workbench-sdd
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseWorkbench SDD
Workbench SDD
Use this skill when a task needs to move from a fuzzy request into executable work.
This is the default workbench planning protocol unless the issue explicitly says the task is a quick fix, emergency repair, or direct verification run.
当任务需要从模糊的需求转化为可执行工作时,使用此Skill。
除非问题明确说明任务是快速修复、紧急修复或直接验证运行,否则这是默认的工作台规划协议。
Pipeline
流程阶段
Move work through these stages:
- Raw Requirement
- Capture the user's literal request.
- Separate confirmed facts from assumptions.
- Name the owner, expected output, and known constraints.
- Product Design
- Define the user-facing behavior, success criteria, non-goals, and edge cases.
- Keep the scope narrow unless the user asks for expansion.
- Technical Design
- Identify the runtime owner, data path, files, commands, integrations, and risk surface.
- Prefer existing stable routes over new infrastructure.
- Task List
- Convert the design into bounded tasks with owner, evidence, and verification criteria.
- Keep tasks executable by one agent unless the issue explicitly asks for parallel work.
- Mark when the owner must keep the objective alive across turns until verified.
GOAL_MODE: yes
- Execution And Verification
- Execute only the assigned slice.
- Verify on the real path.
- Report evidence before claiming done.
按以下阶段推进工作:
- 原始需求
- 捕获用户的字面请求。
- 区分已确认事实与假设。
- 明确负责人、预期输出及已知约束。
- 产品设计
- 定义面向用户的行为、成功标准、非目标及边缘案例。
- 除非用户要求扩展,否则保持范围狭窄。
- 技术设计
- 确定运行时负责人、数据路径、文件、命令、集成点及风险面。
- 优先使用现有稳定路径,而非新建基础设施。
- 任务列表
- 将设计转化为有明确负责人、验证依据及验证标准的独立任务。
- 除非问题明确要求并行工作,否则任务需由单个Agent可执行。
- 当负责人必须在多轮交互中持续推进目标直至验证通过时,标记。
GOAL_MODE: yes
- 执行与验证
- 仅执行分配的任务部分。
- 在实际路径上进行验证。
- 报告验证依据后再宣告完成。
Stage Gate Rules
阶段门规则
- Do not jump from an ambiguous raw request directly to implementation.
- Do not expand blast radius just because more agents or tokens are available.
- If a stage is already answered by issue text or prior comments, cite that evidence and move forward.
- If the task is blocked by a missing decision, state the smallest decision needed.
- If the task is low-risk and obvious, compress the stages into a short SDD card instead of creating ceremony.
- If the user says "go", "continue", or an equivalent approval after a reviewed stage, record that as the gate approval context before proceeding.
- When a stage has both proxy and agent-authored artifacts, use the complete agent-authored artifact as primary and keep the proxy comment as trace evidence.
- If an issue is marked , the Task List must include
GOAL_MODE: yes, closeout gates, and operator-call conditions before execution starts.GOAL_LOCK
- 不得从模糊的原始需求直接跳转到实现环节。
- 不得因可用Agent或令牌数量增加而扩大影响范围。
- 如果某阶段的内容已在问题文本或过往评论中给出,需引用相关依据并推进到下一阶段。
- 如果任务因缺少决策而受阻,说明所需的最小决策内容。
- 如果任务低风险且明确,可将阶段压缩为简短的SDD卡片,无需繁琐流程。
- 如果用户在某阶段审核后表示“开始”“继续”或类似批准,需记录该批准上下文后再推进。
- 当某阶段同时存在代理生成和Agent生成的工件时,以完整的Agent生成工件为主体,保留代理评论作为追溯依据。
- 如果问题标记为,任务列表必须包含
GOAL_MODE: yes、收尾门限及操作员呼叫条件,之后才能开始执行。GOAL_LOCK
Comment Structure
评论结构
Each SDD stage is a structured issue comment, not an issue status. Use this header for every stage artifact:
text
SDD_STAGE: [Raw Requirement / Product Design / Technical Design / Task List / Execution And Verification]
OWNER: [one agent or human owner]
STATUS: READY_FOR_REVIEW / APPROVED / BLOCKED
REVIEWER: Workbench Supervisor or designated reviewer
EVIDENCE: [files, commands, issue/comment IDs, or artifacts checked]
HANDOFF_SUMMARY: [five lines or fewer: what the next agent needs without rereading full history]
SCOPED_EVIDENCE: [exact comment IDs, run IDs, commit hashes, files, or artifact paths to inspect]
ANTI_OVER_READ: [sources to skip unless needed, such as full issue lists or full comment history]Put the stage-specific artifact after the header. Keep discussion replies separate from stage artifacts so the comment history remains scannable.
Every SDD stage comment must literally include the exact uppercase field names , , and in the header. Do not substitute semantic equivalents such as "Evidence", "Skipped", "Context", or prose bullets. Every SDD review comment must literally include . The next agent should be able to start from the handoff summary alone and deep-read only the listed .
HANDOFF_SUMMARYSCOPED_EVIDENCEANTI_OVER_READVERDICT_SUMMARYSCOPED_EVIDENCE每个SDD阶段都是结构化的问题评论,而非问题状态。为每个阶段工件使用以下头部:
text
SDD_STAGE: [Raw Requirement / Product Design / Technical Design / Task List / Execution And Verification]
OWNER: [one agent or human owner]
STATUS: READY_FOR_REVIEW / APPROVED / BLOCKED
REVIEWER: Workbench Supervisor or designated reviewer
EVIDENCE: [files, commands, issue/comment IDs, or artifacts checked]
HANDOFF_SUMMARY: [five lines or fewer: what the next agent needs without rereading full history]
SCOPED_EVIDENCE: [exact comment IDs, run IDs, commit hashes, files, or artifact paths to inspect]
ANTI_OVER_READ: [sources to skip unless needed, such as full issue lists or full comment history]将阶段特定的工件放在头部之后。将讨论回复与阶段工件分开,以便评论历史保持易读性。
每个SDD阶段评论必须在头部中严格包含大写字段名、和。不得使用语义等价的替代词,如“Evidence”“Skipped”“Context”或项目符号列表。每个SDD审核评论必须严格包含。下一个Agent应仅能从交接摘要入手,仅深入阅读列出的内容。
HANDOFF_SUMMARYSCOPED_EVIDENCEANTI_OVER_READVERDICT_SUMMARYSCOPED_EVIDENCEStage Gate Mechanics
阶段门机制
- Supervisor review happens between stages with .
VERDICT: PASS / FLAG / BLOCK - allows the next stage to start.
PASS - means the current stage needs a targeted correction before handoff.
FLAG - means the stage cannot proceed until the smallest blocking decision or missing input is resolved.
BLOCK - Issue statuses stay coarse: before work starts,
todowhile any SDD stage is active,in_progressafter execution is ready for final review, andin_reviewonly after accepted evidence.done
- 阶段之间由主管进行审核,给出。
VERDICT: PASS / FLAG / BLOCK - 允许启动下一阶段。
PASS - 表示当前阶段在交接前需要针对性修正。
FLAG - 表示在解决最小阻塞决策或缺失输入前,该阶段无法推进。
BLOCK - 问题状态保持粗略分类:工作开始前为,任何SDD阶段进行中为
todo,执行完成等待最终审核为in_progress,仅在接受验证依据后标记为in_review。done
Bypass Rules
绕过规则
- Workbench Admin may use for low-risk, obvious changes where a full SDD sequence would add noise.
SDD_BYPASS: quick-fix - Workbench Admin may use for time-critical repair work.
SDD_BYPASS: emergency - Bypass is not allowed for high-risk, ambiguous, multi-system, or public/private boundary changes.
- A bypassed issue still needs one clear owner, explicit scope, verification evidence, and Supervisor review before closure.
- 工作台管理员可对低风险、明确的变更使用,完整SDD流程会增加冗余。
SDD_BYPASS: quick-fix - 工作台管理员可对时间紧迫的修复工作使用。
SDD_BYPASS: emergency - 高风险、模糊、多系统或跨公私边界的变更不允许绕过流程。
- 绕过流程的问题仍需明确负责人、清晰范围、验证依据,并在关闭前经主管审核。
Execution Handoff
执行交接
- The Task List stage names the execution owner, exact files/resources, non-goals, approval gates, and verification commands.
- For , the Task List also names required build/test/help-smoke/docs-or-report/git-status gates and the conditions for calling the operator.
GOAL_MODE: yes - Execution owners implement only their assigned slice and do not create follow-on work unless explicitly requested.
- T8 or smoke-test work remains separate when the task list says it depends on a committed repo batch.
- Completion reports include changed files, verification output, residual risks, commit hash or artifact link, work directory, and branch.
- Live skill or agent-binding changes must have a backup and post-change verification before final PASS.
- If a run reaches evidence-ready state but does not post a stage artifact, the conductor may post a proxy artifact using run-message evidence and mark it as proxy/recovery evidence.
- 任务列表阶段需指定执行负责人、确切文件/资源、非目标、批准门限及验证命令。
- 对于的任务,任务列表还需指定必要的构建/测试/冒烟测试/文档或报告/git状态门限,以及呼叫操作员的条件。
GOAL_MODE: yes - 执行负责人仅实现分配的任务部分,除非明确要求,否则不得创建后续工作。
- 当任务列表显示依赖已提交的仓库批次时,T8或冒烟测试工作需保持独立。
- 完成报告需包含变更文件、验证输出、剩余风险、提交哈希或工件链接、工作目录及分支。
- 实时Skill或Agent绑定变更必须有备份,并在变更后进行验证,才能最终获得PASS。
- 如果运行达到可提交验证依据的状态但未发布阶段工件,指挥者可使用运行消息依据发布代理工件,并标记为代理/恢复依据。
Output Contract
输出约定
For planning or routing, return:
- : current stage.
SDD_STAGE - : facts and constraints.
CONFIRMED - : assumptions that still matter.
ASSUMPTIONS - : owner-scoped tasks.
NEXT_TASKS - : evidence required to close.
VERIFICATION
For execution, return:
- what changed,
- what was verified,
- what remains risky or blocked.
对于规划或路由,返回:
- : 当前阶段。
SDD_STAGE - : 事实与约束。
CONFIRMED - : 仍需关注的假设。
ASSUMPTIONS - : 负责人范围的任务。
NEXT_TASKS - : 关闭任务所需的验证依据。
VERIFICATION
对于执行,返回:
- 变更内容,
- 验证内容,
- 仍存在风险或阻塞的内容。