@webskill/sdk 0.16.0 → 0.18.0

package/dist/agent.d.ts

@@ -1,3 +1,3 @@
-import { E as DEFAULT_MAX_DATA_SOURCE_BYTES } from "./index-DbVqg2V3.js";
-import { $ as NestedFrameRef, $t as TAB_WORKSET_MIN_MAX, A as DelegationRequest, An as withDelegationOrigin, At as PerceptionAuditSink, B as DownloadedFileReader, Bt as SkillGenerationMessages, C as DataSourcePolicyOptions, Cn as createSkillGenerationToolSource, Ct as PagePerceptionPolicy, D as DelegationBudget, Dn as frameSteps, Dt as ParentBudget, E as DataSourceTransport, En as frameLabel, Et as PagePerceptionToolSourceOptions, F as DownloadedFileEntry, Ft as PerceptionScope, G as FrameScope, Gt as SkillGeneratorOptions, H as DownloadedFileToolSourceOptions, Ht as SkillGenerationPolicy, I as DownloadedFileListing, It as PerceptionScopeInput, J as HttpDataSourceOptions, Jt as SubAgentRunInput, K as GENERATE_SKILL_TOOL, Kt as SkillTraceEvidence, L as DownloadedFilePolicy, Lt as READ_DOWNLOADED_FILE_TOOL, M as DelegationToolSourceOptions, Mt as PerceptionPagingBudget, N as DownloadedFileAction, Nt as PerceptionRecord, O as DelegationOrchestrator, On as toActionFrameScopes, Ot as PerceivedImage, P as DownloadedFileConsent, Pt as PerceptionResult, Q as MANAGE_TODO_TOOL, Qt as TAB_WORKSET_MAX_MAX, R as DownloadedFilePolicyOptions, Rt as SKILL_GENERATION_SYSTEM_PROMPT, S as DataSourcePolicy, Sn as createPagePerceptionToolSource, St as PageAuditSink, T as DataSourceRecord, Tn as createTodoToolSource, Tt as PagePerceptionReader, U as FOCUS_PAGE_TAB_TOOL, Ut as SkillGenerationToolSourceOptions, V as DownloadedFileRecord, Vt as SkillGenerationOutcome, W as FrameNote, Wt as SkillGenerator, X as LIST_DOWNLOADED_FILES_TOOL, Xt as SubtreeRejection, Y as ImageCaptureBudget, Yt as SubAgentRunner, Z as LIST_PAGE_TABS_TOOL, Zt as TAB_WORKSET_DEFAULT_MAX, _ as DataSourceCandidatePolicyOptions, _n as clampWorksetMax, _t as PageActionScope, a as ActionTabWindow, an as TabWorksetPort, at as PAGE_DOCUMENT_REF, b as DataSourceDef, bn as createHttpDataSourceTransport, bt as PageActionToolSourceOptions, c as DELEGATE_TASK_TOOL, cn as TodoItem, ct as PERCEPTION_SYSTEM_PROMPT, d as DOWNLOAD_WAIT_MAX_MS, dn as TodoStatus, dt as PageActionKind, en as TAB_WORKSET_SYSTEM_PROMPT, et as OpenedPageInfo, f as DOWNLOAD_WAIT_MIN_MS, fn as TodoStore, ft as PageActionOutcome, g as DataSourceCandidatePolicy, gn as clampDownloadWaitMs, gt as PageActionRequest, h as DataSourceCandidateAuditSink, hn as WorksetTabView, ht as PageActionRecord, i as ActionTabWatcher, in as TabWorksetPolicyOptions, it as PAGE_ACTION_TOOL, j as DelegationResult, jt as PerceptionCaptureOptions, k as DelegationOrchestratorOptions, kn as toFrameScopes, kt as PerceivedNode, l as DELEGATION_SYSTEM_PROMPT, ln as TodoList, lt as PageActionConsent, m as DataSourceCandidate, mn as WorksetTab, mt as PageActionPolicyOptions, n as ActionDownloadWindow, nn as TabAdmission, nt as PAGE_ACTION_KINDS, o as CLOSE_PAGE_TAB_TOOL, on as TabWorksetToolSourceOptions, ot as PERCEIVE_PAGE_TOOL, p as DataSourceAuditSink, pn as TodoToolSourceOptions, pt as PageActionPolicy, q as GeneratedSkillDraft, qt as SkillTraceEvidenceStep, r as ActionFrameScope, rn as TabWorksetPolicy, rt as PAGE_ACTION_SYSTEM_PROMPT, s as ConsentUi, sn as TodoEvent, st as PERCEPTION_PAGING_DEFAULTS, t as ActionDownloadWatcher, tn as TODO_SYSTEM_PROMPT, tt as OpenedTabSignal, u as DOWNLOAD_WAIT_DEFAULT_MS, un as TodoListener, ut as PageActionExecutor, v as DataSourceCandidateRecord, vn as createDelegationToolSource, vt as PageActionScopeInput, w as DataSourceProvenance, wn as createTabWorksetToolSource, wt as PagePerceptionPolicyOptions, x as DataSourceKind, xn as createPageActionToolSource, xt as PageActionUi, y as DataSourceCandidateStore, yn as createDownloadedFileToolSource, yt as PageActionTarget, z as DownloadedFileReadResult, zt as SkillCandidateSink } from "./index-BLfnO7sz.js";
+import { E as DEFAULT_MAX_DATA_SOURCE_BYTES } from "./index-BDyXe-a5.js";
+import { $ as NestedFrameRef, $t as TAB_WORKSET_MIN_MAX, A as DelegationRequest, An as withDelegationOrigin, At as PerceptionAuditSink, B as DownloadedFileReader, Bt as SkillGenerationMessages, C as DataSourcePolicyOptions, Cn as createSkillGenerationToolSource, Ct as PagePerceptionPolicy, D as DelegationBudget, Dn as frameSteps, Dt as ParentBudget, E as DataSourceTransport, En as frameLabel, Et as PagePerceptionToolSourceOptions, F as DownloadedFileEntry, Ft as PerceptionScope, G as FrameScope, Gt as SkillGeneratorOptions, H as DownloadedFileToolSourceOptions, Ht as SkillGenerationPolicy, I as DownloadedFileListing, It as PerceptionScopeInput, J as HttpDataSourceOptions, Jt as SubAgentRunInput, K as GENERATE_SKILL_TOOL, Kt as SkillTraceEvidence, L as DownloadedFilePolicy, Lt as READ_DOWNLOADED_FILE_TOOL, M as DelegationToolSourceOptions, Mt as PerceptionPagingBudget, N as DownloadedFileAction, Nt as PerceptionRecord, O as DelegationOrchestrator, On as toActionFrameScopes, Ot as PerceivedImage, P as DownloadedFileConsent, Pt as PerceptionResult, Q as MANAGE_TODO_TOOL, Qt as TAB_WORKSET_MAX_MAX, R as DownloadedFilePolicyOptions, Rt as SKILL_GENERATION_SYSTEM_PROMPT, S as DataSourcePolicy, Sn as createPagePerceptionToolSource, St as PageAuditSink, T as DataSourceRecord, Tn as createTodoToolSource, Tt as PagePerceptionReader, U as FOCUS_PAGE_TAB_TOOL, Ut as SkillGenerationToolSourceOptions, V as DownloadedFileRecord, Vt as SkillGenerationOutcome, W as FrameNote, Wt as SkillGenerator, X as LIST_DOWNLOADED_FILES_TOOL, Xt as SubtreeRejection, Y as ImageCaptureBudget, Yt as SubAgentRunner, Z as LIST_PAGE_TABS_TOOL, Zt as TAB_WORKSET_DEFAULT_MAX, _ as DataSourceCandidatePolicyOptions, _n as clampWorksetMax, _t as PageActionScope, a as ActionTabWindow, an as TabWorksetPort, at as PAGE_DOCUMENT_REF, b as DataSourceDef, bn as createHttpDataSourceTransport, bt as PageActionToolSourceOptions, c as DELEGATE_TASK_TOOL, cn as TodoItem, ct as PERCEPTION_SYSTEM_PROMPT, d as DOWNLOAD_WAIT_MAX_MS, dn as TodoStatus, dt as PageActionKind, en as TAB_WORKSET_SYSTEM_PROMPT, et as OpenedPageInfo, f as DOWNLOAD_WAIT_MIN_MS, fn as TodoStore, ft as PageActionOutcome, g as DataSourceCandidatePolicy, gn as clampDownloadWaitMs, gt as PageActionRequest, h as DataSourceCandidateAuditSink, hn as WorksetTabView, ht as PageActionRecord, i as ActionTabWatcher, in as TabWorksetPolicyOptions, it as PAGE_ACTION_TOOL, j as DelegationResult, jt as PerceptionCaptureOptions, k as DelegationOrchestratorOptions, kn as toFrameScopes, kt as PerceivedNode, l as DELEGATION_SYSTEM_PROMPT, ln as TodoList, lt as PageActionConsent, m as DataSourceCandidate, mn as WorksetTab, mt as PageActionPolicyOptions, n as ActionDownloadWindow, nn as TabAdmission, nt as PAGE_ACTION_KINDS, o as CLOSE_PAGE_TAB_TOOL, on as TabWorksetToolSourceOptions, ot as PERCEIVE_PAGE_TOOL, p as DataSourceAuditSink, pn as TodoToolSourceOptions, pt as PageActionPolicy, q as GeneratedSkillDraft, qt as SkillTraceEvidenceStep, r as ActionFrameScope, rn as TabWorksetPolicy, rt as PAGE_ACTION_SYSTEM_PROMPT, s as ConsentUi, sn as TodoEvent, st as PERCEPTION_PAGING_DEFAULTS, t as ActionDownloadWatcher, tn as TODO_SYSTEM_PROMPT, tt as OpenedTabSignal, u as DOWNLOAD_WAIT_DEFAULT_MS, un as TodoListener, ut as PageActionExecutor, v as DataSourceCandidateRecord, vn as createDelegationToolSource, vt as PageActionScopeInput, w as DataSourceProvenance, wn as createTabWorksetToolSource, wt as PagePerceptionPolicyOptions, x as DataSourceKind, xn as createPageActionToolSource, xt as PageActionUi, y as DataSourceCandidateStore, yn as createDownloadedFileToolSource, yt as PageActionTarget, z as DownloadedFileReadResult, zt as SkillCandidateSink } from "./index-DQ3-6GFu.js";
 export { type ActionDownloadWatcher, type ActionDownloadWindow, type ActionFrameScope, type ActionTabWatcher, type ActionTabWindow, CLOSE_PAGE_TAB_TOOL, type ConsentUi, DEFAULT_MAX_DATA_SOURCE_BYTES, DELEGATE_TASK_TOOL, DELEGATION_SYSTEM_PROMPT, DOWNLOAD_WAIT_DEFAULT_MS, DOWNLOAD_WAIT_MAX_MS, DOWNLOAD_WAIT_MIN_MS, type DataSourceAuditSink, type DataSourceCandidate, type DataSourceCandidateAuditSink, DataSourceCandidatePolicy, type DataSourceCandidatePolicyOptions, type DataSourceCandidateRecord, type DataSourceCandidateStore, type DataSourceDef, type DataSourceKind, DataSourcePolicy, type DataSourcePolicyOptions, type DataSourceProvenance, type DataSourceRecord, type DataSourceTransport, type DelegationBudget, DelegationOrchestrator, type DelegationOrchestratorOptions, type DelegationRequest, type DelegationResult, type DelegationToolSourceOptions, type DownloadedFileAction, type DownloadedFileConsent, type DownloadedFileEntry, type DownloadedFileListing, DownloadedFilePolicy, type DownloadedFilePolicyOptions, type DownloadedFileReadResult, type DownloadedFileReader, type DownloadedFileRecord, type DownloadedFileToolSourceOptions, FOCUS_PAGE_TAB_TOOL, type FrameNote, type FrameScope, GENERATE_SKILL_TOOL, type GeneratedSkillDraft, type HttpDataSourceOptions, type ImageCaptureBudget, LIST_DOWNLOADED_FILES_TOOL, LIST_PAGE_TABS_TOOL, MANAGE_TODO_TOOL, type NestedFrameRef, type OpenedPageInfo, type OpenedTabSignal, PAGE_ACTION_KINDS, PAGE_ACTION_SYSTEM_PROMPT, PAGE_ACTION_TOOL, PAGE_DOCUMENT_REF, PERCEIVE_PAGE_TOOL, PERCEPTION_PAGING_DEFAULTS, PERCEPTION_SYSTEM_PROMPT, type PageActionConsent, type PageActionExecutor, type PageActionKind, type PageActionOutcome, PageActionPolicy, type PageActionPolicyOptions, type PageActionRecord, type PageActionRequest, type PageActionScope, type PageActionScopeInput, type PageActionTarget, type PageActionToolSourceOptions, type PageActionUi, type PageAuditSink, PagePerceptionPolicy, type PagePerceptionPolicyOptions, type PagePerceptionReader, type PagePerceptionToolSourceOptions, type ParentBudget, type PerceivedImage, type PerceivedNode, type PerceptionAuditSink, type PerceptionCaptureOptions, type PerceptionPagingBudget, type PerceptionRecord, type PerceptionResult, type PerceptionScope, type PerceptionScopeInput, READ_DOWNLOADED_FILE_TOOL, SKILL_GENERATION_SYSTEM_PROMPT, type SkillCandidateSink, type SkillGenerationMessages, type SkillGenerationOutcome, type SkillGenerationPolicy, type SkillGenerationToolSourceOptions, SkillGenerator, type SkillGeneratorOptions, type SkillTraceEvidence, type SkillTraceEvidenceStep, type SubAgentRunInput, type SubAgentRunner, type SubtreeRejection, TAB_WORKSET_DEFAULT_MAX, TAB_WORKSET_MAX_MAX, TAB_WORKSET_MIN_MAX, TAB_WORKSET_SYSTEM_PROMPT, TODO_SYSTEM_PROMPT, type TabAdmission, TabWorksetPolicy, type TabWorksetPolicyOptions, type TabWorksetPort, type TabWorksetToolSourceOptions, type TodoEvent, type TodoItem, type TodoList, type TodoListener, type TodoStatus, TodoStore, type TodoToolSourceOptions, type WorksetTab, type WorksetTabView, clampDownloadWaitMs, clampWorksetMax, createDelegationToolSource, createDownloadedFileToolSource, createHttpDataSourceTransport, createPageActionToolSource, createPagePerceptionToolSource, createSkillGenerationToolSource, createTabWorksetToolSource, createTodoToolSource, frameLabel, frameSteps, toActionFrameScopes, toFrameScopes, withDelegationOrigin };

package/dist/agent.js

@@ -3,7 +3,7 @@ import { b as parseSkillMarkdown, y as isValidSkillName } from "./skill-CAJMsLod
 import { t as assertRemoteUrlAllowed } from "./urlSafety-CiSuCJvX.js";
 import { c as formatAttachmentText, l as UNTRUSTED_LINE_LIMIT, o as XLSX_MIME, s as classifyAttachment, u as sanitizeUntrustedLine } from "./kind-DxgS8LM-.js";
 import { o as toBase64, s as DEFAULT_MAX_DATA_SOURCE_BYTES } from "./linkedDocument-xNXF-z8n.js";
-import { a as frameLabel, i as toFrameScopes, n as PAGE_DOCUMENT_REF, o as frameSteps, r as toActionFrameScopes, t as PAGE_ACTION_KINDS } from "./types-BuacBov_.js";
+import { a as frameLabel, i as toFrameScopes, n as PAGE_DOCUMENT_REF, o as frameSteps, r as toActionFrameScopes, t as PAGE_ACTION_KINDS } from "./types-DOJI5YC3.js";
 
 //#region ../agent/src/todo/store.ts
 const STATUSES = [
@@ -1274,7 +1274,7 @@ const PERCEPTION_SYSTEM_PROMPT = [
 //#endregion
 //#region ../agent/src/perception/toolSource.ts
 const PERCEIVE_PAGE_TOOL = "perceive_page";
-const DESCRIPTION$1 = "Read the parts of the current page the host made readable, as a structured accessibility outline. Read-only: it cannot click, type, submit, navigate or scroll.";
+const DESCRIPTION$1 = "Read the parts of the current page the host made readable, as a structured accessibility outline. Read-only: it cannot click, type, submit or navigate.";
 /**
 * 两个参数，且它们都扩不出一寸新范围（AC-10.8 的判据由 0.14.0 分册 22 §3 修订）。
 *
@@ -2000,7 +2000,7 @@ var PageActionPolicy = class {
 		const where = target.frame === void 0 ? "on the page" : `in the embedded frame "${target.frame}"`;
 		const scopeNote = target.elevated === true ? " This dialog is outside the usual allowlist; it was opened by this task." : "";
 		const scopeLabel = rememberable ? this.#options.consent?.describeScope?.(target, request.action) : void 0;
-		const message = request.action === "back" ? `Allow the assistant to go back to the previous page${target.name === void 0 ? "" : ` on "${target.name}"`}?${scopeNote}` : `Allow the assistant to ${request.action} "${target.name ?? target.role}" ${where}?${scopeNote}`;
+		const message = request.action === "back" ? `Allow the assistant to go back to the previous page${target.name === void 0 ? "" : ` on "${target.name}"`}?${scopeNote}` : request.action === "scroll" ? `Allow the assistant to scroll ${request.value === "up" ? "up" : "down"} the area around "${target.name ?? target.role}" ${where}?${scopeNote}` : `Allow the assistant to ${request.action} "${target.name ?? target.role}" ${where}?${scopeNote}`;
 		const response = await this.#options.ui.request({
 			type: "authorize",
 			id: `page-action-${this.#seq}`,
@@ -2016,7 +2016,8 @@ var PageActionPolicy = class {
 				},
 				...rememberable ? { rememberable: true } : {},
 				...scopeLabel !== void 0 ? { rememberLabel: scopeLabel } : {},
-				...request.action === "fill" && request.value !== void 0 ? { value: target.secret === true ? "••••••" : request.value } : {}
+				...request.action === "fill" && request.value !== void 0 ? { value: target.secret === true ? "••••••" : request.value } : {},
+				...request.action === "scroll" ? { direction: request.value === "up" ? "up" : "down" } : {}
 			}
 		});
 		if (response.cancelled === true || response.value === false) return false;
@@ -2046,6 +2047,10 @@ var PageActionPolicy = class {
 			approved,
 			ok: outcome.ok,
 			...request.action === "fill" && request.value !== void 0 && !secret ? { value: request.value } : {},
+			...request.action === "scroll" && outcome.scrolled !== void 0 ? { scrolled: {
+				direction: request.value === "up" ? "up" : "down",
+				...outcome.scrolled
+			} } : {},
 			...outcome.reason !== void 0 ? { reason: outcome.reason } : {}
 		};
 		this.#records = [record, ...this.#records].slice(0, 50);
@@ -2093,7 +2098,8 @@ const PAGE_ACTION_SYSTEM_PROMPT = [
 	"Every action asks the user to confirm it first, so do not chain actions speculatively: act only when the user asked for something that requires it.",
 	"A reference expires as soon as the page is perceived again; perceive first, then act.",
 	"After an action, the result tells you where it led. \"openedTab\" means a new page opened — switch to it with focus_page_tab to read it. \"navigated\" means this page itself moved to a new address. \"elevatedModal\" means a dialog opened on this page. If none of the three appear, the page did not change.",
-	"Use the \"back\" action to return to the previous page in this page's history. It takes no \"ref\" and no \"value\", and it only affects the page you are currently on. After going back, always perceive_page again: every reference from before the action is dead."
+	"Use the \"back\" action to return to the previous page in this page's history. It takes no \"ref\" and no \"value\", and it only affects the page you are currently on. After going back, always perceive_page again: every reference from before the action is dead.",
+	"Use the \"scroll\" action only when the page itself says there is more than you can see (for example a list that says it has 20 rows but shows 13). Pass the \"ref\" of any element inside the region you want to scroll and a \"value\" of \"down\" or \"up\"; the region is found for you, so do not look for a reference to the scrollbar or the container. The result carries \"scrolled\": when \"atEnd\" is true there is nothing further in that direction — stop scrolling and answer with what you have. Perceive the page again after each scroll: rows that left the screen no longer exist and their references are dead."
 ].join("\n");
 
 //#endregion
@@ -2118,7 +2124,7 @@ const INPUT_SCHEMA = {
 		},
 		value: {
 			type: "string",
-			description: "Text for \"fill\", the option name for \"select\", or \"true\"/\"false\" for \"set\""
+			description: "Text for \"fill\", the option name for \"select\", \"true\"/\"false\" for \"set\", or \"down\"/\"up\" for \"scroll\""
 		}
 	},
 	required: ["action"],
@@ -2209,6 +2215,7 @@ function createPageActionToolSource(options) {
 				"select",
 				"set"
 			].includes(action) && typeof value !== "string") return toolError$2("VALIDATION_FAILED", `act_on_page needs a "value" string when the action is "${action}".`);
+			if (action === "scroll" && value !== "down" && value !== "up") return toolError$2("VALIDATION_FAILED", "act_on_page with \"scroll\" needs a \"value\" of \"down\" or \"up\".");
 			return await run({
 				ref,
 				action,
@@ -2220,11 +2227,11 @@ function createPageActionToolSource(options) {
 
 //#endregion
 //#region ../agent/src/tabs/policy.ts
-/** 工作集容量缺省；「起点页 + 若干详情页」的常见规模 */
-const TAB_WORKSET_DEFAULT_MAX = 8;
+/** 工作集容量缺省（分册 18）；一屏索引视图的行数就是它要盖住的量，撞 `workset-full` 时模型没有任何入口能要更多 */
+const TAB_WORKSET_DEFAULT_MAX = 32;
 /** 地板：2 的含义是「起点页 + 至少一个详情页」，低于它工作集这个概念就没有意义 */
 const TAB_WORKSET_MIN_MAX = 2;
-/** 天花板：宿主写错一个 0 不该让一轮任务铺开几百个标签页 */
+/** 天花板：宿主写错一个 0 不该让一轮任务铺开几百个标签页。与缺省同值，宿主只调得低不调得高 */
 const TAB_WORKSET_MAX_MAX = 32;
 /**
 * 夹取工作集容量（FR-11.5 / D-11-3）。
@@ -2234,7 +2241,7 @@ const TAB_WORKSET_MAX_MAX = 32;
 * @experimental
 */
 function clampWorksetMax(value) {
-	if (value === void 0 || !Number.isFinite(value)) return 8;
+	if (value === void 0 || !Number.isFinite(value)) return 32;
 	return Math.min(32, Math.max(2, Math.trunc(value)));
 }
 const UNKNOWN_MESSAGE = "That page is not in this task's page set.";

package/dist/{approval-BnLLCOG0.js → approval-B7GW86u2.js}

@@ -2,7 +2,7 @@ import { n as messageOf, t as WebSkillError } from "./errors-BDZNpC13.js";
 import { b as parseSkillMarkdown, i as buildCatalog, n as SkillDiscovery, r as SkillReader } from "./skill-CAJMsLod.js";
 import { a as atomicWriteText, r as resolveInsideRoot, t as assertSafePathSegment } from "./pathSecurity-B1owvJAF.js";
 import { t as assertRemoteUrlAllowed } from "./urlSafety-CiSuCJvX.js";
-import { c as findUnpairedToolCalls, d as partsToText, f as promptText, l as interruptedToolResult, p as textParts, u as sealToolCallPairs } from "./llm-B7lLH0ZI.js";
+import { c as findUnpairedToolCalls, d as partsToText, f as promptText, l as interruptedToolResult, p as textParts, u as sealToolCallPairs } from "./llm-eIQNO9tr.js";
 import { a as fromBase64, c as DEFAULT_MAX_DOCUMENT_BYTES, d as TEXT_BUDGETED_CONTENT_TYPES, i as UNSUPPORTED_DOCUMENT_MESSAGE, l as DEFAULT_MAX_DOCUMENT_TEXT_BYTES, n as READ_LINKED_DOCUMENT_TOOL_NAME, o as toBase64, r as SUPPORTED_DOCUMENT_MIME, t as READ_LINKED_DOCUMENT_TOOL } from "./linkedDocument-xNXF-z8n.js";
 import { t as mergeCatalogEntries } from "./external-_ZRQe-V9.js";
 import { i as isArtifactCardHidden, n as extractChartSpec, r as admitArtifactMetadata, t as buildRenderResult } from "./renderResult-D9Q-Vu2x.js";
@@ -1051,6 +1051,7 @@ const DATA_SOURCE_PROVENANCE_NOTE = {
 };
 /** 分片是不是 PDF；mime 允许带参数与大小写差异，与 FR-11.1 的判定同口径 */
 const isPdfPart = (part) => part.type === "file" && part.mimeType.split(";")[0]?.trim().toLowerCase() === SUPPORTED_DOCUMENT_MIME.pdf;
+const isImagePart = (part) => part.type === "image";
 /**
 * LLM 客户端在 HTTP 失败时放进 `details` 的状态码（FR-22.2）。
 * 判不出来返回 undefined——网络层异常没有状态码，回退对它没有意义。
@@ -1227,6 +1228,7 @@ var AgentLoop = class {
 			externalToolSources: /* @__PURE__ */ new Map(),
 			withheldTools: /* @__PURE__ */ new Set(),
 			pdfFallbackDone: false,
+			imageFallbackDone: false,
 			integrityVerdicts: /* @__PURE__ */ new Map(),
 			toolTimeoutMs: this.#config.toolTimeoutMs,
 			now,
@@ -1441,8 +1443,12 @@ var AgentLoop = class {
 						turn--;
 						continue;
 					}
+					if (this.#dropRejectedImages(state, e)) {
+						turn--;
+						continue;
+					}
 					const code = e instanceof WebSkillError && LLM_TURN_CODES.has(e.code) ? e.code : "LLM_REQUEST_FAILED";
-					const hint = state.pdfFallbackDone ? " The attached PDF was re-sent as extracted text after the endpoint rejected the original file, and that attempt failed as well." : messages.some((message) => message.content.some(isPdfPart)) ? " This turn attached a document; the selected model may not accept document input. Try another model, or link a Word or text file instead." : "";
+					const hint = state.pdfFallbackDone ? " The attached PDF was re-sent as extracted text after the endpoint rejected the original file, and that attempt failed as well." : messages.some((message) => message.content.some(isPdfPart)) ? " This turn attached a document; the selected model may not accept document input. Try another model, or link a Word or text file instead." : state.imageFallbackDone ? " The images were dropped after the endpoint rejected them, and the retry failed as well." : "";
 					return finish("failed", "llm-error", `${messageOf(e)}${hint}`, code);
 				}
 				if (response.thinking !== void 0 && response.thinking !== "") {
@@ -1678,6 +1684,7 @@ var AgentLoop = class {
 			externalToolSources: /* @__PURE__ */ new Map(),
 			withheldTools: /* @__PURE__ */ new Set(),
 			pdfFallbackDone: false,
+			imageFallbackDone: false,
 			integrityVerdicts: /* @__PURE__ */ new Map(),
 			toolTimeoutMs: snapshot.config.toolTimeoutMs,
 			now,
@@ -2745,6 +2752,44 @@ var AgentLoop = class {
 		});
 		return converted > 0;
 	}
+	/**
+	* 端点拒收本轮携带的图片后，把 image 分片换成一条说明让本轮可以重发（分册 15 FR-15.1）。
+	* 返回 true 表示已改写、调用方应重试本轮。
+	*
+	* **不去解析端点指了哪一张**：`.messages[64].image[1]` 这类指针是提供方私有的，
+	* 而重试预算只有一次——留一张没换的，重发还是被同一个端点拒。
+	* 判据与 PDF 回退同口径：本轮真的带了图这个**事实**，不是猜错误文案。
+	*
+	* 图是感知结果的补充，结构化大纲仍在上下文里，去掉图这一轮仍然做得下去；
+	* 而现状是一张解不开的图把整次扇出—汇聚打死。
+	*/
+	#dropRejectedImages(state, cause) {
+		if (state.imageFallbackDone) return false;
+		const status = httpStatusOf(cause);
+		if (status === void 0 || status < 400 || status >= 500) return false;
+		if (!state.messages.some((message) => message.content.some(isImagePart))) return false;
+		state.imageFallbackDone = true;
+		let dropped = 0;
+		for (const message of state.messages) {
+			const kept = message.content.filter((part) => !isImagePart(part));
+			const removed = message.content.length - kept.length;
+			if (removed === 0) continue;
+			dropped += removed;
+			kept.push({
+				type: "text",
+				text: `${removed === 1 ? "An image" : `${removed} images`} attached here could not be read by this endpoint and ${removed === 1 ? "was" : "were"} removed. Work from the text description instead, and say so if the answer depends on what ${removed === 1 ? "the picture shows" : "the pictures show"}.`
+			});
+			message.content = kept;
+		}
+		state.trace.record("run.warning", {
+			message: `The endpoint rejected this turn with HTTP ${status} while it carried ${dropped} image(s); they were dropped and the turn was retried. Original error: ${messageOf(cause)}`,
+			data: {
+				dropped,
+				status
+			}
+		});
+		return dropped > 0;
+	}
 	/** 单个 PDF 分片 → 文本；包装形态与 chatEngine 的附件标签同口径（FR-22.3 / FR-22.4） */
 	async #pdfAsText(state, part, extract) {
 		const name = part.name ?? "document.pdf";
@@ -3255,12 +3300,20 @@ var AgentLoop = class {
 		}
 		try {
 			return await this.#interact(state, {
-				type: "confirm",
+				type: "authorize",
 				id: this.#nextInteractionId(state),
+				capability: "readUploadFile",
 				message: `Let skill "${skillName}" read the file you uploaded: ${file.name} (${file.contentType}, ${file.size} bytes)?`,
-				defaultValue: false
+				details: {
+					skill: skillName,
+					file: {
+						name: file.name,
+						contentType: file.contentType,
+						size: file.size
+					}
+				}
 			}, {
-				kind: "confirm",
+				kind: "authorize",
 				uploadFile: file.id
 			}) === true;
 		} catch (e) {

package/dist/browser.d.ts

@@ -1,6 +1,6 @@
-import { Ct as SkillsLockfile, Et as TrustedKeyStore, H as FileStat, Mt as WebSkillError, N as ArchiveLimits, S as UiBridge, U as FileSystemProvider, _t as SkillManifest, b as RenderResultRequest, c as InteractionResponse, jt as VerifyResult, kt as UnsignedPolicy, l as LlmClient, m as LlmStreamEvent, o as InteractionPolicy, p as LlmResponse, pt as SkillInstallSource, s as InteractionRequest, u as LlmCompleteInput, v as MemoryStore } from "./types-C3V6lAeO-BUpcRLER.js";
-import { Kt as ScriptExecutionContext, M as DocumentSurfaceHost, R as ExternalSkillProvider, S as BridgeResponse, Un as WebSkillApi, Yn as bridgeError, ct as NetworkPolicy, gn as ToolDefinition, m as ApprovalScope, ot as LinkedDocumentReader, qt as ScriptExecutor, x as BridgeRequest, y as BridgeCapabilities, yn as ToolResult, yr as parseBridgeRequest, z as ExternalToolSource } from "./index-DbVqg2V3.js";
-import { $ as NestedFrameRef, Ct as PagePerceptionPolicy, It as PerceptionScopeInput, Pt as PerceptionResult, Tt as PagePerceptionReader, Xt as SubtreeRejection, ft as PageActionOutcome, gt as PageActionRequest, jt as PerceptionCaptureOptions, kt as PerceivedNode, ut as PageActionExecutor, vt as PageActionScopeInput, yt as PageActionTarget } from "./index-BLfnO7sz.js";
+import { Ct as SkillsLockfile, Et as TrustedKeyStore, H as FileStat, Mt as WebSkillError, N as ArchiveLimits, S as UiBridge, U as FileSystemProvider, _t as SkillManifest, b as RenderResultRequest, c as InteractionResponse, jt as VerifyResult, kt as UnsignedPolicy, l as LlmClient, m as LlmStreamEvent, o as InteractionPolicy, p as LlmResponse, pt as SkillInstallSource, s as InteractionRequest, u as LlmCompleteInput, v as MemoryStore } from "./types-Btpdd1y--BcxQ10Fa.js";
+import { Kt as ScriptExecutionContext, M as DocumentSurfaceHost, R as ExternalSkillProvider, S as BridgeResponse, Un as WebSkillApi, Yn as bridgeError, ct as NetworkPolicy, gn as ToolDefinition, m as ApprovalScope, ot as LinkedDocumentReader, qt as ScriptExecutor, x as BridgeRequest, y as BridgeCapabilities, yn as ToolResult, yr as parseBridgeRequest, z as ExternalToolSource } from "./index-BDyXe-a5.js";
+import { $ as NestedFrameRef, Ct as PagePerceptionPolicy, It as PerceptionScopeInput, Pt as PerceptionResult, Tt as PagePerceptionReader, Xt as SubtreeRejection, ft as PageActionOutcome, gt as PageActionRequest, jt as PerceptionCaptureOptions, kt as PerceivedNode, ut as PageActionExecutor, vt as PageActionScopeInput, yt as PageActionTarget } from "./index-DQ3-6GFu.js";
 //#region ../browser/dist/index.d.ts
 //#region src/fs/featureDetection.d.ts
 /** 检测当前环境是否可用 OPFS（navigator.storage.getDirectory） */
@@ -382,6 +382,78 @@ declare function checkDictationAvailability(): DictationAvailability;
  */
 declare function startDictation(options: DictationOptions): DictationSession;
 //#endregion
+//#region src/media/camera.d.ts
+/**
+ * 摄像头的可用性（分册 14 FR-14.2），口径与 `checkDictationAvailability` 一致：
+ * 按钮必须在**渲染期**就知道自己能不能用，不允许「点了才失败」。
+ * @experimental
+ */
+type CameraAvailability = {
+  available: true;
+} | {
+  available: false;
+  reason: 'api-missing' | 'insecure-context';
+};
+/** @experimental */
+interface CameraOptions {
+  /**
+   * 期望的最长边像素（FR-14.3）。只是 `ideal`：约束写成 `exact` 时，
+   * 任何拿不到该分辨率的设备会直接 `OverconstrainedError`，等于把相机关死。
+   */
+  maxDimension?: number;
+  /** 优先使用的摄像头；同样只作偏好，设备没有对应朝向时浏览器会给别的 */
+  facingMode?: 'user' | 'environment';
+}
+/** 一次取景会话的句柄；`close()` 后轨道全部停止，指示灯随之熄灭 @experimental */
+interface CameraSession {
+  readonly stream: MediaStream;
+  /** 轨道实际生效的分辨率；浏览器给不到 ideal 时与请求值不同 */
+  resolution(): {
+    width: number;
+    height: number;
+  };
+  close(): void;
+}
+/** @experimental */
+interface CapturePhotoOptions {
+  /** 输出文件名；缺省由调用时刻生成 */
+  fileName?: string;
+  /** 输出 MIME；缺省 `image/jpeg`（同分辨率下比 PNG 小一个量级） */
+  type?: string;
+  /** JPEG/WebP 质量，0~1；缺省 0.92 */
+  quality?: number;
+}
+/**
+ * 默认最长边：宿主要「最大分辨率」，给一个远超当前消费级摄像头的 ideal 让浏览器自己贴顶
+ * @experimental
+ */
+declare const DEFAULT_CAMERA_MAX_DIMENSION = 5000;
+/**
+ * 渲染期调用（FR-14.2）：不可用时按钮置灰并说明原因。
+ *
+ * 顺序与语音一致——`getUserMedia` 压根不存在时谈安全上下文没有意义；
+ * API 在但站点是 http，取流一定会被浏览器拒绝，这条原因用户能自己解决。
+ * @experimental
+ */
+declare function checkCameraAvailability(): CameraAvailability;
+/**
+ * 打开摄像头取景（FR-14.1）。
+ *
+ * **只能由用户手势调用**（FR-14.4）。设计上的保证是它不进工具列表：
+ * agent 拿不到「打开摄像头」这个动作，因此不存在自主开摄像头的路径。
+ * @experimental
+ */
+declare function openCamera(options?: CameraOptions): Promise<CameraSession>;
+/**
+ * 把取景元素的**当前一帧**编码成文件（FR-14.5）。
+ *
+ * 尺寸取 `videoWidth`/`videoHeight`（轨道原生分辨率），不取 CSS 尺寸——
+ * 预览区被布局缩到 320px 宽，照着画出来的就是 320px 的照片。
+ * 编码交给平台的 canvas，不手写编码器（AGENTS.md §6）。
+ * @experimental
+ */
+declare function capturePhoto(video: HTMLVideoElement, options?: CapturePhotoOptions): Promise<File>;
+//#endregion
 //#region src/media/compressImage.d.ts
 /** 图片压缩结果；`originalBytes`/`compressedBytes` 供 UI 显示「3.2 MB → 1.8 MB」（FR-11.3 可见性） */
 interface CompressResult {
@@ -511,6 +583,14 @@ interface DomPageActionOptions {
    * 其余动作不用它——它们的文档从元素自带的 `ownerDocument` 取，跨帧时才不会拿错。
    */
   document?: Document;
+  /**
+   * 滚动后等待页面稳定的上限（毫秒，0.18.0 FR-10.6）。
+   * SDK 夹在 `SCROLL_SETTLE_BOUNDS` 之间；不注入即用缺省值。
+   *
+   * **模型无法影响它**：这是装配期参数，工具入参里没有任何相关字段。
+   * 重 ERP 的首屏渲染可能超过缺省上限，宿主据此上调。
+   */
+  scrollSettleMs?: () => number | undefined;
 }
 /**
  * @experimental
@@ -657,8 +737,10 @@ interface RemoteTargetRegistry {
  * 它**不是** FR-11.5 那个容量闸门：那一条管的是「同时有几个**标签页**可操作」，
  * 触顶硬拒绝，落在 `TabWorksetPolicy`。这里管的是同一个标签页里连续跳页攒下的历史分片，
  * 那些本来就该失效，按最久未使用逐出即可，逐出不构成对模型的拒绝。
+ * 软上限必须宽于 `TAB_WORKSET_MAX_MAX`（分册 18），否则工作集里**还活着**的页
+ * 会因为别的页新发分片而被逐出；64 = 天花板 32 × 每页两代历史。
  */
-declare const SHARD_SOFT_LIMIT = 24;
+declare const SHARD_SOFT_LIMIT = 64;
 /**
  * 把解析结果翻译成给模型看的补充说明（分册 11 / FR-11.6）；
  * 接到 `PageActionPolicyOptions.diagnose` 上即可：
@@ -1514,4 +1596,4 @@ declare function watchBlockedResources(host: ViewerBlockedHost, notice: ViewerBl
 /** 提示文案的单一来源；判据直接引用，不各写一份字面量 */
 declare function blockedMessage(origins: readonly string[]): string;
 //#endregion
-export { type BlockedEventLike, type BridgeCapabilities, type BridgeRequest, type BridgeResponse, type BrowserHostBundle, type BrowserHostOptions, type BrowserHostOverrides, BrowserSkillManager, BrowserWorkerScriptExecutor, type CaptureFailure, type CaptureLevel, type CapturedImage, type ChromeBuiltinAvailability, ChromeBuiltinLlmClient, type CompressResult, type ConfiguredLlmClient, DEFAULT_FRAME_BUDGET, DEFAULT_MAX_VIEWER_PAYLOAD_BYTES, DOCUMENT_SURFACE_AUDIT_EVENT, DOCX_UNEXTRACTED, type DictationAvailability, type DictationOptions, type DictationSession, type DocumentLocationView, type DocumentSurfaceAudit, type DocumentSurfaceOptions, type DomPageActionExecutor, type DomPageActionOptions, type DomPerceptionReader, type DomPerceptionReaderOptions, type FetchLinkedDocumentOptions, type FrameBudget, type FrameDelivery, type FrameDispatcher, type FrameResolveContext, type FrameRouter, type FrameRouterOptions, HOST_PORT_MATRIX, type HandleKind, type HostChatbotAdapterShape, type HostConnectFacadeShape, type HostConsoleBackendShape, type HostKind, type HostSurface, type IframeSandboxOptions, IframeWorkerLike, type InspectHostWiringInput, type LlmClientConfig, type LlmEntryConfig, type MainToWorkerMessage, type OpenDocumentInput, OpfsProvider, type PageAgentHandler, type PageAgentHandlerOptions, type PageAgentReply, type PageAgentRequest, type PageAgentTargetEntry, type PageAgentTransport, type PageDocumentIdentity, type PortApplicability, type PortReport, type PortSpec, type PortStatus, type RemotePageActionExecutorOptions, type RemotePerceptionReaderOptions, type RemoteTargetRegistry, type RemoteTargetRegistryOptions, type RoleHint, SANDBOX_PAGE_SCRIPT_SOURCE, SHARD_SOFT_LIMIT, type SandboxMode, type SerializedError, type TargetResolution, TsTranspiler, type TypeScriptSupportOptions, VIEWER_SANDBOX_TOKENS, type ViewerBlockedHost, type ViewerBlockedNotice, type ViewerBlockedResult, type ViewerCspOptions, type ViewerDocument, type ViewerShellHost, type ViewerShellMount, type ViewerShellResult, WORKER_BOOTSTRAP_SOURCE, type WindowLike, type WorkerEvent, type WorkerFactory, type WorkerLike, type WorkerRequest, WorkerRuntimeClient, type WorkerRuntimeClientDeps, type WorkerRuntimeHostOverrides, type WorkerScopeLike, WorkerUiBridge, XLSX_UNEXTRACTED, blockedMessage, bridgeError, captureElementImage, checkDictationAvailability, compressImageToBudget, createBrowserChatbotHost, createDocumentSurfaceHost, createDomPageActionExecutor, createDomPerceptionReader, createEncryptedMemoryStore, createFetchLinkedDocumentReader, createFrameRouter, createIframeWorker, createLlmClient, createPageAgentHandler, createRemotePageActionExecutor, createRemotePerceptionReader, createRemoteTargetRegistry, deleteMemoryEncryptionKey, documentKey, explainResolution, extractDocxText, extractXlsxText, extractZipWeb, generateMemoryEncryptionKey, inspectHostWiring, installWebSkillNavigator, isCapturableElement, isCaptureFailure, isEncryptedMemoryValue, isOpfsAvailable, missingPorts, openDocumentSurface, openMemoryEncryptionKey, parseBridgeRequest, parseMainMessage, parseWorkerEvent, probeChromeBuiltinAvailability, sha256HexWeb, startDictation, startViewerShell, startWorkerRuntimeHost, viewerCspHeader, watchBlockedResources };
+export { type BlockedEventLike, type BridgeCapabilities, type BridgeRequest, type BridgeResponse, type BrowserHostBundle, type BrowserHostOptions, type BrowserHostOverrides, BrowserSkillManager, BrowserWorkerScriptExecutor, type CameraAvailability, type CameraOptions, type CameraSession, type CaptureFailure, type CaptureLevel, type CapturePhotoOptions, type CapturedImage, type ChromeBuiltinAvailability, ChromeBuiltinLlmClient, type CompressResult, type ConfiguredLlmClient, DEFAULT_CAMERA_MAX_DIMENSION, DEFAULT_FRAME_BUDGET, DEFAULT_MAX_VIEWER_PAYLOAD_BYTES, DOCUMENT_SURFACE_AUDIT_EVENT, DOCX_UNEXTRACTED, type DictationAvailability, type DictationOptions, type DictationSession, type DocumentLocationView, type DocumentSurfaceAudit, type DocumentSurfaceOptions, type DomPageActionExecutor, type DomPageActionOptions, type DomPerceptionReader, type DomPerceptionReaderOptions, type FetchLinkedDocumentOptions, type FrameBudget, type FrameDelivery, type FrameDispatcher, type FrameResolveContext, type FrameRouter, type FrameRouterOptions, HOST_PORT_MATRIX, type HandleKind, type HostChatbotAdapterShape, type HostConnectFacadeShape, type HostConsoleBackendShape, type HostKind, type HostSurface, type IframeSandboxOptions, IframeWorkerLike, type InspectHostWiringInput, type LlmClientConfig, type LlmEntryConfig, type MainToWorkerMessage, type OpenDocumentInput, OpfsProvider, type PageAgentHandler, type PageAgentHandlerOptions, type PageAgentReply, type PageAgentRequest, type PageAgentTargetEntry, type PageAgentTransport, type PageDocumentIdentity, type PortApplicability, type PortReport, type PortSpec, type PortStatus, type RemotePageActionExecutorOptions, type RemotePerceptionReaderOptions, type RemoteTargetRegistry, type RemoteTargetRegistryOptions, type RoleHint, SANDBOX_PAGE_SCRIPT_SOURCE, SHARD_SOFT_LIMIT, type SandboxMode, type SerializedError, type TargetResolution, TsTranspiler, type TypeScriptSupportOptions, VIEWER_SANDBOX_TOKENS, type ViewerBlockedHost, type ViewerBlockedNotice, type ViewerBlockedResult, type ViewerCspOptions, type ViewerDocument, type ViewerShellHost, type ViewerShellMount, type ViewerShellResult, WORKER_BOOTSTRAP_SOURCE, type WindowLike, type WorkerEvent, type WorkerFactory, type WorkerLike, type WorkerRequest, WorkerRuntimeClient, type WorkerRuntimeClientDeps, type WorkerRuntimeHostOverrides, type WorkerScopeLike, WorkerUiBridge, XLSX_UNEXTRACTED, blockedMessage, bridgeError, captureElementImage, capturePhoto, checkCameraAvailability, checkDictationAvailability, compressImageToBudget, createBrowserChatbotHost, createDocumentSurfaceHost, createDomPageActionExecutor, createDomPerceptionReader, createEncryptedMemoryStore, createFetchLinkedDocumentReader, createFrameRouter, createIframeWorker, createLlmClient, createPageAgentHandler, createRemotePageActionExecutor, createRemotePerceptionReader, createRemoteTargetRegistry, deleteMemoryEncryptionKey, documentKey, explainResolution, extractDocxText, extractXlsxText, extractZipWeb, generateMemoryEncryptionKey, inspectHostWiring, installWebSkillNavigator, isCapturableElement, isCaptureFailure, isEncryptedMemoryValue, isOpfsAvailable, missingPorts, openCamera, openDocumentSurface, openMemoryEncryptionKey, parseBridgeRequest, parseMainMessage, parseWorkerEvent, probeChromeBuiltinAvailability, sha256HexWeb, startDictation, startViewerShell, startWorkerRuntimeHost, viewerCspHeader, watchBlockedResources };