# Superpowers → Engineering Harness v0.1 实施手册 > 版本：v0.1 > 目标：在 Superpowers 之上增加一层 Project Engineering Harness，使 Coding Agent 的“完成”由状态、证据、质量门禁和收敛条件决定，而不是由模型自行宣布。 > 适用对象：Codex / Grok / Claude Code 等本地 Coding Agent。 > 实施原则：**State → Evidence → Gate → Convergence** --- ## 1. 项目目标 本项目不是重新实现 Coding Agent，也不是替代 Superpowers。 目标是在 Superpowers 现有开发工作流之上增加一层项目级控制平面： ```text Requirement ↓ Task Contract ↓ Superpowers Brainstorm / Plan / TDD / Development ↓ Deterministic Verification ↓ Adversarial Review ↓ Finding ↓ Reproduce ↓ Regression Test ↓ Fix ↓ Fresh Evidence ↓ Quality Gate ↓ PASS → CONVERGED → DONE BLOCKED → 继续修复 ``` 核心区别： ```text Superpowers: How to build software well. Engineering Harness: How to prove the task is acceptable. ``` --- ## 2. v0.1 非目标 v0.1 暂时不要实现： - Web 控制台 - 数据库 - 独立 Agent Runtime - 模型 Router - 多模型自动调度 - 企业规则中心 - PR Dashboard - 多项目管理 - 复杂评分算法 - LLM 自动质量评分 - 自研 Coding Agent v0.1 只使用： ```text Markdown YAML JSON Python Git Skills Existing Superpowers ``` --- ## 3. v0.1 六条 Iron Laws ### LAW 1 **NO DONE WITHOUT QUALITY GATE PASS.** 没有 Quality Gate PASS，不允许进入 DONE。 ### LAW 2 **NO QUALITY GATE PASS WITHOUT FRESH EVIDENCE.** 所有 Evidence 必须基于当前 Git HEAD。 ### LAW 3 **NO FINDING IS A BUG UNTIL REPRODUCED.** Reviewer 的发现只能是 PROPOSED Finding，必须经过 Reproduce 才能成为 CONFIRMED。 ### LAW 4 **NO CONFIRMED BUG IS CLOSED WITHOUT A REGRESSION TEST.** 确认的 Bug 必须增加 Regression Test。 ### LAW 5 **NO TASK STATE EXISTS ONLY IN MODEL CONTEXT.** 所有任务状态必须持久化在 `.harness/` 中。 ### LAW 6 **NO INFINITE FIX LOOP.** 达到最大迭代次数或长期不收敛时，必须进入 ESCALATED / CONVERGENCE_FAILED。 --- # 4. 总体架构 ```text User Requirement │ ▼ Task Contract Skill │ ▼ Superpowers Brainstorming │ ▼ Design / Spec │ ▼ Invariants / Risks │ ▼ Superpowers Planning │ ▼ Superpowers Implementation / TDD │ ▼ Deterministic Verification │ ▼ Adversarial Review │ ▼ Finding │ ▼ Reproduce / \ REJECTED CONFIRMED │ ▼ Regression Test │ ▼ Fix │ └───────┐ │ ▼ Verification │ ▼ Evidence │ ▼ Quality Gate / \ BLOCKED PASS │ │ └──── loop ▼ CONVERGED │ ▼ DONE ``` --- # 5. Repository 结构 建议创建独立项目： ```text superpowers-engineering-harness/ ├── SKILL.md │ ├── skills/ │ ├── task-contract/ │ │ └── SKILL.md │ │ │ ├── adversarial-review/ │ │ └── SKILL.md │ │ │ ├── reproduce-finding/ │ │ └── SKILL.md │ │ │ ├── collect-evidence/ │ │ └── SKILL.md │ │ │ ├── quality-gate/ │ │ └── SKILL.md │ │ │ └── convergence/ │ └── SKILL.md │ ├── schemas/ │ ├── task.schema.json │ ├── requirement.schema.json │ ├── invariant.schema.json │ ├── finding.schema.json │ └── evidence.schema.json │ ├── scripts/ │ ├── harness_status.py │ ├── collect_evidence.py │ ├── quality_gate.py │ └── validate_state.py │ ├── templates/ │ ├── current-task.yaml │ ├── requirements.yaml │ ├── invariants.yaml │ └── gate.yaml │ ├── tests/ │ ├── test_state_machine.py │ ├── test_quality_gate.py │ ├── test_evidence.py │ └── fixtures/ │ └── README.md ``` 接入业务项目后，在项目根目录增加： ```text .harness/ ├── config.yaml ├── current-task.yaml ├── requirements.yaml ├── invariants.yaml ├── gate.yaml │ ├── findings/ │ ├── F-0001.yaml │ └── ... │ ├── evidence/ │ ├── build.json │ ├── unit-test.json │ ├── integration-test.json │ └── review.json │ └── history/ └── TASK-xxxx/ ``` --- # 6. 状态机 ## 6.1 状态定义 v0.1 仅允许以下状态： ```text CREATED SPECIFYING PLANNED IMPLEMENTING VERIFYING REVIEWING REPRODUCING FIXING GATING BLOCKED CONVERGED DONE ESCALATED ``` ## 6.2 合法转换 ```text CREATED → SPECIFYING SPECIFYING → PLANNED PLANNED → IMPLEMENTING IMPLEMENTING → VERIFYING VERIFYING → IMPLEMENTING VERIFYING → REVIEWING REVIEWING → REPRODUCING REVIEWING → GATING REPRODUCING → REVIEWING REPRODUCING → FIXING FIXING → VERIFYING GATING → BLOCKED GATING → CONVERGED BLOCKED → IMPLEMENTING BLOCKED → REPRODUCING BLOCKED → ESCALATED CONVERGED → DONE ``` ## 6.3 禁止转换 至少显式禁止： ```text IMPLEMENTING → DONE VERIFYING → DONE REVIEWING → DONE FIXING → DONE BLOCKED → DONE ``` 唯一允许进入 DONE 的路径： ```text CONVERGED → DONE ``` --- # 7. current-task.yaml 建议模板： ```yaml task: id: TASK-001 title: "" description: "" state: CREATED iteration: 0 max_iterations: 5 spec: design: null plan: null requirements: total: 0 verified: 0 uncovered: [] invariants: total: 0 verified: 0 violated: [] verification: build: unknown unit_test: unknown integration_test: unknown findings: critical: 0 major: 0 minor: 0 confirmed: 0 proposed: 0 gate: status: unknown blocked_by: [] git: head: null timestamps: created_at: null updated_at: null ``` 要求： 1. 所有状态切换必须更新该文件。 2. 每次执行 Gate 前刷新 `git.head`。 3. Task 不允许只存在于 Agent Conversation Context 中。 4. Session 重启后必须先读取 `current-task.yaml`。 --- # 8. requirements.yaml ```yaml requirements: - id: REQ-001 statement: "" source: user priority: must status: pending evidence: [] - id: REQ-002 statement: "" source: spec priority: should status: pending evidence: [] ``` status： ```text pending implemented verified blocked ``` Quality Gate 默认要求： ```text priority=must 的 Requirement 必须全部 verified ``` --- # 9. invariants.yaml ```yaml invariants: - id: INV-001 statement: "" category: correctness severity: major status: pending verification: [] - id: INV-002 statement: "" category: concurrency severity: critical status: pending verification: [] ``` 推荐 category： ```text correctness transaction concurrency idempotency security authorization state_machine recovery data_consistency architecture ``` status： ```text pending verified violated ``` --- # 10. Finding Schema 每个 Reviewer Finding 单独保存： ```text .harness/findings/F-xxxx.yaml ``` 格式： ```yaml id: F-0001 type: suspected_defect severity: major category: concurrency status: PROPOSED title: "" claim: "" related_requirements: [] related_invariants: - INV-001 location: file: null line: null failure_scenario: - "" expected_behavior: "" possible_actual_behavior: "" recommended_reproducer: type: unit_test description: "" reproduction: command: null exit_code: null evidence: null regression_test: path: null fix: commit: null verification: regression: unknown full_suite: unknown ``` Finding 状态： ```text PROPOSED REPRODUCING CONFIRMED REJECTED FIXING FIXED VERIFIED CLOSED ``` 原则： ```text PROPOSED ≠ BUG CONFIRMED = BUG ``` --- # 11. Evidence Schema Evidence 必须绑定 Git HEAD。 示例： ```json { "type": "unit_test", "timestamp": "2026-08-24T14:30:00+08:00", "command": "pytest", "exit_code": 0, "passed": 142, "failed": 0, "skipped": 2, "commit": "abc123", "stdout_tail": "..." } ``` 最小必填字段： ```text type timestamp command exit_code commit ``` Evidence 类型： ```text build lint typecheck unit_test integration_test contract_test security review custom ``` --- # 12. Fresh Evidence 规则 Quality Gate 必须检查： ```text evidence.commit == git HEAD ``` 如果不一致： ```text Evidence = STALE ``` STALE Evidence 不可用于 Gate。 示例： ```text commit A ↓ pytest PASS ↓ 生成 evidence(commit=A) 代码修改 ↓ commit B / HEAD=B Gate 检查： A != B ↓ unit_test evidence = STALE ↓ BLOCKED ``` --- # 13. gate.yaml 默认模板： ```yaml gate: requirements: must_verified: true uncovered_allowed: 0 invariants: violated_allowed: 0 verification: build: required unit_test: required integration_test: optional findings: critical_allowed: 0 major_allowed: 0 regression: confirmed_finding_without_test: 0 evidence: must_match_head: true convergence: max_iterations: 5 ``` --- # 14. quality_gate.py 这是 v0.1 的核心模块。 不要让 LLM 判断 Gate。 ## 14.1 输入 读取： ```text .harness/current-task.yaml .harness/requirements.yaml .harness/invariants.yaml .harness/gate.yaml .harness/findings/* .harness/evidence/* git HEAD ``` ## 14.2 检查项 依次检查： 1. Harness 文件是否合法。 2. 当前状态是否允许执行 Gate。 3. Requirement 是否全部满足。 4. Invariant 是否存在 violated。 5. Required verification 是否有 Evidence。 6. Evidence 是否属于当前 HEAD。 7. Critical Finding 是否为 0。 8. Major Finding 是否为 0。 9. Confirmed Finding 是否全部有 Regression Test。 10. Regression Test 是否 verified。 ## 14.3 输出 PASS： ```text QUALITY GATE: PASS Requirements PASS Invariants PASS Build PASS Unit Tests PASS Critical Findings 0 Major Findings 0 Regression Debt 0 Evidence HEAD MATCH ``` BLOCKED： ```text QUALITY GATE: BLOCKED Blocking: - REQ-004 not verified - INV-002 violated - Major finding F-0012 is open - Unit-test evidence is stale ``` ## 14.4 Exit Code ```text 0 = PASS 1 = BLOCKED 2 = INVALID_HARNESS_STATE ``` ## 14.5 约束 只有 `quality_gate.py exit=0` 才允许： ```text GATING → CONVERGED ``` --- # 15. collect_evidence.py 作用： ```text 执行确定性命令 → 记录 exit code → 绑定当前 Git HEAD → 保存 Evidence ``` 期望调用： ```bash python scripts/collect_evidence.py \ --type unit_test \ --command "pytest" ``` 生成： ```text .harness/evidence/unit-test.json ``` v0.1 暂时不要求解析所有测试框架报告。 至少保存： ```text command exit_code timestamp commit stdout_tail stderr_tail ``` 命令失败时仍然需要保存 Evidence。 --- # 16. harness_status.py 提供统一状态视图。 调用： ```bash python scripts/harness_status.py ``` 输出： ```text TASK-001 Add resumable tool execution State REVIEWING Iteration 2 / 5 Requirements 8 / 8 Invariants 5 / 6 Build PASS Unit Tests PASS Integration PASS Findings Critical 0 Major 1 Minor 2 Gate BLOCKED Blocking INV-004 F-0021 ``` 每次新的 Agent Session 开始必须优先执行： ```bash python scripts/harness_status.py ``` --- # 17. validate_state.py 作用： ```text validate_state.py CURRENT TARGET ``` 例如： ```bash python scripts/validate_state.py REVIEWING GATING ``` 合法： ```text exit 0 ``` 非法： ```text INVALID TRANSITION: IMPLEMENTING → DONE exit 1 ``` 建议所有状态更新都统一经过该脚本或内部公共模块。 --- # 18. task-contract Skill 路径： ```text skills/task-contract/SKILL.md ``` 职责： ```text User Requirement → Acceptance Criteria → Requirements → Invariants → Risks → Verification Plan ``` 不负责写业务代码。 必须产生： ```text .harness/requirements.yaml .harness/invariants.yaml ``` 并更新： ```text current-task.state: CREATED → SPECIFYING → PLANNED ``` 输出示例： ```yaml requirements: - id: REQ-001 statement: interrupted execution can resume priority: must - id: REQ-002 statement: duplicated recovery must not duplicate side effects priority: must ``` ```yaml invariants: - id: INV-001 statement: one action_id can produce at most one side effect category: idempotency severity: critical - id: INV-002 statement: terminal state cannot transition back to running category: state_machine severity: major ``` --- # 19. Superpowers Integration v0.1 不复制现有 Superpowers Skills。 复用： ```text brainstorming writing-plans test-driven-development systematic-debugging subagent-driven-development requesting-code-review receiving-code-review verification-before-completion finishing-a-development-branch ``` Harness 的定位： ```text Superpowers = Worker Workflow Harness = Controller ``` 主 SKILL.md 只做 orchestration。 伪流程： ```text 1. Load .harness/current-task.yaml 2. CREATED: run task-contract 3. Use Superpowers brainstorming 4. Use Superpowers writing-plans 5. state → IMPLEMENTING 6. Use Superpowers TDD / subagent-driven-development 7. state → VERIFYING 8. Run deterministic verification 9. If verification fails: state → IMPLEMENTING fix repeat 10. state → REVIEWING 11. Run normal Superpowers review 12. Run adversarial-review 13. If blocking finding exists: state → REPRODUCING 14. reproduce-finding 15. Confirmed: regression test state → FIXING fix state → VERIFYING 16. No blocking finding: collect fresh evidence state → GATING 17. Run quality_gate.py 18. PASS: state → CONVERGED state → DONE 19. BLOCKED: continue loop 20. max_iterations reached: state → ESCALATED ``` --- # 20. adversarial-review Skill 路径： ```text skills/adversarial-review/SKILL.md ``` 目标： **不是检查代码风格，而是主动构造失败场景。** 输入： ```text Task Contract Requirements Invariants Diff Public API Relevant State Machine Existing Tests ``` Reviewer 指令核心： ```text Assume the implementation contains defects. Do not optimize the implementation. Do not rewrite the code. Do not produce generic style feedback. Try to construct concrete scenarios that violate: - requirements - invariants - failure guarantees - state transitions - concurrency guarantees - idempotency guarantees - recovery guarantees Every issue must be emitted as a structured PROPOSED finding. ``` 必须输出 Finding Schema。 Reviewer 不允许直接修改代码。 --- # 21. reproduce-finding Skill 路径： ```text skills/reproduce-finding/SKILL.md ``` 流程： ```text PROPOSED ↓ REPRODUCING ↓ construct minimal reproducer ↓ run / \ FAIL cannot reproduce │ │ ▼ ▼ CONFIRMED REJECTED ``` CONFIRMED 后必须： ```text 1. 保留 failing test 2. 更新 finding.status = CONFIRMED 3. 写 regression_test.path 4. state → FIXING 5. 修复 6. regression test GREEN 7. full suite GREEN 8. finding → VERIFIED / CLOSED ``` 禁止： ```text Finding → 直接修改代码 ``` 必须先 Reproduce。 --- # 22. collect-evidence Skill Skill 只负责决定需要哪些 Evidence。 真正执行由： ```text collect_evidence.py ``` 完成。 例如： ```text task modifies backend API ↓ required: - build - unit_test - integration_test ``` Skill 不允许自己写： ```text “测试应该通过” ``` --- # 23. quality-gate Skill 该 Skill 只是调用： ```bash python scripts/quality_gate.py ``` 并根据 Exit Code 更新状态。 禁止 Skill 自己评估： ```text “综合来看质量足够好” ``` --- # 24. convergence Skill 职责： ```text 判断是否继续循环 ``` v0.1 不做复杂评分。 规则： ### PASS ```text Gate PASS → CONVERGED ``` ### Continue ```text Gate BLOCKED + 存在明确可处理 blocker + iteration < max_iterations → continue ``` ### Escalate 以下任一成立： ```text iteration >= max_iterations same confirmed finding repeatedly reappears same invariant repeatedly violated test suite unstable architecture defect suspected spec ambiguity blocks verification ``` 则： ```text state → ESCALATED ``` 并输出原因： ```text SPEC_AMBIGUITY ARCHITECTURE_DEFECT REPEATED_REGRESSION UNSTABLE_TEST REVIEW_DISAGREEMENT MAX_ITERATIONS ``` --- # 25. 主 SKILL.md 主 Skill 的核心角色： ```text You are the Engineering Harness Controller. You do not determine completion from model confidence. You control: - task state - workflow transitions - evidence requirements - findings lifecycle - quality gate execution - convergence Use Superpowers for development execution. The task may reach DONE only through: GATING → quality_gate.py exit 0 → CONVERGED → DONE ``` 启动时必须： ```text 1. Detect whether .harness exists. 2. Load current-task.yaml. 3. Run harness_status.py. 4. Resume from persisted state. ``` --- # 26. v0.1 开发顺序 不要先实现全部 Skills。 按照以下顺序开发。 ## Milestone 1：State Core 实现： ```text schemas/task.schema.json templates/current-task.yaml scripts/validate_state.py scripts/harness_status.py ``` 测试： ```text 合法状态转换 非法状态转换 IMPLEMENTING → DONE 被拒绝 CONVERGED → DONE 被允许 ``` 验收： ```text Harness 状态可以持久化 Session 重启后可以恢复 非法状态不能绕过 ``` --- ## Milestone 2：Evidence + Quality Gate 实现： ```text schemas/evidence.schema.json scripts/collect_evidence.py templates/gate.yaml scripts/quality_gate.py ``` 测试： ```text fresh PASS evidence → Gate PASS stale evidence → Gate BLOCKED missing evidence → Gate BLOCKED major finding exists → Gate BLOCKED unverified requirement → Gate BLOCKED violated invariant → Gate BLOCKED ``` 验收： ```text 错误状态绝对不能 Gate PASS ``` 这是 v0.1 最重要的 Milestone。 --- ## Milestone 3：Finding Lifecycle 实现： ```text schemas/finding.schema.json skills/adversarial-review/ skills/reproduce-finding/ ``` 测试： ```text PROPOSED 不算 confirmed bug PROPOSED → REPRODUCING reproducer fail → CONFIRMED cannot reproduce → REJECTED confirmed bug without regression test → Gate BLOCKED ``` --- ## Milestone 4：Task Contract 实现： ```text schemas/requirement.schema.json schemas/invariant.schema.json skills/task-contract/ ``` 测试一个真实 Feature： ```text requirement → acceptance → invariant → verification plan ``` --- ## Milestone 5：Superpowers Integration 实现： ```text root SKILL.md skills/collect-evidence/ skills/quality-gate/ skills/convergence/ ``` 接入： ```text brainstorming writing-plans test-driven-development subagent-driven-development verification-before-completion finishing-a-development-branch ``` 验收： 一个真实任务可以完整跑： ```text CREATED → SPECIFYING → PLANNED → IMPLEMENTING → VERIFYING → REVIEWING → GATING → CONVERGED → DONE ``` --- # 27. 推荐首个真实 Dogfooding Task 不要使用 Hello World。 建议选一个包含以下特性的 Feature： ```text API Database State Error Handling Idempotency Tests ``` 例如： ```text Implement resumable task execution with idempotent action execution. ``` 至少测试： ```text happy path duplicate request timeout retry process restart duplicate recovery invalid state transition ``` --- # 28. 测试要求 Harness 自身必须测试。 至少： ```text tests/ ├── test_state_machine.py ├── test_quality_gate.py ├── test_evidence.py ├── test_finding_lifecycle.py └── test_convergence.py ``` 禁止只依赖 LLM 验证 Harness 自身。 --- # 29. v0.1 成功指标 Dogfooding 10～20 个真实任务后统计： ## False Done Rate ```text 任务被宣布 DONE 后， 又发现 blocking defect 的比例。 ``` 目标： ```text 显著下降。 ``` ## Finding Confirmation Rate ```text CONFIRMED / PROPOSED ``` 用于衡量 Reviewer 噪声。 ## Regression Capture Rate ```text confirmed bug with regression test / all confirmed bugs ``` 目标： ```text ≈ 100% ``` ## Gate Bypass Rate ```text 未 PASS Gate 却进入 DONE ``` 目标： ```text 0 ``` ## Convergence Rate ```text 在 max_iterations 内成功 Gate PASS 的任务比例 ``` --- # 30. Codex 实施规则 Codex 在实现本项目时必须遵守： 1. 不一次性实现所有模块。 2. 按 Milestone 顺序提交。 3. 每个 Milestone 必须先有测试。 4. 不修改 Superpowers 核心 Skill，除非 Milestone 5 明确需要。 5. v0.1 不引入数据库。 6. v0.1 不引入 Web UI。 7. v0.1 不引入新的 Agent Runtime。 8. 所有 Gate 判断必须是确定性的 Python 逻辑。 9. 所有状态转换必须验证。 10. 所有关键行为必须有测试。 --- # 31. 建议 Git Commit 顺序 ```text feat: bootstrap harness project structure feat: add harness task state schema feat: add deterministic state transition validation feat: add harness status command feat: add evidence collection feat: add deterministic quality gate feat: add finding lifecycle schema feat: add adversarial review skill feat: add finding reproduction workflow feat: add task contract skill feat: add convergence controller feat: integrate superpowers workflow test: add end-to-end harness scenario docs: document harness v0.1 usage ``` --- # 32. 第一阶段立即开始的任务 Codex 首先只实现以下内容： ```text schemas/task.schema.json templates/current-task.yaml templates/gate.yaml scripts/validate_state.py scripts/harness_status.py scripts/quality_gate.py tests/test_state_machine.py tests/test_quality_gate.py ``` 暂时不要实现： ```text adversarial-review reproduce-finding Superpowers integration multi-agent ``` 第一阶段验收条件： ```text 1. CREATED → SPECIFYING 合法 2. IMPLEMENTING → DONE 非法 3. CONVERGED → DONE 合法 4. Requirement 未 verified 时 Gate BLOCKED 5. Invariant violated 时 Gate BLOCKED 6. Major Finding 存在时 Gate BLOCKED 7. Required Evidence 缺失时 Gate BLOCKED 8. Evidence.commit != HEAD 时 Gate BLOCKED 9. 所有条件满足时 Gate PASS 10. Gate PASS exit code = 0 ``` --- # 33. Codex 首轮执行 Prompt 可以直接把下面这段交给本地 Codex： ```text 请按照仓库中的《Superpowers → Engineering Harness v0.1 实施手册》进行实现。 当前只执行 Milestone 1 和 Milestone 2，不要提前实现后续 Skill。 目标： 1. 建立 Harness 状态模型。 2. 实现确定性状态转换验证。 3. 实现 Evidence 数据模型。 4. 实现 collect_evidence。 5. 实现 deterministic quality_gate。 6. 为上述能力补齐自动化测试。 约束： - 不修改 Superpowers 核心代码。 - 不实现 Web UI。 - 不引入数据库。 - 不实现新的 Agent Runtime。 - Gate 判断不得依赖 LLM。 - 状态转换必须由代码验证。 - Evidence 必须绑定当前 Git HEAD。 - IMPLEMENTING/VERIFYING/REVIEWING/FIXING 不允许直接进入 DONE。 - 只有 CONVERGED → DONE 合法。 - 所有实现遵循 TDD。 - 每完成一个 Milestone，运行完整测试并提供实际命令结果。 - 不要仅汇报计划，必须实际修改代码和运行测试。 完成标准： Milestone 1 和 Milestone 2 的所有测试通过，且 quality_gate.py 能正确区分 PASS、BLOCKED 和 INVALID_HARNESS_STATE。 开始前先： 1. 阅读实施手册。 2. 检查当前仓库结构。 3. 输出不超过 10 行的实施计划。 4. 立即开始修改代码。 ``` --- # 34. v0.1 Definition of Done 只有全部满足时，v0.1 才算完成： ```text State persisted outside LLM context State transitions deterministic Evidence collected from real commands Evidence bound to Git HEAD Findings structured Finding reproduction supported Confirmed bugs require regression tests Quality Gate deterministic Gate controls DONE Convergence loop bounded Superpowers reused rather than reimplemented One real project successfully dogfooded ``` --- # 35. 最终原则 Engineering Harness v0.1 的核心不是增加更多 Prompt。 而是把开发过程中的“正确性判断”逐步从模型迁移到： ```text State + Contract + Invariant + Executable Test + Evidence + Deterministic Gate ``` 最终模型的角色应当是： ```text Model = Worker Superpowers = Development Workflow Engineering Harness = Controller Tests / Compiler / Gate = Truth ``` **允许 Agent 犯错，但不允许错误轻易穿过 Harness。**