# agent-opfor - 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 agent-opfor 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

- **正在使用 Claude/Codex/Cursor/Gemini 等宿主 AI 的开发者**：README 或插件配置提到多个宿主 AI。 证据：`README.md` Claim：`clm_0003` supported 0.86
- **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-redteaming/opfor-run/SKILL.md`, `skills/agent-redteaming/opfor-setup/SKILL.md`, `skills/mcp-redteaming/opfor-run/SKILL.md`, `skills/mcp-redteaming/opfor-setup/SKILL.md` Claim：`clm_0004` 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-redteaming/opfor-run/SKILL.md`, `skills/agent-redteaming/opfor-setup/SKILL.md`, `skills/mcp-redteaming/opfor-run/SKILL.md`, `skills/mcp-redteaming/opfor-setup/SKILL.md` 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：`AGENTS.md`, `README.md` Claim：`clm_0002` supported 0.86

## 怎么开始

- `npm install -g @keyvaluesystems/agent-opfor-cli` 证据：`README.md` Claim：`clm_0005` supported 0.86
- `npm install                       # workspaces resolved + Husky pre-commit hooks installed` 证据：`AGENTS.md` Claim：`clm_0006` supported 0.86

## 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**：工具权限边界不能在安装前相信。
- **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：正在使用 Claude/Codex/Cursor/Gemini 等宿主 AI 的开发者**（supported）：有 supported claim 或项目证据支撑，但仍不等于真实安装效果。 证据：`README.md` Claim：`clm_0003` supported 0.86
- **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-redteaming/opfor-run/SKILL.md`, `skills/agent-redteaming/opfor-setup/SKILL.md`, `skills/mcp-redteaming/opfor-run/SKILL.md`, `skills/mcp-redteaming/opfor-setup/SKILL.md` Claim：`clm_0004` 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-redteaming/opfor-run/SKILL.md`, `skills/agent-redteaming/opfor-setup/SKILL.md`, `skills/mcp-redteaming/opfor-run/SKILL.md`, `skills/mcp-redteaming/opfor-setup/SKILL.md` 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：`AGENTS.md`, `README.md` 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：`README.md` Claim：`clm_0005` supported 0.86

### 现在还不能相信

- **工具权限边界不能在安装前相信。**（unverified）：MCP/tool 类项目通常会触碰文件、网络、浏览器或外部 API，必须真实检查权限和日志。
- **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：`AGENTS.md`, `CLAUDE.md`, `skills/agent-redteaming/opfor-run/SKILL.md`, `skills/agent-redteaming/opfor-setup/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：`README.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：`AGENTS.md`, `README.md`
- **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：`AGENTS.md`, `CLAUDE.md`, `skills/agent-redteaming/opfor-run/SKILL.md`, `skills/agent-redteaming/opfor-setup/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：`AGENTS.md`, `README.md`
- **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：`AGENTS.md`, `CONTRIBUTING.md`, `README.md`, `docs/cli.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_0007` 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：`AGENTS.md`, `README.md` Claim：`clm_0008` 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-redteaming/opfor-run/SKILL.md`, `skills/agent-redteaming/opfor-setup/SKILL.md`, `skills/mcp-redteaming/opfor-run/SKILL.md`, `skills/mcp-redteaming/opfor-setup/SKILL.md` 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：`AGENTS.md`, `README.md` Claim：`clm_0002` supported 0.86

### 上下文规模

- 文件总数：979
- 重要文件覆盖：40/979
- 证据索引条目：78
- 角色 / Skill 条目：4

### 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 agent-opfor 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 agent-opfor 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 agent-opfor 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 4 个角色 / Skill / 项目文档条目。

- **opfor-run**（skill）：Run red-team attacks and generate a report for an agent target. 激活提示：当用户任务与“opfor-run”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-redteaming/opfor-run/SKILL.md`
- **opfor-setup**（skill）：Set up an agent or chatbot target for Opfor red-teaming. 激活提示：当用户任务与“opfor-setup”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-redteaming/opfor-setup/SKILL.md`
- **opfor-mcp-run**（skill）：Run red-team attacks and generate a report for an MCP server target. 激活提示：当用户任务与“opfor-mcp-run”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/mcp-redteaming/opfor-run/SKILL.md`
- **opfor-mcp-setup**（skill）：Set up an MCP server target for Opfor red-teaming. 激活提示：当用户任务与“opfor-mcp-setup”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/mcp-redteaming/opfor-setup/SKILL.md`

## 证据索引

- 共索引 78 条证据。

- **Why we built this**（documentation）：Open-source adversary emulation for AI agents, LLM apps, and MCP servers. Test your AI like a real attacker would — from your CLI, your IDE, or a browser extension that anyone on your team can use. 证据：`README.md`
- **Evaluators source of truth**（documentation）：Author evaluator and suite markdown here. The engine, CLI, and tests read these paths directly: 证据：`evaluators/README.md`
- **Suites**（documentation）：Suites reference evaluator IDs the id: in each evaluator's frontmatter , never file paths — so moving an evaluator between family folders never breaks a suite. 证据：`suites/README.md`
- **Access Control & Authorization**（documentation）：Add an evaluator here if the attack crosses an authorization boundary — other users' data, higher privilege, or role/function the requester shouldn't reach. Examples: rbac, bola, bfla, identity-privilege-abuse. 证据：`evaluators/agent/access-control/README.md`
- **Accuracy & Truthfulness**（documentation）：Add an evaluator here if the attack tests fabrication, false-premise acceptance, or over-reliance . Examples: misinformation, hallucination, overreliance. 证据：`evaluators/agent/accuracy/README.md`
- **Bias & Fairness**（documentation）：Add an evaluator here if the attack tests discriminatory or inconsistent treatment by a protected attribute. Examples: bias-age, bias-disability, bias-gender, bias-race one file each . 证据：`evaluators/agent/bias/README.md`
- **Brand & Conduct**（documentation）：Add an evaluator here if the attack pushes the agent off-policy, off-brand, or out of scope . Examples: imitation, competitors, contracts, off-topic. 证据：`evaluators/agent/brand-conduct/README.md`
- **Code & Output Execution**（documentation）：Add an evaluator here if the attack's harm is input or output reaching a code / SQL / shell / markup sink . Examples: improper-output-handling, shell-injection, sql-injection, unexpected-code-execution. 证据：`evaluators/agent/code-execution/README.md`
- **Information Disclosure**（documentation）：Add an evaluator here if the attack's goal is to make the agent EXPOSE protected information. Examples: system-prompt-leakage, sensitive-disclosure, pii-direct, pii-session, pii-api-db, debug-access. 证据：`evaluators/agent/disclosure/README.md`
- **Excessive Agency**（documentation）：Add an evaluator here if the attack makes the agent take an action beyond its remit — ungated, destructive, or out-of-scope tool use. Examples: excessive-agency, tool-misuse. 证据：`evaluators/agent/excessive-agency/README.md`
- **Harmful Content**（documentation）：Add an evaluator here if the attack's success is the agent emitting disallowed/dangerous content in a specific harm category. Examples: harmful-bioweapons, harmful-cybercrime-malicious-code, harmful-self-harm, harmful-violent-crime, harmful-weapons-ied, … one file per category . 证据：`evaluators/agent/harmful/README.md`
- **Prompt Injection & Jailbreak**（documentation）：Add an evaluator here if the attack's mechanism is getting the agent to follow attacker input it should have ignored or to bypass a guardrail . Examples: prompt-injection directory form , jailbreaking, ascii-smuggling, hijacking, agent-goal-hijack. 证据：`evaluators/agent/injection/README.md`
- **MCP Client Safety**（documentation）：Add an evaluator here if the attack tests the agent's behaviour when it uses MCP agent as client . Examples: mcp-tool-description-injection, mcp-shadow-server, mcp-missing-authentication, mcp-scope-escalation, mcp-credential-exposure, mcp-intent-subversion, … the mcp- agent set . 证据：`evaluators/agent/mcp-usage/README.md`
- **Memory & Knowledge Poisoning**（documentation）：Add an evaluator here if the attack corrupts the agent's memory, RAG, or embeddings so it misbehaves later. Examples: memory-poisoning, memory-inject-plant, memory-inject-trigger, data-poisoning, vector-embedding-weaknesses. 证据：`evaluators/agent/memory-rag/README.md`
- **Multi-Agent & Trust**（documentation）：Add an evaluator here if the attack targets agent-to-agent or agent-to-human trust in a multi-party setup. Examples: inter-agent-communication, rogue-agents, cascading-failures, human-agent-trust. 证据：`evaluators/agent/multi-agent/README.md`
- **Resource & Availability**（documentation）：Add an evaluator here if the attack exhausts compute/tokens/cost or degrades availability . Examples: unbounded-consumption, reasoning-dos. 证据：`evaluators/agent/resource/README.md`
- **Source White-box Analysis — skills only**（documentation）：Source White-box Analysis — skills only 证据：`evaluators/agent/source-analysis/README.md`
- **Supply Chain**（documentation）：Add an evaluator here if the attack enters through a dependency or third-party component model, plugin, dataset, package . Examples: supply-chain. 证据：`evaluators/agent/supply-chain/README.md`
- **Authentication & Authorization**（documentation）：Add an evaluator here if the attack tests the MCP server's auth/authz enforcement . Examples: missing-authentication, oauth-token-passthrough, scope-escalation. 证据：`evaluators/mcp/auth/README.md`
- **Information Disclosure**（documentation）：Add an evaluator here if the attack makes the MCP server expose protected data . Examples: secret-exposure, cross-resource-leakage, resource-exposure. 证据：`evaluators/mcp/disclosure/README.md`
- **Injection**（documentation）：Add an evaluator here if the attack drives MCP tool arguments into a shell/exec or network sink . Examples: command-injection, ssrf. 证据：`evaluators/mcp/injection/README.md`
- **Protocol & Telemetry**（documentation）：Add an evaluator here if the attack targets MCP protocol handling or observability . Examples: protocol-abuse, intent-subversion, timing-side-channel, audit-telemetry. 证据：`evaluators/mcp/protocol/README.md`
- **Source White-box Analysis — skills only**（documentation）：Source White-box Analysis — skills only 证据：`evaluators/mcp/source-analysis/README.md`
- **Supply Chain**（documentation）：Add an evaluator here if the attack concerns MCP server provenance/integrity — impersonation or post-approval drift. Examples: mcp-supply-chain, shadow-mcp-server. 证据：`evaluators/mcp/supply-chain/README.md`
- **Tool Poisoning**（documentation）：Add an evaluator here if the attack concerns poisoned MCP tool descriptions, schemas, or return values served by the MCP server. Examples: tool-description-injection, tool-description-scan, content-injection, return-value-injection. 证据：`evaluators/mcp/tool-poisoning/README.md`
- **@keyvaluesystems/agent-opfor-cli**（documentation）：Opfor CLI — adversarial security testing for AI agents and MCP servers. 证据：`runners/cli/README.md`
- **OPFOR Browser Extension**（documentation）：Chrome/Brave MV3 extension for red-teaming embedded web chat agents . 证据：`runners/extension/README.md`
- **@keyvaluesystems/agent-opfor-mcp**（documentation）：Opfor MCP server — expose red-team tools to any MCP-compatible AI agent Cursor, Claude Desktop, etc. . 证据：`runners/mcp/README.md`
- **@keyvaluesystems/agent-opfor-sdk**（documentation）：Opfor SDK — programmatic adversarial testing for AI systems. 证据：`runners/sdk/README.md`
- **SDK Examples in-repo**（documentation）：Runnable examples for @keyvaluesystems/agent-opfor-sdk from the monorepo. 证据：`runners/sdk/examples/README.md`
- **AGENTS.md — Opfor**（documentation）：This file is for AI coding agents Claude Code, Copilot, Cursor, etc. working in this repository. It describes the project structure, build system, key conventions, and how the core subsystems fit together. 证据：`AGENTS.md`
- **Claude**（documentation）：@AGENTS.md 证据：`CLAUDE.md`
- **Evaluator schema**（documentation）：Evaluators are YAML files under evaluators/{agent mcp}/ at repo root. After adding or editing, run npm run build:catalog to rebuild the skill catalogs. 证据：`docs/evaluator-schema.md`
- **Opfor — Evaluators and Suites**（documentation）：An evaluator is a single attack-and-judge pattern e.g. prompt-injection , bola . Each evaluator is a YAML file — the attacker LLM reads it to craft prompts and the judge LLM uses its pass/fail criteria to score responses. 证据：`docs/evaluators.md`
- **Package**（package_manifest）：{ "name": "@keyvaluesystems/agent-opfor-core", "version": "0.10.0", "description": "Opfor core engine — attacker prompt generation, judge, and execution shared by all runners", "license": "Apache-2.0", "private": true, "type": "module", "main": "./dist/index.js", "engines": { "node": " =20" }, "types": "./dist/index.d.ts", "exports": { ".": { "types": "./dist/index.d.ts", "default": "./dist/index.js" }, "./browser": { "types": "./dist/browser.d.ts", "default": "./dist/browser.js" }, "./execute/ .js": { "types": "./dist/execute/ .d.ts", "default": "./dist/execute/ .js" }, "./generate/ .js": { "types": "./dist/generate/ .d.ts", "default": "./dist/generate/ .js" }, "./targets/ .js": { "types":… 证据：`core/package.json`
- **Package**（package_manifest）：{ "name": "agent-opfor", "version": "0.10.0", "description": "Opfor — security testing for AI agents and MCP servers workspace root ", "license": "Apache-2.0", "private": true, "type": "module", "workspaces": "core", "runners/cli", "runners/mcp", "runners/extension", "runners/sdk", "tests/e2e/sdk" , "files": "evaluators/", "suites/", "skills/", "README.md", "LICENSE" , "engines": { "node": " =20" }, "scripts": { "build": "npm run build:catalog && tsc -b core && npm run build -w runners/cli && npm run build -w runners/mcp && npm run build -w runners/sdk && npm run extension:catalog && npm run build:core --workspace=@keyvaluesystems/agent-opfor-extension", "build:ui": "cd runners/cli/ui && np… 证据：`package.json`
- **Contributing to Opfor**（documentation）：Thanks for helping make AI red teaming better. 证据：`CONTRIBUTING.md`
- **Package**（package_manifest）：{ "name": "@keyvaluesystems/agent-opfor-cli", "version": "0.10.0", "description": "Opfor CLI — security testing for AI agents and MCP servers opfor setup run hunt ", "license": "Apache-2.0", "type": "module", "bin": { "opfor": "./dist/index.js" }, "files": "dist/", "data/", "evaluators/", "suites/", "atlas-data/" , "scripts": { "dev": "tsx src/index.ts", "build": "rm -rf dist && npm run build:ui && node scripts/bundle.mjs", "build:ui": "cd ui && npm install && npm run build", "start": "node dist/index.js", "typecheck": "tsc -p tsconfig.json --noEmit", "test": "node --import tsx/esm --test tests/ .test.ts", "prepack": "rm -rf ./evaluators ./suites ./data ./atlas-data ./LICENSE && cp -r ../..… 证据：`runners/cli/package.json`
- **Package**（package_manifest）：{ "name": "@keyvaluesystems/agent-opfor-autonomous-ui", "private": true, "version": "0.10.0", "type": "module", "scripts": { "dev": "vite", "build": "vite build", "preview": "vite preview" }, "dependencies": { "react": "^19.1.0", "react-dom": "^19.1.0" }, "devDependencies": { "@types/react": "^19.1.0", "@types/react-dom": "^19.1.0", "@vitejs/plugin-react": "^4.5.0", "typescript": "^5.8.3", "vite": "^6.3.0" } } 证据：`runners/cli/ui/package.json`
- **Package**（package_manifest）：{ "name": "@keyvaluesystems/agent-opfor-extension", "version": "0.10.0", "description": "Opfor browser extension MV3 — chat UI injector for live testing", "license": "Apache-2.0", "private": true, "type": "module", "scripts": { "build:catalog": "node scripts/build-catalog.mjs", "build:core": "node scripts/bundle-core.mjs", "ensure:core": "node scripts/ensure-core-bundle.mjs", "prebuild:core": "node scripts/ensure-core-bundle.mjs" }, "dependencies": { "yaml": "^2.8.3" }, "devDependencies": { "esbuild": "^0.28.1" } } 证据：`runners/extension/package.json`
- **Package**（package_manifest）：{ "name": "@keyvaluesystems/agent-opfor-mcp", "version": "0.10.0", "description": "Opfor MCP server — expose red team tools to any MCP-compatible AI agent", "license": "Apache-2.0", "type": "module", "bin": { "opfor-agent-mcp": "./dist/index.js" }, "files": "dist/", "evaluators/", "suites/", "atlas-data/" , "scripts": { "dev": "tsx src/index.ts", "build": "rm -rf dist && node scripts/bundle.mjs", "typecheck": "tsc -p tsconfig.json --noEmit", "prepare": "echo 'run npm run build from the repo root'", "start": "node dist/index.js", "prepack": "rm -rf ./evaluators ./suites ./atlas-data ./LICENSE && cp -r ../../evaluators ./evaluators && cp -r ../../suites ./suites && mkdir -p ./atlas-data && cp… 证据：`runners/mcp/package.json`
- **Package**（package_manifest）：{ "name": "@keyvaluesystems/agent-opfor-sdk", "version": "0.10.0", "description": "Opfor SDK — programmatic adversarial testing for AI systems", "license": "Apache-2.0", "type": "module", "main": "./dist/index.js", "types": "./dist/index.d.ts", "exports": { ".": { "types": "./dist/index.d.ts", "import": "./dist/index.js" } }, "files": "dist/", "data/", "evaluators/", "suites/", "atlas-data/" , "scripts": { "dev": "tsx src/index.ts", "build": "tsup", "typecheck": "tsc -p tsconfig.json --noEmit", "test": "node --import tsx/esm --test tests/ .test.ts", "prepack": "rm -rf ./evaluators ./suites ./data ./atlas-data ./LICENSE && cp -r ../../evaluators ./evaluators && cp -r ../../suites ./suites &&… 证据：`runners/sdk/package.json`
- **Opfor — assessment execution**（skill_instruction）：Execute an Opfor assessment using pre-generated attack inputs. The /opfor-setup skill generates all attack variations beforehand; this skill executes them, judges responses, and generates a report. 证据：`skills/agent-redteaming/opfor-run/SKILL.md`
- **Opfor — target configuration**（skill_instruction）：Configure a target for an Opfor assessment. When this skill is installed in the user’s repo , prefer reading that repo endpoints, opfor.config , env examples, telemetry before a long questionnaire; then collect only what is still unknown. 证据：`skills/agent-redteaming/opfor-setup/SKILL.md`
- **Opfor — MCP Assessment Execution**（skill_instruction）：Execute an MCP adversarial assessment using pre-generated attack inputs. The /opfor-mcp-setup skill generates all attack variations beforehand; this skill connects to the MCP server, executes attacks, judges responses, and generates a report. 证据：`skills/mcp-redteaming/opfor-run/SKILL.md`
- **Opfor — MCP Server Target Configuration**（skill_instruction）：Opfor — MCP Server Target Configuration 证据：`skills/mcp-redteaming/opfor-setup/SKILL.md`
- **License**（source_file）：Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ 证据：`LICENSE`
- **Opfor — Browser Extension**（documentation）：The no-code red-team path. Install the extension, open any chat interface, click the icon, pick a suite, watch it run. Aimed at PMs, QA, designers, and security analysts — anyone who can't or won't open a terminal. 证据：`docs/browser-extension.md`
- **Opfor — CLI**（documentation）：The CLI handles everything: interactive setup, attack generation, firing attacks, judging responses, and producing reports. 证据：`docs/cli.md`
- **Opfor Hunt — Autonomous Red-Teaming**（documentation）：Opfor Hunt — Autonomous Red-Teaming 证据：`docs/hunt.md`
- **Opfor — MCP Server**（documentation）：The MCP server exposes Opfor as tools that any MCP-compatible AI agent Cursor, Claude Desktop, Windsurf can call directly. No terminal required — the agent runs the full workflow from your chat. 证据：`docs/mcp.md`
- **Opfor SDK**（documentation）：Adversarial testing for AI systems. TypeScript-first. 证据：`docs/sdk.md`
- **Target Session Handling**（documentation）：How opfor delivers conversation context to an HTTP agent target across turns, and how it threads a session id. This applies to opfor run config file, wizard, SDK , the MCP server tool, and opfor hunt — they share one model. Local-script and browser-extension targets are covered at the end. 证据：`docs/sessions.md`
- **Opfor — Skills**（documentation）：Opfor ships Skills markdown instruction files an AI coding agent reads and follows for both agent and MCP server red-teaming. Install once, then trigger from chat inside your project. 证据：`docs/skills.md`
- **Opfor — Trace-aware testing**（documentation）：Using the SDK? See SDK telemetry sdk.md telemetry . 证据：`docs/telemetry.md`
- **Types**（source_file）：import type { Severity } from "../../evaluators/schema.js"; ⋮---- export interface VulnClass { id: string; name: string; severity: Severity; standards?: Record ; description: string; failRubric: string; passRubric: string; inspiration?: string; } ⋮---- export interface Persona { id: string; name: string; voice: string; traits: string; whenToUse: string; } ⋮---- export interface Strategy { id: string; name: string; mechanics: string; whenToUse: string; escalationNotes: string; } ⋮---- export interface KnowledgeBase { vulnClasses: VulnClass ; personas: Persona ; strategies: Strategy ; } ⋮---- export type KnowledgeKind = "vuln-class" "persona" "strategy"; 证据：`core/src/autonomous/knowledge/types.ts`
- **Types**（source_file）：import type { SessionConfig } from "../../execute/types.js"; ⋮---- export type TargetMode = "stateless" "stateful"; ⋮---- export interface TargetConfig { name: string; endpoint: string; apiKey?: string; headers?: Record ; mode: TargetMode; promptPath?: string; responsePath?: string; sessionField?: string; session?: SessionConfig; model?: string; } ⋮---- export interface HuntOptions { target: TargetConfig; objective: string; commanderModel: string; operatorModel: string; scoutModel: string; maxOperators: number; maxTurns: number; maxThreadTurns: number; maxTotalThreads: number; maxForksPerThread: number; maxTotalSends?: number; maxDepth: number; maxLeadsPerWave: number; budgetUsd?: number; v… 证据：`core/src/autonomous/lib/types.ts`
- **Run**（source_file）：import { randomUUID } from "node:crypto"; import { query, type Options, type AgentDefinition } from "@anthropic-ai/claude-agent-sdk"; import type { HuntOptions } from "../lib/types.js"; import { createTargetClient } from "../target/http.js"; import { loadKnowledge } from "../knowledge/load.js"; import { createRunLog } from "../state/runLog.js"; import { BudgetGuard } from "../lib/budget.js"; import { SessionGate } from "../../lib/sessionGate.js"; import type { RunContext } from "./context.js"; import { buildRedteamServer, REDTEAM SERVER NAME, toolId, TOOL NAMES } from "../tools/server.js"; import { buildHooks, type ProgressReporter } from "../state/hooks.js"; import { threadTreeText, counts… 证据：`core/src/autonomous/orchestrator/run.ts`
- **Types**（source_file）：import type { Severity as Severity } from "../../evaluators/schema.js"; import type { Verdict as Verdict } from "../../lib/judgeTypes.js"; ⋮---- export type Severity = Severity; export type Verdict = Verdict; ⋮---- export interface ReportTurn { turnIndex: number; prompt: string; response: string; persona?: string; strategy?: string; score?: number; } ⋮---- export interface SelfCheckResult { verdict: Verdict; score: number; confidence: number; reasoning: string; } ⋮---- export interface ReportFinding { findingId: string; vulnClassId: string; name: string; severity: Severity; standards?: Record ; threadId: string; strategy: string; personaArc: string ; verdict: Verdict; confidence: number; ev… 证据：`core/src/autonomous/report/types.ts`
- **Schema**（source_file）：import { z } from "zod"; import { PROVIDERS, type ProviderName } from "./types.js"; ⋮---- export type OpforMcpConfig = z.infer ; ⋮---- export type LlmConfig = z.infer ; export type McpServerConfig = z.infer ; ⋮---- export type McpScannerSection = z.infer ; ⋮---- export type OpforConfigFileV3 = z.infer ; ⋮---- export function parseRunConfig raw: unknown : z.infer ⋮---- export function parseAgentTarget raw: unknown : z.infer ⋮---- export function extractMcpScannerConfig raw: unknown : OpforMcpConfig 证据：`core/src/config/schema.ts`
- 其余 18 条证据见 `AI_CONTEXT_PACK.json` 或 `EVIDENCE_INDEX.json`。

## 宿主 AI 必须遵守的规则

- **把本资产当作开工前上下文，而不是运行环境。**：AI Context Pack 只包含证据化项目理解，不包含目标项目的可执行状态。 证据：`README.md`, `evaluators/README.md`, `suites/README.md`
- **回答用户时区分可预览内容与必须安装后才能验证的内容。**：安装前体验的消费者价值来自降低误装和误判，而不是伪装成真实运行。 证据：`README.md`, `evaluators/README.md`, `suites/README.md`

## 用户开工前应该回答的问题

- 你准备在哪个宿主 AI 或本地环境中使用它？
- 你只是想先体验工作流，还是准备真实安装？
- 你最在意的是安装成本、输出质量、还是和现有规则的冲突？

## 验收标准

- 所有能力声明都能回指到 evidence_refs 中的文件路径。
- AI_CONTEXT_PACK.md 没有把预览包装成真实运行。
- 用户能在 3 分钟内看懂适合谁、能做什么、如何开始和风险边界。

---

## Doramagic Context Augmentation

下面内容用于强化 Repomix/AI Context Pack 主体。Human Manual 只提供阅读骨架；踩坑日志会被转成宿主 AI 必须遵守的工作约束。

## Human Manual 骨架

使用规则：这里只是项目阅读路线和显著性信号，不是事实权威。具体事实仍必须回到 repo evidence / Claim Graph。

宿主 AI 硬性规则：
- 不得把页标题、章节顺序、摘要或 importance 当作项目事实证据。
- 解释 Human Manual 骨架时，必须明确说它只是阅读路线/显著性信号。
- 能力、安装、兼容性、运行状态和风险判断必须引用 repo evidence、source path 或 Claim Graph。

- **项目概览与系统架构**：importance `high`
  - source_paths: README.md, core/src/index.ts, core/src/execute/runAll.ts, core/src/execute/runAgentLoop.ts, core/src/report/buildReport.ts
- **核心运行器（CLI / 浏览器扩展 / MCP Server / SDK）**：importance `high`
  - source_paths: runners/cli/src/commands/run.ts, runners/cli/src/commands/setup.ts, runners/cli/src/commands/hunt.ts, runners/cli/ui/src/App.tsx, runners/cli/ui/src/components/ConversationView.tsx
- **评估器体系与攻击覆盖**：importance `high`
  - source_paths: evaluators/README.md, docs/evaluators.md, docs/evaluator-schema.md, core/src/evaluators/schema.ts, core/src/evaluators/judge.ts
- **集成、扩展与可观测性**：importance `high`
  - source_paths: core/src/providers/factory.ts, core/src/llm/openaiCompatible.ts, core/src/targets/agentTarget.ts, core/src/targets/httpClient.ts, core/src/targets/mcpTarget.ts

## Repo Inspection Evidence / 源码检查证据

- repo_clone_verified: true
- repo_inspection_verified: true
- repo_commit: `6b2e2b401712ad5dbecd634108f2964d29ae3fdd`
- inspected_files: `README.md`, `package.json`, `docs/browser-extension.md`, `docs/cli.md`, `docs/evaluator-schema.md`, `docs/evaluators.md`, `docs/hunt.md`, `docs/mcp.md`, `docs/sdk.md`, `docs/sessions.md`, `docs/skills.md`, `docs/telemetry.md`

宿主 AI 硬性规则：
- 没有 repo_clone_verified=true 时，不得声称已经读过源码。
- 没有 repo_inspection_verified=true 时，不得把 README/docs/package 文件判断写成事实。
- 没有 quick_start_verified=true 时，不得声称 Quick Start 已跑通。

## Doramagic Pitfall Constraints / 踩坑约束

这些规则来自 Doramagic 发现、验证或编译过程中的项目专属坑点。宿主 AI 必须把它们当作工作约束，而不是普通说明文字。

### Constraint 1: 失败模式：security_permissions: feat: opfor setup wizard has no prompt for custom HTTP headers on url-transport MCP targets

- Trigger: Developers should check this security_permissions risk before relying on the project: feat: opfor setup wizard has no prompt for custom HTTP headers on url-transport MCP targets
- Host AI rule: Before packaging this project, run the relevant install/config/quickstart check for: feat: opfor setup wizard has no prompt for custom HTTP headers on url-transport MCP targets. Context: Observed during installation or first-run setup.
- Why it matters: Developers may expose sensitive permissions or credentials: feat: opfor setup wizard has no prompt for custom HTTP headers on url-transport MCP targets
- Evidence: failure_mode_cluster:github_issue | https://github.com/KeyValueSoftwareSystems/agent-opfor/issues/166 | feat: opfor setup wizard has no prompt for custom HTTP headers on url-transport MCP targets
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 2: 失败模式：security_permissions: feat: support for evaluating Site-Specific Extension Execution

- Trigger: Developers should check this security_permissions risk before relying on the project: feat: support for evaluating Site-Specific Extension Execution
- Host AI rule: Before packaging this project, run the relevant install/config/quickstart check for: feat: support for evaluating Site-Specific Extension Execution. Context: Source discussion did not expose a precise runtime context.
- Why it matters: Developers may expose sensitive permissions or credentials: feat: support for evaluating Site-Specific Extension Execution
- Evidence: failure_mode_cluster:github_issue | https://github.com/KeyValueSoftwareSystems/agent-opfor/issues/179 | feat: support for evaluating Site-Specific Extension Execution
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 3: 失败模式：installation: v0.10.0

- Trigger: Developers should check this installation risk before relying on the project: v0.10.0
- Host AI rule: Before packaging this project, run the relevant install/config/quickstart check for: v0.10.0. Context: Observed during installation or first-run setup.
- Why it matters: Upgrade or migration may change expected behavior: v0.10.0
- Evidence: failure_mode_cluster:github_release | https://github.com/KeyValueSoftwareSystems/agent-opfor/releases/tag/v0.10.0 | v0.10.0
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 4: 来源证据：feat: Ability to judge and create reports from the conversations between the user and the agent

- Trigger: GitHub 社区证据显示该项目存在一个安装相关的待验证问题：feat: Ability to judge and create reports from the conversations between the user and the agent
- Why it matters: 可能增加新用户试用和生产接入成本。
- Evidence: community_evidence:github | https://github.com/KeyValueSoftwareSystems/agent-opfor/issues/171 | 来源类型 github_issue 暴露的待验证使用条件。
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 5: 可能修改宿主 AI 配置

- Trigger: 项目面向 Claude/Cursor/Codex/Gemini/OpenCode 等宿主，或安装命令涉及用户配置目录。
- Host AI rule: 列出会写入的配置文件、目录和卸载/回滚步骤。
- Why it matters: 安装可能改变本机 AI 工具行为，用户需要知道写入位置和回滚方法。
- Evidence: capability.host_targets | https://github.com/KeyValueSoftwareSystems/agent-opfor | host_targets=mcp_host, claude, cursor, chatgpt
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 6: 失败模式：configuration: bug: Nested .opfor configs

- Trigger: Developers should check this configuration risk before relying on the project: bug: Nested .opfor configs
- Host AI rule: Before packaging this project, run the relevant install/config/quickstart check for: bug: Nested .opfor configs. Context: Observed when using node, linux
- Why it matters: Developers may misconfigure credentials, environment, or host setup: bug: Nested .opfor configs
- Evidence: failure_mode_cluster:github_issue | https://github.com/KeyValueSoftwareSystems/agent-opfor/issues/183 | bug: Nested .opfor configs
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 7: 失败模式：configuration: bug: Significant delay in registering pause button clicks.

- Trigger: Developers should check this configuration risk before relying on the project: bug: Significant delay in registering pause button clicks.
- Host AI rule: Before packaging this project, run the relevant install/config/quickstart check for: bug: Significant delay in registering pause button clicks.. Context: Observed when using node, linux
- Why it matters: Developers may misconfigure credentials, environment, or host setup: bug: Significant delay in registering pause button clicks.
- Evidence: failure_mode_cluster:github_issue | https://github.com/KeyValueSoftwareSystems/agent-opfor/issues/182 | bug: Significant delay in registering pause button clicks.
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 8: 失败模式：configuration: bug: browser extension targets login form instead of chatbot input during red-team assessment

- Trigger: Developers should check this configuration risk before relying on the project: bug: browser extension targets login form instead of chatbot input during red-team assessment
- Host AI rule: Before packaging this project, run the relevant install/config/quickstart check for: bug: browser extension targets login form instead of chatbot input during red-team assessment. Context: Observed when using node, linux
- Why it matters: Developers may misconfigure credentials, environment, or host setup: bug: browser extension targets login form instead of chatbot input during red-team assessment
- Evidence: failure_mode_cluster:github_issue | https://github.com/KeyValueSoftwareSystems/agent-opfor/issues/178 | bug: browser extension targets login form instead of chatbot input during red-team assessment
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 9: 失败模式：configuration: bug:browser extension generates next attack before the agent has completed its response

- Trigger: Developers should check this configuration risk before relying on the project: bug:browser extension generates next attack before the agent has completed its response
- Host AI rule: Before packaging this project, run the relevant install/config/quickstart check for: bug:browser extension generates next attack before the agent has completed its response. Context: Observed when using node, linux
- Why it matters: Developers may misconfigure credentials, environment, or host setup: bug:browser extension generates next attack before the agent has completed its response
- Evidence: failure_mode_cluster:github_issue | https://github.com/KeyValueSoftwareSystems/agent-opfor/issues/180 | bug:browser extension generates next attack before the agent has completed its response
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 10: 失败模式：configuration: feat: add first-class support for additional LLM providers

- Trigger: Developers should check this configuration risk before relying on the project: feat: add first-class support for additional LLM providers
- Host AI rule: Before packaging this project, run the relevant install/config/quickstart check for: feat: add first-class support for additional LLM providers. Context: Source discussion did not expose a precise runtime context.
- Why it matters: Developers may misconfigure credentials, environment, or host setup: feat: add first-class support for additional LLM providers
- Evidence: failure_mode_cluster:github_issue | https://github.com/KeyValueSoftwareSystems/agent-opfor/issues/181 | feat: add first-class support for additional LLM providers
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。