+							kind: "thinking-delta",
+							delta: response.thinking
+						}
+					});
+				}
 				const responseText = partsToText(response.content);
 				if (!llm.stream && responseText !== "") await this.#deps.uiBridge?.onTextDelta?.(state.runId, responseText);
 				trace.record("llm.response", { data: {
@@ -2514,6 +2710,11 @@ var AgentLoop = class {
 					hasToolCalls: Boolean(response.toolCalls?.length),
 					contentLength: responseText.length
 				} });
+				if (response.usage) trace.record("llm.usage", { data: {
+					turn,
+					inputTokens: response.usage.inputTokens,
+					outputTokens: response.usage.outputTokens
+				} });
 				if (!response.toolCalls?.length) {
 					messages.push({
 						role: "assistant",
@@ -2723,6 +2924,7 @@ var AgentLoop = class {
 			surfacePatchCount: 0,
 			processedSurfaceActionNonces: new Set(snapshot.processedSurfaceActionNonces ?? []),
 			emittedToolEvents: /* @__PURE__ */ new Set(),
+			artifactPaths: /* @__PURE__ */ new Set(),
 			unknownToolCalls: 0,
 			startMs,
 			pausedMs: snapshot.pausedMs ?? 0,
@@ -2843,6 +3045,7 @@ var AgentLoop = class {
 	/**
 	* 流式 LLM 调用：text-delta 仅在 EventBus 有订阅者时发射 llm.delta 事件，
 	* 并同步转发 uiBridge.onTextDelta；tool-calls/done 组装为 LlmResponse 走原有分支。
+	* thinking-delta 同理累积成 `response.thinking`（每轮一段，UI-UX5 #32）。
 	*/
 	async #completeStreaming(llm, state, input) {
 		const emitDelta = (delta) => {
@@ -2859,18 +3062,41 @@ var AgentLoop = class {
 			});
 			this.#deps.uiBridge?.onTextDelta?.(state.runId, delta);
 		};
+		const emitThinkingDelta = (delta) => {
+			const bus = this.#deps.eventBus;
+			if (bus && bus.listenerCount() > 0) bus.emit({
+				phase: "execute",
+				runId: state.runId,
+				sessionId: state.run.sessionId,
+				ts: state.now(),
+				data: {
+					kind: "thinking-delta",
+					delta
+				}
+			});
+		};
 		let content = "";
+		let thinking = "";
 		let doneContent;
+		let usage;
 		const toolCalls = [];
 		for await (const event of llm.stream(input)) if (event.type === "text-delta") {
 			content += event.delta;
 			emitDelta(event.delta);
+		} else if (event.type === "thinking-delta") {
+			thinking += event.delta;
+			emitThinkingDelta(event.delta);
 		} else if (event.type === "tool-calls") toolCalls.push(...event.toolCalls);
-		else if (event.type === "done") doneContent = event.content;
+		else if (event.type === "done") {
+			doneContent = event.content;
+			usage = event.usage ?? usage;
+		}
 		const text = doneContent ?? content;
 		return {
 			content: text === "" ? void 0 : textParts(text),
-			toolCalls: toolCalls.length > 0 ? toolCalls : void 0
+			toolCalls: toolCalls.length > 0 ? toolCalls : void 0,
+			...thinking === "" ? {} : { thinking },
+			...usage ? { usage } : {}
 		};
 	}
 	/** 生命周期接线：更新 phase、发事件、跑钩子 */
@@ -3097,10 +3323,13 @@ var AgentLoop = class {
 			});
 			this.#emitTool(state, "failed", call, result.error?.code);
 		}
-		for (const artifact of result.artifacts ?? []) state.trace.record("artifact.created", { data: {
-			artifactId: artifact.id,
-			path: artifact.path
-		} });
+		for (const artifact of result.artifacts ?? []) {
+			state.artifactPaths.add(artifact.path);
+			state.trace.record("artifact.created", { data: {
+				artifactId: artifact.id,
+				path: artifact.path
+			} });
+		}
 		if (state.unknownToolCalls > this.#config.maxUnknownToolRetries) throw new RunTerminated({
 			status: "failed",
 			reason: "tool-resolution-exhausted",
@@ -3116,7 +3345,12 @@ var AgentLoop = class {
 			state.trace.record("run.warning", { message: "UiBridge does not support renderSurface; UI surface was not rendered" });
 			return;
 		}
-		const attributed = this.#attributeSurfaceEvent(state, event);
+		let attributed = this.#attributeSurfaceEvent(state, event);
+		if (attributed.type === "open") {
+			const kept = await this.#stripDuplicateFileLinks(state, attributed);
+			if (kept === void 0) return;
+			attributed = kept;
+		}
 		if (attributed.type === "patch") this.#consumeSurfacePatchBudget(state);
 		await bridge.renderSurface(attributed);
 		state.surfaceEvents.push(structuredClone(attributed));
@@ -3159,6 +3393,44 @@ var AgentLoop = class {
 			}))
 		};
 	}
+	/**
+	* 剥掉指向「本 run 已写出 artifact」的 FileLink 节点（UI-UX5 #5）：
+	* 宿主在 run 完成时会把每个 artifact 自动渲染成下载卡（`buildRenderResult`），
+	* 模型/技能再建一张就是重复。整棵树剥空返回 undefined（放弃这次 open）；
+	* 有剥离就记一条 run.warning，让「少了一张卡」在轨迹里可查。
+	*/
+	async #stripDuplicateFileLinks(state, event) {
+		if (state.artifactPaths.size === 0) {
+			const listed = await this.#deps.artifactStore.listArtifacts(state.runId).catch(() => []);
+			for (const artifact of listed) state.artifactPaths.add(artifact.path);
+		}
+		if (state.artifactPaths.size === 0) return event;
+		const paths = state.artifactPaths;
+		let stripped = 0;
+		const walk = (node) => {
+			if (node.component === "FileLink") {
+				const path = node.props?.["path"];
+				if (typeof path === "string" && paths.has(path)) {
+					stripped += 1;
+					return;
+				}
+			}
+			if (!node.children || node.children.length === 0) return node;
+			const children = node.children.map(walk).filter((child) => child !== void 0);
+			if (children.length === node.children.length) return node;
+			return {
+				...node,
+				children
+			};
+		};
+		const node = walk(event.node);
+		if (stripped === 0) return event;
+		state.trace.record("run.warning", { message: `Dropped ${stripped} FileLink node(s) duplicating auto-rendered artifact download card(s)` });
+		return node === void 0 ? void 0 : {
+			...event,
+			node
+		};
+	}
 	/** Awaits the single action emitted with the most recently persisted tool result. */
 	async #drainSurfaceAction(state) {
 		const request = state.pendingSurfaceAction;
@@ -3437,7 +3709,7 @@ var AgentLoop = class {
 		}))];
 	}
 	/**
-	* 工具结果回喂序列化：超过 toolResultMaxBytes（默认 100KB）时头尾保留截断，
+	* 工具结果回喂序列化：超过 toolResultMaxBytes（默认 256KB）时头尾保留截断，
 	* 完整内容经 artifactStore 落 artifact，回喂摘要含 artifact id。
 	*/
 	async #serializeToolResult(call, result, state) {
@@ -4930,6 +5202,38 @@ var CapabilityApproval = class CapabilityApproval {
 		return "allowed";
 	}
 };
+/**
+* 从 run 的 trace 聚合大模型用量（请求次数 + 输入/输出 token）。
+* 与 `summarizeToolCalls` 同型：消费方不再各自遍历 trace。
+* @stable
+*/
+function summarizeRunUsage(trace) {
+	let llmCalls = 0;
+	let inputTokens = 0;
+	let outputTokens = 0;
+	let hasUsage = false;
+	for (const event of trace) {
+		if (event.type === "llm.request") {
+			llmCalls += 1;
+			continue;
+		}
+		if (event.type !== "llm.usage") continue;
+		const input = event.data?.["inputTokens"];
+		const output = event.data?.["outputTokens"];
+		const hasInput = typeof input === "number" && Number.isFinite(input);
+		const hasOutput = typeof output === "number" && Number.isFinite(output);
+		if (hasInput) inputTokens += input;
+		if (hasOutput) outputTokens += output;
+		if (hasInput || hasOutput) hasUsage = true;
+	}
+	return {
+		llmCalls,
+		...hasUsage ? {
+			inputTokens,
+			outputTokens
+		} : {}
+	};
+}
 const RUN_TRACE_SCHEMA_VERSION = 1;
 /** 终止原因：取最后一条 run.completed/cancelled/failed 事件的 data.reason */
 function extractEndReason(events) {
@@ -4944,6 +5248,7 @@ function extractEndReason(events) {
 function summarize(trace) {
 	const endReason = extractEndReason(trace.events);
 	const durationMs = trace.endedAt === void 0 ? void 0 : Date.parse(trace.endedAt) - Date.parse(trace.startedAt);
+	const runUsage = summarizeRunUsage(trace.events);
 	return {
 		runId: trace.runId,
 		startedAt: trace.startedAt,
@@ -4954,7 +5259,11 @@ function summarize(trace) {
 		...trace.sessionId !== void 0 ? { sessionId: trace.sessionId } : {},
 		...trace.endedAt !== void 0 ? { endedAt: trace.endedAt } : {},
 		...endReason !== void 0 ? { endReason } : {},
-		...durationMs !== void 0 && Number.isFinite(durationMs) && durationMs >= 0 ? { durationMs } : {}
+		...durationMs !== void 0 && Number.isFinite(durationMs) && durationMs >= 0 ? { durationMs } : {},
+		...runUsage.inputTokens !== void 0 ? { usage: {
+			inputTokens: runUsage.inputTokens,
+			outputTokens: runUsage.outputTokens ?? 0
+		} } : {}
 	};
 }
 function parseTraceFile(raw, path) {
@@ -5147,6 +5456,12 @@ function applyFilter(summaries, filter) {
 * @stable
 */
 function summarizeToolCalls(run) {
+	const startedAtByCallId = /* @__PURE__ */ new Map();
+	for (const event of run.trace) {
+		if (event.type !== "tool.started") continue;
+		const callId = event.data?.["callId"];
+		if (typeof callId === "string" && !startedAtByCallId.has(callId)) startedAtByCallId.set(callId, event.ts);
+	}
 	const calls = [];
 	for (const event of run.trace) {
 		if (event.type !== "tool.completed" && event.type !== "tool.failed") continue;
@@ -5156,6 +5471,7 @@ function summarizeToolCalls(run) {
 		const args = event.data?.["args"];
 		const durationMs = event.data?.["durationMs"];
 		const errorCode = event.data?.["code"];
+		const startedAt = startedAtByCallId.get(callId);
 		calls.push({
 			callId,
 			name,
@@ -5163,6 +5479,7 @@ function summarizeToolCalls(run) {
 			...typeof args === "string" ? { args } : {},
 			...typeof durationMs === "number" ? { durationMs } : {},
 			...typeof errorCode === "string" ? { errorCode } : {},
+			...startedAt !== void 0 ? { startedAt } : {},
 			...event.type === "tool.failed" && typeof event.message === "string" ? { errorMessage: event.message } : {}
 		});
 	}
@@ -5386,4 +5703,4 @@ var FsSessionStore = class {
 };
 
 //#endregion
-export { fromVercelStreamPart as $, SerializingMemoryStore as A, bridgeError as B, ProgressiveRouter as C, sealToolCallPairs as Ct, RUN_SNAPSHOT_SCHEMA_VERSION as D, toVercelToolSpecs as Dt, READ_SKILL_FILE_TOOL_NAME as E, toRecordDigests as Et, USER_PROFILE_PROMPT_HEADER as F, exportUserProfile as G, createScriptContext as H, USER_PROFILE_REFINE_PROMPT as I, extractTodoTraceEvents as J, extractChartSpec as K, WebSkillRuntime as L, USER_PROFILE_EXPORT_VERSION as M, USER_PROFILE_KEY as N, RUN_TRACE_SCHEMA_VERSION as O, validateUiSpecEvent as Ot, USER_PROFILE_NO_INVENTION_RULE as P, fromVercelResult as Q, appendBehaviorRecords as R, OpenAiCompatibleClient as S, scriptToolName as St, READ_SKILL_FILE_TOOL as T, toLlmToolSpec as Tt, createWebSkillApi as U, buildRenderResult as V, diffUserProfile as W, findUnpairedToolCalls as X, extractUiSpecEvents as Y, formatSkillScriptManifest as Z, FsRunTraceStore as _, renderUserProfileContext as _t, AgentLoop as a, mergeProfileEntries as at, GoogleGenAiClient as b, schemaSourceLabel as bt, CapabilityApproval as c, normalizeErrorCode as ct, EMPTY_USER_PROFILE as d, parseBridgeRequest as dt, interruptedToolResult as et, EventBus as f, parseUserProfileExport as ft, FsRunSnapshotStore as g, refineUserProfile as gt, FsMemoryStore as h, readUserProfile as ht, ASK_USER_TOOL_NAME as i, mergeCatalogEntries as it, TraceRecorder as j, SESSION_SCHEMA_VERSION as k, validateUiSpecNode as kt, DEFAULT_LOOP_LIMITS as l, normalizeToolContent as lt, FsArtifactStore as m, readProfileEntries as mt, ASK_USER_INPUT_SCHEMA as n, isUnsupportedRunSnapshot as nt, AnthropicClient as o, networkPolicyLibSource as ot, FS_SESSION_PAGE_SIZE as p, readBehaviorRecords as pt, extractSkillCandidate as q, ASK_USER_TOOL as r, listSkillScripts as rt, BEHAVIOR_RECORDS_KEY as s, networkUrlHost as st, ALLOWED_TOOLS_EXCLUSION_REASON as t, isNetworkAllowed as tt, DEFAULT_USER_PROFILE_LIMITS as u, normalizeToolError as ut, FsSessionStore as v, resolveToolName as vt, READ_SKILL_FILE_INPUT_SCHEMA as w, summarizeToolCalls as wt, HookRunner as x, schemaToForm as xt, FullDisclosureRouter as y, sampleBehaviorRecords as yt, applyUserProfileImport as z };
+export { fromVercelStreamPart as $, SerializingMemoryStore as A, validateUiSpecNode as At, bridgeError as B, ProgressiveRouter as C, sealToolCallPairs as Ct, RUN_SNAPSHOT_SCHEMA_VERSION as D, toRecordDigests as Dt, READ_SKILL_FILE_TOOL_NAME as E, toLlmToolSpec as Et, USER_PROFILE_PROMPT_HEADER as F, exportUserProfile as G, createScriptContext as H, USER_PROFILE_REFINE_PROMPT as I, extractTodoTraceEvents as J, extractChartSpec as K, WebSkillRuntime as L, USER_PROFILE_EXPORT_VERSION as M, USER_PROFILE_KEY as N, RUN_TRACE_SCHEMA_VERSION as O, toVercelToolSpecs as Ot, USER_PROFILE_NO_INVENTION_RULE as P, fromVercelResult as Q, appendBehaviorRecords as R, OpenAiCompatibleClient as S, scriptToolName as St, READ_SKILL_FILE_TOOL as T, summarizeToolCalls as Tt, createWebSkillApi as U, buildRenderResult as V, diffUserProfile as W, findUnpairedToolCalls as X, extractUiSpecEvents as Y, formatSkillScriptManifest as Z, FsRunTraceStore as _, renderUserProfileContext as _t, AgentLoop as a, mergeProfileEntries as at, GoogleGenAiClient as b, schemaSourceLabel as bt, CapabilityApproval as c, normalizeErrorCode as ct, EMPTY_USER_PROFILE as d, parseBridgeRequest as dt, interruptedToolResult as et, EventBus as f, parseUserProfileExport as ft, FsRunSnapshotStore as g, refineUserProfile as gt, FsMemoryStore as h, readUserProfile as ht, ASK_USER_TOOL_NAME as i, mergeCatalogEntries as it, TraceRecorder as j, SESSION_SCHEMA_VERSION as k, validateUiSpecEvent as kt, DEFAULT_LOOP_LIMITS as l, normalizeToolContent as lt, FsArtifactStore as m, readProfileEntries as mt, ASK_USER_INPUT_SCHEMA as n, isUnsupportedRunSnapshot as nt, AnthropicClient as o, networkPolicyLibSource as ot, FS_SESSION_PAGE_SIZE as p, readBehaviorRecords as pt, extractSkillCandidate as q, ASK_USER_TOOL as r, listSkillScripts as rt, BEHAVIOR_RECORDS_KEY as s, networkUrlHost as st, ALLOWED_TOOLS_EXCLUSION_REASON as t, isNetworkAllowed as tt, DEFAULT_USER_PROFILE_LIMITS as u, normalizeToolError as ut, FsSessionStore as v, resolveToolName as vt, READ_SKILL_FILE_INPUT_SCHEMA as w, summarizeRunUsage as wt, HookRunner as x, schemaToForm as xt, FullDisclosureRouter as y, sampleBehaviorRecords as yt, applyUserProfileImport as z };