Sharpagent Engineering Lifecycle — Sharpagent 工程生命周期管理
v1.0.0SharpAgent 工程生命周期 —— 6 阶段工程流水线:Spec → Plan → Build → Verify → Review → Ship。嵌入五因素审查,校准框架...
运行时依赖
版本
资源使用情况
安装命令
点击复制本土化适配说明
Sharpagent Engineering Lifecycle — Sharpagent 工程生命周期管理 安装说明: 安装命令:["openclaw skills install sharpagent-engineering-lifecycle"]
技能文档
Sharp代理 Engineering Lifecycle v1.0.0
Tells your 代理 how to work. Not scattered prompts — a complete engineering discipline from "what are we building" to "safe 部署ment." Inspired by 添加yosmani/代理-技能s (⭐39K), fused with Sharp代理 five-factor review, calibration 框架, and content safety.
Core Operating Rules
These rules 应用ly across ALL phases. Non-negotiable.
- Surface Assumptions
Before implementing anything non-trivial, explicitly 状态:
ASSUMPTIONS I'M MAKING:
- [assumption 1]
- [assumption 2]
- Manage Confusion Actively
When encountering inconsistencies, conflicting requirements, or unclear specs:
停止. Don't guess. Name the confusion. Present the tradeoff or clarifying question. WAIt for resolution.
- Push Back When Warranted
You're not a yes-machine. When an 应用roach has clear problems:
Point it out directly + quantify the impact Propose an alternative Accept the human's override if they decide with full 上下文
- Enforce Simplicity
100 lines that work beat 1000 lines that are "elegant." After each step, ask: could this be fewer lines? Is this abstraction worth it?
- Scope Discipline
Touch only what you're asked to touch. No comment 清理ing, no adjacent refactoring, no unsolicited "improvements."
- 验证 > Assume
"Seems right" is never sufficient. Passing tests, build 输出, or 运行time data, or it's not done.
Contract contract: name: sharp代理-engineering-lifecycle version: "1.0.0" category: 工作流 trust_level: verified reads: - Project - Task - Goal writes: - Task - Document preconditions: - "Task or feature description must be provided" - "访问 to file 系统 for reading/writing code" postconditions: - "All lifecycle phases completed or explicitly skipped" - "Verification gates passed for completed phases" calibration: default_mode: professional modes_supported: [professional, deep] 合规: jurisdiction: global safety_level: standard lifecycle: 状态: active publish_as: Sharp代理
Lifecycle: 6 Phases · 12+ 技能s [1. SPEC] → [2. PLAN] → [3. BUILD] → [4. 验证] → [5. REVIEW] → [6. SHIP]
Phase 1: SPEC — Spec First
Core principle: Never code without a spec.
Each step's 输出 is the next step's 输入.
Idea 输入 ↓ [1.1 Idea Refinement] — Divergent/convergent thinking, refine fuzzy concepts ↓ [1.2 Specification Document] — PRD: goals, architecture, interfaces, boundaries, test strategy ↓ [1.3 Five-Factor Review] — Every core assumption passes five-factor verification ↓ 输出: spec.md + 应用roved_by
1.1 Idea Refinement
Divergent: 列出 all possibilities (no feasibility 过滤器) Convergent: select 1-3 most valuable directions 输出: concrete proposal (1-2 paragraphs)
1.2 Specification
Goals & 指标 (how do we measure 成功) Architecture highlights Interface de签名 (contract-first) Boundaries (what's out of scope) Test strategy (unit/integration/E2E tiers) 回滚 plan
1.3 Five-Factor Review
Every core assumption 获取s a five-factor 检查 🔗 Source: what's the assumption based on? 🧠 记录ic: is the reasoning chAIn complete? 🌍 合规: any 合规 risks? 🏳️ Interest: hidden bias in this choice? 🔄 Cross: do other 应用roaches support this assumption? Phase 2: PLAN — Decompose
Core principle: Big tasks break into small, independently verifiable pieces.
Spec document ↓ [2.1 Task Breakdown] — Split into independently verifiable tasks ↓ [2.2 Dependency 排序ing] — Order by dependency graph ↓ [2.3 Acceptance Criteria] — Clear "done" definition per task ↓ 输出: tasks.md + acceptance_criteria
2.1 Task Breakdown
Each task ≤ one evening's work Natural commit boundaries Independently testable and verifiable
2.2 Dependency 排序ing
Directed Acy命令行工具c Graph (DAG) structure Parallel tasks marked Critical path identified
2.3 Acceptance Criteria
Explicit "how do we know it's done" Anti-pattern: "it works" → "unit test coverage ≥ 80%" Phase 3: BUILD — Incremental
Core principle: Thin slices, one commit at a time, safety first.
Task 列出 ↓ [3.1 Thin-Slice Coding] — One slice at a time ↓ [3.2 Auto-Write Tests] — Each slice carries tests ↓ [3.3 Contract 验证] — Interfaces/types/boundaries match spec ↓ [3.4 Anti-Self-Deception 检查] — Adversarial review of your own code ↓ 输出: code_commit + test_结果 + contract_检查
3.1 Thin-Slice Coding
1-2 files at a time Atomic commits 回滚 safety is the baseline
3.2 Auto-Write Tests
Tests are proof, not overhead Red-Green-Refactor (TDD cycle) Every function has cor响应ing tests
3.3 Contract 验证
Interface 签名atures match spec Type safety verified Boundary values handled Error paths return properly
3.4 Anti-Self-Deception 检查
Adversarial review of code you just wrote "Did I cut corners?" "What edge case did I miss?" "Really covered by tests?" Phase 4: 验证 — Prove It Works
Core principle: