@bolloon/bolloon-agent 0.1.35 → 0.1.37

package/src/bollharness/CLAUDE.md

@@ -1,73 +0,0 @@
-# CLAUDE.md — bollharness
-
-bollharness 是一个 **AI Agent Session 治理框架**——通过 Claude Code hooks、guards、上下文路由和安装器，给任意项目加装结构化的 agent 行为约束。
-
-## 架构
-
-```
-bollharness/
-├── .boll/
-│   ├── settings.json        # Hook 注册表（16 hooks，7 stages）
-│   ├── rules/               # Path-scoped rules（Claude Code 按文件路径自动加载）
-│   └── skills/              # Skill 定义（harness-dev-handoff, guardian-fixer 等）
-├── scripts/
-│   ├── hooks/               # 14 个 hook 脚本（session lifecycle + tool guards）
-│   ├── guard-feedback.ts    # PostToolUse 入口 — 上下文路由（机制 A）+ guard 检查（机制 B）
-│   ├── deploy-guard.ts      # PreToolUse Bash 入口 — 部署安全拦截
-│   ├── context_router.ts    # 文件路径 → 上下文片段路由表（ADR-030 §3.4.1）
-│   ├── guard_router.ts      # 文件路径 → guard 脚本映射（ADR-030 §3.3）
-│   ├── context-fragments/   # 17 个上下文片段（guard-feedback.ts 注入用）
-│   ├── checks/              # Guard 脚本（check_*.ts）
-│   ├── install/             # 安装器（phase2_auto.ts + trust token）
-│   └── lib/                 # 共享库
-├── templates/scaffold/      # 目标项目骨架模板（安装时复制）
-├── schemas/                 # YAML/JSON schema 定义
-├── docs/decisions/          # ADR-030, ADR-038, ADR-041
-└── .boll/MANIFEST.yaml  # 物理清单（版本 + 文件注册表）
-```
-
-
-**同步覆盖范围**：
-- Step 1: `scripts/hooks/*`（跳过 boll-only 的 `find-boll-root.sh`）
-- Step 2: 路径修补（`find-boll-root.sh` → `find-project-root.sh`）
-- Step 3: settings.json hook 注册对比
-- Step 4: 共享脚本（guard-feedback.ts, deploy-guard.ts, context_router.ts, guard_router.ts）
-- Step 5: context-fragments/ 目录
-
-## 开发约束
-
-### 必须遵守
-- 所有命令使用 `npx ts-node`
-- Commit message 中英双语 + `Co-Authored-By: Claude Opus 4.6`
-- 审查类 subagent 必须用 opus 模型，不降级 sonnet
-- 审查类 agent 工具白名单必须 schema-level 隔离写权限（ADR-038 D11）
-- Guardian issue 必须先建 `docs/issues/*.md` 再写代码
-
-### 安装器开发
-- `phase2_auto.ts` 所有函数必须**幂等**（重复运行不改变结果）
-- Trust token 使用 HMAC + 30min 滑动窗口 + 6h 绝对窗口
-- 安装步骤顺序不可打乱（trust → bundle → scaffold → gitignore → paths → hooks）
-
-### Hook 开发
-- Hook 必须 fail-open（`|| true`）或有明确的 fail-closed 理由
-- Hook timeout 不超过 30s（guard-feedback.ts 最大）
-- 新增 hook 后必须同时更新 settings.json 注册和 MANIFEST.yaml
-
-### 路径规则
-- `find-project-root.sh` 是 bollharness 的项目根定位器（3 层锚点，fail-closed）
-- `find-boll-root.sh` 是 boll 专用的（不同步到 bollharness）
-- settings.json 中所有 hook command 使用**相对路径**解析器，不含绝对路径
-
-## 接手入口
-
-```bash
-npx ts-node .boll/skills/harness-dev-handoff/scripts/collect_handoff_context.ts
-```
-
-然后读 `.boll/skills/harness-dev-handoff/SKILL.md`。
-
-## 关键 ADR
-
-- **ADR-030**: Guard Signal Protocol — 上下文路由 + guard 检查的设计
-- **ADR-038**: Harness Optimization — 11 决策（metrics、fragment TTL、review agent 隔离等）
-- **ADR-041**: Codex Division of Labor — 判断类 vs 执行类 subagent 分流

package/src/bollharness/LICENSE

@@ -1,21 +0,0 @@
-MIT License
-
-Copyright (c) 2026 bollharness contributors
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.

package/src/bollharness/README.md

@@ -1,143 +0,0 @@
-[中文](README.zh-CN.md) | English
-
-# bollharness
-
-> "How do you let AI handle so much development with so little supervision?"
-
-This is the answer.
-
-bollharness is a governance layer for [Claude Code](https://docs.anthropic.com/en/docs/claude-code). It makes AI agents reliable enough that you can set direction, walk away, and trust the work actually lands — with review gates, completion verification, and mechanical enforcement that no amount of prompting can replicate.
-
-## The Problem
-
-Claude Code is remarkably capable. But left unsupervised, it has structural biases:
-
-- **Claims completion prematurely** — "all tests pass" (didn't run them)
-- **Skips review** — "this change is simple enough" (it wasn't)
-- **Drifts from plans** — starts fixing one bug, ends up refactoring three files
-- **Self-evaluation bias** — asks itself "did I do a good job?", answers "yes"
-
-You end up supervising more than you saved in development time. The 80% it does well makes the 20% it silently drops even harder to catch.
-
-## The Insight
-
-```
-CLAUDE.md instruction compliance:  ~20%
-PreToolUse hook enforcement:       100%
-```
-
-Instructions don't reliably change AI behavior. Mechanical constraints do.
-
-A review agent told "don't modify files" obeys ~70% of the time. A review agent whose tool manifest doesn't list Edit/Write obeys 100% of the time — it physically can't call what isn't there.
-
-bollharness applies this principle everywhere: if it matters, enforce it with a hook, not a sentence.
-
-## What Changes
-
-| Without bollharness | With bollharness |
-|---|---|
-| "Did you run the tests?" → "Yes" (didn't) | Mechanical gate checks `progress.json` — can't fake evidence |
-| AI stops mid-chat, injects completion checklist | Stop hook parses session transcript — only triggers when uncommitted writes exist |
-| Review agent "helpfully" edits what it reviews | Review agent physically cannot call Edit/Write (schema-level isolation) |
-| "This PR is simple, let's skip review" | Gates 2/4/6/8 mechanically require independent review — no exceptions |
-| Parallel AI sessions contaminate each other | Each session's scope is isolated via its own transcript file |
-| Agent drifts into unrelated fixes | Context routing injects domain-specific rules only for files being edited |
-
-## How It Works
-
-### Hooks: enforcement at the moment of action
-
-16 hooks across 7 lifecycle stages. They intercept *as things happen*, not after:
-
-```
-SessionStart  →  Load context, reset risk state, surface tools
-PreToolUse    →  Block unsafe deploys, gate review agents, sanitize reads
-PostToolUse   →  Route context on edit, detect loops, track risk
-Stop          →  Verify completion candidate exists (transcript × git diff)
-SessionEnd    →  Reflect, analyze traces, persist progress
-```
-
-### The 8-Gate State Machine
-
-Every significant change flows through gates. Even-numbered gates require independent review — not the same agent checking its own work:
-
-```
-G0 Problem  →  G1 Design  →  G2 Review*
-  →  G3 Plan  →  G4 Review+Lock*
-  →  G5 Tasks  →  G6 Review*
-  →  G7 Execute+Log  →  G8 Final Review*
-
-* = Independent reviewer (separate context, read-only tools)
-```
-
-### Automated Checks
-
-15 validators run on file changes: API type consistency, doc freshness, security patterns, fragment integrity, hook registration, and more. They catch drift before it compounds.
-
-### Skills
-
-16 specialized behaviors — from architecture design (`arch`) to failure pattern extraction (`crystal-learn`) to structured bug triage (`bug-triage`). Skills install judgment frameworks, not rule lists, so the agent can navigate situations the skill didn't explicitly cover.
-
-Each skill has `{{PLACEHOLDER}}` structural slots designed to be filled with your project's context during installation.
-
-## Install
-
-```
-
-### Three Tiers
-
-| Tier | Trust level | What happens |
-|------|-------------|-------------|
-| **drop-in** | Minimal | Installs hooks + skills as-is. Try it, see what happens. |
-| **adapt** | Medium | Reads your README + docs, customizes skills to your project. |
-| **mine** | Full | Reads your work transcripts, deeply adapts to your patterns. |
-
-### What Gets Installed
-
-```
-your-project/
-├── .boll/
-│   ├── settings.json    # Hook registrations (appends, won't clobber)
-│   ├── skills/          # 16 agent behavior definitions
-│   └── rules/           # Path-scoped context (auto-loaded by file path)
-├── scripts/
-│   ├── hooks/           # 16 lifecycle hooks
-│   └── checks/          # 15 automated validators
-└── CLAUDE.md            # Governance guide (generated, yours to edit)
-```
-
-The installer is idempotent — run it twice, get the same result.
-
-## Design Principles
-
-1. **Hooks over instructions** — If compliance matters, don't ask. Enforce.
-2. **Schema-level isolation** — Review agents' tool manifests exclude write tools. Not "please don't" — *can't*.
-3. **Fail-open where safe** — A hook that can't read its data injects *more* checks, not fewer. The failure mode is always "too cautious," never "silently skipped."
-4. **Session isolation** — Completion detection uses per-session transcript parsing. No shared mutable state between parallel sessions.
-5. **Structural slots over blank space** — Project-specific content becomes `{{PLACEHOLDER}}` with meta-instructions (what to put, why it matters, how to discover it), not empty fields you forget to fill.
-
-## Requirements
-
-- [Claude Code](https://docs.anthropic.com/en/docs/claude-code) CLI
-- Node.js 18+
-- Git
-
-## Origin
-
-Born from 6 months of production use on [boll](https://boll.net), an agent collaboration protocol. The governance layer kept proving independently valuable — every AI-assisted project needs it, not just ours. So we extracted it.
-
-The hooks, gates, and isolation patterns were designed by getting burned first, then building the guard. Every rule in this system exists because an AI agent found a creative way to not follow the previous rule.
-
-## Acknowledgments
-
-This project was originally developed in Python as `wow-harness` and has been migrated to TypeScript for improved type safety, better developer experience, and tighter integration with modern JavaScript/TypeScript tooling.
-
-The migration preserved all original functionality while leveraging TypeScript's:
-- Static type checking for reduced runtime errors
-- Better IDE support and code completion
-- Improved maintainability for larger codebases
-- Native ES modules and modern JavaScript features
-
-## License
-
-MIT

package/src/bollharness/README.zh-CN.md

@@ -1,131 +0,0 @@
-中文 | [English](README.md)
-
-# bollharness
-
-> "你怎么做到让 AI 自己开发，人几乎不用管？"
-
-这就是答案。
-
-bollharness 是 [Claude Code](https://docs.anthropic.com/en/docs/claude-code) 的治理层。它让 AI agent 变得足够可靠——你定方向，AI 交付，工作真的能落地。靠的不是提示词里多写几句"请认真做"，而是机械化的审查门禁、完成验证和工具隔离。
-
-## 问题
-
-Claude Code 能力很强。但不加约束时，它有结构性偏见：
-
-- **假装完成** — "测试全过了"（其实没跑）
-- **跳过审查** — "这个改动很简单"（并不简单）
-- **任务漂移** — 修一个 bug，顺手重构三个文件
-- **自我评价偏差** — 问自己"做得好不好？"，回答"好"
-
-你花在监督上的时间，比省下的开发时间还多。更难受的是，它做对的 80% 让你放松警惕，剩下 20% 的静默遗漏更难发现。
-
-## 核心洞察
-
-```
-CLAUDE.md 指令遵从率:     ~20%
-PreToolUse hook 执行率:   100%
-```
-
-指令改变不了 AI 的行为。机械约束可以。
-
-告诉审查 agent "不要修改文件"，遵从率约 70%。把 Edit/Write 从它的工具清单里删掉，遵从率 100%——它物理上调不了不存在的工具。
-
-bollharness 把这个原则应用到所有该管的地方：重要的事不靠说，靠 hook。
-
-## 装了之后有什么变化
-
-| 没有 bollharness | 有 bollharness |
-|---|---|
-| "跑测试了吗？" → "跑了"（没跑） | 机械化门禁检查 `progress.json` — 伪造不了证据 |
-| 纯聊天时弹出完成检查清单 | Stop hook 解析 session transcript — 只在有未提交的编辑时才触发 |
-| 审查 agent "顺手"改了它审查的代码 | 审查 agent 物理上无法调用 Edit/Write（schema 级隔离） |
-| "这个 PR 简单，跳过审查吧" | Gate 2/4/6/8 机械化强制独立审查 — 没有例外 |
-| 并行 AI session 互相污染 | 每个 session 通过独立 transcript 文件隔离作用域 |
-| Agent 漂移去修不相关的东西 | 上下文路由只对正在编辑的文件注入领域规则 |
-
-## 怎么工作的
-
-### Hook：在动作发生的那一刻介入
-
-16 个 hook 覆盖 7 个生命周期阶段。不是事后审查，是实时拦截：
-
-```
-SessionStart  →  加载上下文、重置风险状态、展示可用工具
-PreToolUse    →  拦截危险部署、门控审查 agent、净化读取内容
-PostToolUse   →  编辑时路由上下文、检测循环、追踪风险
-Stop          →  验证是否存在完成候选品（transcript × git diff）
-SessionEnd    →  反思、分析轨迹、持久化进度
-```
-
-### 8 关状态机
-
-每个重要变更都要过关。偶数关要求独立审查——不是同一个 agent 自己检查自己的工作：
-
-```
-G0 问题  →  G1 设计  →  G2 审查*
-  →  G3 方案  →  G4 审查+锁定*
-  →  G5 任务拆分  →  G6 审查*
-  →  G7 执行+日志  →  G8 终审*
-
-* = 独立审查者（独立上下文、只读工具）
-```
-
-### 自动化检查
-
-15 个验证器在文件变更时运行：API 类型一致性、文档新鲜度、安全模式、片段完整性、hook 注册等。在漂移累积之前就抓住它。
-
-### Skill
-
-16 个专业化行为——从架构设计（`arch`）到失败模式提取（`crystal-learn`）到结构化 bug 分诊（`bug-triage`）。Skill 安装的是判断框架，不是规则清单，所以 agent 能应对 skill 没明确覆盖的情况。
-
-每个 skill 都有 `{{PLACEHOLDER}}` 结构化槽位，安装时会填入你项目的上下文。
-
-```
-
-### 三个层级
-
-| 层级 | 信任度 | 效果 |
-|------|--------|------|
-| **drop-in** | 最低 | 原样安装 hook + skill。先试试看。 |
-| **adapt** | 中等 | 读你的 README + 文档，把 skill 适配到你的项目。 |
-| **mine** | 完全 | 读你的工作 transcript，深度适配到你的模式。 |
-
-### 安装了什么
-
-```
-your-project/
-├── .boll/
-│   ├── settings.json    # Hook 注册（追加模式，不覆盖已有配置）
-│   ├── skills/          # 16 个 agent 行为定义
-│   └── rules/           # 路径作用域上下文规则（按文件路径自动加载）
-├── scripts/
-│   ├── hooks/           # 16 个生命周期 hook
-│   └── checks/          # 15 个自动化验证器
-└── CLAUDE.md            # 治理指南（自动生成，你可以自由编辑）
-```
-
-安装器是幂等的——跑两次，结果一样。
-
-## 设计原则
-
-1. **Hook 优于指令** — 重要的事不靠说，靠执行。
-2. **Schema 级隔离** — 审查 agent 的工具清单里没有写工具。不是"请不要"——是"不能"。
-3. **安全方向 fail-open** — hook 读不到数据时注入*更多*检查，不是更少。失败模式永远是"过于谨慎"，绝不是"静默跳过"。
-4. **Session 隔离** — 完成检测用每个 session 独立的 transcript 解析。并行 session 之间零共享可变状态。
-5. **结构化槽位** — 项目特定内容变成 `{{PLACEHOLDER}}` + 元指令（放什么、为什么重要、怎么发现它），不是你忘记填的空白字段。
-
-## 环境要求
-
-- [Claude Code](https://docs.anthropic.com/en/docs/claude-code) CLI
-- Node.js 18+
-- Git
-
-## 来历
-
-从 [boll](https://boll.net)（流形，一个 Agent 协作协议项目）6 个月的生产使用中提取出来的。治理层不断证明它独立于项目也有价值——每个 AI 辅助开发的项目都需要它，不只是我们的。
-
-这里的每一条 hook、每一个 gate、每一种隔离模式，都是因为 AI agent 找到了创造性的方法来绕过上一条规则，所以才被加上的。不是设计出来的，是被逼出来的。
-
-## License
-
-MIT

package/src/bollharness/package.json

@@ -1,20 +0,0 @@
-{
-  "name": "bollharness",
-  "version": "1.0.0",
-  "description": "AI Agent Session Governance Framework - TypeScript implementation",
-  "main": "dist/index.js",
-  "types": "dist/index.d.ts",
-  "scripts": {
-    "build": "tsc",
-    "watch": "tsc --watch",
-    "lint": "eslint src --ext .ts",
-    "typecheck": "tsc --noEmit"
-  },
-  "keywords": ["claude-code", "agent-governance", "hooks"],
-  "license": "MIT",
-  "dependencies": {},
-  "devDependencies": {
-    "@types/node": "^20.0.0",
-    "typescript": "^5.3.0"
-  }
-}

package/src/bollharness/reference/SOURCE-COMMIT.txt

@@ -1,3 +0,0 @@
-# README
-
-This directory is for reference purposes - it contains a snapshot from the Towow project for historical reference.

package/src/bollharness/reference/boll-reference/scripts/hooks/.sanitize-report.json

@@ -1,12 +0,0 @@
-{
-  "mode": "strict",
-  "files_scanned": 14,
-  "totals": {
-    "SECRET": 0,
-    "TRADE_SECRET": 0,
-    "PII": 0,
-    "NETWORK": 0,
-    "PROTOCOL_INTERNAL": 0
-  },
-  "records": []
-}

package/src/bollharness/reference/boll-reference/scripts/hooks/find-boll-root.sh

@@ -1,27 +0,0 @@
-#!/bin/bash
-# Walks up from $PWD to find the boll repo root.
-# Identifies it by presence of BOTH CLAUDE.md and scripts/guard-feedback.py.
-# This handles nested git repos (e.g. boll-progress/) that would confuse
-# `git rev-parse --show-toplevel`.
-#
-# Prints the absolute path to stdout. Exits 1 if not found.
-set -e
-d="${PWD}"
-while [ "$d" != "/" ] && [ -n "$d" ]; do
-  if [ -f "$d/CLAUDE.md" ] && [ -f "$d/scripts/guard-feedback.py" ]; then
-    echo "$d"
-    exit 0
-  fi
-  d="$(dirname "$d")"
-done
-# Fallback: check if CLAUDE_PROJECT_DIR is set by Claude Code
-if [ -n "${CLAUDE_PROJECT_DIR:-}" ] && [ -f "${CLAUDE_PROJECT_DIR}/scripts/guard-feedback.py" ]; then
-  echo "${CLAUDE_PROJECT_DIR}"
-  exit 0
-fi
-# Final hardcoded fallback (this machine)
-if [ -f "/Users/nature/个人项目/boll/scripts/guard-feedback.py" ]; then
-  echo "/Users/nature/个人项目/boll"
-  exit 0
-fi
-exit 1

package/src/bollharness/reference/boll-reference/scripts/hooks/precompact.sh

@@ -1,57 +0,0 @@
-#!/bin/bash
-# PreCompact hook: 注入必须在 context compaction 中保留的关键信息
-# [来源: ADR-038 D2.3 + D9 (v4 Objective Recitation), CC PreCompact hook — exit 0 追加为 compact 指令]
-
-# 读取当前活跃的 PLAN/TASK/Issue 状态
-REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
-
-cat <<'PRESERVE'
-## 必须保留的关键信息
-
-### 当前工作上下文
-- 保留当前正在进行的 PLAN/TASK/Issue 的完整状态和进度
-- 保留 Sprint Contract（如果有）的全部验收标准
-- 保留当前 WP 的 scope 和验收条件
-
-### 不可降级规则 (ADR-030 核心)
-- Guard > Memory: 能机械化检测的必须写 guard
-- 一个事实一个定义: 同一信息不得在多处重复定义
-- 验证看最后一公里: 验证必须到达用户可观测的终点
-
-### 行为约束
-- 所有 commit message 必须中英双语
-- 所有 subagent 必须使用 claude-opus-4-6
-- 不得跳过 Gate 2/4/6/8 审查
-- Issue 先于代码: 改代码前必须先创建 issue 文档
-PRESERVE
-
-# ── v4 D9: Objective Recitation ──
-# [来源: Manus/IMPACT — recite objectives at end of context after ~50 tool calls]
-# [来源: Cursor — 丢失 reasoning traces 导致 30% 性能下降]
-# 在 compact 后的新上下文末尾追加原始目标，防止注意力漂移。
-# KV cache 友好——只追加到末尾，不重排前文。
-PROGRESS_FILE="$REPO_ROOT/.boll/progress/current.json"
-if [ -f "$PROGRESS_FILE" ]; then
-    echo ""
-    echo "## Objective Recitation (v4 D9)"
-    echo ""
-    echo "**原始目标**（不可漂移，由 D8 Initializer Agent 写入后 read-only）："
-    python3 -c "
-import json, sys
-try:
-    data = json.load(open('$PROGRESS_FILE'))
-    print(f\"  {data.get('objective', '<未设置>')}\")
-    print()
-    pending = [f for f in data.get('features', []) if f.get('status') != 'passing']
-    if pending:
-        print('**未完成 features**：')
-        for f in pending:
-            print(f\"  - [{f.get('status', '?')}] {f.get('id', '?')}: {f.get('subject', '')}\")
-    else:
-        print('**所有 features 已 passing**，可以进入 Stop 流程。')
-except Exception as e:
-    print(f'  <progress.json 解析失败: {e}>')
-" 2>&1
-fi
-
-exit 0

package/src/bollharness/reference/boll-reference/scripts/hooks/stop-evaluator.md

@@ -1,57 +0,0 @@
-# Stop Hook: PreCompletionChecklist + On-Demand Evaluator
-
-你是独立 Evaluator。你的工作是在 agent 完成任务前验证工作质量。
-
-## PreCompletionChecklist
-
-在允许完成前，逐条验证：
-
-1. **Git 状态**: 所有新文件是否已 `git add`？是否有未保存的变更？
-2. **测试**: 如果有代码变更，相关测试是否通过？运行 `backend/venv/bin/pytest -q backend/tests/unit --tb=no -q` 确认。
-3. **文档一致性**: 代码变更是否需要更新文档？路由变更是否反映在路由表中？
-4. **无半成品**: 是否有 TODO/FIXME/HACK 被遗留？是否有 dead code 或未使用的 import？
-5. **契约一致性**: API 类型、路由路径、配置 key 是否在所有引用处一致？
-
-## On-Demand Evaluator
-
-如果当前任务 scope >= 3 files 或涉及 API 契约变更，额外执行 Grading Criteria 评估：
-
-| 维度 | 检查点 |
-|------|--------|
-| **Design Quality** | 架构是否合理？是否利用了现有模式？ |
-| **Originality** | 是否超越模板化实现？ |
-| **Craft** | 代码清洁、命名一致、边界处理、无 AI slop |
-| **Functionality** | 核心路径可走通、契约一致、无断裂 |
-
-每维 1-5 分。**硬阈值: 任何维 < 3 = FAIL**。
-
-## 重要警告
-
-你天然倾向于"识别问题后说服自己不是大问题"。**抵抗这个倾向。**
-
-具体表现：
-- "代码看起来正确" — 不够，要实际验证
-- "测试已经通过了" — 哪些测试？覆盖了变更吗？
-- "这个小问题不影响功能" — 小问题累积成大问题
-
-**如果不确定，判为 FAIL。宁严勿松。**
-
-## 输出格式
-
-```
-## PreCompletionChecklist
-- [x/!] Git 状态: ...
-- [x/!] 测试: ...
-- [x/!] 文档一致性: ...
-- [x/!] 无半成品: ...
-- [x/!] 契约一致性: ...
-
-## Evaluator (如果触发)
-Design Quality: X/5
-Originality: X/5
-Craft: X/5
-Functionality: X/5
-
-## 裁决: PASS / FAIL
-原因: ...
-```

package/src/bollharness/schemas/metrics-jsonl-allowlist.json

@@ -1,67 +0,0 @@
-{
-  "$schema": "https://json-schema.org/draft/2020-12/schema",
-  "$id": "https://bollharness.dev/schemas/metrics-jsonl-allowlist.json",
-  "title": "bollharness metrics jsonl field allowlist",
-  "description": "Allowed top-level fields for every line in .boll/state/metrics/*.jsonl. scan_verify_artifacts.ts --metrics rejects any line containing a key not listed here. Tightening this allowlist is the primary defense against accidental PII / SECRET leakage via metrics sinks (v2 patch_4 / security P1#4).",
-  "type": "object",
-  "additionalProperties": false,
-  "required": ["ts", "event"],
-  "properties": {
-    "ts": {
-      "description": "Event timestamp (ISO-8601 string or epoch-seconds number).",
-      "type": ["string", "number"]
-    },
-    "event": {
-      "description": "Short event name, e.g. guard_fired / hook_failed / rebaseline_required.",
-      "type": "string"
-    },
-    "stage": {
-      "description": "Claude Code hook stage (PreToolUse / PostToolUse / SessionStart / ...).",
-      "type": "string"
-    },
-    "tool": {
-      "description": "Tool name if applicable (Edit / Write / Bash / Task / ...).",
-      "type": "string"
-    },
-    "matcher": {
-      "description": "Hook matcher that routed the event.",
-      "type": "string"
-    },
-    "hook": {
-      "description": "Hook script basename (e.g. guard-feedback.ts).",
-      "type": "string"
-    },
-    "outcome": {
-      "description": "allow / block / warn / noop / error.",
-      "type": "string"
-    },
-    "exit_code": {
-      "type": "integer"
-    },
-    "duration_ms": {
-      "type": "number"
-    },
-    "file_path": {
-      "description": "Relative path of the edited / read file, if applicable.",
-      "type": "string"
-    },
-    "category": {
-      "description": "Sanitize category (SECRET / TRADE_SECRET / PII / NETWORK / PROTOCOL_INTERNAL).",
-      "type": "string"
-    },
-    "reason": {
-      "description": "Short human-readable reason; must not contain raw secret material.",
-      "type": "string"
-    },
-    "session_id": {
-      "description": "Claude Code session id.",
-      "type": "string"
-    },
-    "project_root": {
-      "type": "string"
-    },
-    "schema_version": {
-      "type": "string"
-    }
-  }
-}

package/src/bollharness/scripts/checks/next_decision_number.sh

@@ -1,48 +0,0 @@
-#!/bin/sh
-# 起草 ADR / PLAN 时给出下一个可用号（与 pre-commit numbering check 对齐）.
-#
-# 用法:
-#   scripts/checks/next_decision_number.sh ADR
-#   scripts/checks/next_decision_number.sh PLAN
-#
-# 来源: H 系列收尾 + Nature feedback "plan/adr 序号老是容易重复".
-
-set -e
-
-KIND="${1:-}"
-case "$KIND" in
-    ADR|PLAN) ;;
-    *)
-        echo "Usage: $0 {ADR|PLAN}" >&2
-        exit 2
-        ;;
-esac
-
-REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null)"
-if [ -z "$REPO_ROOT" ]; then
-    echo "Not in a git repo." >&2
-    exit 2
-fi
-
-cd "$REPO_ROOT"
-
-DECISIONS_DIR="docs/decisions"
-if [ ! -d "$DECISIONS_DIR" ]; then
-    echo "No $DECISIONS_DIR/ found." >&2
-    exit 2
-fi
-
-# 取所有非 H 系列 ADR/PLAN-NNN 主号（剥落字母 sub-id），找最大值 +1
-MAX=$(ls "$DECISIONS_DIR" 2>/dev/null \
-    | grep -E "^${KIND}-[0-9]{3}[A-Z]?-" \
-    | grep -vE '^(ADR|PLAN)-H[0-9]+-' \
-    | sed -E "s/^${KIND}-([0-9]{3})[A-Z]?-.*/\1/" \
-    | sort -n | tail -1)
-
-if [ -z "$MAX" ]; then
-    NEXT=1
-else
-    NEXT=$((10#$MAX + 1))
-fi
-
-printf "%s-%03d\n" "$KIND" "$NEXT"