# browser-agent-driver - Doramagic AI Context Pack > 定位：安装前体验与判断资产。它帮助宿主 AI 有一个好的开始，但不代表已经安装、执行或验证目标项目。 ## 充分原则 - **充分原则，不是压缩原则**：AI Context Pack 应该充分到让宿主 AI 在开工前理解项目价值、能力边界、使用入口、风险和证据来源；它可以分层组织，但不以最短摘要为目标。 - **压缩策略**：只压缩噪声和重复内容，不压缩会影响判断和开工质量的上下文。 ## 给宿主 AI 的使用方式 你正在读取 Doramagic 为 browser-agent-driver 编译的 AI Context Pack。请把它当作开工前上下文：帮助用户理解适合谁、能做什么、如何开始、哪些必须安装后验证、风险在哪里。不要声称你已经安装、运行或执行了目标项目。 ## Claim 消费规则 - **事实来源**：Repo Evidence + Claim/Evidence Graph；Human Wiki 只提供显著性、术语和叙事结构。 - **事实最低状态**：`supported` - `supported`：可以作为项目事实使用，但回答中必须引用 claim_id 和证据路径。 - `weak`：只能作为低置信度线索，必须要求用户继续核实。 - `inferred`：只能用于风险提示或待确认问题，不能包装成项目事实。 - `unverified`：不得作为事实使用，应明确说证据不足。 - `contradicted`：必须展示冲突来源，不得替用户强行选择一个版本。 ## 它最适合谁 - **希望把专业流程带进宿主 AI 的用户**：仓库包含 Skill 文档。 证据：`skills/agent-friendly-app-design/SKILL.md`, `skills/browser-agent-driver-testing/SKILL.md`, `skills/design-evolve/SKILL.md`, `skills/domain/amazon.com/SKILL.md` 等 Claim：`clm_0003` supported 0.86 ## 它能做什么 - **AI Skill / Agent 指令资产库**（可做安装前预览）：项目包含可被宿主 AI 读取的 Skill 或 Agent 指令文件，可用于把专业流程带入 Claude、Codex、Cursor 等宿主。 证据：`skills/agent-friendly-app-design/SKILL.md`, `skills/browser-agent-driver-testing/SKILL.md`, `skills/design-evolve/SKILL.md`, `skills/domain/amazon.com/SKILL.md` 等 Claim：`clm_0001` supported 0.86 - **命令行启动或安装流程**（需要安装后验证）：项目文档中存在可执行命令，真实使用需要在本地或宿主环境中运行这些命令。 证据：`docs/COMPETITIVE-EVAL.md`, `docs/guides/design-audit.md`, `scripts/install.sh` Claim：`clm_0002` supported 0.86 ## 怎么开始 - `npm install && npm run dev -- --port $port &` 证据：`docs/guides/design-audit.md` Claim：`clm_0004` supported 0.86 - `pip install browser-use playwright` 证据：`docs/COMPETITIVE-EVAL.md` Claim：`clm_0005` supported 0.86 - `pnpm add -g @browserbasehq/stagehand` 证据：`docs/COMPETITIVE-EVAL.md` Claim：`clm_0006` supported 0.86 - `curl -fSL --progress-bar -o "${TMPDIR}/${TARBALL}" "$URL" \` 证据：`scripts/install.sh` Claim：`clm_0007` unverified 0.25 - `curl -fsSL -o "${TMPDIR}/checksum" "$CHECKSUM_URL" 2>/dev/null && {` 证据：`scripts/install.sh` Claim：`clm_0008` unverified 0.25 - `npx playwright install chromium 2>&1 | tail -1 || warn "Playwright install failed — run manually: npx playwright install chromium"` 证据：`scripts/install.sh` Claim：`clm_0009` unverified 0.25 ## 继续前判断卡 - **当前建议**：需要管理员/安全审批 - **为什么**：继续前可能涉及密钥、账号、外部服务或敏感上下文，建议先经过管理员或安全审批。 ### 30 秒判断 - **现在怎么做**：需要管理员/安全审批 - **最小安全下一步**：先跑 Prompt Preview；若涉及凭证或企业环境，先审批再试装 - **先别相信**：真实输出质量不能在安装前相信。 - **继续会触碰**：命令执行、宿主 AI 配置、本地环境或项目文件 ### 现在可以相信 - **适合人群线索：希望把专业流程带进宿主 AI 的用户**（supported）：有 supported claim 或项目证据支撑，但仍不等于真实安装效果。 证据：`skills/agent-friendly-app-design/SKILL.md`, `skills/browser-agent-driver-testing/SKILL.md`, `skills/design-evolve/SKILL.md`, `skills/domain/amazon.com/SKILL.md` 等 Claim：`clm_0003` supported 0.86 - **能力存在：AI Skill / Agent 指令资产库**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`skills/agent-friendly-app-design/SKILL.md`, `skills/browser-agent-driver-testing/SKILL.md`, `skills/design-evolve/SKILL.md`, `skills/domain/amazon.com/SKILL.md` 等 Claim：`clm_0001` supported 0.86 - **能力存在：命令行启动或安装流程**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`docs/COMPETITIVE-EVAL.md`, `docs/guides/design-audit.md`, `scripts/install.sh` Claim：`clm_0002` supported 0.86 - **存在 Quick Start / 安装命令线索**（supported）：可以相信项目文档出现过启动或安装入口；不要因此直接在主力环境运行。 证据：`docs/guides/design-audit.md` Claim：`clm_0004` supported 0.86 ### 现在还不能相信 - **真实输出质量不能在安装前相信。**（unverified）：Prompt Preview 只能展示引导方式，不能证明真实项目中的结果质量。 - **宿主 AI 版本兼容性不能在安装前相信。**（unverified）：Claude、Cursor、Codex、Gemini 等宿主加载规则和版本差异必须在真实环境验证。 - **不会污染现有宿主 AI 行为，不能直接相信。**（inferred）：Skill、plugin、AGENTS/CLAUDE/GEMINI 指令可能改变宿主 AI 的默认行为。 证据：`CLAUDE.md`, `skills/agent-friendly-app-design/SKILL.md`, `skills/browser-agent-driver-testing/SKILL.md`, `skills/design-evolve/SKILL.md` 等 - **可安全回滚不能默认相信。**（unverified）：除非项目明确提供卸载和恢复说明，否则必须先在隔离环境验证。 - **真实安装后是否与用户当前宿主 AI 版本兼容？**（unverified）：兼容性只能通过实际宿主环境验证。 - **项目输出质量是否满足用户具体任务？**（unverified）：安装前预览只能展示流程和边界，不能替代真实评测。 - **安装命令是否需要网络、权限或全局写入？**（unverified）：这影响企业环境和个人环境的安装风险。 证据：`docs/guides/design-audit.md` ### 继续会触碰什么 - **命令执行**：包管理器、网络下载、本地插件目录、项目配置或用户主目录。 原因：运行第一条命令就可能产生环境改动；必须先判断是否值得跑。 证据：`docs/COMPETITIVE-EVAL.md`, `docs/guides/design-audit.md`, `scripts/install.sh` - **宿主 AI 配置**：Claude/Codex/Cursor/Gemini/OpenCode 等宿主的 plugin、Skill 或规则加载配置。 原因：宿主配置会改变 AI 后续工作方式，可能和用户已有规则冲突。 证据：`CLAUDE.md`, `skills/agent-friendly-app-design/SKILL.md`, `skills/browser-agent-driver-testing/SKILL.md`, `skills/design-evolve/SKILL.md` 等 - **本地环境或项目文件**：安装结果、插件缓存、项目配置或本地依赖目录。 原因：安装前无法证明写入范围和回滚方式，需要隔离验证。 证据：`docs/COMPETITIVE-EVAL.md`, `docs/guides/design-audit.md`, `scripts/install.sh` - **环境变量 / API Key**：项目入口文档明确出现 API key、token、secret 或账号凭证配置。 原因：如果真实安装需要凭证，应先使用测试凭证并经过权限/合规判断。 证据：`.evolve/critical-audit/2026-04-27T08-14-37Z/summary.md`, `CLAUDE.md`, `discovery/tangle-agent.json`, `docs/COMPETITIVE-EVAL.md` 等 - **宿主 AI 上下文**：AI Context Pack、Prompt Preview、Skill 路由、风险规则和项目事实。 原因：导入上下文会影响宿主 AI 后续判断，必须避免把未验证项包装成事实。 ### 最小安全下一步 - **先跑 Prompt Preview**：用安装前交互式试用判断工作方式是否匹配，不需要授权或改环境。（适用：任何项目都适用，尤其是输出质量未知时。） - **只在隔离目录或测试账号试装**：避免安装命令污染主力宿主 AI、真实项目或用户主目录。（适用：存在命令执行、插件配置或本地写入线索时。） - **先备份宿主 AI 配置**：Skill、plugin、规则文件可能改变 Claude/Cursor/Codex 的默认行为。（适用：存在插件 manifest、Skill 或宿主规则入口时。） - **不要使用真实生产凭证**：环境变量/API key 一旦进入宿主或工具链，可能产生账号和合规风险。（适用：出现 API、TOKEN、KEY、SECRET 等环境线索时。） - **安装后只验证一个最小任务**：先验证加载、兼容、输出质量和回滚，再决定是否深用。（适用：准备从试用进入真实工作流时。） ### 退出方式 - **保留安装前状态**：记录原始宿主配置和项目状态，后续才能判断是否可恢复。 - **准备移除宿主 plugin / Skill / 规则入口**：如果试装后行为异常，可以把宿主 AI 恢复到试装前状态。 - **记录安装命令和写入路径**：没有明确卸载说明时，至少要知道哪些目录或配置需要手动清理。 - **准备撤销测试 API key 或 token**：测试凭证泄露或误用时，可以快速止损。 - **如果没有回滚路径，不进入主力环境**：不可回滚是继续前阻断项，不应靠信任或运气继续。 ## 哪些只能预览 - 解释项目适合谁和能做什么 - 基于项目文档演示典型对话流程 - 帮助用户判断是否值得安装或继续研究 ## 哪些必须安装后验证 - 真实安装 Skill、插件或 CLI - 执行脚本、修改本地文件或访问外部服务 - 验证真实输出质量、性能和兼容性 ## 边界与风险判断卡 - **把安装前预览误认为真实运行**：用户可能高估项目已经完成的配置、权限和兼容性验证。 处理方式：明确区分 prompt_preview_can_do 与 runtime_required。 Claim：`clm_0010` inferred 0.45 - **命令执行会修改本地环境**：安装命令可能写入用户主目录、宿主插件目录或项目配置。 处理方式：先在隔离环境或测试账号中运行。 证据：`docs/COMPETITIVE-EVAL.md`, `docs/guides/design-audit.md`, `scripts/install.sh` Claim：`clm_0011` supported 0.86 - **待确认**：真实安装后是否与用户当前宿主 AI 版本兼容？。原因：兼容性只能通过实际宿主环境验证。 - **待确认**：项目输出质量是否满足用户具体任务？。原因：安装前预览只能展示流程和边界，不能替代真实评测。 - **待确认**：安装命令是否需要网络、权限或全局写入？。原因：这影响企业环境和个人环境的安装风险。 ## 开工前工作上下文 ### 加载顺序 - 先读取 how_to_use.host_ai_instruction，建立安装前判断资产的边界。 - 读取 claim_graph_summary，确认事实来自 Claim/Evidence Graph，而不是 Human Wiki 叙事。 - 再读取 intended_users、capabilities 和 quick_start_candidates，判断用户是否匹配。 - 需要执行具体任务时，优先查 role_skill_index，再查 evidence_index。 - 遇到真实安装、文件修改、网络访问、性能或兼容性问题时，转入 risk_card 和 boundaries.runtime_required。 ### 任务路由 - **AI Skill / Agent 指令资产库**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`skills/agent-friendly-app-design/SKILL.md`, `skills/browser-agent-driver-testing/SKILL.md`, `skills/design-evolve/SKILL.md`, `skills/domain/amazon.com/SKILL.md` 等 Claim：`clm_0001` supported 0.86 - **命令行启动或安装流程**：先说明这是安装后验证能力，再给出安装前检查清单。 边界：必须真实安装或运行后验证。 证据：`docs/COMPETITIVE-EVAL.md`, `docs/guides/design-audit.md`, `scripts/install.sh` Claim：`clm_0002` supported 0.86 ### 上下文规模 - 文件总数：578 - 重要文件覆盖：40/578 - 证据索引条目：80 - 角色 / Skill 条目：9 ### 证据不足时的处理 - **missing_evidence**：说明证据不足，要求用户提供目标文件、README 段落或安装后验证记录；不要补全事实。 - **out_of_scope_request**：说明该任务超出当前 AI Context Pack 证据范围，并建议用户先查看 Human Manual 或真实安装后验证。 - **runtime_request**：给出安装前检查清单和命令来源，但不要替用户执行命令或声称已执行。 - **source_conflict**：同时展示冲突来源，标记为待核实，不要强行选择一个版本。 ## Prompt Recipes ### 适配判断 - 目标：判断这个项目是否适合用户当前任务。 - 预期输出：适配结论、关键理由、证据引用、安装前可预览内容、必须安装后验证内容、下一步建议。 ```text 请基于 browser-agent-driver 的 AI Context Pack，先问我 3 个必要问题，然后判断它是否适合我的任务。回答必须包含：适合谁、能做什么、不能做什么、是否值得安装、证据来自哪里。所有项目事实必须引用 evidence_refs、source_paths 或 claim_id。 ``` ### 安装前体验 - 目标：让用户在安装前感受核心工作流，同时避免把预览包装成真实能力或营销承诺。 - 预期输出：一段带边界标签的体验剧本、安装后验证清单和谨慎建议；不含真实运行承诺或强营销表述。 ```text 请把 browser-agent-driver 当作安装前体验资产，而不是已安装工具或真实运行环境。 请严格输出四段： 1. 先问我 3 个必要问题。 2. 给出一段“体验剧本”：用 [安装前可预览]、[必须安装后验证]、[证据不足] 三种标签展示它可能如何引导工作流。 3. 给出安装后验证清单：列出哪些能力只有真实安装、真实宿主加载、真实项目运行后才能确认。 4. 给出谨慎建议：只能说“值得继续研究/试装”“先补充信息后再判断”或“不建议继续”，不得替项目背书。 硬性边界： - 不要声称已经安装、运行、执行测试、修改文件或产生真实结果。 - 不要写“自动适配”“确保通过”“完美适配”“强烈建议安装”等承诺性表达。 - 如果描述安装后的工作方式，必须使用“如果安装成功且宿主正确加载 Skill，它可能会……”这种条件句。 - 体验剧本只能写成“示例台词/假设流程”：使用“可能会询问/可能会建议/可能会展示”，不要写“已写入、已生成、已通过、正在运行、正在生成”。 - Prompt Preview 不负责给安装命令；如用户准备试装，只能提示先阅读 Quick Start 和 Risk Card，并在隔离环境验证。 - 所有项目事实必须来自 supported claim、evidence_refs 或 source_paths；inferred/unverified 只能作风险或待确认项。 ``` ### 角色 / Skill 选择 - 目标：从项目里的角色或 Skill 中挑选最匹配的资产。 - 预期输出：候选角色或 Skill 列表，每项包含适用场景、证据路径、风险边界和是否需要安装后验证。 ```text 请读取 role_skill_index，根据我的目标任务推荐 3-5 个最相关的角色或 Skill。每个推荐都要说明适用场景、可能输出、风险边界和 evidence_refs。 ``` ### 风险预检 - 目标：安装或引入前识别环境、权限、规则冲突和质量风险。 - 预期输出：环境、权限、依赖、许可、宿主冲突、质量风险和未知项的检查清单。 ```text 请基于 risk_card、boundaries 和 quick_start_candidates，给我一份安装前风险预检清单。不要替我执行命令，只说明我应该检查什么、为什么检查、失败会有什么影响。 ``` ### 宿主 AI 开工指令 - 目标：把项目上下文转成一次对话开始前的宿主 AI 指令。 - 预期输出：一段边界明确、证据引用明确、适合复制给宿主 AI 的开工前指令。 ```text 请基于 browser-agent-driver 的 AI Context Pack，生成一段我可以粘贴给宿主 AI 的开工前指令。这段指令必须遵守 not_runtime=true，不能声称项目已经安装、运行或产生真实结果。 ``` ## 角色 / Skill 索引 - 共索引 9 个角色 / Skill / 项目文档条目。 - **Agent-Friendly App Design**（skill）：Use this skill when building product UIs that should be robust for autonomous browser agents while staying human-friendly. 激活提示：当用户任务与“Agent-Friendly App Design”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`skills/agent-friendly-app-design/SKILL.md` - **Agent Browser Driver Testing**（skill）：Use this skill when you need real, non-mocked browser-agent testing with reproducible artifacts. 激活提示：当用户任务与“Agent Browser Driver Testing”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`skills/browser-agent-driver-testing/SKILL.md` - **design-evolve**（skill）：- 激活提示：当用户任务与“design-evolve”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`skills/design-evolve/SKILL.md` - **Skill**（skill）：- Prefer the search box at the top. It's a searchbox role with ref near the page header. After typing the query, press Enter rather than clicking the "Go" button — the button is often hidden behind the keyboard and Enter submits reliably. - Cookie/consent banners only appear on EU/UK domains .co.uk , .de . If present, click the "Accept" or "Customize" button before any other interaction — Amazon's layout shifts afte… 激活提示：当用户任务与“Skill”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`skills/domain/amazon.com/SKILL.md` - **Skill**（skill）：- The top search bar is role="combobox" with aria-label="Search GitHub" . Typing a query and pressing Enter runs the global search. For in-repo search, use the "Go to file" shortcut: press t on a repo page and type the filename. - Repo pages have stable URL structure: /OWNER/REPO , /OWNER/REPO/pulls , /OWNER/REPO/issues , /OWNER/REPO/blob/BRANCH/PATH . Prefer direct URL navigation over clicking through the UI when t… 激活提示：当用户任务与“Skill”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`skills/domain/github.com/SKILL.md` - **Skill**（skill）：- The site aggressively detects automation. If you see an authwall "Join now to see" , the account is not logged in — abort unless the task explicitly expects this. Do not try to sign up programmatically. - Profile pages use dynamic React rendering. Wait for heading role with the person's name to appear before extracting — snapshot on first load is often skeleton UI. - The search bar is role="combobox" with placehol… 激活提示：当用户任务与“Skill”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`skills/domain/linkedin.com/SKILL.md` - **Skill**（skill）：- Question pages have a stable structure: question body, then answers sorted by vote count accepted answer floats to the top . The accepted answer has a green checkmark and aria-label="Accepted answer" . - Answer count appears in the sidebar sub-header "3 Answers" . Extract from that heading, not from counting elements — pinned/deleted answers can throw off the DOM count. - Code blocks inside answers are — their tex… 激活提示：当用户任务与“Skill”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`skills/domain/stackoverflow.com/SKILL.md` - **Skill**（skill）：- Article content is inside mw-content-text . Infoboxes are the table at the top-right of the article — use them for structured facts birth/death dates, coordinates, population, founding year rather than parsing prose. - The article lead first paragraph after the infobox is the highest-density summary; for one-fact queries it usually contains the answer. - When extracting a numeric fact for a goal like "what year wa… 激活提示：当用户任务与“Skill”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`skills/domain/wikipedia.org/SKILL.md` - **stealth-tuning**（skill）：Use when configuring browser stealth profiles, anti-bot evasion, WebDriver detection suppression, JA3 fingerprinting, or CDP leak patches. 激活提示：当用户任务与“stealth-tuning”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`skills/stealth-tuning/SKILL.md` ## 证据索引 - 共索引 80 条证据。 - **Changesets**（documentation）：This directory holds changesets https://github.com/changesets/changesets — small markdown files that describe what changed in a PR. They drive automated versioning and changelog generation. 证据：`.changeset/README.md` - **Skills Pack**（documentation）：This repository includes a versioned skills pack under skills/ . 证据：`skills/README.md` - **Macro candidates**（documentation）：Drop .json candidate files here. Each file is an agent-proposed macro plus its eval plan. pnpm macro:promote --candidate runs it through the eval-gated promotion pipeline: baseline no macro vs treatment macro registered , compared by pass rate / turns / cost / duration across N reps. 证据：`.evolve/candidates/macros/README.md` - **Competitive Benchmark — bad vs the field**（documentation）：Competitive Benchmark — bad vs the field 证据：`bench/competitive/README.md` - **Browser Scenario Suite SWE-bench Style**（documentation）：Browser Scenario Suite SWE-bench Style 证据：`bench/scenarios/README.md` - **WebBench Assets**（documentation）：Place downloaded WebBench CSV files here. 证据：`bench/webbench/README.md` - **Package**（package_manifest）：{ "name": "@tangle-network/browser-agent-driver", "version": "0.33.3", "description": "LLM-driven browser agent and bad CLI for UI automation, testing, and evaluation", "publishConfig": { "access": "public" }, "type": "module", "main": "./dist/index.js", "types": "./dist/index.d.ts", "bin": { "bad": "./dist/cli.js" }, "exports": { ".": { "types": "./dist/index.d.ts", "import": "./dist/index.js" } }, "files": "dist", "discovery", "scripts/postinstall-provider-patches.mjs", "README.md", "LICENSE" , "scripts": { "postinstall": "node ./scripts/postinstall-provider-patches.mjs", "build": "tsc && node ./scripts/copy-static-assets.mjs", "dev": "tsc --watch", "clean": "rm -rf dist", "lint": "tsc --… 证据：`package.json` - **CLAUDE.md**（documentation）：Browser Agent Driver bad CLI — general-purpose agentic browser automation. 证据：`CLAUDE.md` - **Agent-Friendly App Design**（skill_instruction）：Use this skill when building product UIs that should be robust for autonomous browser agents while staying human-friendly. 证据：`skills/agent-friendly-app-design/SKILL.md` - **Agent Browser Driver Testing**（skill_instruction）：Use this skill when you need real, non-mocked browser-agent testing with reproducible artifacts. 证据：`skills/browser-agent-driver-testing/SKILL.md` - **design-evolve — Closed-Loop Design Improvement**（skill_instruction）：design-evolve — Closed-Loop Design Improvement 证据：`skills/design-evolve/SKILL.md` - **Skill**（skill_instruction）：- Prefer the search box at the top. It's a searchbox role with ref near the page header. After typing the query, press Enter rather than clicking the "Go" button — the button is often hidden behind the keyboard and Enter submits reliably. - Cookie/consent banners only appear on EU/UK domains .co.uk , .de . If present, click the "Accept" or "Customize" button before any other interaction — Amazon's layout shifts after consent and @ref values change. - Product search results live under the "Search results" landmark. Each card has a link whose accessible name is the product title. Use click on the title link, not the image, to reach the product detail page. - On product detail pages, the canon… 证据：`skills/domain/amazon.com/SKILL.md` - **Skill**（skill_instruction）：- The top search bar is role="combobox" with aria-label="Search GitHub" . Typing a query and pressing Enter runs the global search. For in-repo search, use the "Go to file" shortcut: press t on a repo page and type the filename. - Repo pages have stable URL structure: /OWNER/REPO , /OWNER/REPO/pulls , /OWNER/REPO/issues , /OWNER/REPO/blob/BRANCH/PATH . Prefer direct URL navigation over clicking through the UI when the owner/repo is known — it's faster and avoids sidebar ambiguity. - PR and issue lists use infinite scroll but the first ~25 fit on one page. Don't scroll unless the task requires more than 25 results. - PR/issue counts: the tab badges "Pull requests 42", "Issues 7" carry the ex… 证据：`skills/domain/github.com/SKILL.md` - **Skill**（skill_instruction）：- The site aggressively detects automation. If you see an authwall "Join now to see" , the account is not logged in — abort unless the task explicitly expects this. Do not try to sign up programmatically. - Profile pages use dynamic React rendering. Wait for heading role with the person's name to appear before extracting — snapshot on first load is often skeleton UI. - The search bar is role="combobox" with placeholder "Search". After typing, press Enter; the dropdown suggestions can lead to unrelated pages if clicked. - Job listings: use extractWithIndex on the search results container class jobs-search-results-list or role list . Each item's title, company, and location are distinct child… 证据：`skills/domain/linkedin.com/SKILL.md` - **Skill**（skill_instruction）：- Question pages have a stable structure: question body, then answers sorted by vote count accepted answer floats to the top . The accepted answer has a green checkmark and aria-label="Accepted answer" . - Answer count appears in the sidebar sub-header "3 Answers" . Extract from that heading, not from counting elements — pinned/deleted answers can throw off the DOM count. - Code blocks inside answers are — their text content is preserved verbatim. If the task is "extract the code from the top answer", grab the first inside the answer body. - Vote counts are in a .js-vote-count or itemprop="upvoteCount" element next to each post. Use the accessible label "23 votes" rather than parsing the cl… 证据：`skills/domain/stackoverflow.com/SKILL.md` - **Skill**（skill_instruction）：- Article content is inside mw-content-text . Infoboxes are the table at the top-right of the article — use them for structured facts birth/death dates, coordinates, population, founding year rather than parsing prose. - The article lead first paragraph after the infobox is the highest-density summary; for one-fact queries it usually contains the answer. - When extracting a numeric fact for a goal like "what year was X founded", emit the number in the exact shape the goal asks for. If the goal implies a JSON object, wrap: {"year": 1815} , not the bare string "1815". Wikipedia extraction goals are the most common site where this formatting detail decides pass/fail. - The search box is at the… 证据：`skills/domain/wikipedia.org/SKILL.md` - **Stealth Tuning**（skill_instruction）：Use this skill when working on browser stealth configuration, anti-bot evasion, or detection resistance for the agent browser driver. 证据：`skills/stealth-tuning/SKILL.md` - **License**（source_file）：Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files the "Software" , to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: 证据：`LICENSE` - **Competitive Eval — head-to-head against the field**（documentation）：Competitive Eval — head-to-head against the field 证据：`docs/COMPETITIVE-EVAL.md` - **Eval Rigor — Canonical Validation Protocol**（documentation）：Eval Rigor — Canonical Validation Protocol 证据：`docs/EVAL-RIGOR.md` - **Gen 11 — Master Comparison Report**（documentation）：Date : 2026-04-09T06:07:43.489Z Generated by : scripts/run-master-comparison.mjs Output dir : agent-results/master-comparison-1775710102 Cost cap : $25 cumulative across tiers 证据：`docs/GEN11-MASTER-COMPARISON.md` - **bad extensions**（documentation）：Customize bad's agent loop and design audits without forking the codebase. 证据：`docs/extensions.md` - **Benchmarks & Experiments**（documentation）：Tier Scope Gate threshold ------ ------- ---------------- Tier 1 Deterministic local fixtures 100% required Tier 2 Authenticated staging flows Push to 100% Tier 3 Open web WebBench-50 Track separately, no Tier 1/2 regression 证据：`docs/guides/benchmarks.md` - **Browser Bridges & CDP Connection**（documentation）：Connect bad to existing browser instances — use their logged-in sessions, extensions, and AI agent capabilities. 证据：`docs/guides/browser-bridges.md` - **CLI Reference**（documentation）：bash single task bad run --goal "Sign up for account" --url http://localhost:3000 证据：`docs/guides/cli.md` - **Configuration Reference**（documentation）：Create browser-agent-driver.config.ts in your project root: 证据：`docs/guides/configuration.md` - **Custom Drivers**（documentation）：Implement the Driver interface to use a non-Playwright browser backend: 证据：`docs/guides/custom-drivers.md` - **Design Audit Guide**（documentation）：bad design-audit is a vision-powered product and design quality analysis system. It loads pages in a real browser, captures screenshots, extracts design tokens from the DOM, infers the page's audience/job/stakes, and uses an LLM with vision to produce structured findings with concrete fixes. 证据：`docs/guides/design-audit.md` - **Memory System**（documentation）：Trajectory memory stores successful run recordings, domain-scoped knowledge, and session history, then injects them into subsequent runs as context for the LLM. Enabled by default. 证据：`docs/guides/memory.md` - **Providers**（documentation）：Default provider is openai with gpt-5.4 . 证据：`docs/guides/providers.md` - **Reporters & Sinks**（documentation）：typescript import { generateReport } from '@tangle-network/browser-agent-driver' 证据：`docs/guides/reporters.md` - **Wallet & EVM Application Testing**（documentation）：Test wallet-connected flows on EVM dApps — DeFi swaps, token approvals, lending, NFT mints — using a real browser extension MetaMask or Rabby against a local Anvil fork. 证据：`docs/guides/wallet.md` - **Browser-Agent Competitive Analysis 2026-03**（documentation）：Browser-Agent Competitive Analysis 2026-03 证据：`docs/research/competitor-analysis-2026-03.md` - **RFC-002: World-Class Design Audit — 8-Layer Architecture**（documentation）：RFC-002: World-Class Design Audit — 8-Layer Architecture 证据：`docs/rfc/design-audit-world-class.md` - **Browser Agent Ops**（documentation）：Canonical operating roadmap for browser-agent-driver . 证据：`docs/roadmap/browser-agent-ops.md` - **bad / bad-app — World-Class Spec & Priority Checklist**（documentation）：bad / bad-app — World-Class Spec & Priority Checklist 证据：`docs/roadmap/world-class-checklist.md` - **Cli Design Audit**（source_file）：import chalk from 'chalk' import { chromium, type Page } from 'playwright' import { Brain } from './brain/index.js' import type { DesignFinding, DesignEvolveResult } from './types.js' import { PlaywrightDriver } from './drivers/playwright.js' import { resolveDefaultProvider, resolveProviderApiKey, resolveProviderModelName, isSupportedProvider, SUPPORTED PROVIDERS, type SupportedProvider } from './provider-defaults.js' import { loadLocalEnvFiles } from './env-loader.js' import { cliError } from './cli-ui.js' import { auditOnePage } from './design/audit/pipeline.js' import type { PageAuditResult as Gen2PageAuditResult, EthicsViolation } from './design/audit/types.js' import { extractDesignTok… 证据：`src/cli-design-audit.ts` - **Cli Preview**（source_file）：import chalk from 'chalk' import type { Plan } from './types.js' export class PreviewError extends Error ⋮---- constructor message: string ⋮---- export interface PreviewOptions { goal: string url: string model?: string provider?: string apiKey?: string baseUrl?: string output?: string json?: boolean maxSteps?: number headed?: boolean } export interface PreviewResult { goal: string url: string plan: Plan null raw: string parseError?: string tokensUsed?: number durationMs: number } async function runPreview opts: PreviewOptions : Promise function renderPreview result: PreviewResult : void export async function handlePreviewCommand opts: PreviewOptions : Promise function indent text: string, p… 证据：`src/cli-preview.ts` - **Cli Showcase**（source_file）：import { runShowcase, quickCapture } from './showcase/index.js' import type { ShowcaseConfig } from './showcase/types.js' import { cliError } from './cli-ui.js' export interface ShowcaseCliArgs { url?: string script?: string capture?: string crop?: string highlight?: string format?: string viewport?: string output?: string headless: boolean colorScheme?: 'dark' 'light' scale?: number storageState?: string quality?: number } export async function handleShowcase args: ShowcaseCliArgs : Promise function printResult result: import './showcase/types.js' .ShowcaseResult : void function parseViewport vp?: string : 证据：`src/cli-showcase.ts` - **Design Audit**（source_file）：import type { Driver } from './drivers/types.js'; import { Brain } from './brain/index.js'; import { BrowserAgent } from './runner.js'; import type { AgentConfig, AuditFlow, DesignFinding, FlowAuditResult, DesignAuditReport, } from './types.js'; export class DesignAuditor ⋮---- constructor driver: Driver, config: AgentConfig = async auditFlow flow: AuditFlow : Promise async audit flows: AuditFlow , options?: { onFlowComplete?: flowName: string, index: number, total: number = void } : Promise ⋮---- export function generateDesignAuditReport report: DesignAuditReport : string ⋮---- // Summary ⋮---- // Per-flow results ⋮---- const esc = s: string 证据：`src/design-audit.ts` - **Preview**（source_file）：import type { Page } from 'playwright'; import { AriaSnapshotHelper } from './drivers/snapshot.js'; import type { PreviewVerification } from './types.js'; async function extractPreviewUrl page: Page, customSelector?: string : Promise export async function verifyPreview page: Page, snapshot: AriaSnapshotHelper, options?: { captureScreenshot?: boolean; screenshotQuality?: number; previewUrl?: string; iframeSelector?: string; } : Promise ⋮---- function isVisible el: HTMLElement : boolean ⋮---- // Vite error overlay uses a custom element with Shadow DOM — // textContent is empty, so just detect its presence 证据：`src/preview.ts` - **Run**（source_file）：import { fileURLToPath } from 'node:url' import { evaluateCalibration } from './calibration.js' import { evaluateReproducibility } from './reproducibility.js' import { evaluatePatches } from './patches.js' import { emptyScorecard, summarize, type DesignAuditScorecard } from './scorecard.js' import type { Corpus } from './calibration.js' ⋮---- interface CliArgs { tier?: string reps: number urls?: string calibrationOnly: boolean reproOnly: boolean patchesOnly: boolean roots: string outDir: string generation: number provider?: string model?: string baseUrl?: string writeScorecardPath?: string } function parseArgs argv: string : CliArgs async function main : Promise function appendToProjectScor… 证据：`bench/design/eval/run.ts` - **Run**（source_file）：import { parseArgs } from 'node:util' ⋮---- import { randomUUID } from 'node:crypto' ⋮---- import { Brain } from '../../../src/brain/index.js' import { loadLocalEnvFiles } from '../../../src/env-loader.js' import { resolveProviderApiKey, resolveProviderModelName, type SupportedProvider } from '../../../src/provider-defaults.js' import { setCliVersion, setInvocation, getTelemetry } from '../../../src/telemetry/index.js' import { loadFixtures, selectFixtures } from './fixtures/loader.js' import { runGepaLoop, type GepaConfig } from './loop.js' import { AuditScoreAdapter } from './score-adapter.js' import { DeterministicMutator, ReflectiveMutator } from './mutators.js' import { KNOWN TARGETS,… 证据：`bench/design/gepa/run.ts` - **Design Audit**（source_file）：import type { ModelMessage, SystemModelMessage } from 'ai'; import type { PageState, DesignFinding } from '../../types.js'; import { DESIGN AUDIT PROMPT } from '../prompts.js'; import type { UserContent } from '../types.js'; import type { ModelSelection, GenerateResult } from '../model-client.js'; export interface BrainDesignAuditHost { debug: boolean; buildUserContent text: string, screenshot?: string, forceVision?: boolean : UserContent; generate system: string SystemModelMessage , messages: ModelMessage , selection?: ModelSelection, maxOutputTokens?: number, : Promise ; } ⋮---- buildUserContent text: string, screenshot?: string, forceVision?: boolean : UserContent; generate system: strin… 证据：`src/brain/tasks/design-audit.ts` - **Evaluate**（source_file）：import type { ModelMessage, SystemModelMessage } from 'ai'; import type { PageState } from '../../types.js'; import { EVALUATE PROMPT } from '../prompts.js'; import type { QualityEvaluation, UserContent } from '../types.js'; import type { ModelSelection, GenerateResult } from '../model-client.js'; export interface BrainEvaluateHost { debug: boolean; buildUserContent text: string, screenshot?: string, forceVision?: boolean : UserContent; generate system: string SystemModelMessage , messages: ModelMessage , selection?: ModelSelection, maxOutputTokens?: number, : Promise ; } ⋮---- buildUserContent text: string, screenshot?: string, forceVision?: boolean : UserContent; generate system: string S… 证据：`src/brain/tasks/evaluate.ts` - **Goal Verification**（source_file）：import type { ModelMessage, SystemModelMessage } from 'ai'; import type { PageState, GoalVerification } from '../../types.js'; import { buildFirstPartyBoundaryNote } from '../../domain-policy.js'; import { budgetSnapshot } from '../snapshot-budget.js'; import type { UserContent } from '../types.js'; import type { BrainProvider, ModelSelection, GenerateResult } from '../model-client.js'; export interface BrainGoalVerificationHost { provider: BrainProvider; debug: boolean; adaptiveModelRouting: boolean; navProvider?: BrainProvider; navModelName?: string; verifierProvider?: string; verifierModel?: string; buildUserContent text: string, screenshot?: string, forceVision?: boolean : UserContent;… 证据：`src/brain/tasks/goal-verification.ts` - **Knowledge**（source_file）：import type { ModelMessage, SystemModelMessage } from 'ai'; import type { ModelSelection, GenerateResult } from '../model-client.js'; export interface BrainKnowledgeHost { generate system: string SystemModelMessage , messages: ModelMessage , selection?: ModelSelection, maxOutputTokens?: number, : Promise ; } ⋮---- generate system: string SystemModelMessage , messages: ModelMessage , selection?: ModelSelection, maxOutputTokens?: number, : Promise ; ⋮---- export async function extractKnowledgeImpl self: BrainKnowledgeHost, trajectoryText: string, domain: string, : Promise