feat(browser): complete camofox BrowserProvider with tabs, interaction, download capture and control lease

Contracts for browser API and browser.* agent capabilities; adapter implements
close/list/restore sessions, tabs + reconcile, history, click/type/select/scroll/press,
extract, waitForDownload (staged via DownloadSink, sha256), stale-ref guard and
automation/human actor check. testkit mock camofox gains an interactive site
(login redirect, captcha, click-driven page changes, downloads, PNG screenshots).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
This commit is contained in:
rootandClaude Opus 5.5 committed 2026-09-28 20:17:46 +08:00
1 parent 62290d4477
commit c17d4b90bd
12 files changed
+1832 -101

No files matched your search

+162
View File
@@ -2,6 +2,7 @@ import { z } from "zod";
import * as agent from "./agent.ts";
import * as assets from "./assets.ts";
import * as auth from "./auth.ts";
import * as browser from "./browser.ts";
import { CapabilityDescriptor, KnownCapabilityId } from "./capability.ts";
import { type ApiScope, CursorQuery, cursorPage, HealthStatus, Id, IdParam } from "./common.ts";
import * as content from "./content.ts";
@@ -2008,6 +2009,167 @@ export const Api = {
}),
},
// ── 浏览器(C1-E):站点浏览器会话、原站视图、人工接管;写动作需人工持有控制权(接管) ──
browser: {
providers: ep({
method: "get",
path: "/browser/providers",
scope: "admin",
summary: "浏览器服务列表(地址、凭证引用、启用、能力 supported/verified);probe=true 现场探测并写回",
query: browser.BrowserProvidersQuery,
response: z.object({ items: z.array(browser.BrowserProviderView) }),
}),
sessions: ep({
method: "get",
path: "/browser/sessions",
scope: "admin",
summary: "浏览器会话列表(以数据库为准;默认只列未关闭的)",
query: browser.BrowserSessionsQuery,
response: z.object({ items: z.array(browser.BrowserSessionView) }),
}),
open: ep({
method: "post",
path: "/browser/sessions",
scope: "admin",
summary: "打开站点浏览器会话(同一站点/账号已有未关闭会话时直接返回它)",
body: browser.BrowserSessionOpen,
response: browser.BrowserSessionView,
}),
get: ep({
method: "get",
path: "/browser/sessions/{id}",
scope: "admin",
summary: "浏览器会话详情(控制者、标签页、租约、是否需重新读取页面)",
params: browser.BrowserSessionParams,
response: browser.BrowserSessionView,
}),
close: ep({
method: "delete",
path: "/browser/sessions/{id}",
scope: "admin",
summary: "关闭浏览器会话(关闭上游标签页,撤销租约)",
params: browser.BrowserSessionParams,
response: browser.BrowserSessionView,
}),
openTab: ep({
method: "post",
path: "/browser/sessions/{id}/tabs",
scope: "admin",
summary: "新建标签页并设为当前(需持有控制权)",
params: browser.BrowserSessionParams,
body: browser.BrowserTabOpen,
response: browser.BrowserSessionView,
}),
closeTab: ep({
method: "delete",
path: "/browser/sessions/{id}/tabs/{tabId}",
scope: "admin",
summary: "关闭标签页(需持有控制权)",
params: browser.BrowserTabParams,
response: browser.BrowserSessionView,
}),
navigate: ep({
method: "post",
path: "/browser/sessions/{id}/navigate",
scope: "admin",
summary: "地址栏导航或后退/前进/刷新(需持有控制权;限站点域名)",
params: browser.BrowserSessionParams,
body: browser.BrowserNavigate,
response: browser.BrowserNavigation,
}),
snapshot: ep({
method: "get",
path: "/browser/sessions/{id}/snapshot",
scope: "admin",
summary: "结构化页面快照(凭证已遮蔽)",
params: browser.BrowserSessionParams,
query: browser.BrowserSnapshotQuery,
response: browser.BrowserSnapshot,
}),
screenshot: ep({
method: "get",
path: "/browser/sessions/{id}/screenshot",
scope: "admin",
summary: "当前画面截图(base64,只给原站视图,不进模型)",
params: browser.BrowserSessionParams,
query: browser.BrowserScreenshotQuery,
response: browser.BrowserScreenshotView,
}),
click: ep({
method: "post",
path: "/browser/sessions/{id}/click",
scope: "admin",
summary: "点击(快照引用 / 选择器 / 截图坐标;需持有控制权)",
params: browser.BrowserSessionParams,
body: browser.BrowserClick,
response: browser.BrowserActionResult,
}),
type: ep({
method: "post",
path: "/browser/sessions/{id}/type",
scope: "admin",
summary: "输入文本 / 站点账号凭证 / 按键(需持有控制权;凭证由服务端注入)",
params: browser.BrowserSessionParams,
body: browser.BrowserType,
response: browser.BrowserActionResult,
}),
scroll: ep({
method: "post",
path: "/browser/sessions/{id}/scroll",
scope: "admin",
summary: "滚动页面(需持有控制权)",
params: browser.BrowserSessionParams,
body: browser.BrowserScroll,
response: browser.BrowserActionResult,
}),
select: ep({
method: "post",
path: "/browser/sessions/{id}/select",
scope: "admin",
summary: "下拉框选择(需持有控制权)",
params: browser.BrowserSessionParams,
body: browser.BrowserSelect,
response: browser.BrowserActionResult,
}),
extract: ep({
method: "post",
path: "/browser/sessions/{id}/extract",
scope: "admin",
summary: "按 JSON Schema(x-ref)提取页面字段",
params: browser.BrowserSessionParams,
body: browser.BrowserExtract,
response: browser.BrowserExtractResult,
}),
download: ep({
method: "post",
path: "/browser/sessions/{id}/download",
scope: "organize.execute",
summary:
"捕获浏览器下载 → 后端暂存并核验 infohash、匹配本地资源;给出 submit 时经 POST /acquisition-plans 同一流程提交",
params: browser.BrowserSessionParams,
body: browser.BrowserDownload,
response: browser.BrowserDownloadResult,
}),
acquire: ep({
method: "post",
path: "/browser/sessions/{id}/control/acquire",
scope: "admin",
summary: "人工接管:先停止该会话新的自动化输入并等进行中的操作收尾,再发放人工租约",
params: browser.BrowserSessionParams,
body: browser.BrowserControlAcquire,
response: browser.BrowserControlResult,
}),
release: ep({
method: "post",
path: "/browser/sessions/{id}/control/release",
scope: "admin",
summary: "交还控制权:Agent 恢复前必须重新读取页面,旧元素引用失效",
params: browser.BrowserSessionParams,
body: browser.BrowserControlRelease,
response: browser.BrowserControlResult,
}),
},
// ── 定时任务看板(W7-C):统一注册表 + Agent 巡检;全部只接受控制台会话 ──
schedules: {
list: ep({
+219
View File
@@ -1,5 +1,6 @@
import { z } from "zod";
import { EntityBase, Id, JsonObject, NonNegInt, Timestamp } from "./common.ts";
import { ManualAcquisitionResult } from "./execution.ts";
import { InstanceHealth } from "./instances.ts";
// 浏览器能力合同(REQ §4.4、§4.7.2–4.7.4、§12 浏览器实体)。
@@ -192,6 +193,8 @@ export const DownloadArtifact = z.object({
bytes: BlobRef,
/** 十六进制 sha256 */
sha256: z.string().min(1),
/** 已脱敏的下载来源地址(上游提供时) */
sourceUrl: z.string().nullable().optional(),
capturedAt: Timestamp,
});
export type DownloadArtifact = z.infer<typeof DownloadArtifact>;
@@ -219,5 +222,221 @@ export const BrowserErrorKind = z.enum([
"budget_exceeded",
"control_conflict",
"lock_timeout",
/** 元素引用来自导航/交接之前的快照,已失效:先重新读取页面 */
"stale_ref",
]);
export type BrowserErrorKind = z.infer<typeof BrowserErrorKind>;
// ── C1-E:浏览器 API 与 Agent 能力的请求/响应(docs/runbook/browser.md) ──
/** 查询串布尔 */
const QueryBool = z.enum(["true", "false"]).optional();
/** 浏览器服务(GET /browser/providers);凭证只给引用是否存在,不给明文 */
export const BrowserProviderView = z.object({
id: Id,
name: z.string(),
kind: BrowserProviderKind,
baseUrl: z.string().nullable(),
/** 访问凭据的 secret 引用(明文永不出 API) */
secretRef: Id.nullable(),
enabled: z.boolean(),
/** 最近一次声明/探测到的能力;verified 只由人工核验写入 config.verified */
capabilities: BrowserCapabilities.nullable(),
verifiedAt: Timestamp.nullable(),
healthStatus: InstanceHealth,
lastHealthAt: Timestamp.nullable(),
/** probe=true 时的即时健康结果 */
health: BrowserHealth.nullable(),
});
export type BrowserProviderView = z.infer<typeof BrowserProviderView>;
export const BrowserProvidersQuery = z.object({
/** true = 现场探测 /health 与 /openapi.json 并写回 */
probe: QueryBool,
});
export type BrowserProvidersQuery = z.infer<typeof BrowserProvidersQuery>;
/** 会话视图:会话 + 标签页 + 当前有效租约 + 界面可用能力 */
export const BrowserSessionView = BrowserSession.extend({
tabs: z.array(BrowserTab),
currentTabId: z.string().nullable(),
lease: BrowserControlLease.nullable(),
/** 控制权交还后到重新读取页面之前为 true:旧元素引用一律失效 */
needsSnapshot: z.boolean(),
/** 等待处理/接管原因(login_required / captcha / human_takeover / lease_expired …) */
waitingReason: z.string().nullable(),
});
export type BrowserSessionView = z.infer<typeof BrowserSessionView>;
export const BrowserSessionOpen = z.object({
siteId: Id,
accountId: Id.nullable().optional(),
/** 缺省取第一个启用的浏览器服务 */
providerId: Id.optional(),
/** 缺省打开站点首页;必须在站点域名内 */
url: z.url().optional(),
});
export type BrowserSessionOpen = z.infer<typeof BrowserSessionOpen>;
export const BrowserSessionsQuery = z.object({
siteId: Id.optional(),
includeClosed: QueryBool,
});
export type BrowserSessionsQuery = z.infer<typeof BrowserSessionsQuery>;
export const BrowserSessionParams = z.object({ id: Id });
export const BrowserTabParams = z.object({ id: Id, tabId: z.string().min(1).max(200) });
export const BrowserTabOpen = z.object({ url: z.url().optional() });
/** 导航:地址栏(url)或导航按钮(action),二选一(url 优先) */
export const BrowserNavigate = z.object({
tabId: z.string().min(1).optional(),
url: z.url().max(4000).optional(),
action: z.enum(["back", "forward", "reload"]).optional(),
});
export type BrowserNavigate = z.infer<typeof BrowserNavigate>;
export const BrowserNavigation = z.object({
/** 已脱敏的最终地址 */
url: z.string(),
httpStatus: z.number().int().nullable(),
ok: z.boolean(),
attempts: z.number().int().min(0),
/** 落在登录页(登录失效) */
loginRequired: z.boolean(),
session: BrowserSessionView,
});
export type BrowserNavigation = z.infer<typeof BrowserNavigation>;
export const BrowserSnapshotQuery = z.object({
tabId: z.string().min(1).optional(),
offset: z.coerce.number().int().min(0).optional(),
});
export const BrowserScreenshotQuery = z.object({ tabId: z.string().min(1).optional() });
/** 截图:base64 直接给原站视图;不落盘、不进模型(modelSafe 标注) */
export const BrowserScreenshotView = z.object({
mimeType: z.string(),
dataBase64: z.string(),
/** 图片像素尺寸(PNG/JPEG 头解析;解析不出为 null) */
width: z.number().int().nullable(),
height: z.number().int().nullable(),
url: z.string().nullable(),
modelSafe: z.boolean(),
capturedAt: Timestamp,
controller: BrowserController,
});
export type BrowserScreenshotView = z.infer<typeof BrowserScreenshotView>;
/** 元素定位:快照引用 / CSS 选择器 / 截图像素坐标(原站视图点击) */
export const BrowserTarget = z.union([
/** 快照元素引用,形如 e12 */
z.object({ ref: z.string().min(1).max(40) }),
z.object({ selector: z.string().min(1).max(500) }),
z.object({ x: z.number().min(0), y: z.number().min(0) }),
]);
export type BrowserTarget = z.infer<typeof BrowserTarget>;
export const BrowserClick = z.object({ tabId: z.string().min(1).optional(), target: BrowserTarget });
export type BrowserClick = z.infer<typeof BrowserClick>;
/** 输入:按键(key)> 站点账号凭证字段(credential,服务端解析后直接下发)> 明文(text);target 缺省 = 当前焦点 */
export const BrowserType = z.object({
tabId: z.string().min(1).optional(),
target: BrowserTarget.optional(),
text: z.string().max(4000).optional(),
credential: z.enum(["username", "password", "totp"]).optional(),
key: z.string().min(1).max(40).optional(),
submit: z.boolean().optional(),
clear: z.boolean().optional(),
});
export type BrowserType = z.infer<typeof BrowserType>;
export const BrowserScroll = z.object({
tabId: z.string().min(1).optional(),
direction: z.enum(["up", "down", "left", "right"]),
amount: z.number().int().min(1).max(10_000).default(600),
});
export type BrowserScroll = z.infer<typeof BrowserScroll>;
export const BrowserSelect = z.object({
tabId: z.string().min(1).optional(),
target: BrowserTarget,
value: z.string().max(500),
});
export type BrowserSelect = z.infer<typeof BrowserSelect>;
export const BrowserActionResult = z.object({ ok: z.literal(true), session: BrowserSessionView });
export type BrowserActionResult = z.infer<typeof BrowserActionResult>;
export const BrowserExtract = z.object({
tabId: z.string().min(1).optional(),
/** JSON Schema;属性用 x-ref 指向快照元素(camofox /extract) */
schema: JsonObject,
});
export const BrowserExtractResult = z.object({ data: JsonObject });
/** 捕获下载:可先点击 target 触发;submit 给出时经 POST /acquisition-plans 同一流程提交 */
export const BrowserDownload = z.object({
tabId: z.string().min(1).optional(),
target: BrowserTarget.optional(),
timeoutMs: z.number().int().min(500).max(120_000).default(15_000),
submit: z
.object({ slotId: Id.optional(), entityId: Id.optional(), note: z.string().max(2000).optional() })
.optional(),
});
export type BrowserDownload = z.infer<typeof BrowserDownload>;
/** 暂存核验结果:种子解析出的 infohash、与本地资源的匹配 */
export const BrowserDownloadVerification = z.object({
valid: z.boolean(),
infohashV1: z.string().nullable(),
infohashV2: z.string().nullable(),
name: z.string().nullable(),
totalSize: NonNegInt.nullable(),
private: z.boolean().nullable(),
error: z.string().nullable(),
});
export type BrowserDownloadVerification = z.infer<typeof BrowserDownloadVerification>;
export const BrowserDownloadResult = z.object({
artifact: DownloadArtifact,
verification: BrowserDownloadVerification,
/** 按 infohash / 站点资源 ID 匹配到的本地资源;null = 不在本地索引 */
sourceReleaseId: Id.nullable(),
/** 同一会话已捕获过同一文件(sha256 相同):不重复提交 */
duplicate: z.boolean(),
/** submit 给出且匹配到资源时的提交结果(与 POST /acquisition-plans 相同) */
submission: ManualAcquisitionResult.nullable(),
note: z.string().nullable(),
});
export type BrowserDownloadResult = z.infer<typeof BrowserDownloadResult>;
export const BrowserControlAcquire = z.object({
reason: z.string().min(1).max(200).default("人工接管"),
/** 租约时长;到期未交还 → 会话转"等待处理",不自动恢复自动化 */
ttlMs: z
.number()
.int()
.min(60_000)
.max(4 * 3_600_000)
.default(30 * 60_000),
});
export type BrowserControlAcquire = z.infer<typeof BrowserControlAcquire>;
export const BrowserControlRelease = z.object({
leaseId: Id.optional(),
reason: z.string().min(1).max(200).default("人工交还"),
/** 交还给自动化(缺省)或留在等待处理 */
to: z.enum(["automation", "waiting"]).default("automation"),
});
export type BrowserControlRelease = z.infer<typeof BrowserControlRelease>;
export const BrowserControlResult = z.object({
lease: BrowserControlLease.nullable(),
session: BrowserSessionView,
});
export type BrowserControlResult = z.infer<typeof BrowserControlResult>;
+100
View File
@@ -1,5 +1,6 @@
import { z } from "zod";
import { MediaAsset } from "./assets.ts";
import { BrowserController, BrowserDownloadVerification, BrowserSnapshot } from "./browser.ts";
import { CAPABILITY_CATALOG, type CapabilitySpec, type KnownCapabilityId } from "./capability.ts";
import { Id, JsonObject, NonNegInt } from "./common.ts";
import { MediaEntity, MediaKind } from "./content.ts";
@@ -318,6 +319,93 @@ export const SiteUserDataOutput = z.object({
history: z.array(SiteUserDataPoint),
});
// C1-E 浏览器能力:会话按站点/账号隔离;元素引用只来自最近一次 browser.snapshot,
// 控制权交还或导航后旧引用失效;凭证用 credential 字段由服务端注入,不进入模型。
const BrowserAgentTarget = z.union([
z.object({ ref: z.string().min(1).max(40) }),
z.object({ selector: z.string().min(1).max(500) }),
]);
export const BrowserSnapshotInput = z
.object({
/** 会话 ID;缺省用 siteId 对应的未关闭会话 */
sessionId: Id.optional(),
siteId: Id.optional(),
tabId: z.string().min(1).optional(),
offset: z.number().int().min(0).optional(),
})
.refine((v) => !!v.sessionId || !!v.siteId, { message: "sessionId 与 siteId 至少给一个" });
export const BrowserSnapshotOutput = z.object({
snapshot: BrowserSnapshot,
controller: BrowserController,
/** 会话在等待处理(验证码 / 二次认证 / 登录失效)时的原因;此时自动化动作被拒,需人工在原站视图处理 */
waitingReason: z.string().nullable(),
loginRequired: z.boolean(),
});
export const BrowserNavigateInput = z
.object({
sessionId: Id.optional(),
/** 无会话时按站点(和账号)打开一个 */
siteId: Id.optional(),
accountId: Id.optional(),
tabId: z.string().min(1).optional(),
/** 必须在站点域名内 */
url: z.url().max(4000),
})
.refine((v) => !!v.sessionId || !!v.siteId, { message: "sessionId 与 siteId 至少给一个" });
export const BrowserNavigateOutput = z.object({
sessionId: Id,
url: z.string(),
httpStatus: z.number().int().nullable(),
ok: z.boolean(),
attempts: z.number().int(),
loginRequired: z.boolean(),
controller: BrowserController,
});
export const BrowserInteractInput = z.object({
sessionId: Id,
tabId: z.string().min(1).optional(),
/** request_human:转"等待处理",由人在原站视图接管(验证码、二次认证等) */
action: z.enum(["click", "type", "select", "scroll", "press", "request_human"]),
target: BrowserAgentTarget.optional(),
text: z.string().max(4000).optional(),
/** 站点账号凭证字段:服务端解析后直接下发到输入框 */
credential: z.enum(["username", "password", "totp"]).optional(),
key: z.string().min(1).max(40).optional(),
value: z.string().max(500).optional(),
direction: z.enum(["up", "down", "left", "right"]).optional(),
amount: z.number().int().min(1).max(10_000).optional(),
submit: z.boolean().optional(),
reason: z.string().max(200).optional(),
});
export const BrowserInteractOutput = z.object({
ok: z.literal(true),
controller: BrowserController,
/** true = 页面可能已变化,下一步先 browser.snapshot 再用新引用 */
needsSnapshot: z.boolean(),
});
export const BrowserDownloadInput = z.object({
sessionId: Id,
tabId: z.string().min(1).optional(),
/** 触发下载的元素(快照引用);缺省只等待已触发的下载 */
target: BrowserAgentTarget.optional(),
timeoutMs: z.number().int().min(500).max(120_000).default(15_000),
});
export const BrowserDownloadOutput = z.object({
artifactId: Id,
filename: z.string(),
sizeBytes: NonNegInt,
sha256: z.string(),
verification: BrowserDownloadVerification,
/** 匹配到的本地资源;下一步用 download.plan / download.submit 走同一提交流程 */
sourceReleaseId: Id.nullable(),
duplicate: z.boolean(),
note: z.string().nullable(),
});
/** 能力 ID → 输入/输出 schema(server 注册处理器、agent 生成工具定义时共用) */
export const CAPABILITY_IO = {
"catalog.lookup": { input: CatalogLookupInput, output: CatalogLookupOutput },
@@ -341,6 +429,10 @@ export const CAPABILITY_IO = {
"site.health_check": { input: SiteHealthCheckInput, output: SiteHealthCheckOutput },
"site.refresh_login": { input: SiteLoginRefreshInput, output: SiteLoginRefreshOutput },
"site.user_data": { input: SiteUserDataInput, output: SiteUserDataOutput },
"browser.snapshot": { input: BrowserSnapshotInput, output: BrowserSnapshotOutput },
"browser.navigate": { input: BrowserNavigateInput, output: BrowserNavigateOutput },
"browser.interact": { input: BrowserInteractInput, output: BrowserInteractOutput },
"browser.download": { input: BrowserDownloadInput, output: BrowserDownloadOutput },
} as const satisfies Record<KnownCapabilityId, { input: z.ZodType; output: z.ZodType }>;
type CapabilityIo = typeof CAPABILITY_IO;
@@ -404,3 +496,11 @@ export type SiteLoginRefreshOutput = z.infer<typeof SiteLoginRefreshOutput>;
export type SiteUserDataInput = z.infer<typeof SiteUserDataInput>;
export type SiteUserDataOutput = z.infer<typeof SiteUserDataOutput>;
export type SubagentBudget = z.infer<typeof SubagentBudget>;
export type BrowserSnapshotInput = z.infer<typeof BrowserSnapshotInput>;
export type BrowserSnapshotOutput = z.infer<typeof BrowserSnapshotOutput>;
export type BrowserNavigateInput = z.infer<typeof BrowserNavigateInput>;
export type BrowserNavigateOutput = z.infer<typeof BrowserNavigateOutput>;
export type BrowserInteractInput = z.infer<typeof BrowserInteractInput>;
export type BrowserInteractOutput = z.infer<typeof BrowserInteractOutput>;
export type BrowserDownloadInput = z.infer<typeof BrowserDownloadInput>;
export type BrowserDownloadOutput = z.infer<typeof BrowserDownloadOutput>;
+50
View File
@@ -72,6 +72,10 @@ export const CAPABILITY_IDS = [
"site.health_check",
"site.refresh_login",
"site.user_data",
"browser.snapshot",
"browser.navigate",
"browser.interact",
"browser.download",
] as const;
export const KnownCapabilityId = z.enum(CAPABILITY_IDS);
export type KnownCapabilityId = z.infer<typeof KnownCapabilityId>;
@@ -348,6 +352,52 @@ export const CAPABILITY_CATALOG: Record<KnownCapabilityId, CapabilityDescriptor>
true,
false,
),
// C1-E 浏览器能力(REQ §4.4、§4.7.3):经 BrowserProvider 访问用户已有的 camofox-browser;
// 导航限站点域名,凭证由服务端注入,人工接管期间自动化动作被拒;下载产物只进后端暂存核验
"browser.snapshot": d(
"browser.snapshot",
"读取原站页面",
"读取站点浏览器会话当前页面的结构化快照(可访问性文本 + eN 元素引用,凭证已遮蔽);控制权交还或导航后必须重新读取",
"read",
"BrowserSnapshot",
"read",
"low",
true,
false,
),
"browser.navigate": d(
"browser.navigate",
"打开原站页面",
"在站点浏览器会话里打开站点域名内的地址(没有会话时按站点打开);受站点限流、超时与重试预算约束",
"execute",
"BrowserNavigate",
"admin",
"low",
false,
false,
),
"browser.interact": d(
"browser.interact",
"操作原站页面",
"在站点浏览器会话里点击 / 输入 / 选择 / 滚动 / 按键;登录字段用 credential 由服务端注入凭证;验证码或二次认证用 request_human 转人工",
"execute",
"BrowserInteract",
"admin",
"medium",
false,
false,
),
"browser.download": d(
"browser.download",
"捕获原站下载",
"捕获浏览器会话里的种子下载,交后端暂存并核验 infohash、匹配本地资源;提交下载仍用 download.plan / download.submit",
"execute",
"BrowserDownload",
"organize.execute",
"medium",
false,
false,
),
};
function d(