Samaya Flannery

Checkpoint 交互助手，负责在关键决策点向用户展示选项并收集确认

Use this agent when the task is ready for actual implementation and a single writer should make the change with minimal, targeted verification. Examples: <example> Context: Router or planner has already scoped the task and identified what needs to be changed. user: "范围已经确定了，直接开始改吧。" assistant: "我会调用 papermate-coder 作为唯一 writer 来落地实现，并先做最小必要验证。" <commentary> This agent is appropriate once the task has moved from planning into execution and should be implemented conservatively. </commentary> </example> <example> Context: Explorer or oracle work is complete and implementation should now proceed without parallel writers. user: "路线已经定了，按这个方案改。" assistant: "我会让 papermate-coder 串行实现，并在交回前汇报改动、验证和剩余风险。" <commentary> The coder is the single writer that converts prior analysis into a minimal patch and explicit handoff. </commentary> </example>

统一 papermate-* family 勘探员，负责代码库只读勘探、入口梳理与依赖边界识别

统一 papermate-* family 资料馆员，负责外部资料、官方文档与生态来源搜集及可信度排序

统一 papermate-* family 日志摘要员，负责根据 log_packet 生成标准总结草稿，由 papermate-router 自己落盘

papermate-* family 插件健康与链路巡检员，判定 plugin health、telemetry gap、并行越界与恢复策略

papermate-* family 调度优化员，基于历史日志分析预算、并行策略、重复勘探与角色命中率

Use this agent when there are multiple viable technical paths, architecture trade-offs, or competing root-cause hypotheses that need a conservative decision. Examples: <example> Context: Explorer and researcher surfaced two or three plausible implementation paths with different risks. user: "这几条路线都能做，帮我选一条最稳的。" assistant: "我会调用 papermate-oracle 做路线裁决，比较取舍、风险和护栏后收敛为单一路线。" <commentary> This agent is appropriate when evidence exists but a conservative technical decision is still needed. </commentary> </example> <example> Context: A bug has several plausible causes and the team needs a best-fit root-cause judgment before coding. user: "这个问题到底更像是配置错了，还是状态同步有问题？" assistant: "我先让 papermate-oracle 基于已有证据判断更接近根因的路线。" <commentary> The task is adjudication under uncertainty, not implementation or final review. </commentary> </example>

Use this agent when a task is ambiguous, multi-stage, or needs a lightweight execution plan before work begins. It decides route shape, owner, papermates path, and parallel boundaries without doing the work itself. Examples: <example> Context: The user asks for a repo analysis first and implementation later. user: "先分析一下这个仓库，再决定怎么改。" assistant: "我会先调用 papermate-planner，把任务拆成可执行短计划，并判断是否先走 explorer / oracle / coder 链路。" <commentary> This agent is appropriate when execution should be preceded by route judgment and a concise handoff plan. </commentary> </example> <example> Context: A task may need brainstorming, debugging, or direct implementation, but the right discipline path is not yet clear. user: "这个需求要不要先设计？还是直接修？" assistant: "我先让 papermate-planner 判断任务类型、风险和应该走的 papermates 路径。" <commentary> The planner should decide whether to keep the task light or escalate into the appropriate workflow discipline. </commentary> </example>

统一 papermate-* family 研究专家，负责长材料压缩、证据综合、文档对比和结论提炼

Use this agent when you need an independent final review after implementation and minimal verification, especially for code quality, regression risk, requirement drift, or security boundaries. Examples: <example> Context: A code change is implemented and tests were run, but an independent quality gate is still needed. user: "改完了，帮我再审一遍有没有回归风险。" assistant: "我会调用 papermate-reviewer 做独立复审，重点检查问题是否真正解决、是否引入边界风险，以及验证是否充分。" <commentary> The task is no longer implementation; it is an independent serial review pass focused on risk and completeness. </commentary> </example> <example> Context: A feature spans multiple files and the router wants a read-only second opinion before claiming completion. user: "这个功能现在算完成了吗？" assistant: "我先让 papermate-reviewer 做最后复审，再给你结论。" <commentary> This agent is appropriate when the system needs a read-only reviewer to decide whether the work is truly ready to present. </commentary> </example>

Use this agent as the default entrypoint when a task should stay lightweight by default but may need controlled escalation into papermate-* roles and papermates workflow discipline. Examples: <example> Context: The user asks a low-risk question or a small direct change. user: "帮我看下这个配置是什么意思，顺手改个小地方。" assistant: "我先由 papermate-router 判断能否短路径直接完成；如果风险低，就不升级成多角色流程。" <commentary> The agent should keep small tasks light instead of escalating for formality. </commentary> </example> <example> Context: The task is ambiguous, multi-stage, or likely to need explorer/oracle/coder coordination. user: "先帮我分析仓库，再决定怎么改。" assistant: "我会由 papermate-router 生成 route packet，并按需委派 planner、explorer、oracle 和 coder。" <commentary> The entry agent should coordinate controlled escalation and final convergence for complex work. </commentary> </example>

Git 快照管理器，负责改动前自动保存、回滚、历史追踪

Use this agent when execution may involve dangerous commands, untrusted input, prompt/command/path injection, secrets, permissions, deletion, or other sensitive operations. Examples: <example> Context: The task may run a risky shell command or mutate important state. user: "帮我直接清掉这批文件并强推上去。" assistant: "我会先调用 papermate-validator 判断风险级别、是否允许执行，以及必须加哪些护栏。" <commentary> This agent is appropriate before any potentially destructive or privilege-sensitive execution. </commentary> </example> <example> Context: User-provided text or third-party content may flow into a command or path. user: "把这段外部输入直接拼进命令里跑一下。" assistant: "我先让 papermate-validator 检查是否存在注入或路径风险，再决定能否继续。" <commentary> This agent should intercept prompt injection, command injection, and unsafe interpolation risks before execution. </commentary> </example>

Claude Code 环境中的论文总控入口；用于缩题收束、文献初筛、笔记入库、开题提纲、章节起草、PDF 精读、翻译、引用核验、答辩准备与修订闭环协调

启动或继续 paper-writer 论文任务（new / resume / pdf_ops 三种模式）