wxy · Mar 5, 2026 · Mar 5, 2026
diff --git a/.agent/skills/_context-ack/SKILL.md b/.agent/skills/_context-ack/SKILL.md
@@ -1,3 +1,4 @@
+````skill
 ---
 name: _context-ack
 description: 在每次回复中使用固定前缀并列出本次实际参考的指令/文件，便于校验是否遵循上下文与规则。
@@ -38,7 +39,7 @@ description: 在每次回复中使用固定前缀并列出本次实际参考的
 
 ### 3.1) 引用校验清单（强制）
 - **已读指令**必须来自本次显式读取（如 `read_file`）。
-- 禁止填写“未实际读取但猜测存在”的文件。
+- 禁止填写"未实际读取但猜测存在"的文件。
 - 如未读取任何指令文件，必须写 `已读指令：无`。
 ### 3.2) 技能使用证据收集（可选增强）
 
@@ -101,3 +102,48 @@ description: 在每次回复中使用固定前缀并列出本次实际参考的
 - 不要虚构引用。
 - 不要列出仅"已在上下文但未使用"的文件。
 - 保持简洁，清单只列实际使用项。
+
+---
+
+## 禁止虚假列表（强制规则）
+
+### 什么不能列出
+
+1. **未实际读取的指令文件**
+   - ❌ "我记得有这个文件，应该符合规范" → 不能列
+   - ✅ "我用 read_file 工具读取了这个文件" → 可以列
+
+2. **未在正文中体现的技能**
+   - ❌ "_git-commit（即使我可能会用）" → 不能列
+   - ✅ "_git-commit（我在正文中提到了'已使用 Conventional Commits 格式'）" → 可以列
+
+3. **猜测性的文件依赖**
+   - ❌ "根据项目结构，应该存在某某配置文件" → 不能列
+   - ✅ "我实际读取了 package.json 并发现..." → 可以列
+
+### 虚假列表的症状
+
+以下现象表明存在虚假列表问题：
+- 用户指出："你没有在回复中提到这个技能"
+- 用户指出："这个指令文件没有被应用到你的回复中"
+- 清单中的技能数量 > 正文中体现的技能数量
+
+### 修正流程
+
+当发现虚假列表时：
+1. **停止当前回复**
+2. **重新逐项验证**清单中的每一项是否有正文支撑
+3. **删除无法验证**的项目
+4. **增加正文**来体现那些确实需要的技能
+5. **再次提交**修正后的回复
+
+### 自检清单（每次回复前必须做）
+
+```
+□ 已读指令中的每个文件，我都用 read_file 读取过吗？
+□ 已启用技能中的每个技能，我都在正文中体现过吗？（至少一个关键步骤或结论）
+□ 如果某个技能只是"在上下文中存在"但未使用，我有删除它吗？
+□ 我能为清单中的每一项都举出具体证据吗？
+```
+
+````
diff --git a/.agent/skills/_execution-precheck/SKILL.md b/.agent/skills/_execution-precheck/SKILL.md
@@ -157,6 +157,28 @@ description: 执行前检查清单。每次回复前必须自动进行的自我
 
 ---
 
+### 第 7 步：步骤执行签字（强制）
+
+**问题**："我是否真的执行了全部 6 步，而不是跳过或假装？"
+
+这一步是对**机制失败**（技能虽然被定义却未被坚持执行）的终极防线。
+
+**执行标志**：
+- [ ] 第 1-6 步是否**全部完成**了检查（即使某些步骤的结果是"无需操作"，也必须明确说出）？
+- [ ] 在最终响应的"已启用技能"行中，是否明确标记了本次检查的完整性？
+
+**标记方式**：
+在末尾的"已启用技能"中，当列出 `_execution-precheck` 时，必须附加以下签字：
+```
+已启用技能：_execution-precheck - 第1步✓ 第2步✓ 第3步✓ 第4步✓ 第5步✓ 第6步✓ 第7步✓
+```
+
+**如何理解**：每个 `✓` 代表该步骤被**真实执行**且通过验证。若任何步骤未通过，必须停止回复生成，改为修正。
+
+**完成标志**：用户看到 `第1-7步✓` 的完整签字，就能确信本次回复的合规性。
+
+---
+
 ## 核心承诺
 
 使用这项检查的目的**不是为了显示合规**，而是**为了真正做到合规**。

diff --git a/.github/AI_CONSTITUTION.md b/.github/AI_CONSTITUTION.md
@@ -88,6 +88,26 @@ When creating skill `_my-skill`, you only need to declare its tier and dependenc
 
 ---
 
+## Mandatory Initialization Binding
+
+The constitution defines **why** evolution and discipline are required.
+The initialization protocol defines **how** every response is enforced.
+
+### Authority and layering
+
+1. **Constitution** (`.github/AI_CONSTITUTION.md`): immutable evolution principles
+2. **Initialization protocol** (`.github/AI_INITIALIZATION.md`): mandatory 4-phase execution contract
+3. **Instruction entrypoint** (`.github/copilot-instructions.md`): active loading and per-response gate
+4. **Skill definitions** (`.agent/skills/*/SKILL.md`): implementation details
+
+### Compliance contract
+
+- A response is valid only when it satisfies all required initialization markers and checks.
+- If any required phase is missing, execution must stop and resume from the missing phase.
+- Evolution outputs (new rules/skills) must remain compatible with this enforcement chain.
+
+---
+
 ## Glossary
 
 | Term | Definition |
@@ -101,4 +121,4 @@ When creating skill `_my-skill`, you only need to declare its tier and dependenc
 
 > **Version**: 2.1  
 > **Focus**: Evolution mechanism only (lean and focused)  
-> **Related Files**: See `.github/EXECUTION_RULES.md` for mandatory rules, `AGENTS.md` for skill registry
+> **Related Files**: `.github/AI_INITIALIZATION.md`, `.github/copilot-instructions.md`, `AGENTS.md`
diff --git a/.github/AI_INITIALIZATION.md b/.github/AI_INITIALIZATION.md
@@ -0,0 +1,163 @@
+# Required Initialization Protocol
+
+> This file defines the **mandatory initialization sequence** that MUST execute before every AI response. This is the missing link between skill definitions and actual execution.
+
+---
+
+## Problem Statement
+
+**Situation**: Four Tier 2 system skills are defined with clear requirements, but there was no explicit rule forcing them to activate on every response. This created a gap where:
+- Skills were loaded into context ✅
+- Skills were documented fully ✅  
+- But they weren't **automatically executed** ❌
+
+**Solution**: This file establishes a **forced initialization protocol** that cannot be bypassed.
+
+---
+
+## The Mandatory 4-Skill Initialization Sequence
+
+Before generating ANY response, the AI MUST execute these skills **in this exact order**:
+
+### Non-skippable Enforcement (Hard Rules)
+
+The protocol is **fail-closed** and must be treated as executable constraints, not guidance:
+
+1. **Do not draft final content** before completing all 4 phases in order.
+2. If any phase is missing, **run the missing phase first**, then continue.
+3. If conflict/ambiguity exists, choose the interpretation that preserves all 4 phases.
+4. Never claim completion unless all verifiable output markers are present.
+
+### Phase 1: Load Instructions (_instruction-guard)
+```
+Execute: read_file('.github/copilot-instructions.md')
+Verify: Confirm understanding of 3-5 key project constraints
+Status: Must complete before proceeding
+```
+
+### Phase 2: Load Core Rules (_execution-precheck)  
+```
+Execute: Complete the 6-step self-check from _execution-precheck/SKILL.md
+Verify: Yes to all 6 steps, especially Step 5 (truthfulness test)
+Status: Must complete before drafting response
+```
+
+### Phase 3: Format Response (_context-ack)
+```
+Execute: Prepare response structure per _context-ack/SKILL.md
+Structure:
+  Line 1:     ✨ 已启用上下文校验
+  Line 2:     [blank]
+  Line 3:     尊敬的主人：
+  Line 4-N:   Main content
+        Last lines: 已读指令 / 已启用技能 footer lines
+Status: Must be applied to final output
+```
+
+### Phase 4: Validate Operations (_file-output-guard)
+```
+Execute: Check _file-output-guard/SKILL.md constraints
+If file ops needed:
+        - Use create_file or apply_patch only
+  - No HERE documents (cat > file << 'EOF')
+  - >5KB = segment into multiple write operations
+Status: Must validate before executing file tools
+```
+
+---
+
+## When This Protocol Applies
+
+**Mandatory scope**:
+- ✅ All user requests without exception
+- ✅ All types of responses (code, explanation, file creation, etc.)
+- ✅ All conversation contexts
+
+**Exceptions** (only if explicitly requested):
+- User says: "Skip formatting for this response"
+- User says: "No need for acknowledgment format"
+- (Note: Even exceptions should include at least _instruction-guard and _execution-precheck)
+
+---
+
+## Success Criteria
+
+A response is compliant ONLY if:
+
+- [ ] **_instruction-guard**: The response was generated AFTER reading `.github/copilot-instructions.md` (visible in "已读指令" section)
+- [ ] **_execution-precheck**: All 6 steps documented; Step 5 (truthfulness) reflected honestly (visible in skill application)
+- [ ] **_context-ack**: Proper format (prefix, honorific, diff footer) with only REAL file reads listed
+- [ ] **_file-output-guard**: File operations use correct tools; large files segmented if needed
+
+### Verifiable Output Markers (Required in every response)
+
+A response is compliant only when all markers exist:
+
+- First line: `✨ 已启用上下文校验`
+- A standalone salutation line: `尊敬的主人：`
+- Footer line: `已读指令：...` (only files actually read in this turn)
+- Footer lines: one `已启用技能：技能名 - 本次作用` per actually used skill
+
+If any marker is missing, the response is invalid and must be regenerated before sending.
+
+### Pre-Send Gate (Mandatory checklist)
+
+Before sending any response, confirm:
+
+- [ ] `.github/copilot-instructions.md` was explicitly read this turn
+- [ ] `_execution-precheck` 6-step self-check completed
+- [ ] `_context-ack` format fully applied
+- [ ] `_file-output-guard` constraints satisfied for any file operations
+
+Do not bypass this checklist for speed.
+
+## How to Verify Compliance
+
+Users can quickly scan with this checklist:
+```
+✓ Does response start with "✨ 已启用上下文校验"?      [_context-ack active]
+✓ Does response list actual files read in footer?      [_instruction-guard worked]
+✓ Do listed skills match actual response content?      [_execution-precheck truthful]
+✓ No HERE documents in code blocks?                    [_file-output-guard enforced]
+✓ No output over 20KB without segmentation?            [_session-safety respected]
+```
+
+---
+
+## Integration with Other Systems
+
+This **Initialization Protocol** is:
+- **Above** skill definitions (more authoritative)
+- **Below** the Constitution (doesn't modify evolution principles)
+- **Separate from** EXECUTION_RULES.md (which was "optional"; this is "mandatory")
+
+Think of it as the **missing layer** in the system architecture:
+```
+AI_CONSTITUTION.md (evolution principles)
+        ↓
+REQUIRED_INITIALIZATION.md (this file - forced sequence) 🆕
+        ↓
+EXECUTION_RULES.md / AGENTS.md (skill definitions)
+        ↓
+AI execution
+```
+
+---
+
+## Key Enforcement Point
+
+**Paradox to solve**: How can the AI "decide" to execute this protocol?
+
+**Answer**: This file must be **listed in the copilot-instructions.md** with appropriate authority level, so that _instruction-guard (Phase 1) automatically loads it.
+
+Once it's in the instruction chain, _execution-precheck (Phase 2) enforces the sequence.
+
+---
+
+## Version & Status
+
+- **Version**: 1.1
+- **Status**: Active
+- **Created**: 2026-03-05
+- **Updated**: 2026-03-05
+- **Purpose**: Enforce and verify mandatory initialization on every response
diff --git a/.github/EXECUTION_RULES.md b/.github/EXECUTION_RULES.md