16gagent/AGENTS.md

AGENTS.md - Your Workspace

This folder is home. Treat it that way.

Cache Stability ⚡

ALL files in this workspace are ABOVE the OpenClaw prompt cache boundary. Every edit invalidates the DeepSeek KV cache prefix → cache MISS → full token cost. Prefer writing to memory/daily/ or memory/vault.md (NOT injected into context). Only edit above-boundary files for critical corrections.

Session Startup

Use runtime-provided startup context first. Do not reread startup files unless explicitly asked, context is missing, or you need a deeper follow-up.

---

Execution Protocol — Planning Layer 🔵

完整协议文档（§0-7 + Learning Loop + Context Budget）: docs/AGENTS-PROTOCOL-FULL.md 以下为规则摘要。详细规则、表格、代码示例、Checklist → 读完整文档。

0. Complexity Classification

简单任务 → 直接执行（单文件/单命令/纯查询/对话） 复杂任务 → 必须用 Task Tree（3+文件/多步骤/跨项目/新功能/bug分析/预估>45min） 不确定 → 视为复杂。 if uncertain → treat as Complex

详见 docs/AGENTS-PROTOCOL-FULL.md §0。

1. Task Tree Construction

复杂任务必须先建 Task Tree（执行树，非功能树）。含 depends_on、并行分组、注入风险。质量门见完整文档 §1f。

详见 docs/AGENTS-PROTOCOL-FULL.md §1。

1f. Task Tree Quality Gate

详见 docs/AGENTS-PROTOCOL-FULL.md §1f（7 checks: Root/Orphan/Circular/XL/Serial/Risks/Human-dep）。

2. Planning Protocol

## PLAN
### Goal / Subtasks / Risks / Verification Strategy

详见 docs/AGENTS-PROTOCOL-FULL.md §2。

3. Pre-flight Check

3a. Dependency Scan

FK/imports/types/blast radius → docs/AGENTS-PROTOCOL-FULL.md §3a

3b. Learning Context Injector

3 次搜索（反射/最佳实践/Playbook）→ docs/AGENTS-PROTOCOL-FULL.md §3b

3c. Domain Checklist Injection

SQL/JS/Next.js/Backend/Generator/Release → docs/AGENTS-PROTOCOL-FULL.md §3c

4. Failure Replanning Protocol

局部失败→修复继续。结构失败→REPLAN。同错3次→暂停问用户。修复后必查 Auto-Capture（§7）。

详见 docs/AGENTS-PROTOCOL-FULL.md §4。

5. Verification Protocol

Build/Tests/Lint/Regression/Manual 全检。完成后必须通过 Pre-flight Compliance Gate。

详见 docs/AGENTS-PROTOCOL-FULL.md §5。

6. Governance Compliance 🔒

• 状态机/心跳/blocker/replan/健康评分/自动停止 → GOVERNANCE.md
• 多项目管理 → PORTFOLIO.md
• Governance Quality Gate 在输出前必须运行（GOVERNANCE.md §9）

7. Auto-Capture Trigger

失败后必查 T1-T7 触发条件（build/test/lint 失败、耗时>30min、同类错误再现、路径不匹配、运行时崩溃）。Reflection 必须含 7 字段。详见 docs/AGENTS-PROTOCOL-FULL.md §7。

---

Learning Loop 🔄

晋升路径： 反射(1次) → 模式(2次) → 最佳实践(3次) → 清单(4次) → Hard Rule(5次) 存储： reflections/ → patterns/ → docs/AGENTS-PROTOCOL-FULL.md §3c

详见 docs/AGENTS-PROTOCOL-FULL.md Learning Loop 章节。

Promotion Rules: 1st → Reflection, 2nd → Pattern, 3rd → Best Practice, 4th → Checklist, 5th → Hard Rule

---

Agent Package Context Budget ⚠️ HARD RULE

Size	Token Budget	读文件	改文件	新代码
Small	≤30K	≤3	≤2	≤300行
Medium	≤80K	≤8	≤5	≤800行
Large	≤200K	≤15	≤10	≤1500行
XL	❌ 禁止	—	—	—

每次分发 Agent Package 前必须声明 Context Budget。Input Size Pre-check 必须执行。 详见 docs/AGENTS-PROTOCOL-FULL.md。

7. Auto-Capture Trigger

失败后必查 T1-T7 触发条件（build/test/lint 失败、耗时>30min、同类错误再现、路径不匹配、运行时崩溃）。Reflection 必须含 7 字段。 Injected Risks: 每个任务必须注入学习上下文中的风险。详见 docs/AGENTS-PROTOCOL-FULL.md §7。