feat(subtitles): 字幕库验证码由智能体视觉接管;字幕站原站视图下载当前字幕

Co-Authored-By: Claude Opus 5.5 <[email protected]>

apps/server/deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/server",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "exports": { ".": "./src/main.ts" },
   "tasks": {
     "dev": "deno run -A --watch --env-file=../../.env src/main.ts",

apps/server/src/modules/browser/routes.ts

@@ -75,7 +75,7 @@ function route<D extends EndpointDef>(app: ModuleApp, def: D, handler: (input: I
 }
 
 /** 人工操作者:控制台会话 / API Key 的身份(租约 holder) */
-function human(c: Context<AuthEnv>): Actor {
+export function human(c: Context<AuthEnv>): Actor {
   const p = c.get("principal");
   return { kind: "human", holder: `${p?.kind ?? "unknown"}:${p?.name ?? p?.id ?? "?"}` };
 }

apps/server/src/modules/browser/service.ts

@@ -217,6 +217,12 @@ export type ChallengeVision = (req: {
 
 const visions = new WeakMap<object, ChallengeVision>();
 
+/** 页面验证码接管结果:passed 带回站点 Cookie;needs_human 已开待处理事项;unavailable 没有接管条件(自动化关 / 没有浏览器能力) */
+export type PageCaptchaOutcome =
+  | { status: "passed"; by: "browser" | "vision"; cookies: { name: string; value: string }[] }
+  | { status: "needs_human"; reason: string; decisionId: string | null }
+  | { status: "unavailable"; reason: string };
+
 export function registerChallengeVision(ctx: object, fn: ChallengeVision): void {
   visions.set(ctx, fn);
 }
@@ -2012,7 +2018,7 @@ export class BrowserService {
    */
   private async cfSession(
     site: SiteRow,
-    accountId: string,
+    accountId: string | null,
     url: string,
   ): Promise<{ id: string } | { busy: BrowserSessionRow }> {
     const live = await this.liveSessionFor(site.id, accountId);
@@ -2194,6 +2200,111 @@ export class BrowserService {
     }
   }
 
+  /**
+   * 站点自己的页面验证码(字幕库的云锁)代码识别没过时,由智能体视觉接管(docs/plan/subtitles.md):
+   * 匿名会话打开 url → 已不是验证码页就直接算过 → 否则派视觉子流程(系统发起,读图、输入、提交,≤ 6 步,最多 1 次)
+   * → 按页面核实 marker 消失 → 读回站点 Cookie,交调用方并入它自己的 Cookie 罐后重试原请求。
+   * 没过:会话转等待,待处理事项写明去原站视图手动过一次;人过完交还后,下一次站点请求再走这里时就直接读回 Cookie。
+   */
+  async solvePageCaptcha(req: {
+    siteId: string;
+    url: string;
+    marker: string;
+    goal: string;
+  }): Promise<PageCaptchaOutcome> {
+    const site = await this.liveSite(req.siteId);
+    const vision = visions.get(this.ctx);
+    const auto = await this.ctx.settings.get("automation").catch(() => null);
+    if (!vision || !auto?.enabled || auto.readOnlyReconcile)
+      return { status: "unavailable", reason: "自动化未开启，智能体不接管验证码" };
+    const profileKey = profileKeyFor(site.id, null);
+    return this.cfLock.run(profileKey, async () => {
+      const got = await this.cfSession(site, null, req.url);
+      if ("busy" in got)
+        return {
+          status: "needs_human" as const,
+          reason: "已在等你在原站视图里过验证码",
+          decisionId: (obj(got.busy.recovery_state) as RecoveryState).waitingDecisionId ?? null,
+        };
+      const id = got.id;
+      let page = await this.readPage(id).catch(() => null);
+      if (!page) return { status: "unavailable" as const, reason: "浏览器服务不支持读取页面内容" };
+      let by: "browser" | "vision" = "browser";
+      if (page.text.includes(req.marker)) {
+        const r = await vision({ sessionId: id, goal: req.goal, textAbsent: req.marker, maxSteps: 6 }).catch((e) => ({
+          status: "failed",
+          reason: String((e as Error)?.message ?? e),
+          childRunId: null,
+        }));
+        this.ctx.logger.info(
+          { component: "browser", sessionId: id, siteId: site.id, status: r.status },
+          "页面验证码视觉接管结束",
+        );
+        page = await this.readPage(id).catch(() => null);
+        if (!page || page.text.includes(req.marker)) return this.captchaNeedsHuman(id, site, r.reason);
+        by = "vision";
+      }
+      const cookies = await this.siteCookies(profileKey, site.base_url);
+      return { status: "passed" as const, by, cookies };
+    });
+  }
+
+  /** profile 里作用于站点主机、未过期的 Cookie(只取名和值) */
+  private async siteCookies(profileKey: string, siteUrl: string): Promise<{ name: string; value: string }[]> {
+    const prow = await this.providerRow().catch(() => null);
+    const provider = prow ? this.providerFor(prow) : null;
+    const host = new URL(siteUrl).hostname;
+    const now = this.now().getTime();
+    return ((await provider?.getCookies?.(profileKey, [siteUrl]).catch(() => [])) ?? [])
+      .filter((c) => {
+        const d = c.domain.replace(/^\./, "");
+        return (
+          (host === d || host.endsWith(`.${d}`)) &&
+          !(c.expires !== undefined && c.expires > 0 && c.expires * 1000 <= now)
+        );
+      })
+      .map((c) => ({ name: c.name, value: c.value }));
+  }
+
+  /** 原站视图当前页的地址与站点 Cookie:字幕站从人工打开的字幕页直接下载(带上人在浏览器里过的验证) */
+  async currentPage(
+    id: string,
+    actor: Actor,
+  ): Promise<{ siteId: string; url: string; cookies: { name: string; value: string }[] }> {
+    const row = await this.sessionRow(id);
+    const page = await this.run(
+      id,
+      actor,
+      (p, ref) => (p.pageContent ? p.pageContent(ref, { currentVisible: true }) : Promise.resolve(null)),
+      {
+        write: false,
+      },
+    );
+    if (!page?.url) throw new BrowserServiceError("precondition", "无法读取当前页面，请重新连接浏览器");
+    const site = await this.liveSite(row.site_id);
+    return { siteId: site.id, url: page.url, cookies: await this.siteCookies(row.profile_key, site.base_url) };
+  }
+
+  private async captchaNeedsHuman(id: string, site: SiteRow, reason: string): Promise<PageCaptchaOutcome> {
+    const text = {
+      title: `${site.name}需要你在原站视图里过一次验证码`,
+      summary: `站点要求输入验证码，代码识别和智能体都没能通过（${reason.slice(0, 200) || "未说明原因"}）。在站点详情的原站视图接管会话，输完验证码后交还，再搜索一次即可。`,
+    };
+    let row = await this.sessionRow(id);
+    const rs = obj(row.recovery_state) as RecoveryState;
+    if (row.controller === "automation" || !rs.waitingDecisionId)
+      row = await this.enterWaiting(row, "challenge_unsolved", text);
+    else
+      await this.ctx.db
+        .updateTable("decision")
+        .set({ title: text.title, summary: text.summary })
+        .where("id", "=", rs.waitingDecisionId)
+        .execute();
+    const decisionId = (obj(row.recovery_state) as RecoveryState).waitingDecisionId ?? null;
+    this.ctx.logger.info({ component: "browser", sessionId: id, siteId: site.id }, "页面验证码转人工");
+    return { status: "needs_human", reason, decisionId };
+  }
+
   /** 人过完验证交还后:站点在等人过验证、会话是那一个 → 读回 cf_clearance / UA,通知站点请求继续 */
   private async afterHumanChallenge(id: string): Promise<void> {
     const row = await this.sessionRow(id);

apps/server/src/modules/subtitles/index.ts

@@ -1,9 +1,16 @@
 import { Api, type SubtitleApplyRequest, type SubtitleFetchRequest, type SubtitleSource } from "@umt/contracts";
 import { ApiException } from "../../http/errors.ts";
 import { route } from "../agent/http.ts";
+import { human } from "../browser/routes.ts";
 import { sourcesIdentityModule } from "../sources-identity/index.ts";
 import type { ServerModule } from "../types.ts";
-import { downloadOnline, SubtitleSiteError, searchOnline, subtitleSearchQuery } from "./online/index.ts";
+import {
+  downloadFromSession,
+  downloadOnline,
+  SubtitleSiteError,
+  searchOnline,
+  subtitleSearchQuery,
+} from "./online/index.ts";
 import { applySubtitles, parseSubtitles, subtitleTargets, subtitleWrites, undoSubtitleWrite } from "./service.ts";
 
 /** 补字幕(docs/plan/subtitles.md):上传 / 在线下载的字幕解析、配对、写入、历史与撤销 */
@@ -32,10 +39,12 @@ export function createSubtitlesModule(): ServerModule {
       route(
         app,
         Api.subtitles.fetch,
-        ({ body }) =>
+        ({ body, c }) =>
           siteErr(async () => {
             const b = body as SubtitleFetchRequest;
-            const f = await downloadOnline(ctx, rt(), b.siteId, b.ref);
+            const f = b.sessionId
+              ? await downloadFromSession(ctx, rt(), b.sessionId, human(c))
+              : await downloadOnline(ctx, rt(), b.siteId!, b.ref!);
             const parsed = await parseSubtitles(ctx, [{ name: f.name, base64: f.base64 }], b.entityId);
             // 在线来源:每项带上字幕站与条目地址,写入时记进历史
             return {

apps/server/src/modules/subtitles/online/index.ts

@@ -6,6 +6,8 @@ import type { SiteRow } from "@umt/db";
 import { subtitleEpisode } from "@umt/domain";
 import { runSandbox } from "../../../sandbox/client.ts";
 import { b64encode } from "../../../sandbox/jobs.ts";
+import { browserServiceFor } from "../../browser/index.ts";
+import type { Actor } from "../../browser/service.ts";
 import type { SourcesRuntime } from "../../sources-identity/runtime.ts";
 import type { ServerContext } from "../../types.ts";
 import type { PageEntry, PageKind, PageResult } from "./parse.ts";
@@ -22,6 +24,7 @@ export class SubtitleSiteError extends Error {
 const jars = new Map<string, Map<string, string>>();
 
 interface Env {
+  ctx: ServerContext;
   rt: SourcesRuntime;
   site: SiteRow;
 }
@@ -206,22 +209,53 @@ async function subhdDownload(env: Env, ref: string): Promise<{ name: string; byt
 
 // ── 字幕库 ──
 
-/** 打开页面;遇到云锁验证码页就在沙箱里识别、提交验证后重开(最多 3 次) */
+/** 云锁验证码页上的字样(浏览器里 innerText 能读到) */
+const ZIMUKU_CAPTCHA_MARKER = "网站访问认证页面";
+
+/**
+ * 打开页面;遇到云锁验证码页:先在沙箱里识别、提交验证后重开(最多 3 次);
+ * 还没过 → 智能体视觉在浏览器里接管(读图输码),通过后把浏览器里的站点 Cookie 并进 Cookie 罐再开一次。
+ */
 async function zimukuPage(env: Env, url: string, referer?: string): Promise<string> {
-  for (let i = 0; i < 3; i++) {
+  /** 开一次页面:正常页返回 HTML,验证码页返回 BMP */
+  const open = async (): Promise<{ html: string } | { bmp: string }> => {
     const res = await request(env, url, referer ? { referer } : {});
     const html = await res.text();
-    const cap = await page("zimuku.captcha", html);
+    const bmp = (await page("zimuku.captcha", html)).bmpBase64;
-    if (!cap.bmpBase64) {
+    if (bmp) return { bmp };
-      if (!res.ok) throw new SubtitleSiteError(`${env.site.name} 返回 HTTP ${res.status}`);
+    if (!res.ok) throw new SubtitleSiteError(`${env.site.name} 返回 HTTP ${res.status}`);
-      return html;
+    return { html };
-    }
+  };
-    const { code } = await runSandbox("captcha.solve", { kind: "zimuku", bmpBase64: cap.bmpBase64 });
+  for (let i = 0; i < 3; i++) {
+    const o = await open();
+    if ("html" in o) return o.html;
+    const { code } = await runSandbox("captcha.solve", { kind: "zimuku", bmpBase64: o.bmp });
     if (!code) break;
     const hex = [...code].map((c) => c.charCodeAt(0).toString(16)).join("");
     await (await request(env, `${url}${url.includes("?") ? "&" : "?"}security_verify_img=${hex}`)).body?.cancel();
   }
-  throw new SubtitleSiteError("字幕库的验证码没能自动通过：可在站点的原站视图里打开一次再试");
+  const browser = browserServiceFor(env.ctx);
+  const r = browser
+    ? await browser.solvePageCaptcha({
+        siteId: env.site.id,
+        url,
+        marker: ZIMUKU_CAPTCHA_MARKER,
+        goal: "这是字幕库的网站防火墙验证页：截图看清验证码图片里的字符，在下面的输入框里输入，再点「点击继续访问网站」，等页面跳转。只做这些；看不清就刷新一次图片再试，仍不行就停下说明需要人工。",
+      })
+    : { status: "unavailable" as const, reason: "浏览器模块未启用" };
+  if (r.status === "passed") {
+    const jar = jars.get(env.site.id) ?? new Map<string, string>();
+    jars.set(env.site.id, jar);
+    for (const c of r.cookies) jar.set(c.name, c.value);
+    const o = await open();
+    if ("html" in o) return o.html;
+    throw new SubtitleSiteError("字幕库的验证码在浏览器里过了，但验证结果带不回站点请求：稍后再试");
+  }
+  if (r.status === "needs_human")
+    throw new SubtitleSiteError(
+      "字幕库的验证码代码识别和智能体都没过：已开待处理事项，去站点的原站视图输一次验证码后再搜",
+    );
+  throw new SubtitleSiteError(`字幕库的验证码没能自动通过（${r.reason}）：可在站点的原站视图里打开一次再试`);
 }
 
 async function zimukuSearch(env: Env, q: SearchQuery): Promise<PageEntry[]> {
@@ -326,7 +360,7 @@ export async function searchOnline(
   const errors: SubtitleSearchResponse["errors"] = [];
   await Promise.all(
     sites.map(async (site) => {
-      const env = { rt, site };
+      const env = { ctx, rt, site };
       try {
         const entries = await PROVIDERS[site.key]!.search(env, q);
         const seen = new Set<string>();
@@ -350,10 +384,40 @@ export async function downloadOnline(
   rt: SourcesRuntime,
   siteId: string,
   ref: string,
+  cookies: { name: string; value: string }[] = [],
 ): Promise<{ name: string; base64: string; siteKey: string; pageUrl: string }> {
   const site = (await subtitleSites(ctx)).find((s) => s.id === siteId);
   if (!site || !PROVIDERS[site.key]) throw new SubtitleSiteError("字幕站不存在或已停用");
-  const env = { rt, site };
+  const env = { ctx, rt, site };
+  if (cookies.length) {
+    const jar = jars.get(site.id) ?? new Map<string, string>();
+    jars.set(site.id, jar);
+    for (const c of cookies) jar.set(c.name, c.value);
+  }
   const f = await PROVIDERS[site.key]!.download(env, ref);
   return { name: f.name, base64: b64encode(f.bytes), siteKey: site.key, pageUrl: abs(env, ref) };
 }
+
+/** 字幕详情页地址 → 条目 ref(SubHD /a/<sid>、字幕库 /detail/<id>.html);不是详情页为 null */
+const DETAIL_PAGE: Record<string, RegExp> = {
+  subhd: /^\/a\/[A-Za-z0-9_-]+$/,
+  zimuku: /^\/detail\/\d+\.html$/,
+};
+
+/** 原站视图里人工打开的字幕页:取当前页地址与浏览器里的 Cookie,按同一流程下载 */
+export async function downloadFromSession(
+  ctx: ServerContext,
+  rt: SourcesRuntime,
+  sessionId: string,
+  actor: Actor,
+): Promise<{ name: string; base64: string; siteKey: string; pageUrl: string }> {
+  const browser = browserServiceFor(ctx);
+  if (!browser) throw new SubtitleSiteError("浏览器模块未启用");
+  const p = await browser.currentPage(sessionId, actor);
+  const site = (await subtitleSites(ctx)).find((s) => s.id === p.siteId);
+  if (!site || !PROVIDERS[site.key]) throw new SubtitleSiteError("这个会话不是已启用的字幕站");
+  const url = new URL(p.url);
+  if (url.hostname !== new URL(site.base_url).hostname || !DETAIL_PAGE[site.key]?.test(url.pathname))
+    throw new SubtitleSiteError("先在原站视图里打开一条字幕的详情页，再点下载");
+  return downloadOnline(ctx, rt, site.id, url.pathname, p.cookies);
+}

apps/server/test/browser/harness.ts

@@ -69,6 +69,7 @@ export async function createBrowserHarness(
     /** CF-2:mock 站点的 Cloudflare 挑战与页面 HTML */
     cloudflare?: MockCamofoxSite["cloudflare"];
     html?: MockCamofoxSite["html"];
+    pageText?: MockCamofoxSite["pageText"];
     /** 浏览器模块的额外选项(挑战轮询间隔等) */
     browserOptions?: Partial<BrowserModuleOptions>;
   } = {},
@@ -82,6 +83,7 @@ export async function createBrowserHarness(
       ...(opts.sessionCookie ? { sessionCookie: opts.sessionCookie } : {}),
       ...(opts.cloudflare ? { cloudflare: opts.cloudflare } : {}),
       ...(opts.html ? { html: opts.html } : {}),
+      ...(opts.pageText ? { pageText: opts.pageText } : {}),
       origin: ORIGIN,
       username: "fixture-user",
       password: PASSWORD,

apps/server/test/browser/page-captcha.test.ts

@@ -0,0 +1,141 @@
+import { afterAll, beforeAll, beforeEach, describe, expect, it } from "vitest";
+import { browserServiceFor } from "../../src/modules/browser/index.ts";
+import { type BrowserService, registerChallengeVision } from "../../src/modules/browser/service.ts";
+import { type BrowserHarness, createBrowserHarness, type Loose, ORIGIN, pgAvailable } from "./harness.ts";
+
+// 站点自己的页面验证码(字幕库云锁)代码识别没过 → 智能体视觉接管:
+// 过了带回站点 Cookie;没过会话转等待、开待处理事项;自动化关闭时不接管。
+
+const up = await pgAvailable();
+const MARKER = "网站访问认证页面";
+
+describe.skipIf(!up)("页面验证码视觉接管", () => {
+  let h: BrowserHarness;
+  let svc: BrowserService;
+  let captchaUp = true;
+  let visionSolves = true;
+  const visionCalls: string[] = [];
+  const url = `${ORIGIN}/search?q=dune`;
+  const req = { url, marker: MARKER, goal: "输入验证码" };
+
+  const sessions = async () => (await h.req("/browser/sessions")).body.data.items as Loose[];
+
+  beforeAll(async () => {
+    h = await createBrowserHarness("pagecap", {
+      pageText: (u) => (captchaUp && u.includes("/search") ? `网站防火墙\n${MARKER}\n请输入验证码后继续访问：` : null),
+      browserOptions: { challengePollMs: 5, challengeAutoWaitMs: 200, settleMs: 0 },
+    });
+    await sessions();
+    svc = browserServiceFor(h.ctx)!;
+    // 视觉子流程(假):输对验证码 → 站点种下验证 Cookie、页面放行
+    registerChallengeVision(h.ctx, async (r) => {
+      visionCalls.push(r.sessionId);
+      if (visionSolves) {
+        captchaUp = false;
+        h.mock.cookieJars.set(`umt-site-${h.siteId}-acct-anon`, [
+          {
+            name: "security_session_verify",
+            value: "ok-from-browser",
+            domain: new URL(ORIGIN).hostname,
+            path: "/",
+            expires: Math.floor(Date.now() / 1000) + 3600,
+            httpOnly: true,
+            secure: true,
+          },
+        ]);
+      }
+      return {
+        status: visionSolves ? "solved" : "needs_human",
+        reason: visionSolves ? "" : "看不清",
+        childRunId: null,
+      };
+    });
+  }, 120_000);
+  afterAll(async () => {
+    await h?.close();
+  });
+  beforeEach(async () => {
+    for (const s of await sessions()) await h.req(`/browser/sessions/${s.id}`, { method: "DELETE" });
+    h.mock.cookieJars.clear();
+    captchaUp = true;
+    visionSolves = true;
+    visionCalls.length = 0;
+    await h.db
+      .insertInto("settings")
+      .values({ key: "automation", value: JSON.stringify({ enabled: true }) })
+      .onConflict((oc) => oc.column("key").doUpdateSet({ value: JSON.stringify({ enabled: true }) }))
+      .execute();
+  });
+
+  it("视觉接管通过:匿名会话,带回站点 Cookie", async () => {
+    const r = await svc.solvePageCaptcha({ siteId: h.siteId, ...req });
+    expect(r).toEqual({
+      status: "passed",
+      by: "vision",
+      cookies: [{ name: "security_session_verify", value: "ok-from-browser" }],
+    });
+    expect(visionCalls).toHaveLength(1);
+    const [s] = await sessions();
+    expect(s.accountId ?? null).toBeNull();
+    expect(s.controller).toBe("automation");
+  });
+
+  it("浏览器里已经没有验证码页(人工过过):不派子流程,直接带回 Cookie", async () => {
+    captchaUp = false;
+    const r = await svc.solvePageCaptcha({ siteId: h.siteId, ...req });
+    expect(r).toMatchObject({ status: "passed", by: "browser" });
+    expect(visionCalls).toEqual([]);
+  });
+
+  it("视觉接管没过:会话转等待,待处理事项写明去原站视图输验证码;再来一次不打扰", async () => {
+    visionSolves = false;
+    const r = await svc.solvePageCaptcha({ siteId: h.siteId, ...req });
+    expect(r).toMatchObject({ status: "needs_human", reason: "看不清" });
+    const d = await h.db
+      .selectFrom("decision")
+      .selectAll()
+      .where("id", "=", (r as Loose).decisionId)
+      .executeTakeFirstOrThrow();
+    expect(d.title).toContain("需要你在原站视图里过一次验证码");
+    expect(d.status).toBe("awaiting_user");
+    const [s] = await sessions();
+    expect(s.controller).toBe("waiting");
+    const again = await svc.solvePageCaptcha({ siteId: h.siteId, ...req });
+    expect(again).toMatchObject({ status: "needs_human", decisionId: (r as Loose).decisionId });
+    expect(visionCalls).toHaveLength(1);
+  });
+
+  it("原站视图当前页:人工打开的页面地址与浏览器里的站点 Cookie(字幕站从当前页下载)", async () => {
+    captchaUp = false;
+    await svc.solvePageCaptcha({ siteId: h.siteId, ...req });
+    h.mock.cookieJars.set(`umt-site-${h.siteId}-acct-anon`, [
+      {
+        name: "security_session_verify",
+        value: "from-human",
+        domain: new URL(ORIGIN).hostname,
+        path: "/",
+        expires: Math.floor(Date.now() / 1000) + 3600,
+        httpOnly: true,
+        secure: true,
+      },
+    ]);
+    const [s] = await sessions();
+    const p = await svc.currentPage(s.id, { kind: "human", holder: "user:tester" });
+    expect(p).toEqual({
+      siteId: h.siteId,
+      url,
+      cookies: [{ name: "security_session_verify", value: "from-human" }],
+    });
+  });
+
+  it("自动化总开关关闭:不接管", async () => {
+    await h.db
+      .updateTable("settings")
+      .set({ value: JSON.stringify({ enabled: false }) })
+      .where("key", "=", "automation")
+      .execute();
+    const r = await svc.solvePageCaptcha({ siteId: h.siteId, ...req });
+    expect(r).toMatchObject({ status: "unavailable" });
+    expect(visionCalls).toEqual([]);
+  });
+});

apps/web/app/components/browser/desktop-view.tsx

@@ -11,6 +11,7 @@ import {
 } from "~/api/browser.ts";
 import { LoadingState } from "~/components/console/empty-state.tsx";
 import { StatusBadge, type StatusTone } from "~/components/console/status-badge.tsx";
+import { SubtitleFromPage } from "~/components/subtitles/subtitle-from-page.tsx";
 import { Button } from "~/components/ui/button.tsx";
 import { apiData, errorMessage } from "~/lib/api.ts";
 import { qk } from "~/lib/query-keys.ts";
@@ -209,7 +210,7 @@ function DesktopView({
 
   return (
     <div className="space-y-3">
-      <ResourceDownload session={s} />
+      {site.kind === "subtitle" ? <SubtitleFromPage session={s} /> : <ResourceDownload session={s} />}
       <div className="flex flex-wrap items-center gap-2">
         <StatusBadge tone={conn.tone}>{conn.text}</StatusBadge>
         <StatusBadge tone={CONTROLLER_TONE[s.controller] ?? "neutral"}>

apps/web/app/components/browser/site-origin-view.tsx

@@ -31,6 +31,7 @@ import {
 } from "~/api/browser.ts";
 import { EmptyState, ErrorState, LoadingState } from "~/components/console/empty-state.tsx";
 import { StatusBadge, type StatusTone } from "~/components/console/status-badge.tsx";
+import { SubtitleFromPage } from "~/components/subtitles/subtitle-from-page.tsx";
 import { Button } from "~/components/ui/button.tsx";
 import { Input } from "~/components/ui/input.tsx";
 import { errorMessage } from "~/lib/api.ts";
@@ -154,7 +155,7 @@ export function SiteOriginView({ site }: { site: SiteListItem }) {
       <p className="text-caption text-muted-foreground">验证码等需本人操作在此完成。</p>
       <ReloadNotice session={s} />
 
-      <ResourceDownload session={s} />
+      {site.kind === "subtitle" ? <SubtitleFromPage session={s} /> : <ResourceDownload session={s} />}
 
       <div className="flex flex-wrap items-center gap-2">
         <StatusBadge tone={conn.tone}>{conn.text}</StatusBadge>

apps/web/app/components/subtitles/subtitle-dialog.tsx

@@ -6,7 +6,7 @@ import type {
   SubtitleWriteView,
 } from "@umt/contracts";
 import { Download, ExternalLink, FileUp, LoaderCircle, Search, Undo2 } from "lucide-react";
-import { useId, useMemo, useRef, useState } from "react";
+import { useEffect, useId, useMemo, useRef, useState } from "react";
 import { toast } from "sonner";
 import {
   useApplySubtitles,
@@ -68,11 +68,14 @@ export function SubtitleDialog({
   title,
   open,
   onOpenChange,
+  fromSession,
 }: {
   entityId: string;
   title: string;
   open: boolean;
   onOpenChange: (o: boolean) => void;
+  /** 原站视图会话:打开时自动下载它当前页的字幕并配对 */
+  fromSession?: string;
 }) {
   const inputId = useId();
   const input = useRef<HTMLInputElement>(null);
@@ -86,7 +89,20 @@ export function SubtitleDialog({
   const [dragging, setDragging] = useState(false);
   const [tab, setTab] = useState<"upload" | "online">("upload");
   const list = targets.data?.targets ?? [];
-  const busy = parse.isPending || apply.isPending;
+  const fetchPage = useFetchSubtitle();
+  const busy = parse.isPending || apply.isPending || fetchPage.isPending;
+  const fetchedFrom = useRef<string | null>(null);
+  useEffect(() => {
+    if (!open || !fromSession || fetchedFrom.current === fromSession) return;
+    fetchedFrom.current = fromSession;
+    fetchPage
+      .mutateAsync({ entityId, sessionId: fromSession })
+      .then((r) => {
+        setRows((prev) => [...prev, ...r.items.map((it) => ({ ...it, target: it.targetLocationId ?? SKIP }))]);
+        setSkipped((prev) => [...prev, ...r.skipped]);
+      })
+      .catch((err) => toast.error(errorMessage(err, "下载当前字幕失败")));
+  }, [open, fromSession, entityId, fetchPage]);
 
   const load = async (files: File[]) => {
     if (!files.length) return;
@@ -207,6 +223,12 @@ export function SubtitleDialog({
               />
             )}
 
+            {fetchPage.isPending ? (
+              <p className="flex items-center gap-2 text-caption text-muted-foreground">
+                <LoaderCircle className="size-4 animate-spin" />
+                正在下载原站视图当前页的字幕…
+              </p>
+            ) : null}
             {rows.length ? (
               <ul className="space-y-2">
                 {rows.map((r) => (

apps/web/app/components/subtitles/subtitle-from-page.tsx

@@ -0,0 +1,64 @@
+import type { BrowserSessionView } from "@umt/contracts";
+import { Download } from "lucide-react";
+import { useState } from "react";
+import { EntityPicker, type PickedEntity } from "~/components/library/identity-entity-picker.tsx";
+import { Button } from "~/components/ui/button.tsx";
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from "~/components/ui/dialog.tsx";
+import { SubtitleDialog } from "./subtitle-dialog.tsx";
+
+/**
+ * 字幕站的原站视图:人在浏览器里打开一条字幕的详情页(验证码已在浏览器里过)→「下载当前字幕」→ 选作品 →
+ * 补字幕对话框自动下载这条字幕、按季集配对到视频,可改写到哪个视频后写入。
+ */
+export function SubtitleFromPage({ session }: { session: BrowserSessionView }) {
+  const [picking, setPicking] = useState(false);
+  const [entity, setEntity] = useState<PickedEntity | null>(null);
+  const [writing, setWriting] = useState<PickedEntity | null>(null);
+  return (
+    <>
+      <Button size="sm" disabled={session.controller === "closed"} onClick={() => setPicking(true)}>
+        <Download />
+        下载当前字幕
+      </Button>
+      <Dialog open={picking} onOpenChange={setPicking}>
+        <DialogContent className="sm:max-w-lg">
+          <DialogHeader>
+            <DialogTitle>下载当前字幕</DialogTitle>
+            <DialogDescription>
+              先在原站视图里打开一条字幕的详情页。选这条字幕属于哪部作品，下一步再选写到哪个视频。
+            </DialogDescription>
+          </DialogHeader>
+          <EntityPicker value={entity} onChange={setEntity} />
+          <DialogFooter>
+            <Button
+              disabled={!entity}
+              onClick={() => {
+                setWriting(entity);
+                setPicking(false);
+              }}
+            >
+              下载并配对
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+      {writing ? (
+        <SubtitleDialog
+          key={writing.id}
+          entityId={writing.id}
+          title={writing.title}
+          fromSession={session.id}
+          open
+          onOpenChange={(o) => !o && setWriting(null)}
+        />
+      ) : null}
+    </>
+  );
+}

apps/web/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/web",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "private": true,
   "type": "module",
   "scripts": {

docs/plan/README.md

@@ -31,7 +31,7 @@
 
 ### 进行中
 
-- 补字幕：[subtitles.md](subtitles.md)，第 1、2 步已完成；待办：以后接进整理 / 订阅工作流；站点验证码识别真出现失败时再接智能体视觉兜底。
+- 补字幕：[subtitles.md](subtitles.md)，第 1、2 步已完成；待办：以后接进整理 / 订阅工作流。
 - 沙箱第二批：站点网页、RSS、网页抓取的 HTML / XML 解析迁入常驻子进程池（ADR-005）。
 
 ### 业务（需要真实环境或决定）

docs/plan/subtitles.md

@@ -26,15 +26,19 @@
 ## 在线（SubHD、字幕库）
 
 - 搜索词：作品中文名、原名、年份，剧集带季号；结果按标题与季集匹配度排序，对不上的不列。
-- 验证码三层兜底：① 代码识别（SubHD SVG 路径长度表、字幕库 BMP 采样模板，插件里各 80 行左右）；② 浏览器服务打开页面（与 PT 站过人机验证同一流程，可人工接管）；③ 智能体视觉处理页面挑战。
+- 验证码：① 代码识别（SubHD SVG 路径长度表、字幕库 BMP 采样模板，在沙箱里）；② 没过就由智能体视觉在浏览器里接管；③ 智能体也没过就转人工。
+  - 何时接管：字幕库请求遇到云锁验证码页，沙箱识别提交 3 次仍是验证码页（或识别不出字符），且自动化总开关开着。SubHD 现在的下载页前端已没有验证码这一步（2026-10-03 抓页确认），只保留旧 SVG 识别，不接管。
+  - 怎么接管：用该字幕站的匿名浏览器会话打开同一地址；页面已不是验证码页（例如人刚在原站视图里过过）就直接算过；否则派视觉子流程（系统发起，截图读码、输入、点「点击继续访问网站」，≤ 6 步、1 次，预算取智能体设置），按页面上「网站访问认证页面」字样消失核实。
+  - 结果：过了 → 读回浏览器里该站的 Cookie 并进字幕库请求的 Cookie 罐，重开原页面，搜索 / 下载照常继续，用户无感；没过 → 会话转等待处理，待处理事项「字幕库需要你在原站视图里过一次验证码」（附子流程每步截图），本次搜索报错并提示去原站视图；人过完交还后再搜，接管时发现页面已放行，直接带回 Cookie。
+- 原站视图手动下载（与 PT 站「下载当前资源」同一思路）：字幕站的原站视图工具栏是「下载当前字幕」。人在原站视图里打开一条字幕详情页（SubHD `/a/<sid>`、字幕库 `/detail/<id>.html`）→ 点按钮 → 选作品 → 补字幕对话框自动带着浏览器里的 Cookie 下载这一条、解压、按季集配对到视频，可逐个改写到哪个视频后写入，历史记来源。
 - 下载结果走与上传相同的解析 → 预览 → 写入。
 
 ## 分步
 
 1. 手动上传：解析、配对预览、写入、历史与撤销、作品页入口。（v0.3.22 完成）
 2. 字幕站类型 + SubHD / 字幕库搜索下载（第 ① 层验证码），站点页与添加站点；rar / 7z。（v0.3.24 完成；解压、页面解析、验证码都在沙箱里，见 ADR-005）
-3. 验证码第 ② ③ 层兜底：Cloudflare 挑战已由站点出口自动走浏览器；站点自己的验证码（云锁 / SubHD）识别失败时，目前提示去原站视图手动处理，智能体视觉处理待接。
+3. 验证码智能体视觉接管 + 原站视图「下载当前字幕」。（v0.3.25 完成）
 
 实测记录（2026-10-03）：字幕库搜索、验证码、下载、解压、配对全链路通；SubHD 搜索通，下载流程与单独实测一致，但当晚测试请求较多后 SubHD 对本机限流（直连与代理都超时），需要隔天复测。
 
-复测（2026-10-03 晚）：SubHD 按豆瓣 ID 搜索、prepare-download、下载全通，连续下载 7 个未触发验证码。验证码第 ② ③ 层先不做：第 ① 层失败时已提示去原站视图手动下载再上传，等实际出现识别失败再接。
+复测（2026-10-03 晚）：SubHD 按豆瓣 ID 搜索、prepare-download、下载全通，连续下载 7 个未触发验证码。

packages/adapters/deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/adapters",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "exports": { ".": "./src/index.ts", "./fsworker": "./src/fs/fsworker.ts" },
   "tasks": {
     "test": "deno run -A npm:vitest run",

packages/agent/deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/agent",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "exports": { ".": "./src/index.ts" },
   "tasks": {
     "test": "deno run -A npm:vitest run",

packages/contracts/deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/contracts",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "exports": { ".": "./src/index.ts" },
   "tasks": {
     "test": "deno run -A npm:vitest run",

packages/contracts/src/subtitles.ts

@@ -188,10 +188,14 @@ export const SubtitleSearchResponse = z.object({
 });
 export type SubtitleSearchResponse = z.infer<typeof SubtitleSearchResponse>;
 
-export const SubtitleFetchRequest = z.object({
+/** 下载一条在线字幕:搜索结果给 siteId + ref;原站视图里人工打开的字幕页给 sessionId(取当前页与浏览器里的 Cookie) */
-  entityId: Id,
+export const SubtitleFetchRequest = z
-  siteId: Id,
+  .object({
-  ref: z.string().min(1).max(500),
+    entityId: Id,
-  title: z.string().max(500).default(""),
+    siteId: Id.optional(),
-});
+    ref: z.string().min(1).max(500).optional(),
+    sessionId: Id.optional(),
+    title: z.string().max(500).default(""),
+  })
+  .refine((v) => !!v.sessionId || (!!v.siteId && !!v.ref), { message: "给 sessionId，或 siteId 加 ref" });
 export type SubtitleFetchRequest = z.input<typeof SubtitleFetchRequest>;

packages/db/deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/db",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "exports": { ".": "./src/index.ts" },
   "tasks": {
     "test": "deno run -A npm:vitest run",

packages/domain/deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/domain",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "exports": { ".": "./src/index.ts" },
   "tasks": {
     "test": "deno run -A npm:vitest run",

packages/mapping/deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/mapping",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "exports": { ".": "./src/index.ts" },
   "tasks": {
     "test": "deno run -A npm:vitest run",

packages/tasks/deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/tasks",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "exports": { ".": "./src/index.ts" },
   "tasks": {
     "test": "deno run -A npm:vitest run",

packages/testkit/deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@umt/testkit",
-  "version": "0.3.24",
+  "version": "0.3.25",
   "exports": { ".": "./src/index.ts" },
   "tasks": {
     "test": "deno run -A npm:vitest run",

packages/testkit/src/browser/mock-camofox.ts

@@ -48,6 +48,8 @@ export interface MockCamofoxSite {
   cloudflare?: { mode: "auto" | "interactive" | "captcha"; autoPolls?: number; clearance?: string };
   /** 页面 HTML(umt-page 读页面内容时返回);缺省按快照行拼一段简单 HTML */
   html?: (url: string) => string;
+  /** 页面 innerText(缺省按页面行拼);返回 null 时用缺省 */
+  pageText?: (url: string) => string | null;
 }
 
 /** mock 的 Cookie 接口形态:plugin = umt-cookies 插件(读写);upstream = 只有上游导入端点(要求 CAMOFOX_API_KEY);none = 都没有 */
@@ -819,7 +821,7 @@ export async function createMockCamofox(opts: MockCamofoxOptions = {}): Promise<
                   url: t.url,
                   title,
                   html,
-                  text: lines.map((l) => l.text).join("\n"),
+                  text: site?.pageText?.(t.url) ?? lines.map((l) => l.text).join("\n"),
                   ua: MOCK_CAMOFOX_IDENTITY.userAgent,
                 },
               },