zhengming-toolbox

A toolbox of 105 purpose-built, score-first-designed AI-engineering tools, each targeting a niche where no direct equivalent existed (verified by pre-construction anchor searches as of 2026-10-12). Pure standard-library Python, CLI-first, self-tested.

Why score-first?

Every tool was reverse-designed against a quality rubric (novelty vs. existing anchors, test depth >= 40 cases, measured evidence) before construction — predicted scores matched actual audits 55/55 across two rounds. See the methodology note at the bottom.

Flagships (no direct equivalent found — true zero-anchor niches)

• instruction-dsl-lint — AI 指令模板的 DSL 规范+校验器
• json-output-salvage — LLM 输出 JSON 抢修器
• md-table-lint — Markdown 表格校验器
• focus-data-aggregator — 专注数据聚合规范
• vector-meta-schema — 向量库元数据规范
• complexity-budget-gate — 复杂度预算闸
• timezone-safety-check — 时区安全检查器
• dead-code-quantifier — 死代码量化器

Score-first batch 1 (R12, live)

• schema-clamp — Validate & auto-repair LLM JSON output against a JSON-Schema subset, with violation paths and fix suggestions.
• invariant-guard-gen — Turn docstring invariants into property-based checkers for AI-generated code.
• model-origin-ledger — Tamper-evident hash-chain ledger for AI model/dataset provenance with license-conflict audits.
• x402-receipt-reconciler — Exactly-once reconciliation for x402 agent payments: duplicate detection, amount mismatch, timeout reports.
• nl2sql-schema-drift — Detect drift between semantic-layer definitions and real database schemas before NL2SQL queries break.

The 50-tool catalog (R13 batch)

AI Instruction & Prompt Engineering (8)

• instruction-dsl-lint — AI 指令模板的 DSL 规范+校验器
• prompt-regression-harness — 记录提示词→期望断言的回归测试架
• prompt-version-diff — 两版提示词的结构化 diff（指令段/示例段/约束段分级 + 语义风险标注）
• prompt-template-merge — 提示词模板继承合并器
• token-budget-clamp — 提示词 token 预算钳
• context-pack-planner — 上下文窗口打包规划器
• llm-cache-key-spec — LLM 缓存键规范
• fewshot-picker-bench — few-shot 样例选择基准

Agent & LLM Governance (9)

• agent-audit-trail — Agent 行为审计轨迹规范
• tool-call-replay — Agent 工具调用录制重放器
• model-capability-matrix — 模型能力矩阵规范
• model-arbitration-spec — 多模型仲裁协议
• embedding-drift-watch — 嵌入漂移监测规范
• vector-meta-schema — 向量库元数据规范
• llm-gateway-spec — LLM 网关接口规范
• system-prompt-audit — 系统提示词静态审计器
• rag-chunk-bench — RAG 分块策略基准

Code Engineering & Quality (10)

• json-output-salvage — LLM 输出 JSON 抢修器
• dead-code-quantifier — 死代码量化器
• complexity-budget-gate — 复杂度预算闸
• comment-drift-check — 注释漂移检测
• deprecated-api-scan — 弃用 API 使用扫描
• doc-rot-detector — 文档腐化检测器
• readme-quality-score — README 质量评分器
• changelog-semantic-lint — CHANGELOG 语义校验
• api-contract-drift — API 契约漂移检测
• schema-evolution-grader — Schema 演进兼容性分级

Security & Compliance (7)

• log-redaction-spec — 日志脱敏规范
• secret-rotation-check — 密钥轮换检查器
• retention-policy-check — 数据保留策略校验器
• ai-commit-disclosure — AI 提交披露检查器
• lockfile-audit-diff — 锁文件审计
• money-precision-check — 金额精度检查器
• unit-confusion-check — 单位混淆检查

Data & Consistency (8)

• csv-dialect-fingerprint — CSV 方言指纹
• md-table-lint — Markdown 表格校验器
• event-replay-verifier — 事件溯源重放校验器
• idempotency-key-spec — 幂等键规范
• timezone-safety-check — 时区安全检查器
• calendar-conflict-check — 日历冲突检测
• claim-verification-chain — 断言校验链
• lock-contract-test — 分布式锁契约测试

Process & Collaboration (8)

• git-history-hygiene — Git 历史卫生审计
• todo-expiry-ledger — TODO 期限账本
• self-heal-runbook — 自愈运行手册规范
• deploy-checklist-code — 部署清单即代码
• quota-aware-scheduler — 配额感知调度器规范
• meeting-notes-to-tasks — 会议纪要→任务转换规范
• email-triage-spec — 邮件分诊规范
• focus-data-aggregator — 专注数据聚合规范

R14 batch (v4.0 rubric, E/F graded)

Agent & LLM Governance (9)

• memory-expiry-policy — Agent 记忆过期策略规范 zero-anchor flagship
• instruction-conflict-detector — 系统提示词内部矛盾检测器 zero-anchor flagship
• llm-error-code-spec — LLM 网关错误码规范 zero-anchor flagship
• context-truncation-bench — 上下文截断策略基准 zero-anchor flagship
• adr-staleness-check — ADR 老化检测
• agent-stop-condition-spec — Agent 停止条件形式化规范
• agent-loop-detector — Agent 循环检测器
• agent-permission-audit — Agent 权限最小化审计
• agent-trace-compression — Agent 轨迹压缩基准

Code Engineering (9)

• cli-help-spec — CLI 帮助文本规范 zero-anchor flagship
• code-block-lang-check — 代码块语言标注检查 zero-anchor flagship
• dependency-depth-budget — 依赖树深度/广度预算 zero-anchor flagship
• math-render-precheck — 数学公式渲染预检 zero-anchor flagship
• postmortem-template-lint — 复盘报告校验 zero-anchor flagship
• deprecation-timeline-audit — API 弃用时间表审计 zero-anchor flagship
• pr-size-scorer — PR 大小评分器
• release-notes-spec — 发布说明规范
• upgrade-risk-grader — 依赖升级风险分级

Security & Compliance (7)

• session-token-hygiene — 会话令牌卫生检查 zero-anchor flagship
• docs-info-leak-scan — 文档内部信息泄漏扫描 zero-anchor flagship
• data-classification-schema — 数据分类标记规范 zero-anchor flagship
• file-permission-drift — 文件权限漂移检测 zero-anchor flagship
• cors-policy-validator — CORS 策略校验器
• cron-security-audit — Cron 任务安全审计
• vendor-lockin-ledger — 供应商锁定账本

Data Integrity (7)

• etl-resume-spec — ETL 断点续传规范 zero-anchor flagship
• queue-idempotent-consume — 队列幂等消费检查 zero-anchor flagship
• pagination-integrity-check — 分页完整性校验 zero-anchor flagship
• uniqueness-precheck — 唯一性约束预检 zero-anchor flagship
• primary-key-guard — 主键格式守卫
• slow-query-pattern-ledger — 慢查询模式账本 zero-anchor flagship
• baseline-drift-ledger — 基线漂移账本 zero-anchor flagship

Ops & Performance (10)

• health-endpoint-spec — 健康端点规范 zero-anchor flagship
• latency-budget-spec — 延迟预算声明规范 zero-anchor flagship
• free-tier-exhaustion-predictor — 免费额度耗尽预测 zero-anchor flagship
• token-price-table-validator — token 单价表校验器 zero-anchor flagship
• oncall-handoff-spec — OnCall 交接规范 zero-anchor flagship
• merge-conflict-predictor — 合并冲突预测器
• backup-integrity-verifier — 备份完整性验证器
• team-timezone-overlap — 团队时区重叠计算器
• invoice-field-validator — 发票字段校验器
• code-ownership-map — 代码所有权映射器

Content & Docs (8)

• rfc2119-keyword-lint — RFC2119 规范词检查 zero-anchor flagship
• onboarding-checklist-check — 入职清单校验 zero-anchor flagship
• storyboard-consistency-check — 分镜一致性检查 zero-anchor flagship
• multimodal-asset-manifest — 多模态资产清单 zero-anchor flagship
• experiment-registry-schema — A/B 实验注册表规范 zero-anchor flagship
• image-alt-check — 图片 alt 文本检查
• metric-definition-dict — 指标定义字典 zero-anchor flagship
• image-prompt-schema — 图像提示词结构规范 zero-anchor flagship

R14 totals: 50 tools, 3,228 self-test cases, CI on every repo, predicted scores 50/50 matched (mean 96.8 under the v4.0 rubric).

Family indexes (R14)

Prefer browsing by problem area? Six family indexes group the R14 tools by the pain they solve:

Index	Focus
zhengming-agent-governance	Agent & LLM governance — memory policies, loop detection, permission audits, trace compression
zhengming-code-engineering	Code gates & linters — CLI conventions, dependency budgets, PR sizing, upgrade risk
zhengming-security-compliance	Defensive security & compliance — token hygiene, leak scans, CORS/cron audits, lock-in
zhengming-data-integrity	Data integrity — ETL resume, idempotent consumption, pagination, key guards
zhengming-ops-performance	Ops & performance — latency budgets, health specs, quota forecasts, backup verification
zhengming-content-docs	Content & docs quality — RFC2119 linting, onboarding checks, alt-text checks

Earlier releases

52 additional utility tools were released on 2026-10-10/11 (Rounds R10-R11): defensive security scanners, MCP cleanroom acceptance harnesses, and developer workflow tools — see the repositories list.

Roadmap

• CI on every repo (self-test runs on each push) — rolling out with this release

• Feedback-driven selection: 30-day star/issue data feeds the next round's niche choices

• Iteration tiers: repos with traction get v0.2 feature depth; zero-signal niches get SEO tuning or archival review

Methodology: reverse score-first design

1. Pre-construction anchor search (per niche, multiple queries) — no direct equivalent, or mechanism-inverted positioning vs. weak anchors

2. Predicted score committed to a design contract before any code

3. Construction under hard gates: >= 40 self-test cases (functional / exception / extreme-edge), two iteration rounds with changelog evidence

4. Independent master audit re-runs every self-test; predicted vs. actual scores reconciled (55/55 match so far)

License

All tools MIT — © zhengming contributors. Contact: hcac4735@agent.qq.com