tech-lead-skill 5.4.0

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 240xu
+
+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/README.md

@@ -0,0 +1,128 @@
+# Tech Lead Skill
+
+English | [简体中文](README.zh-CN.md)
+
+An evidence-driven planning and delivery skill for software, infrastructure, research, reverse-engineering, and operations work.
+
+## What It Solves
+
+Many plans fail in one of two ways: they are vague and cannot guide execution, or they are over-detailed and become false after the first environmental change. This skill keeps the plan coarse until evidence justifies detail, then continuously revises it against observed reality.
+
+The control loop is:
+
+```text
+goal -> constraints/assumptions -> L0/L1/L2 plan -> execute -> observe evidence
+     -> revise -> choose CONTINUE / PAUSE / SCOPE-DOWN / PIVOT / STOP
+```
+
+## Core Features
+
+- PLAN and EXECUTE modes with a clear boundary between planning and side effects.
+- Progressive disclosure: L0 architecture, L1 milestones, L2 executable focus.
+- Goal, metric, fact, assumption, decision, risk, dependency, and evidence ledgers.
+- Protected asset handling for source, user data, configuration, secrets, runtime state, and generated artifacts.
+- Minimal mutation protocol: `READ -> CLASSIFY -> PROTECT -> CHANGE -> VERIFY -> RECONCILE -> ROLLBACK/RECORD`.
+- Evidence levels E0-E4, from model inference to observed user outcomes.
+- Failure re-planning, stagnation circuit breaker, rollback discipline, and real-state reconciliation.
+- Optional adversarial review for high-impact or irreversible changes.
+- Git-backed and non-Git project-state recovery.
+- Release-readiness checks for public documentation and skill publishing.
+- Runtime discipline rules for external dependency health, silent failure classes, automation guardrails, idempotent batch jobs, and fallback ladders.
+
+## Install
+
+### npm (GitHub-hosted, no registry account needed)
+
+```bash
+npm i -g github:240xu/tech-lead-skill
+tech-lead-skill          # installs into ~/.config/opencode/skills/tech-lead
+```
+
+One-shot without a global install:
+
+```bash
+npx github:240xu/tech-lead-skill
+```
+
+The installer is idempotent: repeated runs back up existing files as `*.bak-<timestamp>` first. Use `--target <dir>` to choose another destination, `--check` to verify an installed copy against the package (hash + version drift), `--dry-run` to preview without writing, and `--uninstall` to remove. Uninstall removes only manifest-managed files — user files and `*.bak-*` backups are kept.
+
+### DeepSeek Harness plugin (read-only tools)
+
+The repository also ships an opt-in DSH bundle exposing 21 read-only lifecycle tools. The original nine audit tools remain compatible, and the strengthened surface adds context validation, evidence graph/freshness analysis, progress decisions, critical-path and impact analysis, resume reconciliation, gate planning/aggregation/reopen checks, and mutation preview. The tools compute over caller-supplied JSON only — no filesystem writes, no subprocesses, no network access.
+
+```bash
+dsh plugin --profile headless add /path/to/tech-lead-skill/packages/dsh-tech-lead-bundle
+dsh --profile headless --dump-config   # verify the tech-lead-tools row is present
+```
+
+See [`docs/superpowers/specs/2026-08-25-dsh-tech-lead-system.md`](https://github.com/240xu/tech-lead-skill/blob/main/docs/superpowers/specs/2026-08-25-dsh-tech-lead-system.md) for the architecture and permission matrix.
+
+### Manual
+
+Copy `skill/SKILL.md` and the `skill/templates/` directory into the skills directory of your OpenCode-compatible environment:
+
+```text
+~/.config/opencode/skills/tech-lead/
+```
+
+The skill is triggered by project construction, system building, implementation plans, deployment, migration, release, recovery, restructuring, operations work, and other substantial planning requests. It can also be loaded explicitly.
+
+## Operating Modes
+
+### PLAN
+
+Use for intake, goal definition, architecture, decomposition, risk analysis, and verification design. PLAN does not edit files or run side-effecting commands.
+
+### EXECUTE
+
+Use after the L2 scope is clear. Execute only the smallest approved mutation, record the change, verify behavior, reconcile actual state, and update the plan.
+
+## Protected Assets
+
+| Class | Default handling |
+|---|---|
+| `SOURCE` | Inspectable diff, tests, and restore point |
+| `USER_DATA` | Read-only by default; write only to an explicit target with recovery |
+| `CONFIG` | Read current state, make the smallest change, reload, verify |
+| `SECRET` | Never place in plans, logs, ordinary backups, or diffs |
+| `RUNTIME` | Inspect live state before restart, kill, replace, or migration |
+| `GENERATED` | Prefer regeneration; do not treat as the source of truth |
+
+## Evidence Levels
+
+- `E0`: model inference; hypothesis only.
+- `E1`: static inspection, grep, or configuration inspection.
+- `E2`: local command or unit test; local behavior only.
+- `E3`: integration test, real process, or real endpoint.
+- `E4`: user acceptance, real business result, or production observation.
+
+## Templates
+
+- `templates/intake.md`: goal, constraints, assets, risk, and completion level.
+- `templates/plan.md`: L0/L1/L2 plan and current focus.
+- `templates/change-record.md`: one EXECUTE mutation and its reconciliation.
+- `templates/round.md`: one planning iteration and its outcome.
+- `templates/state.json`: canonical resumable state projection.
+- `templates/gate-review.md` and `templates/gate-verdict.md`: independent review and decision records.
+- `templates/release-check.md`: publication inventory, scans, remote verification, and limitations.
+
+## Scope
+
+This is an engineering planning skill. It focuses on correctness of planning, safe handling of user files and code changes, evidence quality, rollback, and real-environment reconciliation. It intentionally stays focused on project-level engineering work rather than organization-wide process design.
+
+## Limitations
+
+- The judgment layer remains prose by design; the optional DSH bundle provides a machine-checkable, composable read-only runtime for context, evidence, progress, gates, release/install audits, recovery, and mutation preview.
+- Evidence freshness and state reconciliation are operator/agent responsibilities until dedicated tooling exists.
+- The skill does not provide a sandbox; untrusted code must not be executed unless an actual isolated execution environment is already available.
+- MCP candidates should be selected only after observing repeated real-project violations.
+
+## Notes
+
+The executable skill body (`SKILL.md`) is authored in Simplified Chinese; coding agents execute it correctly regardless of the conversation language. Templates are English and agent-facing. Documentation translations cover this README and the technical guide.
+
+## Version
+
+Current version: `v5.4.0`.
+
+See [`docs/TECHNICAL_GUIDE.md`](docs/TECHNICAL_GUIDE.md) for the full operating model and [`docs/AUDIT_REPORT.md`](docs/AUDIT_REPORT.md) for the publication audit.

package/README.zh-CN.md

@@ -0,0 +1,130 @@
+# Tech Lead Skill
+
+[English](README.md) | 简体中文
+
+一套以证据驱动的规划与交付技能，适用于软件、基础设施、研究、逆向工程与运维类工作。
+
+## 解决什么问题
+
+计划失败通常有两种方式：要么模糊到无法指导执行，要么过度细化而在第一次环境变化后就失真。本技能让计划在证据支撑之前保持粗粒度，然后依据观察到的现实持续修订。
+
+控制回路：
+
+```text
+目标 -> 约束/假设 -> L0/L1/L2 计划 -> 执行 -> 观察证据
+     -> 修订 -> 选择 CONTINUE / PAUSE / SCOPE-DOWN / PIVOT / STOP
+```
+
+## 核心特性
+
+- PLAN 与 EXECUTE 双模式，划清规划与副作用的边界。
+- 渐进展开：L0 架构、L1 里程碑、L2 可执行焦点。
+- 目标、指标、事实、假设、决策、风险、依赖与证据账本。
+- 受保护资产处置：源码、用户数据、配置、秘密、运行态与生成物。
+- 最小变更协议：`READ -> CLASSIFY -> PROTECT -> CHANGE -> VERIFY -> RECONCILE -> ROLLBACK/RECORD`。
+- E0-E4 证据分级，从模型推断到真实用户结果。
+- 失败重规划、停滞断路器、回滚纪律与真实状态对账。
+- 面向高影响或不可逆变更的对抗式评审（可选）。
+- 支持 Git 与非 Git 两种项目状态恢复方式。
+- 面向公开文档与技能发布的收尾验证检查。
+- 运行期纪律规则：外部依赖健康、静默失败类别、自动化护栏、幂等批处理与降级阶梯。
+
+## 安装
+
+### npm 方式（GitHub 源，无需注册账号）
+
+```bash
+npm i -g github:240xu/tech-lead-skill
+tech-lead-skill          # 安装到 ~/.config/opencode/skills/tech-lead
+```
+
+一次性运行（不全局安装）：
+
+```bash
+npx github:240xu/tech-lead-skill
+```
+
+安装器幂等，重复执行会先把旧文件备份为 `*.bak-<时间戳>`；`--target <目录>` 可自定义目标；`--check` 校验已装副本与包的哈希/版本漂移；`--dry-run` 预览不写入；`--uninstall` 卸载时只删除清单受管文件，用户文件与 `*.bak-*` 备份全部保留。
+
+`--check` 退出码：`0` 正常，`1` 检出漂移，`2` 用法/拒绝错误。
+
+### DeepSeek Harness 插件（只读工具）
+
+本仓库同时提供可选 DSH bundle，暴露 21 个只读生命周期工具。原有九个审计工具保持兼容，并新增上下文校验、证据图与新鲜度分析、推进决策、关键路径与影响分析、续跑对账、Gate 计划/聚合/重开检查，以及变更预览。工具仅对调用方传入的 JSON 做计算——无文件写入、无子进程、无网络访问。
+
+```bash
+dsh plugin --profile headless add /path/to/tech-lead-skill/packages/dsh-tech-lead-bundle
+dsh --profile headless --dump-config   # 确认 tech-lead-tools 行已注入
+```
+
+架构与权限矩阵见 [spec](https://github.com/240xu/tech-lead-skill/blob/main/docs/superpowers/specs/2026-08-25-dsh-tech-lead-system.md)。
+
+### 手动方式
+
+将 `skill/SKILL.md` 和 `skill/templates/` 目录复制到你所用 OpenCode 兼容环境的技能目录：
+
+```text
+~/.config/opencode/skills/tech-lead/
+```
+
+该技能由"构建项目/搭系统/制定执行计划/部署/迁移/发布/恢复/重构/运维"等实质性任务触发，即使用户没有明确说“项目”或“Tech Lead”也应加载；也可以显式加载。
+
+## 工作模式
+
+### PLAN
+
+用于需求接入、目标定义、架构设计、任务分解、风险分析与验证设计。PLAN 不改文件、不执行有副作用的命令。
+
+### EXECUTE
+
+在 L2 范围明确后使用。只执行最小已批准变更，记录改动，验证行为，对账实际状态，并更新计划。
+
+## 受保护资产
+
+| 类别 | 默认处置 |
+|---|---|
+| `SOURCE` | 可审查 diff、测试与恢复点 |
+| `USER_DATA` | 默认只读；仅在目标明确且有可恢复副本时写入 |
+| `CONFIG` | 先读现状，最小修改，重载后验证 |
+| `SECRET` | 绝不进入计划、日志、普通备份或 diff |
+| `RUNTIME` | 重启/杀进程/替换/迁移前必须先查实时状态 |
+| `GENERATED` | 优先重新生成；不作为事实源 |
+
+## 证据分级
+
+- `E0`：模型推断；仅可用于提出假设。
+- `E1`：静态阅读、grep 或配置检查。
+- `E2`：本地命令或单元测试；仅证明局部行为。
+- `E3`：集成测试、真实进程或真实端点。
+- `E4`：用户验收、真实业务结果或生产观察。
+
+## 模板
+
+- `templates/intake.md`：目标、约束、资产、风险与完成层级。
+- `templates/plan.md`：L0/L1/L2 计划与当前焦点。
+- `templates/change-record.md`：单次 EXECUTE 变更及其对账。
+- `templates/round.md`：单轮规划迭代及其结果。
+- `templates/state.json`：可续跑的规范状态投影。
+- `templates/gate-review.md` 与 `templates/gate-verdict.md`：独立评审与裁决记录。
+- `templates/release-check.md`：发布清单、扫描、远程验证与限制记录。
+
+## 范围
+
+这是一套工程规划技能，聚焦规划正确性、用户文件与代码改动的安全处置、证据质量、回滚以及真实环境对账。它有意不涉及组织级流程设计或通用项目管理方法论。
+
+## 局限
+
+- 目前规则是文本形式，不是机器强制的状态机。
+- 在专用工具出现之前，证据时效与状态对账依赖执行环境与操作者。
+- 本技能不提供沙箱；除非已有真实隔离执行环境，否则不得运行不可信代码。
+- MCP 工具化候选应在观察到真实项目中反复违例之后再筛选。
+
+## 说明
+
+可执行的技能正文 `SKILL.md` 以简体中文编写；编码代理无论对话语言为何均可正确执行。模板为面向代理的英文文件。文档翻译覆盖本 README 与技术指南。
+
+## 版本
+
+当前版本：`v5.4.0`。
+
+完整运行模型见[技术指南](https://github.com/240xu/tech-lead-skill/blob/main/docs/TECHNICAL_GUIDE.zh-CN.md)，发布审计见 [docs/AUDIT_REPORT.md](docs/AUDIT_REPORT.md)。

package/bin/install.js

@@ -0,0 +1,245 @@
+#!/usr/bin/env node
+'use strict';
+const fs = require('fs');
+const path = require('path');
+const os = require('os');
+const crypto = require('crypto');
+
+const pkgRoot = path.resolve(__dirname, '..');
+const pkg = JSON.parse(fs.readFileSync(path.join(pkgRoot, 'package.json'), 'utf8'));
+const srcSkill = path.join(pkgRoot, 'skill');
+const markerName = '.tech-lead-skill.json';
+const args = process.argv.slice(2);
+const defaultTarget = path.resolve(path.join(os.homedir(), '.config', 'opencode', 'skills', 'tech-lead'));
+const knownFlags = new Set([
+  '--help', '-h', '--version', '-v',
+  '--uninstall', '--target', '--check', '--dry-run',
+]);
+
+for (const arg of args) {
+  if (arg.startsWith('-') && !knownFlags.has(arg)) {
+    console.error('unknown option: ' + arg);
+    process.exit(2);
+  }
+}
+
+function argValue(name) {
+  const i = args.indexOf(name);
+  if (i === -1) return undefined;
+  const value = args[i + 1];
+  if (!value || value.startsWith('-')) {
+    console.error(name + ' requires a directory argument');
+    process.exit(2);
+  }
+  return value;
+}
+
+if (args.includes('--help') || args.includes('-h')) {
+  console.log(
+    `tech-lead-skill ${pkg.version}\n` +
+    'Usage: tech-lead-skill [options]\n' +
+    '  (no args)         install into ~/.config/opencode/skills/tech-lead\n' +
+    '  --target <dir>    install into <dir>\n' +
+    '  --check           verify installed files match this package\n' +
+    '  --dry-run         print planned actions without writing anything\n' +
+    '  --uninstall       remove only manifest-managed files (keeps user files)\n' +
+    '  --version         print version\n' +
+    '  --help            show this help\n' +
+    '\n' +
+    'Exit codes: 0 ok, 1 drift detected (--check), 2 usage/refusal error'
+  );
+  process.exit(0);
+}
+
+if (args.includes('--version') || args.includes('-v')) {
+  console.log(pkg.version);
+  process.exit(0);
+}
+
+function walkFiles(root) {
+  const out = [];
+  (function walk(rel) {
+    const abs = path.join(root, rel);
+    let entries;
+    try { entries = fs.readdirSync(abs, { withFileTypes: true }); }
+    catch (err) { console.warn('warn: cannot read ' + abs + ' (' + err.code + '), skipping'); return; }
+    for (const entry of entries) {
+      const relPath = rel ? rel + '/' + entry.name : entry.name;
+      if (entry.isDirectory()) walk(relPath);
+      else out.push(relPath.split(path.sep).join('/'));
+    }
+  })('');
+  return out.sort();
+}
+
+function isBackup(rel) {
+  return /\.bak-\d+(?:-\d+)?$/.test(rel);
+}
+
+function readMarker(target) {
+  try {
+    return JSON.parse(fs.readFileSync(path.join(target, markerName), 'utf8'));
+  } catch (_) {
+    return null;
+  }
+}
+
+// A manifest entry may only address paths INSIDE target: blocks traversal via
+// tampered marker.files ("../x", absolute paths resolve under target anyway).
+function containedRel(target, rel) {
+  if (typeof rel !== 'string' || !rel || rel.includes('\\0')) return null;
+  const abs = path.resolve(target, rel);
+  const base = path.resolve(target);
+  const relCheck = path.relative(base, abs);
+  if (relCheck.startsWith('..') || path.isAbsolute(relCheck)) return null;
+  return relCheck.split(path.sep).join('/');
+}
+
+function sha256File(file) {
+  return crypto.createHash('sha256').update(fs.readFileSync(file)).digest('hex');
+}
+
+function copyWithBackup(src, dest) {
+  if (fs.existsSync(dest) && fs.statSync(dest).isDirectory()) {
+    console.error('refusing: destination exists as a directory: ' + dest);
+    console.error('remove or rename it, then re-run the installer');
+    process.exit(2);
+  }
+  if (fs.existsSync(dest)) {
+    const bak = dest + '.bak-' + Date.now() + '-' + process.pid;
+    fs.copyFileSync(dest, bak);
+    console.log('backup : ' + bak);
+  }
+  fs.copyFileSync(src, dest);
+}
+
+function isHomeDir(dir) {
+  try { return fs.realpathSync(dir) === fs.realpathSync(os.homedir()); }
+  catch (_) { return false; }
+}
+
+let target = argValue('--target') || defaultTarget;
+target = path.resolve(target);
+
+const dryRun = args.includes('--dry-run');
+
+// ── uninstall ────────────────────────────────────────────────────────────────
+if (args.includes('--uninstall')) {
+  const marker = readMarker(target);
+  if (!marker || !Array.isArray(marker.files)) {
+    console.error('refusing to remove an unmarked target: ' + target);
+    console.error('install this package there first, or remove it manually after checking the path');
+    process.exit(2);
+  }
+  let removed = 0;
+  const dirsTouched = new Set();
+  let wouldRemove = 0;
+  for (const raw of marker.files) {
+    const rel = containedRel(target, raw);
+    if (!rel) { console.warn('skipped unsafe manifest entry: ' + JSON.stringify(raw)); continue; }
+    const abs = path.join(target, rel);
+    if (fs.existsSync(abs) && fs.statSync(abs).isDirectory()) {
+      console.warn('skipped directory entry (not managed): ' + rel);
+      continue;
+    }
+    dirsTouched.add(path.dirname(abs));
+    if (dryRun) { console.log('would remove: ' + abs); wouldRemove++; continue; }
+    if (fs.existsSync(abs)) {
+      fs.unlinkSync(abs);
+      removed++;
+      console.log('removed : ' + rel);
+    }
+  }
+  const markerAbs = path.join(target, markerName);
+  if (dryRun) console.log('would remove: ' + markerAbs);
+  else { fs.unlinkSync(markerAbs); removed++; }
+
+  // Try to prune now-empty directories touched by the manifest (bottom-up).
+  const sortedDirs = [...dirsTouched].sort((a, b) => b.length - a.length);
+  for (const dir of sortedDirs) {
+    if (dir === target) continue;
+    if (!dryRun) { try { fs.rmdirSync(dir); } catch (_) { /* non-empty: keep */ } }
+  }
+
+  // Report leftovers so nothing disappears silently.
+  const leftovers = [];
+  (function scan(rel) {
+    const abs = path.join(target, rel);
+    for (const entry of fs.readdirSync(abs, { withFileTypes: true })) {
+      const relPath = rel ? rel + '/' + entry.name : entry.name;
+      if (entry.isDirectory()) scan(relPath);
+      else leftovers.push(relPath.split(path.sep).join('/'));
+    }
+  })('');
+  if (leftovers.length) {
+    console.log(dryRun ? 'still present (dry-run; includes managed files):' : 'kept (not managed):');
+    for (const rel of leftovers) console.log('  kept   : ' + rel);
+  } else if (!dryRun && !isHomeDir(target)) {
+    try { fs.rmdirSync(target); console.log('removed empty target directory'); } catch (_) { /* keep */ }
+  }
+  const count = dryRun ? wouldRemove : removed;
+  console.log((dryRun ? '[dry-run] would remove ' : 'removed ') + count + ' managed entr' + (count === 1 ? 'y' : 'ies'));
+  process.exit(0);
+}
+
+// ── check ────────────────────────────────────────────────────────────────────
+if (args.includes('--check')) {
+  const marker = readMarker(target);
+  if (!marker || !Array.isArray(marker.files)) {
+    console.error('not installed here (no valid manifest): ' + target);
+    process.exit(2);
+  }
+  let hasError = false;
+  if (marker.version !== pkg.version) {
+    console.log(`DRIFT  version  : installed ${marker.version} != package ${pkg.version}`);
+    hasError = true;
+  }
+  const missing = [];
+  const drifted = [];
+  for (const raw of marker.files) {
+    const rel = containedRel(target, raw);
+    if (!rel) { console.warn('skipped unsafe manifest entry: ' + JSON.stringify(raw)); continue; }
+    const abs = path.join(target, rel);
+    const src = path.join(srcSkill, rel);
+    if (!fs.existsSync(abs)) { missing.push(rel); continue; }
+    if (fs.existsSync(src) && sha256File(abs) !== sha256File(src)) drifted.push(rel);
+  }
+  if (missing.length) hasError = true;
+  if (drifted.length) hasError = true;
+
+  const managedSet = new Set(marker.files);
+  const unmanaged = walkFiles(target).filter((rel) => !managedSet.has(rel) && !isBackup(rel) && rel !== markerName);
+  if (missing.length) console.log('MISSING managed files:');
+  for (const rel of missing) console.log('  missing: ' + rel);
+  if (drifted.length) console.log('MODIFIED since install:');
+  for (const rel of drifted) console.log('  drift  : ' + rel);
+  if (unmanaged.length) {
+    console.log('NOTE unmanaged files kept in place (informational):');
+    for (const rel of unmanaged) console.log('  extra  : ' + rel);
+  }
+  if (hasError) process.exit(1);
+  console.log(`ok: ${marker.files.length} managed files match tech-lead-skill v${pkg.version}`);
+  process.exit(0);
+}
+
+// ── install ──────────────────────────────────────────────────────────────────
+const files = walkFiles(srcSkill);
+console.log((dryRun ? '[dry-run] would install ' : 'installing ') +
+  `${pkg.name} v${pkg.version} -> ${target} (${files.length} files)`);
+for (const rel of files) {
+  const dest = path.join(target, rel);
+  if (dryRun) { console.log('  would write: ' + rel); continue; }
+  fs.mkdirSync(path.dirname(dest), { recursive: true });
+  copyWithBackup(path.join(srcSkill, rel), dest);
+}
+if (!dryRun) {
+  fs.writeFileSync(path.join(target, markerName), JSON.stringify({
+    package: pkg.name,
+    version: pkg.version,
+    installedAt: new Date().toISOString(),
+    files,
+  }, null, 2) + '\n');
+  console.log('installed: tech-lead-skill v' + pkg.version + ' -> ' + target);
+} else {
+  console.log('[dry-run] no changes written');
+}

package/docs/AUDIT_REPORT.md

@@ -0,0 +1,58 @@
+# Publication Audit Report
+
+Date: 2026-08-25
+Scope: public release of the `tech-lead` planning skill and its eight templates.
+
+## Result
+
+The publication set contains only the following intended artifacts:
+
+- `package.json` / `bin/install.js` (npm installer)
+- `skill/SKILL.md`
+- `skill/templates/intake.md`
+- `skill/templates/plan.md`
+- `skill/templates/change-record.md`
+- `skill/templates/round.md`
+- `skill/templates/release-check.md`
+- `skill/templates/state.json`
+- `skill/templates/gate-review.md`
+- `skill/templates/gate-verdict.md`
+- `README.md` (English) / `README.zh-CN.md` (简体中文)
+- `docs/TECHNICAL_GUIDE.md` (English) / `docs/TECHNICAL_GUIDE.md (zh guide ships via GitHub only)
+- `docs/AUDIT_REPORT.md`
+- `LICENSE`
+
+Documentation is bilingual (English + Simplified Chinese) by design; additional languages were intentionally excluded to limit maintenance surface. The skill body (`SKILL.md`) is authored in Simplified Chinese and is executed correctly by coding agents regardless of conversation language.
+
+No unrelated home-directory files, project histories, server notes, credentials, or private operational documents are included.
+
+## Checks Performed
+
+### Sensitive-content scan
+
+The source set was scanned for local absolute paths, passwords, API-key prefixes, SSH host details, private domains, and server IP patterns. The only address-like match is the explicitly documented loopback placeholder `127.0.0.1`; no real operational endpoint or credential was found in the publication set.
+
+### Structural scan
+
+- YAML frontmatter has `name: tech-lead` and a trigger description.
+- The skill version is `v5.4.0`.
+- The seven templates referenced by the skill exist.
+- The installer validates options, requires a marker before removing any target, and records the managed package/version after installation.
+- The release-check template covers allowlist inventory, sensitive-content scanning, reference checks, scope checks, publication results, and remote verification.
+- The plan includes PLAN/EXECUTE, L0/L1/L2, protected assets, mutation protocol, E0-E4 evidence, completion levels, stagnation control, state recovery, and real-state reconciliation.
+- Version markers verified against package.json at audit time (see CHANGELOG history for prior releases)
+
+### Safety review
+
+The skill defaults user data to read-only, requires a recovery path before writes, excludes secrets from normal plans/logs/diffs/backups, requires live-state inspection before runtime operations, and prevents untrusted code execution without an actual isolated environment.
+
+## Known Limitations
+
+- Core judgment stays prose-first; a mechanical subset (state schema, evidence provenance, gate prechecks, release/install audits) is machine-checkable via @240xu/dsh-tech-lead-core and the optional DSH bundle tools.
+- Freshness, reconciliation, and rollback verification depend on the executing environment.
+- No claim is made that the skill has completed a multi-project effectiveness trial; that is the next validation phase.
+- The audit validates publication content, not the security of the hosting platform or every consumer's local OpenCode installation.
+
+## Release Decision
+
+Publication is appropriate as a public documentation and skill repository. Consumers should review the skill before applying it to production or user-data changes, and should treat its templates as guidance until project-specific validation is complete.

package/docs/TECHNICAL_GUIDE.md

@@ -0,0 +1,133 @@
+# Technical Guide
+
+[English](TECHNICAL_GUIDE.md) | [简体中文](https://github.com/240xu/tech-lead-skill/blob/main/docs/TECHNICAL_GUIDE.zh-CN.md)
+
+## 1. Planning Model
+
+The skill treats a project as a controlled loop rather than a static checklist.
+
+### 1.1 Planning Objects
+
+Every important statement should be typed:
+
+| Object | Question |
+|---|---|
+| Goal | What user-visible result matters? |
+| Metric | How will that result be measured? |
+| Constraint | What must not be violated? |
+| Fact | What has been observed and where? |
+| Assumption | What is temporarily believed but unverified? |
+| Decision | Which option was selected and why? |
+| Risk | What could make the goal fail? |
+| Dependency | What external state or prerequisite blocks progress? |
+| Evidence | What observation supports or falsifies a claim? |
+
+The minimum goal chain is:
+
+```text
+goal -> metric -> baseline -> target -> measurement -> deadline/stop -> non-goal
+```
+
+### 1.2 Plan Resolution
+
+- L0 answers what the system or project is.
+- L1 turns L0 into milestones with goals, DoD, gates, dependencies, triggers, and status.
+- L2 expands only the current one or two milestones into the smallest executable actions.
+
+The plan should be detailed near the current focus and deliberately coarse in the distance.
+
+## 2. Iteration Cycle
+
+Each meaningful round has five beats:
+
+1. Display the full plan snapshot and `Delta vs previous round`.
+2. Execute the smallest current action.
+3. Observe hard evidence rather than inferred success.
+4. Revise affected plan items and assumptions.
+5. Ask whether the goal, evidence, blockers, and falsification strategy still hold.
+
+The round ends with exactly one outcome:
+
+- `CONTINUE`: the path remains supported and the next action has a testable result.
+- `PAUSE`: a missing resource or external fact blocks safe progress.
+- `SCOPE-DOWN`: preserve value by reducing scope and rewriting DoD and Non-Goals.
+- `PIVOT`: a core assumption or decision was falsified.
+- `STOP`: the goal is achieved, risk is unacceptable, or further work is not worth its cost.
+
+## 3. Change Safety
+
+Every state-changing action uses the mutation protocol:
+
+```text
+READ -> CLASSIFY -> PROTECT -> CHANGE -> VERIFY -> RECONCILE -> ROLLBACK/RECORD
+```
+
+The protocol is intentionally asset-aware. A source edit, a user database migration, a secret-file update, and a process restart are not equivalent changes and must not share the same default procedure.
+
+Before changing a runtime or deployment target, compare the plan with current reality: file version or hash, service state, effective configuration, health endpoint, port state, database migration version, and deployed version.
+
+## 4. Completion
+
+Completion is layered:
+
+1. `artifact-complete`: the intended artifact exists.
+2. `functional-verified`: the behavior passes applicable functional checks.
+3. `operational-verified`: the real process, deployment, or integration path works.
+4. `outcome-validated`: the user goal or key metric is observed.
+
+A stable plan is not enough. Two rounds with no plan delta are useful convergence evidence, but they cannot replace the completion level required by the project.
+
+## 5. Review and Stagnation
+
+Adversarial review is required when a gate is irreversible, affects several downstream milestones, or follows a disputed decision. Reviewers inspect the same artifact snapshot and must cite concrete evidence or counterexamples. Repeated wording does not count as independent findings.
+
+The stagnation circuit breaker activates when there is no new E2+ evidence for two rounds, the critical path does not shrink, assumptions are reused without a validation design, or scope reduction is not re-baselined. The next action must be a falsifying experiment, pause, pivot, or stop.
+
+## 6. State and Recovery
+
+Git repositories use commits and tags for source/configuration lineage. Non-Git environments use a manifest describing paths, versions, backups, service state, and verification commands. `state.json` stores the compact machine-readable summary, while human-readable ledgers live in the project-state files.
+
+On resume, reconcile saved state with the real environment. If they conflict, reality wins and the affected plan items become `[待重估]`.
+
+## 7. Future Toolization
+
+Core judgment stays prose-first; a mechanical subset is machine-checkable: `bin/install.js --check|--dry-run` for install drift, and the optional DSH bundle (`packages/`) exposing 21 composable read-only tools. The original nine audit tools remain compatible; the strengthened surface adds context, evidence graph/freshness, progress, critical path, impact, resume reconciliation, gate orchestration, and mutation preview.
+
+## 8. Release Readiness
+
+When publishing a skill or technical document, do not publish the mixed working directory. Build a clean staging directory from an explicit allowlist, then run:
+
+```text
+allowlist -> content scan -> reference check -> scope check
+          -> publication check -> publish -> remote verify
+```
+
+The content scan covers local paths, personal data, credentials, cookies, tokens, internal hosts, deployment details, private repository links, and real test data. The reference check covers versions, section numbers, repository-relative paths, template references, placeholders, and the README entry point. The scope check removes project-specific operational details and retains only mechanisms that can be reused elsewhere.
+
+After publication, independently check the remote visibility, default branch, commit, file inventory, and README access. A local commit is not evidence that a remote publication succeeded. If repository creation or push fails, stop the dependent verification steps and report the actual remote state.
+
+The release check must record limitations. A clean scan is evidence that the defined checks passed, not proof that unknown sensitive information does not exist.
+
+## 9. Runtime Discipline
+
+Projects that produce long-running services, scheduled jobs, batch pipelines, or third-party dependencies add five rules:
+
+1. **External dependency health**: register probes for critical upstreams. Placeholder data (such as all-`127.0.0.1` nodes), dead domains, and upstream end-of-life mean the dependency is dead — take the alternative path instead of retrying as a transient error.
+2. **Silent failure class**: HTTP 200 with an empty body or early EOF, placeholder content, and silent truncation are failure classes distinct from explicit errors; verify content validity rather than status codes, and count repeated silent failures in the re-planning tally.
+3. **Automation trio**: long-running loops carry a hard timeout, a consecutive-failure circuit breaker, and a concurrency cap. Recovery is staged as bounded retry, isolate a connection or instance, controlled restart, then controlled rebuild; rebuilding needs idempotency, a recovery point, rate limits, a maximum count, and a human escalation condition. Systems that live across days add scheduled health checks, and the monitor itself needs a liveness check.
+4. **Batch idempotency**: batch tasks are re-runnable with dry-run preview and incremental reconciliation; repeated runs never duplicate data.
+5. **Fallback ladder**: degradation chains are designed up front with per-level triggers; improvised fallbacks during incidents are treated as directional defects and re-planned.
+
+## 10. State and Review Artifacts
+
+The bundled templates are records, not decorative checklists:
+
+- `state.json` is the canonical resumable projection. If the project is read-only, record `state_persistence: unavailable` elsewhere instead of creating files.
+- `intake.md` establishes the goal chain and typed planning ledgers.
+- `plan.md` carries L0/L1/L2 lineage, milestone gates, completion levels, and evidence anchors.
+- `round.md` records delta counts, failure classification, rework level, stagnation checks, and outcome details.
+- `change-record.md` proves each step of the mutation protocol and records requirement changes separately from implementation changes.
+- `gate-review.md` and `gate-verdict.md` record independent review, anchored findings, and adopted or rejected decisions.
+- `release-check.md` records the seven release phases and stops publication when a prerequisite or remote verification is unresolved.
+
+Use the smallest applicable artifact set, but do not omit a required record merely because the work is inconvenient to document. A template entry without an evidence anchor is an open item, not proof of completion.

package/package.json

@@ -0,0 +1,41 @@
+{
+  "name": "tech-lead-skill",
+  "version": "5.4.0",
+  "description": "Evidence-driven planning and safe change-control skill for OpenCode-compatible agents",
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/240xu/tech-lead-skill.git"
+  },
+  "homepage": "https://github.com/240xu/tech-lead-skill#readme",
+  "bugs": "https://github.com/240xu/tech-lead-skill/issues",
+  "bin": {
+    "tech-lead-skill": "bin/install.js"
+  },
+  "files": [
+    "skill/",
+    "bin/",
+    "README.md",
+    "README.zh-CN.md",
+    "docs/TECHNICAL_GUIDE.md",
+    "docs/AUDIT_REPORT.md",
+    "LICENSE"
+  ],
+  "engines": {
+    "node": ">=16"
+  },
+  "keywords": [
+    "opencode",
+    "skill",
+    "planning",
+    "agent",
+    "tech-lead"
+  ],
+  "workspaces": [
+    "packages/*"
+  ],
+  "devDependencies": {
+    "@deepseek-ai/dsh-tools": "0.1.0-rc.7",
+    "@deepseek-ai/dsh-llm": "0.1.0-rc.7"
+  }
+}

package/skill/SKILL.md

@@ -0,0 +1,214 @@
+---
+name: tech-lead
+description: 用于软件、基础设施、研究和运维中的多步、跨边界或有状态工程交付：项目规划、架构选择、跨模块实现、部署、迁移、发布、恢复、重构、长期运行与跨会话续跑。涉及不可逆/生产变更、回滚验证、真实环境对账、发布安全或项目级交付治理时优先加载，即使用户没有说“项目”或“Tech Lead”。不要仅因一次性小修复、单文件编辑、普通解释、单次测试或只读代码审查而自动加载，除非用户明确要求项目级计划或本规范。
+---
+
+# Tech Lead · 自主演进规划与交付规范 v5.4.0
+
+> 双层架构声明：本文件是**判断层**（自由裁量域）。其中可机械判定的不变量（Gate 前置校验、
+> 证据锚点存在性、分级判定器、状态机转换有效性）为未来 **MCP 工具化候选**——
+> 由真实项目违例统计驱动（见附录A），被违反最多的条款才值得固化成工具。
+
+## §0 铁律与工作模式
+1. **PLAN 模式**只负责目标、约束、方案、风险、任务分解和验证设计；不得改文件或执行副作用命令，可给出文件路径和命令意图。
+2. **EXECUTE 模式**在已确认的 L2 范围内执行最小变更、运行验证并记录结果；涉及受保护资产或副作用时必须先过 §3.2 变更协议。
+3. 每条 PLAN 回复末行自检：`code-free: yes`。EXECUTE 回复可包含必要的代码、diff 和命令结果，但必须说明当前模式。
+4. 冲突时序：系统/平台安全与授权边界 > 用户现场明示指令 > 项目安全政策与受控变更流程 > 项目文档和真实状态 > 行业惯例。用户可改变目标和范围，但不能授权越权访问、泄露秘密或绕过安全控制。
+5. 用户已明确授权实现、修改、部署或迁移时，先输出最小范围、风险、恢复点和验证计划，并把该授权记录为确认；目标、资产、权限或不可逆影响仍不明确时才 `PAUSE` 提问。
+
+## §1 任务分级与双向升降档
+1. T0（<30min·单文件·零依赖·可丢弃·无受保护资产）＝直通给方案＋一句风险提示；铁律自检不可省。
+2. T1（单模块·1~3 天·有测试边界）＝骨架+DoD，跳过盲评。
+3. T2（≥2 模块/跨周/多方依赖/含不可逆操作）＝全流程。
+4. 复杂度与改动风险取高者定档；普通源码/配置小改可为 T1，但涉及用户数据、秘密、生产运行态、删除/覆盖/迁移、公共接口、CI/CD、权限、网络策略、计费、安全策略或不可轻易回滚的操作直接 T2；风险不确定时升档。
+5. `[T-DOWN|T-UP]` 双向重判＋一句理由留痕；降档准入：连续 2 个 Gate 全绿且假设清单清空；
+   升档触发：命中 §4.4 影响“高”判据、触及不可逆操作，或失败暴露出边界/契约/选型问题。高档已完成成果冻结复用不返工。
+
+## §2 需求前置澄清（T1/T2 硬门）
+M0 前列 ≤5 个关键假设请用户确认；目标、Non-Goals 和不可逆影响不明确时不得冻结蓝图，可先给出标注 `[草案]` 的范围选项与只读调研结果；
+无回应时仅允许无副作用的只读调研并标注 [ASSUMPTION] 写入 state.json；涉及改动、删除、覆盖、部署、运行态操作或目标不明确时必须 `PAUSE`，不得自动推进；
+后续任何 Gate 引用到该假设时必须先复查确认状态；被推翻则受影响条目全部标记 [待重估]。
+
+## §3 骨架蓝图（L0 层）
+目录结构｜选型 ADR（理由+被否决备选）｜模块边界｜数据流拓扑｜接口契约草案｜Non-Goals。
+
+## §3.1 系统规划对象（先分清，再推进）
+重要陈述必须归入一种类型，禁止把假设伪装成事实：
+- **Goal**：用户要得到的结果；必须绑定至少一个 Metric、基线、目标值、测量方法和截止/停止条件。
+- **Constraint**：不可违反的资源、平台、时间、权限或环境边界；Constraint 与偏好分开记录。
+- **Fact**：有来源、有时间、有适用范围的已观察事实；没有证据的内容只能写成 Assumption。
+- **Assumption**：当前为推进而暂接受、尚未验证的前提；写明验证方式、过期条件和受影响条目。
+- **Decision**：在候选方案之间做出的选择；记录备选、理由、可逆性和复查触发器。
+- **Risk**：可能导致目标失败的事件；记录影响、概率/不确定性、缓解措施和触发阈值。
+- **Dependency**：一个条目对外部资源、前置产物或环境状态的依赖；记录阻塞关系和替代路径。
+- **Evidence**：用于支持或推翻结论的证据；记录来源、时间、范围、复现方法和可信度。
+
+最小目标账本：`goal → metric → baseline → target → measure → deadline/stop → non-goal`。
+局部测试全绿只能证明局部行为，不得直接宣称系统目标达成。
+
+### §3.2 受保护资产与最小变更协议
+每个 L2 动作先标注资产类别和爆炸半径：
+- `SOURCE`：源码、测试、文档；通过 diff、测试和版本恢复点保护。
+- `USER_DATA`：用户文件、工作成果、业务数据库内容；默认只读，写入前明确目标并建立可恢复副本。
+- `CONFIG`：环境变量、服务配置、编辑器/部署配置；先读现状，最小修改，重载后独立验证。
+- `SECRET`：密钥、令牌、Cookie、凭据文件；不输出、不进入普通备份、日志、diff 或计划正文。
+- `RUNTIME`：进程、端口、队列、缓存、线上服务状态；先查真实状态，禁止依据旧快照直接 kill/restart/replace。
+- `GENERATED`：构建产物、缓存、临时文件；优先重新生成，不把生成物当事实源。
+
+凡涉及状态改变，严格按 `READ → CLASSIFY → PROTECT → CHANGE → VERIFY → RECONCILE → ROLLBACK/RECORD` 执行：先读实际状态，分类资产/可逆性/影响面，建立合适恢复点，只做最小变更，验证目标行为，对账实际状态与计划状态，最后记录 diff、证据、恢复点和残余风险。任何一步失败都不得标记完成。
+
+备份按资产分类：源码/配置保留可审查 diff；用户数据和数据库保留可恢复副本并验证恢复；秘密只使用受控秘密存储或安全权限文件，不复制到普通备份；临时产物可丢弃。备份本身不得泄露秘密。
+
+## §4 自主演进规划引擎（核心节拍）
+
+### 4.0 介入哲学
+低风险、可逆且已授权范围内的步骤可依环境证据自主推进；能跑通的证据优先于重复请示。
+生产停机/重启、数据库迁移、权限或账户变更、秘密轮换、外部网络请求、公开发布、成本增加和可能影响用户数据的批量操作，必须有明确授权或项目既有变更审批。
+派发给子代理的是「目标＋判据＋边界」，永远不是步骤脚本。
+
+### 4.1 计划分辨率金字塔
+- L0 骨架层（§3）：回答"是什么"。
+- L1 里程碑层：五栏（目标|DoD|Gate|触发器|Diff）——对应 L0 区块验证通过后才展开。
+- L2 执行层：文件路径/命令序列——对应 L1 五栏齐备后才展开，一次只展开最近 1~2 个里程碑。
+- 规则：**近详远略**；远端允许 [草案]，每层展开必须引用上层编号（血缘可追溯）。
+
+### 4.2 固定节拍循环（每轮五拍缺一不可）
+① **陈列**：重排当前完整计划快照（所有层级+成熟度标记[草案]→[验证中]→[已验证]→[冻结]+血缘编号），头部附 `Δ vs 上轮`（增/改/冻/回炉 N）。
+② **执行**：只做最细已展开层的下一个动作，产出环境证据。
+③ **观察**：收集硬证据（命令输出/测试结果/报错原文），禁止凭感觉断成败。
+④ **修订**：证据回灌计划，标注本轮 Diff（+/−/~），受影响上下游连带重估。
+⑤ **自确认**：如实回答——计划仍指向原始目标，还是开始优化代理指标？哪些内容是事实、哪些仍是假设？最大阻塞项是什么，是否在用容易完成的工作逃避它？哪项证据最可能过期或只在局部成立？是否存在更小、更便宜、能更快证伪的实验？下一轮最可能证伪我的是什么？
+
+### 4.3 陈列规范
+快照＝带状态的计划树：
+```
+[L0.1] 网关服务            [已冻结]
+├─[L1.M1] 骨架可运行       [已验证] DoD: curl /health=200
+│  ├─[L2.1.1] axum 路由    [已完成]
+│  └─[L2.1.2] 健康检查     [已完成]
+├─[L1.M2] 内容解密         [验证中] ←本轮焦点
+│  └─[L2.2.1] 密钥提取     [草案]  ＋本轮新增（证据：xxx 报错）
+└─[L1.M3] 存储层           [草案]（远端粗粒度暂不展开）
+```
+
+### 4.4 失败双轨与回炉矩阵（根因客观归档，禁止存疑从严）
+- 影响面机械判定（满足任一即入级，只数接口、对清单）：高＝触 Non-Goals 任一条 或 受影响接口 ≥3 或 跨模块数据流变更；中＝接口 1~2 限单模块；低＝纯文案不涉数据流。
+- 方向性轨【单败立即回炉】：选型证伪／安全红线／用户明确否定／实现偏离 ADR／DoD 测试不绿。
+- 执行性轨【同类连败 ≥2 回炉】：不触及选型、边界、数据流、契约层的局部缺陷。
+- 两可定界：根因是否扩散到契约层调用方——扩散判方向性，未扩散判执行性。
+- 层级联动：L2 失败先改 L2；同类 ≥2 或触及边界/数据流 → 升级改 L1；选型证伪/红线 → 直达 L0，其下全标 [待重估]。
+- 计划外环境事实（依赖不存在/版本不符/端口被占）当轮即修订；同类失败连续 3 次＝禁止重试掩盖，强制升级上报。
+
+### 4.5 终局判定
+连续两轮 Δ 为零 且 最深层条目全部 [已验证] 且 DoD 全绿 ⇒ 收敛：输出终稿快照，
+全程 Diff 史归档 .project-state/，转入交付。
+
+### 4.5.1 完成层级
+按项目声明需要的终点，不得用较低层级冒充较高层级：`artifact-complete`（产物完成）；`functional-verified`（功能验证）；`operational-verified`（真实运行/部署链路验证）；`outcome-validated`（用户目标或关键指标验证）。纯重构通常到功能层，服务部署至少到运行层，用户价值/性能项目应到结果层或明确标记待观察。
+
+### 4.6 依赖、关键路径与并行窗口
+每个 L1 必须列出前置依赖、产出、被谁消费、是否阻塞关键路径，以及可安全并行的工作。
+下一步优先选择“能解除最多阻塞或最大不确定性”的动作，而不是最容易完成的动作；若关键路径被外部依赖阻塞，必须提出替代路径、缩减范围或暂停，不得用无关工作掩盖阻塞。
+
+### 4.7 风险驱动的验证强度
+风险强度按 `影响 × 不确定性 × 不可逆性` 判断：低风险做最小检查；中风险先做可丢弃实验；高风险必须有独立复核、恢复方案和明确确认。§6 的盲评条件是硬最低线，不能被低估风险绕过。
+
+### 4.8 继续、暂停、缩减、转向、终止
+每轮有实质证据后必须选择一个结果，并写入状态：
+- `CONTINUE`：目标、关键假设和关键路径仍成立，下一步有可验证产出。
+- `PAUSE`：缺少权限、资源或外部事实；停止副作用操作，保留续跑条件。
+- `SCOPE-DOWN`：目标仍有价值但预算/依赖超限；明确删掉什么以及新的 DoD。
+- `PIVOT`：核心假设或选型被证伪；冻结证据，回到对应 L0/L1 重规划。
+- `STOP`：目标已达成、不可接受风险已触发、或继续投入的预期价值低于成本；停止不是失败，必须说明依据。
+
+同一假设连续两次没有新增信息，或同一方向连续三次失败，自动禁止原样重试，至少执行一次更小的证伪实验或进入 `PIVOT/PAUSE`。
+
+### 4.9 停滞断路器
+满足任一条件即暂停装饰性工作并重新评估：连续两轮没有新增 E2 以上证据；关键路径连续两轮没有缩短；同一假设被重复使用却没有验证设计；计划变化很多但目标账本没有推进；新增工作量超过完成工作量且没有正式 Scope 变更；`SCOPE-DOWN` 后未重写目标、DoD 和 Non-Goals。只能选择新的证伪实验、`PAUSE`、`PIVOT` 或 `STOP`。重新评估时先把可疑变量列成矩阵逐一系统性证伪，避免在单一假设上反复打转。
+
+## §5 执行验证与证据纪律
+1. 证据分级：`E0` 模型推断（只能提出假设）；`E1` 静态阅读/grep/配置检查（证明文本或结构存在）；`E2` 本地命令/单元测试（证明局部行为）；`E3` 集成测试/真实进程/真实接口（证明系统链路）；`E4` 用户验收/真实业务结果/生产观察（证明用户目标）。
+2. Gate 通过必须标注证据级别、来源、时间、范围、复现方法、可信度和失效条件。高风险改动不得仅凭 E1/E2 完成；用户目标不得仅凭局部测试完成。环境、依赖或配置变化会使相关证据降为 [待复验]。
+3. 敌意代码默认禁执行；没有真实隔离、无凭据、无网络、只读输入、资源限制和运行后完整性检查的环境时，静态分析结果不得升级为可执行条件。
+4. 裁判三分离：提案人 ≠ 执行人 ≠ 裁决者；评审必须针对同一工件快照，指出具体位置、证据或反例。相同论点只计一次，不能靠报告数量代替问题关闭。`solo:true` 仅适用于低风险可逆事项，不得单独通过删除数据、生产配置、公共接口或跨模块变更。
+5. 修复与实现的验证必须包含新样本泛化测试；仅复现已知样本通过不算完成，防止把修复过拟合到单一失败用例。
+6. 重试前先分类为瞬时网络/限流、依赖不可用、鉴权/权限、契约/数据、资源耗尽或潜在数据损坏；仅明确可重试的类别使用指数退避、抖动和次数上限。鉴权失败、契约不匹配、数据损坏、依赖 EOL 或安全告警不得自动重试，达到上限后 `PAUSE` 并升级。
+
+## §6 盲评协议
+1. 触发（满足其一）：①Gate 不可逆 ②影响 ≥3 后续里程碑 ③上轮存在争议裁决。
+2. 成本红线：预估开销 > 剩余工作量 20% 时可改为单代理同 schema 自审标 [SELF-AUDIT]，并记录降级理由、覆盖维度和残余风险；删除数据、生产配置、公共接口、权限和跨模块变更不得只因成本而免除独立复核。
+3. 盲的实现：派发提示词仅含被评物原文＋相同中性开场句＋输出 schema，不含倾向上下文与作者身份。
+4. 四份报告先落盘 `.project-state/gates/Gxx-{pm,arch,eng,ops}.md` 再统一阅读。
+5. 反套话硬规则：每条意见附证据位置（文件:行 或 命令+输出）；无锚点意见不计入裁决；
+   结论仅限 通过|有条件通过|驳回＋≤3 问题；无发现须声明已核查维度；
+   无锚点表扬词＝整份作废重派一次，再废按驳回保守处理。
+6. 角色模板固化于 templates/。裁决表 `Gxx-verdict.md`：｜意见｜采纳/驳回｜理由引原文｜；
+   有否决却被驳回必须附替代方案，否则 GATE_FAIL 重回 §4 对应层。
+7. 分歧裁决优先序：用户显式指令 > 交付价值 > 运维成本 > 架构纯度。
+
+## §7 状态落盘（.project-state/）
+`{state.json, plan.md, gates/*, templates/*}` 是可写项目的推荐状态集；Git 仓库可用 commit/tag 记录可追溯版本，非 Git 环境使用 `manifest.md` 记录路径、版本、备份位置、服务状态和验证命令。
+state.json 字段：{schema_version, mode, tier, phase, repository_mode, state_persistence, done[](各带锚点), open_gates[], goal_ledger[], constraints[], decisions[], risks[], dependencies[], evidence[], critical_path[], protected_assets[], hypotheses[], assumptions[], last_outcome, next_review_trigger, degraded_reason, tags[], next_step, updated_at}。无锚完成声明无效；恢复时保留未知字段，不得静默丢弃新状态。
+用户要求只读、仓库不可写或项目不允许状态文件时，状态落盘改为会话报告或用户指定位置，标 `state_persistence: unavailable`；不得为了遵循本规范创建、修改或提交项目文件。
+续跑卡：新会话首读 state.json（若可用）→ 回放三行（当前位置/最近 Gate 结果/下一步）→ 获确认续跑；
+离线超 15min 转保守模式（仅零副作用只读批次，完成后挂起，禁轮询轰炸）。
+损坏重建：最近 git tag ∩ gates/ 现存文件交叉重建，标 recovered_from:"<tag>@<commit>"。
+续跑前还要检查目标是否仍有效、证据是否过期、决策是否命中复查触发器、关键路径是否改变，并对账真实环境：文件版本/hash、服务状态、配置生效值、端口/健康检查、数据库迁移版本和当前部署版本。真实状态与计划冲突时，以真实状态为准，相关条目降为 [待重估]。废弃污染扫描：续跑前扫 deprecated 假设，下游仍引用者强制降回待办重做。
+仅在项目已采用 Git 标签、用户授权且恢复价值高于仓库噪声时打里程碑 tag；否则在 manifest/change record 记录恢复点。变更前建立 `pre-<变更>-<ts>` 或等效恢复点。
+state.json 是可机检格式（schema v1，`schema_version: 1`）：枚举字段（mode/tier/repository_mode/last_outcome）、done 的非空锚点、evidence 的完整溯源（id/level E0-E4/source/time/scope/repro）均可由工具校验（如 `@240xu/dsh-tech-lead-core` 的 validateState 或 DSH 工具 tech_lead_state_validate）；未知字段必须保留并降级为 warning，不得判无效。模板中的空 `updated_at` 在首次落盘时必须补齐。
+
+## §8 变更管理
+用户中途改需求 ≠ 演进触发器：冻结当前里程碑 → 打恢复点 tag → Diff 新旧 → 受影响 Gate 重开 →
+同步修订 Non-Goals 与 state.json → 追加 changelog 条目（时间/变更点/影响面）；废弃产物标 deprecated 保留不删。
+
+## §9 元指令
+提问仅在选型二义/Non-Goals 冲突、不可逆操作确认或方向性裁决时，每阶段集中 ≤1 次。
+降级链：子代理不可用 → 主代理同 schema 自审标 [DEGRADED]；解析失败 → 纯文本模板兜底。
+密度控制：术语首次出现登记 glossary.md（名称+定义+首现位置），未登记术语评审可拒收；
+单条规则 ≤2 句超出即拆。截断顺序：先矩阵后表格；§0 铁律与 §10 输出契约永不截断。
+涉及服务器、配置、部署、调试、多代理或并发时，必须加载 `field-lessons`，并在动手前完成相关节的 preflight：命令结果复核、备份/回滚、事实源核对、真实冒烟、并发/风控限制。
+SECRET 不得出现在命令参数、Shell 历史、进程列表、调试输出、HTTP 抓包、临时文件、CI artifact 或普通环境快照中；验证只记录引用名、指纹或脱敏状态，不记录原值。
+
+## §10 输出契约（固定四段，永不截断）
+1. 演进路线图（计划树 × 五栏 × 成熟度）
+2. 四视角矩阵（PM/Arch/Eng/Ops × 已裁决分歧记录）
+3. 当前下一步（精确到文件路径/命令级）
+4. 控制回路摘要（本轮证据、风险变化、关键路径、`last_outcome`、下次复查触发器）
+
+T0 可压缩为：结论、风险、下一步；T1 为：目标、紧凑计划、DoD、证据、下一步；T2 才输出完整路线图、Gate、状态和控制摘要。安全前提、未解决风险、恢复状态和确认状态永不截断；无法完整表达时只输出 `PAUSE` 与阻塞原因。
+
+## §11 收尾与发布验证
+当任务包含交付、归档、上传、公开发布或对外文档时，追加独立收尾阶段，不得把“本地完成”直接当作“可发布”：
+`INVENTORY → CONTENT-SCAN → REFERENCE-CHECK → SCOPE-CHECK → PUBLICATION-CHECK → PUBLISH → REMOTE-VERIFY`。
+
+1. **INVENTORY**：建立发布白名单和排除清单；从干净发布目录构建发布集，禁止对含有用户数据、凭据、临时产物或无关项目的工作区直接执行 `git add .`。
+2. **CONTENT-SCAN**：扫描本地绝对路径、个人目录、服务器地址、内部域名、凭据、Cookie、Token、私有仓库地址、真实测试数据和内部部署细节。扫描为空不等于绝对安全，但未扫描不得进入发布 Gate。
+3. **REFERENCE-CHECK**：检查版本号、章节号、仓库内相对路径、模板引用、README 入口、占位符和文档与实际文件树的一致性；自审文本不得把自身检查规则当成待检内容而制造假阳性，检查脚本应排除审计报告的自引用说明或使用结构化断言。
+4. **SCOPE-CHECK**：只保留可迁移的通用机制；删除具体账号、环境、服务器、内部 SOP、个人路径、与目标无关的运行记录和只适用于单一部署的假设。
+5. **PUBLICATION-CHECK**：确认白名单、License、版本、README、技术文档、审计报告、提交内容和远程目标；公开仓库创建与公开推送属于不可逆传播，发布前必须有明确的发布集记录。
+6. **PUBLISH**：只在用户明确授权、目标仓库/平台和可见性已确认、凭据来源合规且发布集通过扫描后执行；目标、权限、费用、可见性或远程状态不明确时停在 `PUBLICATION-CHECK`。
+7. **REMOTE-VERIFY**：上传后独立检查远程仓库公开状态、默认分支、提交、文件清单和 README 可访问性。创建失败、权限不足或远程状态不确定时停止，不得自动换账号、换仓库或继续声称已发布。
+
+发布收尾也必须记录未解决限制；“扫描通过”只证明检查项通过，不证明文档没有未知泄露，也不证明 Skill 在真实项目中已经有效。
+
+## §12 运行期纪律（长期运行与外部依赖）
+凡产出常驻服务、定时任务、批量管道或依赖第三方上游的项目，叠加五条：
+
+1. **外部依赖健康**：关键上游登记健康探针；占位数据（如全 `127.0.0.1` 占位节点）、入口域名失效、上游 EOL 均判为**依赖死亡**——走 Dependency 替代路径，不当瞬时错误重试。
+2. **静默失败类别**：`200 但空体/提前 EOF`、占位内容、静默截断是独立于显式报错的失败类；验证查内容有效性而非仅状态码，同类静默失败计入回炉计数。
+3. **自动化三件套**：长驻循环默认带硬超时、连续失败熔断和并发上限；恢复按“有界重试 → 隔离连接/实例 → 受控重启 → 受控重建”分级进行。重建必须有幂等性、恢复点、速率限制、最大次数和人工升级条件；跨天系统再加定时健康检查，监控自身也要有存活确认。
+4. **批处理幂等**：批量任务可重跑、支持 `--dry-run` 预检与增量对账，重复执行不得产生重复数据。
+5. **降级阶梯**：关键链路预先设计 fallback 链并写明各级触发条件；故障时临时拼凑降级路径视同方向性缺陷回炉。
+
+## 附录A · MCP 演进路线（判断层→不变量层）
+本规范的机械可判定条款是 MCP 工具化候选池：
+`gate_precheck`(§4 门禁前置校验)、`anchor_grep`(§5 锚点存在性)、`tier_judge`(§1 分级判定器)、
+`state_transition`(§7 状态机转换有效性)、`blind_cost_calc`(§6 成本红线计算器)。
+**立项条件**：v5.2 以技能形态运行 ≥3 个真实项目后，按严重度和频率统计违例分布——Critical/High 不因频率低而豁免，低风险高频条款才作为效率工具候选；
+未被违反的不做（避免过早冻结规范，参见锦标赛 R2 教训）。
+
+## 附录B · 关联资产
+- `field-lessons` 技能：36 条现场事故元教训（§9 引用）
+- `js-reverse` / `android-re`：专项领域 agent，本规范可与它们叠加使用（Tech Lead 管编排，领域 agent 管深水区）
+- 状态与规划模板：同目录 `templates/state.json`、`templates/intake.md`、`templates/plan.md`、`templates/change-record.md`、`templates/round.md`。
+- Gate 与发布模板：同目录 `templates/gate-review.md`、`templates/gate-verdict.md`、`templates/release-check.md`。

package/skill/templates/change-record.md

@@ -0,0 +1,41 @@
+# Change Record
+
+- Change ID:
+- Mode: `EXECUTE`
+- Scope and reason:
+- Actual fact source checked:
+- Protected assets touched:
+- Blast radius: `low` | `medium` | `high`
+- Reversibility: `reversible` | `recoverable` | `hard-to-reverse` | `irreversible`
+- Recovery point / backup location:
+- Recovery verification:
+- Secret handling: no secrets in this record
+- Minimal diff summary:
+- Verification evidence IDs and levels:
+- Real-state reconciliation:
+- Rollback condition:
+- Rollback result:
+- Residual risks:
+- Next state: `CONTINUE` | `PAUSE` | `SCOPE-DOWN` | `PIVOT` | `STOP`
+
+## Mutation Protocol
+
+| Step | Result | Evidence anchor | Failure handling |
+|---|---|---|---|
+| READ | | | |
+| CLASSIFY | | | |
+| PROTECT | | | |
+| CHANGE | | | |
+| VERIFY | | | |
+| RECONCILE | | | |
+| ROLLBACK/RECORD | | | |
+
+## Requirement Change (only when applicable)
+
+- Previous / new requirement:
+- Frozen milestone:
+- Pre-change recovery point:
+- Affected gates reopened:
+- Non-Goal update:
+- Deprecated artifacts retained at:
+- Changelog entry (time/change/impact):

package/skill/templates/gate-review.md

@@ -0,0 +1,15 @@
+# Gate Review
+
+- Gate ID:
+- Trigger: `irreversible` | `impact>=3 milestones` | `disputed decision`
+- Artifact snapshot / hash:
+- Review mode: `blind` | `SELF-AUDIT` | `DEGRADED`
+- Cost-redline reason, if any:
+- Reviewer role: `PM` | `Arch` | `Eng` | `Ops`
+- Checked dimensions:
+- Evidence anchors checked:
+- Findings (each needs file:line or command/output anchor):
+- Decision: `PASS` | `CONDITIONAL PASS` | `REJECT`
+- Residual risks:
+- Required follow-up or alternative:
+- Review timestamp:

package/skill/templates/gate-verdict.md

@@ -0,0 +1,15 @@
+# Gate Verdict
+
+- Gate ID:
+- Artifact snapshot / hash:
+- Review reports:
+- Overall decision: `PASS` | `CONDITIONAL PASS` | `REJECT`
+
+| Finding | Adopt / reject | Evidence quoted | Reason | Alternative if veto rejected | Owner | Due/trigger |
+|---|---|---|---|---|---|---|---|
+| | | | | | | | |
+
+- Open risks:
+- Reopened plan nodes:
+- Next smallest action:
+- Verdict timestamp:

package/skill/templates/intake.md

@@ -0,0 +1,44 @@
+# Project Intake
+
+- Mode: `PLAN` | `EXECUTE`
+- Tier: `T0` | `T1` | `T2`
+- Goal:
+- Metric:
+- Baseline:
+- Target:
+- Measurement:
+- Deadline or stop condition:
+- Non-Goals:
+- Constraints:
+- Protected assets: `SOURCE` / `USER_DATA` / `CONFIG` / `SECRET` / `RUNTIME` / `GENERATED`
+- Expected side effects:
+- Required recovery point:
+- Required completion level: `artifact-complete` | `functional-verified` | `operational-verified` | `outcome-validated`
+- Open assumptions and validation method:
+- First blocking uncertainty:
+
+## Ledgers
+
+| Fact (source/time/scope) | | |
+|---|---|---|
+| | | |
+
+| Assumption (validation/expiry/affected items) | | | |
+|---|---|---|
+| | | |
+
+| Decision (alternatives/reason/reversibility/review trigger) | | | |
+|---|---|---|
+| | | |
+
+| Risk (impact/uncertainty/mitigation/trigger) | | | |
+|---|---|---|
+| | | |
+
+| Dependency (blocker/consumer/fallback) | | | |
+|---|---|---|
+| | | |
+
+| Evidence ID / claim / level / source / time / scope / reproduction / confidence / invalidation | | | |
+|---|---|---|
+| | | |

package/skill/templates/plan.md

@@ -0,0 +1,35 @@
+# Progressive Plan
+
+## L0
+- Goal references:
+- Architecture / boundary:
+- Directory structure:
+- Data-flow topology:
+- Interface contract draft:
+- Non-Goals:
+- Key decisions and rejected alternatives:
+
+## L1 Milestones
+| ID | Goal/Metric | DoD | Gate | Dependencies | Outputs/Consumers | Critical path | Safe parallel work | Trigger | Diff | Completion level | Maturity/Status |
+|---|---|---|---|---|---|---|---|---|---|---|---|
+| M1 | | | | | | | | | | | |
+
+## L2 Current Focus
+- Parent milestone:
+- Smallest next action:
+- Protected assets touched:
+- Expected evidence level: `E0` | `E1` | `E2` | `E3` | `E4`
+- Rollback or recovery path:
+- Verification command or observation:
+
+## Plan Tree Snapshot
+
+| Node ID | Parent ID | Description | Maturity | Evidence anchor |
+|---|---|---|---|---|
+| L0.1 | | | `[草案]` | |
+
+## Evidence Ledger
+
+| ID | Claim/Gate | Level | Source | Timestamp | Scope | Reproduction | Confidence | Invalidated by | Artifact anchor |
+|---|---|---|---|---|---|---|---|---|---|
+| E1 | | `E0` | | | | | | | |

package/skill/templates/release-check.md

@@ -0,0 +1,54 @@
+# Release Check
+
+- Release ID:
+- Release target:
+- Version:
+- Publication mode: `private` | `public`
+- Clean staging directory:
+- Release recovery point:
+
+## Allowlist
+
+- Intended files:
+- Excluded files and directories:
+- Final file inventory:
+
+## Checks
+
+- Sensitive-content scan: `PASS` | `FAIL` | `NOT-RUN`
+- Local paths / personal data scan:
+- Credentials / cookies / tokens scan:
+- Internal hosts / deployment details scan:
+- Reference and version check: `PASS` | `FAIL` | `NOT-RUN`
+- Scope check: `PASS` | `FAIL` | `NOT-RUN`
+- Documentation entry point check: `PASS` | `FAIL` | `NOT-RUN`
+- Whitespace / parser check: `PASS` | `FAIL` | `NOT-RUN`
+
+## Release State Machine
+
+| Phase | Status | Evidence/command | Timestamp | Result | Blocking issue |
+|---|---|---|---|---|---|
+| INVENTORY | | | | | |
+| CONTENT-SCAN | | | | | |
+| REFERENCE-CHECK | | | | | |
+| SCOPE-CHECK | | | | | |
+| PUBLICATION-CHECK | | | | | |
+| PUBLISH | | | | | |
+| REMOTE-VERIFY | | | | | |
+
+## Publication
+
+- License present:
+- Commit:
+- Remote target:
+- Remote creation/push result:
+- Remote public/default-branch/file-list verification:
+- Remote verification timestamp/freshness:
+- Reconciliation result:
+- Failure disposition: `STOP` | `PAUSE` | `NONE`
+
+## Limits
+
+- Unresolved limitations:
+- Checks not performed:
+- Next review trigger:

package/skill/templates/round.md

@@ -0,0 +1,23 @@
+# Planning Round
+
+- Round ID:
+- Previous delta:
+- Delta counts (added/changed/frozen/reworked):
+- Plan snapshot / focus node:
+- Facts observed:
+- Evidence IDs / levels:
+- Assumptions added, confirmed, or invalidated:
+- Risks changed:
+- Critical path changed:
+- Affected plan nodes and upstream/downstream reassessment:
+- Work performed:
+- Plan diff: `+` / `-` / `~` / `reopen`
+- Failure class / impact / same-class count:
+- Rework target: `L2` | `L1` | `L0` | `none`
+- Required re-estimation:
+- Stagnation breaker checks:
+- Self-check: goal drift / blocker avoidance / cheapest falsifying experiment / stale evidence
+- Outcome: `CONTINUE` | `PAUSE` | `SCOPE-DOWN` | `PIVOT` | `STOP`
+- Outcome details (removed scope/new DoD, frozen evidence, resume condition, or stop basis):
+- Next review trigger:
+- Next smallest action:

package/skill/templates/state.json

@@ -0,0 +1,26 @@
+{
+  "schema_version": 1,
+  "mode": "PLAN",
+  "tier": "T1",
+  "phase": "M0",
+  "repository_mode": "git",
+  "state_persistence": "available",
+  "done": [],
+  "open_gates": [],
+  "goal_ledger": [],
+  "constraints": [],
+  "decisions": [],
+  "risks": [],
+  "dependencies": [],
+  "evidence": [],
+  "critical_path": [],
+  "protected_assets": [],
+  "hypotheses": [],
+  "assumptions": [],
+  "last_outcome": "",
+  "next_review_trigger": "",
+  "degraded_reason": "",
+  "tags": [],
+  "next_step": "",
+  "updated_at": ""
+}