chore: archive development research artifacts and debug scripts

This commit is contained in:
fujie
2026-02-07 12:49:42 +08:00
parent 89a12a4fe8
commit 795ac34cd9
25 changed files with 2855 additions and 0 deletions

View File

@@ -0,0 +1,68 @@
{"index": 1, "question": "Provide a python code example for configuring `CopilotClient` to connect to a local MCP server (e.g. Brave Search) via `CopilotClient` config.", "answer": "SDK 源码显示 MCP 服务器通过 `create_session`/`resume_session` 的 `mcp_servers` 配置传入并映射为 `mcpServers`,不是 CopilotClient 构造参数(源:本地源码 copilot/types.py / copilot/client.py。下面是本地 stdio MCP以 Brave Search 为例,命令/密钥请替换)的 Python 示例。 \n```python\nimport asyncio\nfrom copilot import CopilotClient\n\nasync def main():\n client = CopilotClient()\n await client.start()\n\n session = await client.create_session({\n \"model\": \"gpt-5\",\n \"mcp_servers\": {\n \"brave_search\": {\n \"type\": \"local\", # 或 \"stdio\"\n \"command\": \"/path/to/brave-search-mcp\", # 替换为你的 MCP server 启动命令\n \"args\": [],\n \"env\": {\"BRAVE_API_KEY\": \"YOUR_KEY\"},\n \"tools\": [\"*\"],\n \"timeout\": 30000, # 毫秒\n }\n },\n })\n\n await session.send_and_wait({\"prompt\": \"用 Brave Search 查询 OpenAI 最新公告\"})\n await client.stop()\n\nasyncio.run(main())\n```\n\n要点/坑:\n- `tools`: `[\"*\"]` 表示全量,`[]` 表示不启用\n- `timeout`: 毫秒\n- `command/args/env`: 按你的 MCP server 实际启动方式替换\n\n建议/落地:\n- pipe 内复用同一 session结束时统一 `client.stop()`", "model": "gpt-5.2-codex"}
{"index": 2, "question": "Explain how to programmatically handle `tool.confirmation_required` events in a non-interactive stream using `session.on()`.", "answer": "SDK 源码未定义 `tool.confirmation_required` 事件,权限确认由 JSONRPC `permission.request` 触发并交给 `on_permission_request` 回调处理,`session.on()` 只能订阅 `tool.user_requested`/`tool.execution_start` 等事件做记录关联(来源:本地源码 copilot/client.py、copilot/session.py、copilot/types.py。 \n在非交互流里必须提供 `on_permission_request` 自动审批/拒绝,否则默认拒绝且回调异常也会被拒绝。\n\n要点/坑 \n- 配置键: `on_permission_request`SessionConfig \n- 权限类型: `shell` / `write` / `mcp` / `read` / `url` \n- 回调异常或缺失 -> `denied-no-approval-rule-and-could-not-request-from-user`\n\n```python\nimport asyncio\nfrom copilot import CopilotClient\nfrom copilot.generated.session_events import SessionEventType\n\nasync def main():\n client = CopilotClient()\n await client.start()\n\n def on_event(ev):\n if ev.type == SessionEventType.TOOL_USER_REQUESTED:\n print(f\"tool requested: {ev.data.tool_name} {ev.data.tool_call_id}\")\n elif ev.type == SessionEventType.TOOL_EXECUTION_START:\n print(f\"tool start: {ev.data.tool_name}\")\n\n async def on_permission_request(req, ctx):\n kind = req.get(\"kind\")\n if kind in (\"read\", \"mcp\"):\n return {\"kind\": \"approved\"}\n return {\"kind\": \"denied-by-rules\", \"rules\": [{\"kind\": \"deny-non-read\"}]}\n\n session = await client.create_session({\n \"model\": \"gpt-5\",\n \"on_permission_request\": on_permission_request,\n })\n session.on(on_event)\n\n await session.send_and_wait({\"prompt\": \"请调用需要权限的工具\"})\n await client.stop()\n\nasyncio.run(main())\n```\n\n建议/落地 \n- pipe 启动时注册审批策略 + `session.on` 记录 `tool_call_id` 做审计", "model": "gpt-5.2-codex"}
{"index": 3, "question": "Show how to implement a 'Device Flow' login helper using SDK primitives (if available) or raw HTTP showing how to persist credentials.", "answer": "SDK 本身没有 deviceflow 登录 API可用鉴权原语只有 CopilotClient 的 `github_token`/`use_logged_in_user` 与 `get_auth_status`(源:本地源码 copilot/client.py、copilot/types.py。 \n下面用 GitHub OAuth Device Flow 原始 HTTP 获取 token、持久化本地并注入 `github_token`端点说明https://docs.github.com/en/apps/oauth-apps/building-oauth-apps/authorizing-oauth-apps#device-flow。\n\n```python\nimport asyncio\nimport json\nimport os\nimport time\nimport urllib.parse\nimport urllib.request\nfrom pathlib import Path\n\nfrom copilot import CopilotClient\n\nCLIENT_ID = \"YOUR_OAUTH_APP_CLIENT_ID\"\nSCOPE = \"read:user\" # 按你的 OAuth App 需求调整\nTOKEN_PATH = Path.home() / \".config\" / \"myapp\" / \"copilot_token.json\"\n\ndef http_post(url, data):\n body = urllib.parse.urlencode(data).encode()\n req = urllib.request.Request(url, data=body, headers={\"Accept\": \"application/json\"})\n with urllib.request.urlopen(req) as resp:\n return json.loads(resp.read().decode())\n\ndef load_token():\n if TOKEN_PATH.exists():\n return json.loads(TOKEN_PATH.read_text()).get(\"access_token\")\n return None\n\ndef save_token(token):\n TOKEN_PATH.parent.mkdir(parents=True, exist_ok=True)\n TOKEN_PATH.write_text(json.dumps(token))\n os.chmod(TOKEN_PATH, 0o600)\n\ndef device_flow():\n code = http_post(\n \"https://github.com/login/device/code\",\n {\"client_id\": CLIENT_ID, \"scope\": SCOPE},\n )\n print(f\"Open {code['verification_uri']} and enter {code['user_code']}\")\n interval = int(code.get(\"interval\", 5))\n\n while True:\n token = http_post(\n \"https://github.com/login/oauth/access_token\",\n {\n \"client_id\": CLIENT_ID,\n \"device_code\": code[\"device_code\"],\n \"grant_type\": \"urn:ietf:params:oauth:grant-type:device_code\",\n },\n )\n if \"access_token\" in token:\n return token\n if token.get(\"error\") == \"slow_down\":\n interval += 5\n elif token.get(\"error\") not in (\"authorization_pending\",):\n raise RuntimeError(token)\n time.sleep(interval)\n\nasync def main():\n access_token = load_token()\n if not access_token:\n token = device_flow()\n save_token(token)\n access_token = token[\"access_token\"]\n\n client = CopilotClient({\"github_token\": access_token})\n await client.start()\n # ... create_session / send ...\n await client.stop()\n\nif __name__ == \"__main__\":\n asyncio.run(main())\n```\n\n要点/坑 \n- `cli_url` 与 `github_token` 互斥 \n- SDK 不保存凭据 需自行持久化 \n- 轮询需处理 `authorization_pending` / `slow_down`\n\n建议/落地 \n- pipe 启动时读取 token 文件 不存在则走 device flow \n- 用 `github_token` 注入 保持非交互运行", "model": "gpt-5.2-codex"}
{"index": 4, "question": "Compare the pros and cons of 'Session Replay' (fast-forwarding history) vs 'Session Resume' (stateful ID) for a stateless web backend like OpenWebUI.", "answer": "对无状态后端(OpenWebUI)而言SDK 只支持通过 `resume_session(session_id)` 的状态续接所谓“Session Replay”需要你自行重放历史 prompt来源本地 SDK 源码 copilot/client.py、copilot/session.py。 \n对比要点短语 \n- Replay无状态/易扩展;缺点是耗时耗 tokens、易截断/丢工具状态、可能重复副作用 \n- Resume省 tokens、保留工具/compaction/`workspace_path` 状态;缺点是需持久化 session_id、依赖会话存活/跨实例一致性。 \n建议/落地:优先持久化 chat_id→session_id 并 resume_session失效时回退为重放最近 N 轮并重建工具上下文。", "model": "gpt-5.2-codex"}
{"index": 5, "question": "Detail the exact protocol for 'Virtual Workspace': how to implement a file system provider that feeds content to Copilot without physical files.", "answer": "SDK 源码里没有“Virtual Workspace”协议或文件系统 provider 接口;会话只识别真实路径(`SessionConfig.working_directory`、`MessageOptions.attachments`)。要在无物理文件情况下供给内容,只能用自定义工具或 MCP 服务器把“读写文件”变成 tool 调用(来源:本地 SDK 源码 copilot/types.py / copilot/client.py。\n\n可用原语源码\n- `SessionConfig.tools` / `available_tools`:自定义工具与白名单\n- `MessageOptions.attachments`:仅支持真实路径文件/目录\n\n方案 A自定义工具实现 VFS推荐\n```python\nimport asyncio\nfrom pydantic import BaseModel, Field\nfrom copilot import CopilotClient, define_tool\n\nVFS = {\"README.md\": \"# Virtual README\\n内容来自内存\\n\"}\n\nclass VfsReadParams(BaseModel):\n path: str = Field(description=\"virtual path\")\n\n@define_tool(name=\"vfs_read\", description=\"Read a virtual file\")\ndef vfs_read(params: VfsReadParams) -> str:\n return VFS.get(params.path, \"\")\n\n@define_tool(name=\"vfs_list\", description=\"List virtual files\")\ndef vfs_list() -> list[str]:\n return sorted(VFS.keys())\n\nasync def main():\n client = CopilotClient()\n await client.start()\n session = await client.create_session({\n \"model\": \"gpt-5\",\n \"tools\": [vfs_read, vfs_list],\n \"available_tools\": [\"vfs_read\", \"vfs_list\"],\n \"system_message\": {\"content\": \"Use vfs_* tools for all file access.\"},\n })\n await session.send_and_wait({\"prompt\": \"打开 README.md\"})\n await client.stop()\n\nasyncio.run(main())\n```\n\n方案 BMCP 服务器(本地 stdioVFS 由你实现)\n```python\nsession = await client.create_session({\n \"model\": \"gpt-5\",\n \"mcp_servers\": {\n \"vfs\": {\n \"type\": \"local\",\n \"command\": \"python\",\n \"args\": [\"-m\", \"my_vfs_mcp_server\"],\n \"tools\": [\"*\"],\n }\n },\n})\n```\n\n注意/坑 \n- `available_tools` 只保留你的 VFS 工具以强制使用 \n- 内置文件工具只读真实文件;要兼容只能落地临时文件 \n- 大文件需分块返回,避免上下文爆炸 \n\n建议/落地 \n- pipe 内维护 VFS 存储DB/Redis/内存) \n- 启动 session 时注入 `system_message` + `available_tools` 强制 VFS 入口", "model": "gpt-5.2-codex"}
{"index": 6, "question": "Create an 'Agent Manifest' example: how to define an Agent capable of specific high-privileged tools via SDK.", "answer": "SDK 没有独立“Agent Manifest”文件格式等价于 `create_session` 的 `custom_agents`CustomAgentConfig并用 `tools`/`available_tools` 约束高权限工具,权限由 `on_permission_request` 处理(来源:本地 SDK 源码 copilot/types.py、copilot/client.py。 \n示例如下manifest 为 Python dict可序列化为 JSON 配置)。\n\n```python\nimport asyncio\nfrom pydantic import BaseModel, Field\nfrom copilot import CopilotClient, define_tool\n\nclass AdminShellParams(BaseModel):\n cmd: str = Field(description=\"approved shell command\")\n\n@define_tool(name=\"admin_shell\", description=\"Run an approved shell command\")\ndef admin_shell(params: AdminShellParams) -> str:\n # TODO: 执行受控命令(白名单/沙箱)\n return f\"approved: {params.cmd}\"\n\nclass AdminWriteParams(BaseModel):\n path: str = Field(description=\"target path\")\n content: str = Field(description=\"content\")\n\n@define_tool(name=\"admin_write\", description=\"Write to privileged store\")\ndef admin_write(params: AdminWriteParams) -> str:\n # TODO: 受控写入\n return \"ok\"\n\nasync def on_permission_request(req, _ctx):\n if req.get(\"kind\") in (\"shell\", \"write\"):\n return {\"kind\": \"approved\"}\n return {\"kind\": \"denied-by-rules\", \"rules\": [{\"kind\": \"deny\"}]}\n\nasync def main():\n client = CopilotClient()\n await client.start()\n\n agent_manifest = {\n \"name\": \"ops-admin\",\n \"display_name\": \"Ops Admin\",\n \"description\": \"Privileged ops agent\",\n \"prompt\": \"Use only admin_* tools for privileged ops.\",\n \"tools\": [\"admin_shell\", \"admin_write\"],\n \"infer\": True,\n }\n\n session = await client.create_session({\n \"model\": \"gpt-5\",\n \"tools\": [admin_shell, admin_write],\n \"custom_agents\": [agent_manifest],\n \"available_tools\": [\"admin_shell\", \"admin_write\"],\n \"on_permission_request\": on_permission_request,\n })\n\n await session.send_and_wait({\"prompt\": \"使用 ops-admin 执行受控操作\"})\n await client.stop()\n\nasyncio.run(main())\n```\n\n要点/坑 \n- session.tools 未注册 -> agent.tools 不可用 \n- custom_agents.tools 仅写名称 \n- 非交互缺少 on_permission_request -> 默认拒绝 \n- SDK 无显式 agent 选择参数 -> 需在 prompt 指定 \n\n建议/落地 \n- pipe 启动时加载 manifest 注入 custom_agents + available_tools + on_permission_request", "model": "gpt-5.2-codex"}
{"index": 7, "question": "List all 'hidden' `SessionConfig` parameters relevant to Agent behavior and personality.", "answer": "基于本地 SDK 源码 `copilot/types.py` 的 SessionConfig影响 Agent 行为/人格的相关参数如下(来源:本地 SDK 源码)。 \n- `system_message`: `append`/`replace``replace` 会移除 SDK guardrails风险 \n- `custom_agents`: `prompt`/`description`/`display_name`/`tools`/`mcp_servers`/`infer``tools` 仅名称且必须在 `tools` 注册 \n- `tools` / `available_tools` / `excluded_tools`: 能力与行为边界;`available_tools` 优先于 `excluded_tools` \n- `hooks`: `on_user_prompt_submitted`/`on_pre_tool_use`/`on_post_tool_use`/`on_session_start`/`on_session_end`/`on_error_occurred`,可改写 prompt/工具参数/结果 \n- `on_permission_request`: 非交互审批策略;缺失或异常 => 拒绝 \n- `on_user_input_request`: 启用/处理 `ask_user` \n- `mcp_servers`: 注入外部工具集 \n- `skill_directories` / `disabled_skills`: 技能启用/禁用 \n- `model` / `provider`: 模型与 BYOK 影响输出风格与能力 \n- `working_directory`: 上下文范围 \n- `infinite_sessions`: compaction 阈值影响记忆保留 \n- `config_dir`: CLI 配置/状态目录,行为取决于 CLI 侧配置 \n\n建议/落地:在 pipe 启动时固定 `system_message` + `available_tools/excluded_tools` + `hooks`,并持久化 session_id 以便 `resume_session`。", "model": "gpt-5.2-codex"}
{"index": 8, "question": "Give a concise overview of MCP Server Integration: JSON-RPC config and SDK hooks.", "answer": "SDK 中 MCP 通过 create_session/resume_session 的 mcp_servers 传入(映射到 JSONRPC 的 mcpServers支持 local/stdio 与 http/sse两类配置字段分别为 type/command/args/env/cwd/timeout/tools 或 type/url/headers/timeout/tools来源本地 SDK 源码)。 \nSDK hooks 使用 hooks(on_pre_tool_use/on_post_tool_use/on_user_prompt_submitted/on_session_start/on_session_end/on_error_occurred) 进行拦截与附加上下文,权限需 on_permission_request 对 kind='mcp' 明确批准,且 session.on 可审计 tool.execution_* 事件中的 mcp_server_name/mcp_tool_name。 \n示例见下建议/落地pipe 中启用 on_permission_request、限制 tools 白名单并记录 MCP 事件。\n\n```python\nimport asyncio\nfrom copilot import CopilotClient\n\nasync def on_permission_request(req, _):\n return {\"kind\": \"approved\"} if req.get(\"kind\") == \"mcp\" else {\n \"kind\": \"denied-by-rules\",\n \"rules\": [{\"kind\": \"deny\"}],\n }\n\nasync def on_pre_tool_use(input, _):\n if input.get(\"toolName\", \"\").startswith(\"mcp.\"):\n return {\"additionalContext\": \"MCP tool call observed\"}\n\nasync def main():\n client = CopilotClient()\n await client.start()\n\n session = await client.create_session({\n \"model\": \"gpt-5\",\n \"mcp_servers\": {\n \"brave\": {\n \"type\": \"local\",\n \"command\": \"/path/to/brave-mcp\",\n \"args\": [],\n \"env\": {\"BRAVE_API_KEY\": \"YOUR_KEY\"},\n \"tools\": [\"*\"],\n \"timeout\": 30000,\n }\n },\n \"hooks\": {\"on_pre_tool_use\": on_pre_tool_use},\n \"on_permission_request\": on_permission_request,\n })\n\n await session.send_and_wait({\"prompt\": \"用 Brave Search 查询 Copilot SDK 新闻\"})\n await client.stop()\n\nasyncio.run(main())\n```", "model": "gpt-5.2-codex"}
{"index": 9, "question": "Provide best practices and common pitfalls for MCP Server Integration: JSON-RPC config and SDK hooks.", "answer": "基于本地 SDK 源码MCP 配置通过 `create_session`/`resume_session` 的 `mcp_servers` 进入 JSONRPC `mcpServers`,并通过 `on_permission_request` 与 `hooks` 控制工具审批与行为(来源:本地 SDK 源码) \n- 最佳实践local/stdio 用 `type/command/args/env/cwd/timeout/tools`remote http/sse 用 `type/url/headers/timeout/tools``tools` 尽量白名单、必要时 `[\"*\"]``timeout` 毫秒;`hooks.on_pre_tool_use` 做改参/拒绝,`session.on` 记录 `tool.execution_*` 与 `mcp_server_name`/`mcp_tool_name` \n- 常见坑:把 MCP 配置放在 CopilotClient缺少 `on_permission_request`kind=\"mcp\")导致默认拒绝;`tools: []` 等于禁用全部;设置 `available_tools` 却未包含 MCP 工具名导致被过滤 \n\n建议/落地pipe 中固定审批策略 + 统一 MCP 配置模板 + 审计日志", "model": "gpt-5.2-codex"}
{"index": 10, "question": "Show a minimal example snippet for MCP Server Integration: JSON-RPC config and SDK hooks.", "answer": "基于本地 SDK 源码MCP 通过 `mcp_servers` 传入并用 `hooks` 处理工具调用(来源:本地 SDK 源码)。最小示例: \n```python\nimport asyncio\nfrom copilot import CopilotClient\n\nasync def main():\n client = CopilotClient()\n await client.start()\n\n session = await client.create_session({\n \"model\": \"gpt-5\",\n \"mcp_servers\": {\n \"brave\": {\n \"type\": \"local\",\n \"command\": \"/path/to/brave-mcp\",\n \"args\": [],\n \"env\": {\"BRAVE_API_KEY\": \"YOUR_KEY\"},\n \"tools\": [\"*\"],\n }\n },\n \"hooks\": {\n \"on_pre_tool_use\": lambda input, _: {\"additionalContext\": \"MCP tool call\"},\n },\n \"on_permission_request\": lambda req, _: {\"kind\": \"approved\"},\n })\n\n await session.send_and_wait({\"prompt\": \"用 Brave Search 搜索 Copilot SDK\"})\n await client.stop()\n\nasyncio.run(main())\n```\n\n建议/落地:固定 `on_permission_request`kind=\"mcp\")并限制 `tools` 白名单。", "model": "gpt-5.2-codex"}
{"index": 11, "question": "List recommended configuration defaults for MCP Server Integration: JSON-RPC config and SDK hooks.", "answer": "推荐默认值MCP JSONRPC 配置,来源:本地 SDK 源码local/stdio 服务器用 `type=\"local\"/\"stdio\"`(远程用 `http`/`sse`)、`tools` 白名单(无则 `\"*\"`)、`timeout=30000` 毫秒、`env` 仅放密钥、`cwd` 指向受控目录。 \nHooks 默认:`on_pre_tool_use` 追加上下文并可设 `permissionDecision`ask/allow、必要时改参`on_post_tool_use` 记录结果,`on_error_occurred` 设置 retry/skip。 \n建议/落地:非交互场景配置 `on_permission_request` 仅放行 `kind=\"mcp\"`,否则拒绝;工具名不确定时先不设 `available_tools`,避免误过滤。", "model": "gpt-5.2-codex"}
{"index": 12, "question": "How does MCP Server Integration: JSON-RPC config and SDK hooks relate to building a custom Agent?", "answer": "SDK 源码中,自定义 Agent 通过 `SessionConfig.custom_agents` 定义,`CustomAgentConfig` 支持 `mcp_servers`因此“Agent 绑定 MCP”就是在 agent config 里声明 `mcp_servers`,底层在 `create_session` 时转成 JSONRPC 的 `mcpServers`(来源:本地 SDK 源码 copilot/types.py、copilot/client.py。 \nHooks 是会话级拦截器(`on_pre_tool_use`/`on_post_tool_use` 等),对 MCP 工具同样生效;权限确认需 `on_permission_request` 审批 `kind=\"mcp\"`。 \n建议/落地pipe 中为目标 Agent 绑定 MCP + 审批钩子,示例: \n```python\nsession = await client.create_session({\n \"model\": \"gpt-5\",\n \"custom_agents\": [{\n \"name\": \"search\",\n \"prompt\": \"Use MCP for search.\",\n \"mcp_servers\": {\n \"brave\": {\n \"type\": \"local\",\n \"command\": \"/path/to/brave-mcp\",\n \"args\": [],\n \"env\": {\"BRAVE_API_KEY\": \"YOUR_KEY\"},\n \"tools\": [\"*\"],\n }\n },\n }],\n \"hooks\": {\"on_pre_tool_use\": lambda i, _: {\"additionalContext\": \"mcp\"}},\n \"on_permission_request\": lambda r, _: {\"kind\": \"approved\"},\n})\n```", "model": "gpt-5.2-codex"}
{"index": 13, "question": "Give a concise overview of Agent Manifests: Defining capabilities and permissions programmatically.", "answer": "SDK 没有独立的 Agent Manifest 文件,等价做法是用 `SessionConfig.custom_agents` 描述 agent并用 `tools`/`available_tools`/`excluded_tools` 限定能力。 \n权限通过 `on_permission_request` 与 `hooks`(如 `on_pre_tool_use`)在会话级控制,且 `custom_agents.tools` 只能写工具名并需在 `SessionConfig.tools` 注册。 \n建议/落地pipe 启动时加载 manifest→注入 `custom_agents` + `available_tools` + `on_permission_request` 固化权限边界。", "model": "gpt-5.2-codex"}
{"index": 14, "question": "Provide best practices and common pitfalls for Agent Manifests: Defining capabilities and permissions programmatically.", "answer": "SDK 没有独立 Manifest 文件,等价于 `SessionConfig.custom_agents` + `tools/available_tools/excluded_tools` + `on_permission_request` + `hooks` 来定义能力与权限(来源:本地 SDK 源码)。 \n最佳实践`custom_agents.tools` 只写已在 `SessionConfig.tools` 注册的名称,配合 `available_tools` 白名单与 `on_permission_request` 精准放行,必要时用 `hooks.on_pre_tool_use` 做参数校验/审计,`system_message` 选 append 以保留 guardrails。 \n常见坑工具未注册或名称不一致、缺少 `on_permission_request` 导致默认拒绝、误用 `system_message` replace 去掉安全限制、期待 agent 自动被选中或跨实例不持久 `session_id`;建议/落地pipe 启动时加载 manifest→注入 `custom_agents/available_tools/on_permission_request` 并持久化 `session_id`。", "model": "gpt-5.2-codex"}
{"index": 15, "question": "Show a minimal example snippet for Agent Manifests: Defining capabilities and permissions programmatically.", "answer": "最小示例(本地 SDK 源码:`custom_agents` + `tools` + `available_tools` + `on_permission_request` \n```python\nimport asyncio\nfrom pydantic import BaseModel, Field\nfrom copilot import CopilotClient, define_tool\n\nclass PingParams(BaseModel):\n text: str = Field(description=\"text to echo\")\n\n@define_tool(name=\"priv_ping\", description=\"Privileged echo\")\ndef priv_ping(params: PingParams) -> str:\n return f\"ok: {params.text}\"\n\nasync def main():\n client = CopilotClient()\n await client.start()\n\n agent_manifest = {\n \"name\": \"priv-agent\",\n \"prompt\": \"Only use priv_ping for actions.\",\n \"tools\": [\"priv_ping\"],\n \"infer\": True,\n }\n\n session = await client.create_session({\n \"model\": \"gpt-5\",\n \"tools\": [priv_ping],\n \"custom_agents\": [agent_manifest],\n \"available_tools\": [\"priv_ping\"],\n \"on_permission_request\": lambda req, _: {\"kind\": \"approved\"},\n })\n\n await session.send_and_wait({\"prompt\": \"调用 priv_ping\"})\n await client.stop()\n\nasyncio.run(main())\n```\n\n建议/落地:保持工具名一致并白名单化,非交互务必提供 `on_permission_request`。", "model": "gpt-5.2-codex"}
{"index": 16, "question": "List recommended configuration defaults for Agent Manifests: Defining capabilities and permissions programmatically.", "answer": "基于本地 SDK 源码copilot/types.py / copilot/client.py推荐默认值如下 \n- custom_agentsname 必填、prompt 必填、tools 仅名称、infer=Truedisplay_name/description 可选 \n- available_tools=白名单、excluded_tools=空、system_message=append \n- on_permission_request除明确允许的 kind 其余拒绝hooks.on_pre_tool_use 做审计/参数校验(其余 hooks 按需) \n建议/落地pipe 启动时加载 manifest→注册 tools + available_tools + on_permission_request并持久化 session_id。", "model": "gpt-5.2-codex"}
{"index": 17, "question": "How does Agent Manifests: Defining capabilities and permissions programmatically relate to building a custom Agent?", "answer": "“Agent Manifest”在 SDK 里就是 `SessionConfig.custom_agents` 的配置对象,用它定义自定义 Agent 的 `name/prompt/tools/mcp_servers/infer`,并在同一 session 里用 `tools` 注册实际工具、用 `available_tools/excluded_tools` 限定能力(来源:本地 SDK 源码)。 \n权限不在 manifest 内,而是由会话级 `on_permission_request` 和 hooks如 `on_pre_tool_use`)统一控制,且对所有 Agent 生效。 \n建议/落地pipe 启动时把 manifest 转成 `custom_agents` 并同步注册工具 + 白名单,再配置 `on_permission_request` 作为权限策略。", "model": "gpt-5.2-codex"}
{"index": 18, "question": "Give a concise overview of Headless Auth: Device Code Flow and credential persistence.", "answer": "SDK 仅提供 `CopilotClient` 的 `github_token/use_logged_in_user` 与 `get_auth_status()`,没有内置 deviceflow API源码copilot/client.py、copilot/types.py。 \nHeadless Device Flow 需走 GitHub OAuth Device Flow HTTP 端点获取 tokenhttps://docs.github.com/en/apps/oauth-apps/building-oauth-apps/authorizing-oauth-apps#device-flow将 token 持久化(权限 600并通过 `CopilotClient({\"github_token\": token})` 注入,且与 `cli_url` 互斥。 \n建议/落地:启动时读 token 文件,不存在则执行 device flow 并保存,再启动客户端。\n\n```python\ntoken = load_token() or device_flow() # 调用 device code + access_token 端点\nsave_token(token) # chmod 600\nclient = CopilotClient({\"github_token\": token[\"access_token\"]})\nawait client.start()\n```", "model": "gpt-5.2-codex"}
{"index": 19, "question": "Provide best practices and common pitfalls for Headless Auth: Device Code Flow and credential persistence.", "answer": "SDK 不内置 device flow只能用 GitHub OAuth Device Flow HTTP 换取 token 并通过 CopilotClient({github_token}) 注入,且与 cli_url 互斥源码copilot/client.py、copilot/types.py。 \n最佳实践/坑:- token 文件 chmod 600 或 OS keychain- 记录到期与刷新策略,启动时用 get_auth_status 验证;- 轮询需处理 authorization_pending/slow_down- 未持久化、use_logged_in_user 与 github_token 同时使用或与 cli_url 混用会失败端点说明https://docs.github.com/en/apps/oauth-apps/building-oauth-apps/authorizing-oauth-apps#device-flow。 \n建议/落地pipe 启动时读 token 文件,不存在则走 device flow 并保存,失败时回退重新授权。", "model": "gpt-5.2-codex"}
{"index": 20, "question": "Show a minimal example snippet for Headless Auth: Device Code Flow and credential persistence.", "answer": "最小示例:用 GitHub Device Flow端点说明见 https://docs.github.com/en/apps/oauth-apps/building-oauth-apps/authorizing-oauth-apps#device-flow拿 token持久化后注入 CopilotClient({github_token}). \n```python\nimport asyncio\nimport json\nimport os\nimport time\nimport urllib.parse\nimport urllib.request\nfrom pathlib import Path\nfrom copilot import CopilotClient\n\nCLIENT_ID = \"YOUR_OAUTH_APP_CLIENT_ID\"\nTOKEN_PATH = Path.home() / \".config\" / \"myapp\" / \"copilot_token.json\"\n\ndef post(url, data):\n body = urllib.parse.urlencode(data).encode()\n req = urllib.request.Request(url, data=body, headers={\"Accept\": \"application/json\"})\n with urllib.request.urlopen(req) as resp:\n return json.loads(resp.read().decode())\n\ndef device_flow():\n code = post(\"https://github.com/login/device/code\", {\"client_id\": CLIENT_ID, \"scope\": \"read:user\"})\n print(f\"Open {code['verification_uri']} and enter {code['user_code']}\")\n interval = int(code.get(\"interval\", 5))\n while True:\n tok = post(\n \"https://github.com/login/oauth/access_token\",\n {\n \"client_id\": CLIENT_ID,\n \"device_code\": code[\"device_code\"],\n \"grant_type\": \"urn:ietf:params:oauth:grant-type:device_code\",\n },\n )\n if \"access_token\" in tok:\n return tok\n if tok.get(\"error\") == \"slow_down\":\n interval += 5\n elif tok.get(\"error\") != \"authorization_pending\":\n raise RuntimeError(tok)\n time.sleep(interval)\n\ndef load_token():\n return json.loads(TOKEN_PATH.read_text()).get(\"access_token\") if TOKEN_PATH.exists() else None\n\ndef save_token(tok):\n TOKEN_PATH.parent.mkdir(parents=True, exist_ok=True)\n TOKEN_PATH.write_text(json.dumps(tok))\n os.chmod(TOKEN_PATH, 0o600)\n\nasync def main():\n access_token = load_token()\n if not access_token:\n tok = device_flow()\n save_token(tok)\n access_token = tok[\"access_token\"]\n\n client = CopilotClient({\"github_token\": access_token})\n await client.start()\n # ... create_session / send ...\n await client.stop()\n\nasyncio.run(main())\n```\n\n建议/落地 \n- 启动先读 token 文件;缺失再 device flow \n- token 文件权限 600 \n- 失效则重新授权", "model": "gpt-5.2-codex"}
{"index": 21, "question": "List recommended configuration defaults for Headless Auth: Device Code Flow and credential persistence.", "answer": "基于本地 SDK 源码(`github_token` 与 `cli_url` 互斥)与 GitHub Device Flow 文档https://docs.github.com/en/apps/oauth-apps/building-oauth-apps/authorizing-oauth-apps#device-flow推荐默认值 \n- Token 持久化:`~/.config/<app>/copilot_token.json` 或系统密钥链,文件权限 600 \n- Client 选项:仅 `github_token``use_logged_in_user` 留空/False禁用 `cli_url` \n- 轮询:使用 device code 返回的 `interval`(缺省 5s处理 `authorization_pending`/`slow_down` \n- 启动校验:`get_auth_status()` 失败即重新授权 \n\n建议/落地:启动先读 token 文件 → 无则 device flow → 保存 → `client.start()`", "model": "gpt-5.2-codex"}
{"index": 22, "question": "How does Headless Auth: Device Code Flow and credential persistence relate to building a custom Agent?", "answer": "Headless Auth 只负责拿到并持久化 `github_token`,让 `CopilotClient` 能在无交互环境启动会话并加载 `custom_agents`;因此它是自定义 Agent 能稳定运行的前置条件,而不是 Agent 定义本身(来源:本地 SDK 源码)。 \n建议/落地:在无状态后端启动时读取 token 文件(不存在就走 device flow→ `CopilotClient({\"github_token\": token})` → `create_session(custom_agents=...)`。", "model": "gpt-5.2-codex"}
{"index": 23, "question": "Give a concise overview of Session Replay vs Resume: Handling stateless frontend history.", "answer": "基于本地 SDK 源码copilot/client.py、copilot/session.pyResume 是 `resume_session(session_id)` 保留会话状态Replay 只能由应用重放历史 prompt。优缺点Resume 省 tokens/保工具与 compaction 状态但依赖持久化 session_idReplay 无状态易扩展但成本高且易丢副作用控制。建议/落地:无状态前端请保存 session_id 并优先 Resume失效时回放最近 N 轮。", "model": "gpt-5.2-codex"}
{"index": 24, "question": "Provide best practices and common pitfalls for Session Replay vs Resume: Handling stateless frontend history.", "answer": "最佳实践:在无状态前端每次请求携带 session_id 并优先 `resume_session`,同时恢复 tools/hooks/on_permission_request/system_message 以保持行为一致(来源:本地 SDK 源码)。 \n常见坑Replay 会重复执行工具且耗 tokensResume 未重新注册工具或权限处理器会出现 tool not supported/默认拒绝;会话失效未回退。 \n建议/落地:持久化 chat_id→session_idresume 失败时仅回放最近 N 轮并禁用/幂等化有副作用的工具。", "model": "gpt-5.2-codex"}
{"index": 25, "question": "Show a minimal example snippet for Session Replay vs Resume: Handling stateless frontend history.", "answer": "来源:本地 SDK 源码copilot/client.py, copilot/session.py最小示例如下优先 Resume失败则 Replay。 \n```python\nimport asyncio\nfrom copilot import CopilotClient\n\nasync def handle_request(session_id: str | None, history: list[str], latest: str):\n client = CopilotClient()\n await client.start()\n\n try:\n # Resume: 复用会话状态\n if not session_id:\n raise RuntimeError(\"no session_id\")\n session = await client.resume_session(session_id)\n except Exception:\n # Replay: 逐条重放历史\n session = await client.create_session({\"model\": \"gpt-5\"})\n session_id = session.session_id\n for prompt in history:\n await session.send_and_wait({\"prompt\": prompt})\n\n response = await session.send_and_wait({\"prompt\": latest})\n await client.stop()\n return response.data.content, session_id\n\n# asyncio.run(handle_request(session_id, history, latest_prompt))\n```\n建议/落地:持久化 session_idresume 失败时仅回放最近 N 轮。", "model": "gpt-5.2-codex"}
{"index": 26, "question": "List recommended configuration defaults for Session Replay vs Resume: Handling stateless frontend history.", "answer": "基于本地 SDK 源码推荐默认配置如下Resume / Replay \n- Resume持久化 session_id`resume_session({...})` 传回 tools、hooks、on_permission_request、working_directory、streaming、provider、mcp_servers、custom_agents、skill_directories、disabled_skills`disable_resume` 保持默认 False \n- Replay`create_session({...})` 传回 model、system_message、tools、available_tools/excluded_tools、hooks、on_permission_request、working_directory、provider、mcp_servers、custom_agents、infinite_sessions(默认启用);仅回放最近 N 轮;默认拒绝 write/shell 或用只读白名单", "model": "gpt-5.2-codex"}
{"index": 27, "question": "How does Session Replay vs Resume: Handling stateless frontend history relate to building a custom Agent?", "answer": "Resume 会保留 session_id 关联的 custom_agents/工具状态Replay 需要每次 create_session 时重新注入 custom_agents + tools否则 persona/能力会丢失(来源:本地 SDK 源码。对于无状态前端Resume 更可靠且省 token但要持久化 session_id 并在 resume 时重新注册 tools/hooksReplay 则需回放历史并保证相同 custom_agents 配置。建议/落地pipe 优先 resume_session失败再 replay 最近 N 轮并重建 custom_agents/tools/权限策略。", "model": "gpt-5.2-codex"}
{"index": 28, "question": "Give a concise overview of Advanced Session Hooks: Intercepting and modifying user prompts.", "answer": "拦截/修改用户提示要用 `SessionHooks.on_user_prompt_submitted`(不是 `session.on` 事件订阅),`input.prompt` 可改为 `output.modifiedPrompt` 并附加 `additionalContext`/`suppressOutput`,且需在 `create_session`/`resume_session` 传入 `hooks` 才会触发 `hooks.invoke`(来源:本地 SDK 源码 copilot/types.py、copilot/session.py、copilot/client.py。 \n```python\nfrom copilot import CopilotClient\n\nasync def on_user_prompt_submitted(input, _ctx):\n return {\"modifiedPrompt\": input[\"prompt\"].strip(), \"additionalContext\": \"normalize whitespace\"}\n\nsession = await client.create_session({\n \"model\": \"gpt-5\",\n \"hooks\": {\"on_user_prompt_submitted\": on_user_prompt_submitted},\n})\n```\n建议/落地:在 pipe 入口统一做 prompt 清洗/拼接上下文,并记录修改前后用于审计。", "model": "gpt-5.2-codex"}
{"index": 29, "question": "Provide best practices and common pitfalls for Advanced Session Hooks: Intercepting and modifying user prompts.", "answer": "最佳实践:在 create_session/resume_session 传入 hooks.on_user_prompt_submitted返回 {modifiedPrompt, additionalContext, suppressOutput},修改保持幂等、最小化、可审计(来源:本地 SDK 源码 copilot/types.py、copilot/client.py、copilot/session.py。 \n常见坑忘记传 hooks 或 resume 时未重新传、handler 抛异常会被吞并返回 None、返回非 dict 或过度改写提示导致无效/偏航。 \n建议/落地:在 pipe 入口统一注册该 hook 并记录修改前后,保持 system_message 为 append。", "model": "gpt-5.2-codex"}
{"index": 30, "question": "Show a minimal example snippet for Advanced Session Hooks: Intercepting and modifying user prompts.", "answer": "最小示例(来源:本地 SDK 源码): \n```python\nimport asyncio\nfrom copilot import CopilotClient\n\nasync def on_user_prompt_submitted(hook_input, _ctx):\n return {\n \"modifiedPrompt\": hook_input[\"prompt\"].strip(),\n \"additionalContext\": \"normalized\",\n }\n\nasync def main():\n client = CopilotClient()\n await client.start()\n session = await client.create_session({\n \"model\": \"gpt-5\",\n \"hooks\": {\"on_user_prompt_submitted\": on_user_prompt_submitted},\n })\n await session.send_and_wait({\"prompt\": \" hello \"})\n await client.stop()\n\nasyncio.run(main())\n```\n建议/落地:在 pipe 入口统一注册该 hook并记录修改前后以便审计。", "model": "gpt-5.2-codex"}
{"index": 31, "question": "List recommended configuration defaults for Advanced Session Hooks: Intercepting and modifying user prompts.", "answer": "- 推荐默认值(来源:本地 SDK 源码 copilot/types.py`hooks={\"on_user_prompt_submitted\": handler}`handler 默认返回 `{\"modifiedPrompt\": input[\"prompt\"], \"additionalContext\": \"\", \"suppressOutput\": False}` 或 `None`(不改写)。 \n- `create_session`/`resume_session` 都要传同一 hooks`system_message` 保持 append 以保留 guardrails。 \n- 建议/落地pipe 入口统一注册该 hook 并记录原/改提示用于审计。", "model": "gpt-5.2-codex"}
{"index": 32, "question": "How does Advanced Session Hooks: Intercepting and modifying user prompts relate to building a custom Agent?", "answer": "Hooks 是会话级拦截器,对所有自定义 Agent 都生效;因此它是定义 Agent 行为边界和输入规范化的“外层控制”,而不是 Agent Manifest 本身(来源:本地 SDK 源码 copilot/types.py、copilot/session.py。 \n在自定义 Agent 场景下,`on_user_prompt_submitted` 可统一清洗/增强用户指令,再由 custom_agents 的 prompt 决定具体风格。 \n建议/落地pipe 启动时同时注入 custom_agents + hooks确保 resume 时同样注册 hooks。", "model": "gpt-5.2-codex"}
{"index": 33, "question": "Give a concise overview of Workspace Virtualization: Handling CWD for remote/virtual files.", "answer": "SDK 只有 `SessionConfig.working_directory` 来约束 CWD`attachments` 也要求真实路径,没有“虚拟工作区/FS provider”协议来源本地 SDK 源码 copilot/types.py、copilot/client.py。要处理远程/虚拟文件,需用自定义工具或 MCP 服务器实现 VFS 读写,或把内容落地到 working_directory 下的临时文件再交给模型。建议/落地pipe 固定 working_directory 映射到会话工作区,统一走 vfs_* 工具/MCP且在 on_permission_request 限制 write/shell。", "model": "gpt-5.2-codex"}
{"index": 34, "question": "Provide best practices and common pitfalls for Workspace Virtualization: Handling CWD for remote/virtual files.", "answer": "SDK 没有虚拟工作区/FS provider仅支持 `SessionConfig.working_directory` 与真实路径 `attachments`(来源:本地 SDK 源码),因此远程/虚拟文件需落地或用自定义工具/MCP 提供内容。 \n最佳实践固定 working_directory 到可控目录/会话 workspaceVFS 工具分块返回,配合 `available_tools`/`on_permission_request` 限权;常见坑:直接传虚拟路径/URL、working_directory 不存在、resume 未重注册工具/权限、一次性返回过大内容导致上下文截断。 \n建议/落地pipe 入口做 remote→临时文件或 vfs_* 映射,并持久化路径索引。", "model": "gpt-5.2-codex"}
{"index": 35, "question": "Show a minimal example snippet for Workspace Virtualization: Handling CWD for remote/virtual files.", "answer": "SDK 仅提供 `SessionConfig.working_directory` 与真实路径附件,没有虚拟 FS/provider来源本地 SDK 源码 copilot/types.py。最小做法是把远程内容落地到 CWD 下再作为附件传入: \n```python\nimport asyncio\nimport pathlib\nfrom copilot import CopilotClient\n\nasync def main():\n client = CopilotClient()\n await client.start()\n\n workdir = \"/tmp/owui\"\n pathlib.Path(workdir).mkdir(parents=True, exist_ok=True)\n remote_content = \"...\"\n path = f\"{workdir}/remote.txt\"\n pathlib.Path(path).write_text(remote_content)\n\n session = await client.create_session({\"model\": \"gpt-5\", \"working_directory\": workdir})\n await session.send_and_wait({\n \"prompt\": \"Summarize remote.txt\",\n \"attachments\": [{\"type\": \"file\", \"path\": path}],\n })\n await client.stop()\n\nasyncio.run(main())\n```\n建议/落地pipe 入口固定 working_directory落地临时文件后立即处理并用 `on_permission_request` 限制 write/shell。", "model": "gpt-5.2-codex"}
{"index": 36, "question": "List recommended configuration defaults for Workspace Virtualization: Handling CWD for remote/virtual files.", "answer": "推荐默认值(来源:本地 SDK 源码 copilot/types.py \n- `working_directory`: 指向可控的会话目录(如 session.workspace_path 或 /tmp/<app>/<session_id> \n- `available_tools`: 只允许 vfs_* 或只读工具,避免任意 shell/write \n- `on_permission_request`: 默认拒绝 write/shell仅放行 read/mcp \n- `attachments`: 仅真实路径文件;远程内容先落地 \n\n建议/落地pipe 入口统一映射 remote→临时文件或 VFS 工具,并持久化映射表。", "model": "gpt-5.2-codex"}
{"index": 37, "question": "How does Workspace Virtualization: Handling CWD for remote/virtual files relate to building a custom Agent?", "answer": "SDK 没有虚拟文件系统协议custom agent 只能依赖 `working_directory` 和真实路径附件,因此处理远程/虚拟文件需通过自定义工具/MCP 并在 agent prompt/available_tools 中强制使用(来源:本地 SDK 源码 copilot/types.py、copilot/client.py。 \n这使“工作区虚拟化”成为 agent 能力与权限设计的一部分,需要 `available_tools`/`on_permission_request`/hooks 来限定访问范围。 \n建议/落地pipe 中固定 working_directory 到会话目录,注入 vfs_* 工具白名单并在 prompt 明确使用 VFS。", "model": "gpt-5.2-codex"}
{"index": 38, "question": "Give a concise overview of Error Recovery: Handling session disconnects and re-auth.", "answer": "本地 SDK 源码中没有显式“重连”API断线通常表现为请求异常或 `SessionEventType.SESSION_ERROR`,需要 `stop()/force_stop()` 后重新 `start()` 并 `resume_session`(失败则 replay来源本地 SDK 源码)。 \n认证可用 `get_auth_status()` 检测,未认证时重新创建 `CopilotClient({\"github_token\": token})`;若 `cli_url` 连接外部服务器则认证由外部处理(来源:本地 SDK 源码)。\n\n```python\nimport asyncio\nfrom copilot import CopilotClient\nfrom copilot.generated.session_events import SessionEventType\n\nasync def safe_send(session_id, prompt, token):\n client = CopilotClient({\"github_token\": token})\n await client.start()\n try:\n session = await (client.resume_session(session_id) if session_id else client.create_session({\"model\": \"gpt-5\"}))\n session.on(lambda ev: ev.type == SessionEventType.SESSION_ERROR and print(ev.data.message))\n resp = await session.send_and_wait({\"prompt\": prompt})\n return resp, session.session_id\n except Exception:\n await client.stop()\n client = CopilotClient({\"github_token\": token})\n await client.start()\n session = await client.create_session({\"model\": \"gpt-5\"})\n resp = await session.send_and_wait({\"prompt\": prompt})\n return resp, session.session_id\n finally:\n await client.stop()\n```\n\n建议/落地:持久化 session_id 与 token优先 resume失败再回放最近 N 轮。", "model": "gpt-5.2-codex"}
{"index": 39, "question": "Provide best practices and common pitfalls for Error Recovery: Handling session disconnects and re-auth.", "answer": "最佳实践(基于本地 SDK 源码):监听 `SESSION_ERROR`/捕获请求异常后 `stop()/force_stop()`→重新 `start()`→`resume_session`(失败再 replay启动时用 `get_auth_status()` 校验并在失效时重建 `CopilotClient({\"github_token\": ...})`,同时在 resume 时重新注册 tools/hooks/权限处理器。 \n常见坑把 `cli_url` 与 `github_token` 混用、忘记重注册工具导致 `tool not supported`、缺少 `on_permission_request` 导致默认拒绝、以及不回退到 replay 造成会话中断。 \n建议/落地:无状态后端持久化 `session_id+token`、实现幂等的重试/回放策略并记录错误上下文用于告警。", "model": "gpt-5.2-codex"}
{"index": 40, "question": "Show a minimal example snippet for Error Recovery: Handling session disconnects and re-auth.", "answer": "最小示例(本地 SDK 源码copilot/client.py、copilot/session.py \n```python\nimport asyncio\nfrom copilot import CopilotClient\n\nasync def send_with_recovery(session_id, prompt, token):\n client = CopilotClient({\"github_token\": token})\n await client.start()\n try:\n session = await (client.resume_session(session_id) if session_id else client.create_session({\"model\": \"gpt-5\"}))\n resp = await session.send_and_wait({\"prompt\": prompt})\n return resp.data.content, session.session_id\n except Exception:\n await client.stop()\n client = CopilotClient({\"github_token\": token})\n await client.start()\n session = await client.create_session({\"model\": \"gpt-5\"})\n resp = await session.send_and_wait({\"prompt\": prompt})\n return resp.data.content, session.session_id\n finally:\n await client.stop()\n```\n建议/落地:持久化 token+session_idresume 失败再重建并回放必要历史。", "model": "gpt-5.2-codex"}
{"index": 41, "question": "List recommended configuration defaults for Error Recovery: Handling session disconnects and re-auth.", "answer": "推荐默认值(字段来源:本地 SDK 源码CopilotClientOptions—auto_start=True、log_level='info'、use_stdio=True非 cli_url、github_token=持久化 tokenSessionConfig/ResumeSessionConfig—hooks.on_error_occurred=handler、on_permission_request=非交互审批、working_directory 固定、mcp_servers/custom_agents/streaming 与首次一致、可固定 config_dir 并持久化 session_id。 \n常见坑cli_url 与 github_token 互斥、resume 时未重传 tools/hooks 导致 tool not supported 或默认拒绝。 \n建议/落地pipe 保存 token+session_id断线后先 resume_session失败再回放最近 N 轮。", "model": "gpt-5.2-codex"}
{"index": 42, "question": "How does Error Recovery: Handling session disconnects and re-auth relate to building a custom Agent?", "answer": "错误恢复与自定义 Agent 强相关Agent 定义与工具/权限处理器都是 session 级配置,断线或重认证后必须重建客户端并重新注入 `custom_agents`、`tools`、`hooks`、`on_permission_request`,否则能力/人格会丢失(来源:本地 SDK 源码)。 \n建议/落地:持久化 session_id+token重连优先 `resume_session` 并重新注册上述配置,失败再回放最近 N 轮。", "model": "gpt-5.2-codex"}
{"index": 43, "question": "Give a concise overview of Confirmation Events: programmatic handling of 'confirmation_required'.", "answer": "SDK 源码里没有独立的 `confirmation_required` 事件;权限确认是 JSONRPC `permission.request`,通过 `SessionConfig.on_permission_request` 回调处理(缺失或异常会默认拒绝)。 \n`session.on()` 只能用于审计/关联工具事件(如 `tool.user_requested`、`tool.execution_start`),不能替代确认逻辑。", "model": "gpt-5.2-codex"}
{"index": 44, "question": "Provide best practices and common pitfalls for Confirmation Events: programmatic handling of 'confirmation_required'.", "answer": "SDK 源码未定义 `confirmation_required` 事件,确认流程必须用 `SessionConfig.on_permission_request` 处理,缺失或异常会默认拒绝(来源:本地 SDK 源码)。 \n最佳实践/坑: \n- ✅ 对 `kind` 建立白名单策略(如仅允许 `read/mcp`),记录决策并用 `session.on` 仅做审计关联 `tool.user_requested` \n- ✅ `resume_session` 时重新注册 `on_permission_request`/hooks \n- ❌ 误以为 `session.on()` 可完成确认、返回非规范 dict 或抛异常导致全部拒绝", "model": "gpt-5.2-codex"}
{"index": 45, "question": "Show a minimal example snippet for Confirmation Events: programmatic handling of 'confirmation_required'.", "answer": "SDK 里没有 `confirmation_required` 事件,需用 `SessionConfig.on_permission_request` 处理权限确认。 \n最小示例 \n```python\nfrom copilot import CopilotClient\n\nclient = CopilotClient()\nawait client.start()\n\nsession = await client.create_session({\n \"model\": \"gpt-5\",\n \"on_permission_request\": lambda req, _: {\"kind\": \"approved\"},\n})\n```\n建议/落地:在 pipe 中统一注册 on_permission_request并在 resume_session 时重复注册。", "model": "gpt-5.2-codex"}
{"index": 46, "question": "List recommended configuration defaults for Confirmation Events: programmatic handling of 'confirmation_required'.", "answer": "推荐默认值SDK 无 confirmation_required 事件,改用 on_permission_request来源本地 SDK 源码):- 对 kind 建白名单(常见只放行 read/mcp- 未命中默认返回 denied-by-rules 并给 rules - resume_session 时重注册 on_permission_request/hooks。 \n建议/落地pipe 启动即注入 on_permission_request 并记录审批日志。", "model": "gpt-5.2-codex"}
{"index": 47, "question": "How does Confirmation Events: programmatic handling of 'confirmation_required' relate to building a custom Agent?", "answer": "SDK 没有 `confirmation_required` 事件,权限确认由 `create_session`/`resume_session` 的 `on_permission_request` 处理,且对同一 session 的所有 custom agent 生效(来源:本地 SDK 源码)。 \n建议/落地:在 pipe 启动时注入 `on_permission_request` 白名单策略并在 resume 时重注册,以约束高权限工具。", "model": "gpt-5.2-codex"}
{"index": 48, "question": "Give a concise overview of Skills: Conflict resolution and precedence defaults.", "answer": "在本地 SDK 源码中Skills 只通过 SessionConfig 的 `skill_directories` 与 `disabled_skills` 暴露,未定义冲突解析/优先级规则,因此行为取决于 CLI 运行时实现(来源:本地 SDK 源码 copilot/types.py。 \n为可控性建议确保技能名唯一、仅配置单一目录或由你自行按顺序拼接目录并用 `disabled_skills` 明确屏蔽同名/不应启用的技能。 \n建议/落地:在无状态后端每次 `create_session`/`resume_session` 时显式传入同一份 `skill_directories` 与 `disabled_skills`,避免因默认加载顺序差异导致行为漂移。", "model": "gpt-5.2-codex"}
{"index": 49, "question": "Provide best practices and common pitfalls for Skills: Conflict resolution and precedence defaults.", "answer": "SDK 源码仅暴露 SessionConfig/ResumeSessionConfig 的 `skill_directories` 和 `disabled_skills`,未定义冲突解析或优先级,行为由 CLI 侧决定(来源:本地 SDK 源码 copilot/types.py。 \n最佳实践技能名全局唯一、只配置单一目录或显式排序并用 `disabled_skills` 禁用冲突项;常见坑是依赖默认加载顺序或 resume 未重传这两个字段导致行为漂移。 \n建议/落地:在 pipe 启动与 resume 时统一注入同一份 `skill_directories`/`disabled_skills` 并记录变更。", "model": "gpt-5.2-codex"}
{"index": 50, "question": "Show a minimal example snippet for Skills: Conflict resolution and precedence defaults.", "answer": "SDK 里技能只通过 `skill_directories` 与 `disabled_skills` 配置,冲突/优先级未在源码定义,需自行避免同名。 \n最小示例 \n```python\nsession = await client.create_session({\n \"model\": \"gpt-5\",\n \"skill_directories\": [\"/opt/skills\"],\n \"disabled_skills\": [\"legacy_skill\"],\n})\n``` \n建议技能名唯一并在 `resume_session` 时重传这两项。", "model": "gpt-5.2-codex"}
{"index": 51, "question": "List recommended configuration defaults for Skills: Conflict resolution and precedence defaults.", "answer": "推荐默认值(基于 SDK 仅暴露 `skill_directories`/`disabled_skills` \n- `skill_directories`: 明确设置为单一目录(或你自己控制顺序的列表),不要依赖 CLI 默认搜索顺序 \n- `disabled_skills`: 默认空列表;如存在同名/旧技能则显式禁用 \n- Resume/Replay始终复用同一份配置以避免“隐式优先级”漂移", "model": "gpt-5.2-codex"}
{"index": 52, "question": "How does Skills: Conflict resolution and precedence defaults relate to building a custom Agent?", "answer": "Skills 是会话级配置(`skill_directories`/`disabled_skills`),会影响所有自定义 Agent 的可用能力与提示行为SDK 不定义冲突优先级,因此需要你在构建自定义 Agent 时确保技能命名唯一并显式禁用冲突项(来源:本地 SDK 源码)。 \n建议/落地:为 custom agent 固定同一份 skills 配置,并在 resume 时重复传入,避免行为漂移。", "model": "gpt-5.2-codex"}
{"index": 53, "question": "Give a concise overview of Debugging: Tracing JSON-RPC traffic in the SDK.", "answer": "本地 SDK 源码里 JsonRpcClient 没有内置 JSONRPC trace只能通过 `CopilotClient({\"log_level\":\"debug\"})` 打开 CLI 侧日志(非原始包)(来源:本地 SDK 源码)。 \n要抓实际 JSONRPC 流量,可 monkeypatch `JsonRpcClient._send_message`/`_read_message` 或包装 stdio 进行 tee最小示例如下。 \n建议/落地:在 pipe 中开启 patch 并对日志做脱敏与采样。 \n```python\nfrom copilot.jsonrpc import JsonRpcClient\n\norig_send = JsonRpcClient._send_message\nasync def _send(self, message):\n print(\">>\", message)\n return await orig_send(self, message)\nJsonRpcClient._send_message = _send\n\norig_read = JsonRpcClient._read_message\ndef _read(self):\n msg = orig_read(self)\n if msg:\n print(\"<<\", msg)\n return msg\nJsonRpcClient._read_message = _read\n```", "model": "gpt-5.2-codex"}
{"index": 54, "question": "Provide best practices and common pitfalls for Debugging: Tracing JSON-RPC traffic in the SDK.", "answer": "源码里 JsonRpcClient 没有内置 trace只能用 `CopilotClient({\"log_level\":\"debug\"})` 查看 CLI 侧日志或 monkeypatch `JsonRpcClient._send_message/_read_message` 做 tee来源本地 SDK 源码)。 \n最佳实践仅在调试启用、对 github_token/API key/内容脱敏、采样/限流、只读不改消息并注意线程安全;常见坑:记录完整 prompt/工具参数导致泄露,或同步打印阻塞 IO 造成超时/断连。", "model": "gpt-5.2-codex"}
{"index": 55, "question": "Show a minimal example snippet for Debugging: Tracing JSON-RPC traffic in the SDK.", "answer": "最小示例monkeypatch `JsonRpcClient` 打印收发): \n```python\nfrom copilot.jsonrpc import JsonRpcClient\n\norig_send = JsonRpcClient._send_message\nasync def _send(self, msg):\n print(\">>\", msg)\n return await orig_send(self, msg)\nJsonRpcClient._send_message = _send\n\norig_read = JsonRpcClient._read_message\ndef _read(self):\n msg = orig_read(self)\n if msg:\n print(\"<<\", msg)\n return msg\nJsonRpcClient._read_message = _read\n```\n建议仅在调试启用并脱敏日志。", "model": "gpt-5.2-codex"}
{"index": 56, "question": "List recommended configuration defaults for Debugging: Tracing JSON-RPC traffic in the SDK.", "answer": "推荐默认值(基于 SDK 源码):- 生产环境关闭 trace- 调试时 `CopilotClient({\"log_level\":\"debug\"})`- 如需 JSONRPC 报文,采用临时 monkeypatch 并启用脱敏/采样;- 不修改消息内容,避免线程阻塞。 \n建议/落地pipe 中通过环境变量开关调试与采样率。", "model": "gpt-5.2-codex"}
{"index": 57, "question": "How does Debugging: Tracing JSON-RPC traffic in the SDK relate to building a custom Agent?", "answer": "JSONRPC trace 能看到 custom_agents 在 `session.create/session.resume` 的实际 payload以及后续 `tool.call`/`permission.request`/`hooks.invoke` 是否按预期触发,从而验证自定义 Agent 的能力与权限边界(来源:本地 SDK 源码)。 \n建议/落地:仅在开发环境开启并脱敏日志;用它核对 customAgents/tools/权限决策与 MCP 配置。", "model": "gpt-5.2-codex"}
{"index": 58, "question": "Give a concise overview of Billing & Policies: How seat management affects SDK features.", "answer": "- 访问与政策SDK 依赖 Copilot CLI需有效 Copilot 订阅且组织/企业未禁用 CLI否则无法使用https://raw.githubusercontent.com/github/copilot-sdk/main/README.mdhttps://raw.githubusercontent.com/github/copilot-cli/main/README.md \n- 计费与认证:每次 prompt 计入 Copilot premium requests 配额PAT 需 Copilot Requests 权限https://raw.githubusercontent.com/github/copilot-sdk/main/README.mdhttps://raw.githubusercontent.com/github/copilot-cli/main/README.md\n\n建议/落地:启动先 `get_auth_status()`,失败时提示分配席位/启用 CLI 或更新 token", "model": "gpt-5.2-codex"}
{"index": 59, "question": "Provide best practices and common pitfalls for Billing & Policies: How seat management affects SDK features.", "answer": "最佳实践:确认用户/组织席位允许 Copilot CLI/SDK 使用,启动前用 `get_auth_status()` 预检,使用 PAT 时确保具备 Copilot Requests 权限https://raw.githubusercontent.com/github/copilot-sdk/main/README.mdhttps://raw.githubusercontent.com/github/copilot-cli/main/README.md。 \n常见坑未分配席位或组织/企业策略禁用 CLI 导致 SDK 失效、PAT 权限不足或 premium requests 用尽引发认证/配额错误https://raw.githubusercontent.com/github/copilot-sdk/main/README.mdhttps://raw.githubusercontent.com/github/copilot-cli/main/README.md。 \n建议/落地:在 pipe 启动时做席位与鉴权检测并给出明确提示/自助链接,记录每次请求计费与配额告警阈值以便限流。", "model": "gpt-5.2-codex"}
{"index": 60, "question": "Show a minimal example snippet for Billing & Policies: How seat management affects SDK features.", "answer": "基于 SDK README最小示例是在启动时用 `get_auth_status()` 检查是否已分配席位/可用认证https://raw.githubusercontent.com/github/copilot-sdk/main/README.md。 \n```python\nimport asyncio\nfrom copilot import CopilotClient\n\nasync def main():\n client = CopilotClient()\n await client.start()\n\n auth = await client.get_auth_status()\n if not auth.isAuthenticated:\n raise RuntimeError(\"Copilot 未认证/未分配席位,请先登录或分配权限\")\n\n # ...正常创建 session...\n await client.stop()\n\nasyncio.run(main())\n```\n建议/落地pipe 启动先做该检查,失败时提示分配席位或更新 PAT 权限。", "model": "gpt-5.2-codex"}
{"index": 61, "question": "List recommended configuration defaults for Billing & Policies: How seat management affects SDK features.", "answer": "推荐默认值CopilotClient 用 `use_logged_in_user=True`、`auto_start=True`、`log_level=\"info\"`;如用 PAT 则用 `github_token` 注入并确保具备 Copilot Requests 权限,且启动前 `get_auth_status()` 失败即按席位/策略限制处理来源https://raw.githubusercontent.com/github/copilot-sdk/main/README.md https://raw.githubusercontent.com/github/copilot-cli/main/README.md。 \n建议/落地pipe 启动先做 auth 预检并给出续订/分配席位提示,避免会话中途才触发权限或配额错误。", "model": "gpt-5.2-codex"}
{"index": 62, "question": "How does Billing & Policies: How seat management affects SDK features relate to building a custom Agent?", "answer": "席位与组织策略决定 Copilot CLI/SDK 是否可用,因此自定义 Agent 只是同一会话上的配置,前提仍是有有效订阅与允许的策略,否则 Agent 无法运行https://raw.githubusercontent.com/github/copilot-sdk/main/README.md https://raw.githubusercontent.com/github/copilot-cli/main/README.md。 \n建议/落地:在创建自定义 Agent 前先 `get_auth_status()` 预检并对配额/政策限制做降级提示。", "model": "gpt-5.2-codex"}
{"index": 63, "question": "How to configure MCP servers in ~/.copilot/config.json for SDK usage?", "answer": "我在 CLI 文档里未看到 `~/.copilot/config.json` 的 MCP 配置结构CLI 仅提供 `/mcp add|edit|delete|show` 命令管理配置来源Copilot CLI 文档)。 \n因此建议用 CLI 写入配置,并在 SDK 中用 `config_dir` 指向该配置目录;或直接在 SDK 的 `create_session(mcp_servers=...)` 传入。 \n如果你确认存在 `~/.copilot/config.json` 的具体 schema请提供示例或路径我可按你的文件格式给出精确配置。", "model": "gpt-5.2-codex"}
{"index": 64, "question": "What CLI environment variables force 'Agent' mode vs 'Generic' mode?", "answer": "CLI 文档未提到任何“通过环境变量强制 Agent 模式/Generic 模式”的开关Copilot CLI 文档)。 \n已知与模式相关的是命令/参数,例如 `--agent` 选择自定义 agent或交互内 `/agent` 切换;但不是环境变量。 \n如果你指的是内部实验开关请提供具体变量名或版本线索我可继续追索。", "model": "gpt-5.2-codex"}
{"index": 65, "question": "Explain the 'confirmation' flow in CLI and how it maps to SDK events.", "answer": "- CLI工具权限弹窗确认可用 `/allow-all` 或 `/yolo` 自动批准。(源: https://raw.githubusercontent.com/github/copilot-cli/main/README.md \n- SDK对应 JSONRPC `permission.request`→`on_permission_request` 返回 `approved/denied-*``session.on()` 只用于审计 tool.execution_*,无独立 confirmation 事件。(源: 本地 SDK 源码 copilot/client.py、copilot/session.py、copilot/types.py\n\n示例\n```python\nfrom copilot import CopilotClient\nfrom copilot.generated.session_events import SessionEventType\n\nclient = CopilotClient()\nawait client.start()\nsession = await client.create_session({\n \"model\": \"gpt-5\",\n \"on_permission_request\": lambda req, _: {\"kind\": \"approved\"},\n})\nsession.on(lambda ev: ev.type == SessionEventType.TOOL_EXECUTION_START and print(ev.data.tool_name))\n```\n\n建议/落地:非交互模式实现 `on_permission_request` 并记录 `tool.execution_*` 事件。", "model": "gpt-5.2-codex"}
{"index": 66, "question": "Does the CLI support 'dry-run' permission checks for tools?", "answer": "CLI 文档未提到“dry-run”权限检查只有正常确认与 `/allow-all`、`/yolo` 自动批准https://raw.githubusercontent.com/github/copilot-cli/main/README.md。SDK 侧仅有 `permission.request`→`on_permission_request` 与 `hooks.on_pre_tool_use` 的拦截/拒绝(本地 SDK 源码),不提供不执行的试运行;建议/落地pipe 中用 `on_permission_request` 记录/拒绝高风险工具并在 hooks 做审计。", "model": "gpt-5.2-codex"}
{"index": 67, "question": "What are the undocumented requirements for 'workspace' context updates?", "answer": "CLI 文档与本地 SDK 源码都未提到“workspace context updates”的未公开要求SDK 只暴露 `working_directory` 与 `session.workspace_path`,以及 `attachments` 真实路径(来源:本地 SDK 源码)。 \n如果你指的是 CLI 的“工作区/会话文件”更新规则,请提供具体页面或日志片段,我可以进一步定位。", "model": "gpt-5.2-codex"}
{"index": 68, "question": "How does the CLI handle 'device code' re-authentication automatically?", "answer": "CLI 文档未说明“自动 device code 重新认证”的具体机制;只提到未登录时使用 `/login`PAT 可通过 `GH_TOKEN/GITHUB_TOKEN` 提供Copilot CLI 文档)。因此无法确认是否有自动 re-auth 或缓存刷新策略。 \n若需无交互场景请使用 SDK 的 `github_token` 注入并自行实现 device flow 轮询与持久化。", "model": "gpt-5.2-codex"}

File diff suppressed because it is too large Load Diff