venhdev/zenforge

perimeter

Detect raw runtime API bypasses circumventing internal abstractions and codify enforceable boundary constraints.

查看源码

使用方法

  1. 复制安装命令,在自己的智能体环境安装前先检查源码包。
  2. 从原始文档核对所需工具、凭据和支持的宿主。
  3. 说明输入材料与期望交付物,并对照来源检查输出。
npx skills add https://github.com/venhdev/zenforge --skill perimeter

开始前需了解

  • 安装不包含模型订阅、API 额度或访问文件的权限。
  • 仓库星数和安装量反映活跃情况,不等于独立质量或安全测试。
仓库原始内容

按源仓库内容呈现,保留标题、案例、代码、表格、链接以及原文引用的演示图片。

perimeter — Architectural Boundary & Bypass Detection Engine

Domain Rubric

1. Boundary Audit Dimensions

  • Sanctioned Primitives vs. Raw Bypass (Horizontal): Identify internal wrappers governing I/O, network clients, resources, or lifecycles; flag call sites bypassing them for raw runtime APIs.
  • Layer & Dependency Directionality (Vertical): Trace import graphs across architectural layers (Domain, App, Infra, UI); flag inverted or cross-boundary imports.
  • Resource & Sandbox Hygiene: Audit storage paths, concurrency pools, and memory buffers; flag unanchored filesystem access or unthrottled resources.
  • Micro-Snippet:
text
  Banned:   fetch("https://api.domain/orders") // Raw runtime bypass
  Required: appHttpClient.get("/orders")       // Sanctioned primitive with auth & telemetry

2. The Load-Bearing Qualification Gate

Qualify candidate constraints strictly against 3 conditions:

  1. Existing Sanctioned Primitive: Codebase defines an internal wrapper, base class, or canonical path addressing the boundary.
  2. Empirical Bypass Evidence: At least one active call site violates the boundary by calling raw APIs directly.
  3. Quantifiable Failure Risk: Bypass induces architectural drift, resource leaks, memory exhaustion, or security exposure.
  • Negative Boundary: Confine constraints strictly to load-bearing architectural boundaries with existing internal primitives; never police cosmetic styling, standard utility methods (Math.*, collections), or speculative rules lacking active wrappers.

Canonical Output Contract

markdown
# Perimeter Report: <Target Subsystem / Scope>

## 1. Executive Verdict
- **Target Inspected**: `<directory_or_package>`
- **Verdict**: <HARDENED | PARTIAL | UNPROTECTED> (<N> active boundaries)
- **Primary Risk**: <Single-sentence technical judgment on highest boundary exposure>

## 2. Qualified Boundary Declarations
*(If HARDENED: 'All architectural boundaries are enforced by active test suites; zero raw bypasses detected.')*

### Boundary: <Constraint Name, e.g., Direct Network Client Bypass>
- **Offending Evidence**: [<file>#L<N>](file:///path/to/file#L<N>) ── <Bypass snippet or call site>
- **Failure Risk**: <Concrete risk: memory leak / auth bypass / architectural drift>
- **Rule**: Required: `<Sanctioned Primitive>` ── Banned: `<Forbidden Raw API>`
- **Enforcement**: Tier: `<Arch Unit Test | AST Linter | Script Guard>` | Whitelist: `<Exempt files | 'None'>`

## 3. Disqualified Invariants (Pruned) *(omit if no candidate patterns pruned)*
- **[PRUNED] <Candidate Pattern>**: <Reason for rejection, e.g., No existing codebase primitive / Trivial style issue>

Routing

  • Qualified boundaries need a concrete implementation plan → (via /changeset).
  • Boundary tests have an evidence-backed changeset and authorized execution scope → (via /forge).
  • Boundary ownership, whitelist exceptions, or architectural trade-offs require human deliberation → (via /clarify).

Execution Protocol

Phase 1: Macro Anchor Grounding

  1. Ingest target directory or subsystem.
  2. Scout architectural anchors (internal wrappers, sanctioned clients, resolvers, base interfaces) to ground boundary baselines.

Phase 2: Scale-Adaptive Boundary Audit

  • In-Turn Execution (< 10 files or localized module): Audit call sites against discovered anchors directly in-turn.
  • Subagent Fan-Out (broad codebase or cross-layer scopes): Launch concurrent research subagents dynamically scaled to partitioned slices or caller-declared budget (Role: Perimeter Reviewer (<Slice>)), seeded with target slice directory, discovered architectural anchors, Load-Bearing Qualification Gate, path to references/fitness-functions.md, and Canonical sub-report contract requiring offending code pointers (file:///...#L<N>).

Phase 3: Codification & Turn-Halt Gate (Fan-in)

  1. Reconcile findings into the Canonical Output Contract, moving ungrounded rules to Section 3 (Pruned; omit Section 3 if none).
  2. If boundary ownership or whitelist exemptions contain unresolved trade-offs: halt and deliberate (via /clarify).
  3. Emit the report into the conversation stream and halt turn immediately.
来自同一仓库

更多 Skills

全部 Skills