360 Secure SkillHub
360 安全技能中心

Concept Guardrails

该技能需通过`$concept-guardrails`显式调用,提供audit、concept、drift等6种模式,核心功能包括创建/维护概念规范文档、检测规范与实现的漂移、生成审计报告及派生地图等。典型场景覆盖项目概念规范搭建、现有规范审计、代码与规范一致性校验等。输入为带模式及可选路径/描述的指令,输出对应模式的文档或报告,操作遵循只读审计、授权修改等约束。

agenticweb4
agenticweb4
热度 26.08
最近更新 2026-09-15 10:07:13
分类:知识依赖:读写文件 / 代码执行应用场景:概念解释 / 报告撰写

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add concept-guardrails
Metadata
name
concept-guardrails
description
Runs concept-spec guardrails in audit, concept, drift, pipeline, sync, or map mode only when the user explicitly invokes $concept-guardrails. Use to create, backfill, compare, or map colocated concept specifications.
compatibility
File read/write and Grep/Glob; optional boundary-injection runtime requires Claude Code hooks and jq
metadata
{"language":"en","translation_of":"concept-guardrails-cn","author":"ontology-of-everything","version":"0.32.0","openclaw":{"homepage":"https://github.com/ontology-of-everything/concept-skills/tree/main/skills/en/concept-guardrails"}}

Concept Guardrails

Focus: do existing specifications state module contracts clearly and correspond to facts and derived maps? Run only on explicit $concept-guardrails invocation. The workflow adapts wyx v0.26.0 under MIT (LICENSE.upstream); specifications use this repository's Jackson dialect while runtime scripts retain upstream behavior.

Authority Model

  • CONCEPT.md defines purpose, state, actions, and operational principle; optional notes hold application context.
  • SYNCS.md records app/include/sync coordination for application purposes and scenarios; when/where/then define rules and // flow: groups them.
  • PIPELINE.md records sources, stages, outputs, invariants, triggers, and data boundaries without replacing synchronizations.
  • ARCHITECTURE.md is a derived navigation map. Product dependencies come from the overall PRD, not synchronization edges.
  • Backfill reconstructs evidenced contracts from implementation; drift compares implementation with contract. Neither proves fitness for purpose.

Specification coverage does not prove design correctness, and an accurate map does not prove code conformance. Label inferred purposes; never use an ideal OP to justify an observed defect.

Modes

Use $concept-guardrails <mode> [path-or-description]. A path means backfill, a description means create, and no argument means discover candidates. Start with audit when the mode is absent or the project has no specifications.

Mode Output and required reference
audit Read-only coverage and command plan: references/audit.md
concept Concept candidates or CONCEPT: references/concept.md
drift Read-only drift report: references/drift-detection.md
pipeline Pipeline candidates or PIPELINE: references/pipeline.md
sync Coordination candidates or SYNCS: references/sync.md
map Derived maps: references/map.md

Route complex modeling to concept-design, confirmed-model transcription to concept-prd, and fitness/independence/composition review to concept-audit.

Constraints

  • Read specifications, implementation, and project conventions first. Mark current defects during backfill; update specifications before implementation for confirmed new behavior.
  • Present a draft or diff before unapproved writes. Complete already-authorized creation or repair; ask only for unresolved model choices or expanded scope. Discovery lists candidates; read-only audits do not write history or fixes.
  • Keep one authoritative specification per concept, its PIPELINE beside it, cross-concept pipelines in the composition layer, and one SYNCS per synchronization package grouped by flow. Rebuild derived maps after specification changes when authorized.
  • Before creating or reading CONCEPT/SYNCS or checking pipeline interfaces, read references/spec-format.md. Use independent sections, named inputs/outputs, result-case matching, and query bindings. Write operational principle; read legacy principle and signatures compatibly.
  • Mark legacy wyx interactions, dependencies, known coupling, dispatching, coordination graphs, and sync: sections for migration rather than mixing dialects. When authorized, rewrite CONCEPT/SYNCS/PIPELINE together while preserving triggers, bindings, effects, and failure semantics.

Completion means the selected mode's definitions, relations, and references are evidenced; backfill distinguishes facts, inference, and open decisions; reports name scan scope and repair direction; maps are rebuilt from content rather than timestamps. For automatic injection read references/hooks-runtime.md: hooks list specifications, flag staleness, and inject PIPELINE data boundaries but do not validate semantic seams.