qingye-lab/hengmu

Specialized architecture audit for AI-agent systems and AI-enabled products. Use for agent runtimes, tool-using workflows, memory systems, RAG, MCP integrations, long-running tasks, model routing, evidence tracking, human approval, evaluations, or systems exposed to prompt injection and model uncertainty. Extends rather than replaces the general project architecture audit.

Independently verifies candidate architecture findings against current code and repository evidence. Use after any project, AI-agent, mobile, or portfolio architecture audit; when a finding may be based on weak heuristics; or before remediation and quality gates. Filters false positives, deduplicates findings, recalibrates severity and confidence, and preserves an auditable decision trail without changing code.

Deterministic architecture quality gate for verified review artifacts. Use in local checks or CI to validate architecture configuration, block configured confirmed severities, enforce review freshness, and honor explicit baselines or time-limited waivers. Never treats raw model findings as gate failures and never replaces finding verification.

Converts an accepted architecture decision into an executable, dependency-aware migration roadmap for confirmed remediation findings or an accepted open/constrained Greenfield target. Use for implementation slices, compatibility, data migration, deployment sequencing, observability, rollback, stop conditions, effort, and acceptance criteria. Produces a plan only; it does not choose architecture or implement changes.

Compare, design, specify, or constrain an architecture for a verified remediation Review or an approved Design Brief. Use for open or constrained target architecture, technology and pattern selection, ADR-like decisions, trade-off analysis, and bounded technology-evolution assessment. Required, preferred, and prohibited constraints are challenged inputs; they are not proof. Produce a decision and target architecture, never an implementation plan or code change.

Unified public entry point for Hengmu. Use when the user invokes $hengmu, cannot remember a focused Hengmu Skill name, asks what Hengmu can do, wants a read-only lifecycle or status explanation, or wants Hengmu to route a natural-language request for repository, AI-agent, mobile, or portfolio audit, finding verification, open or constrained design/specification, emerging technology-evolution assessment, solution comparison, remediation planning, or a deterministic quality gate. Chinese design, specification, and constraint requests route the same way.

Specialized architecture audit for mobile applications, especially iOS and SwiftUI. Use when assessing local persistence, offline behavior, synchronization and conflicts, migrations, background execution, notifications, networking and caches, state ownership, privacy permissions, battery efficiency, or excessive client-side business logic. Extends rather than replaces the general project architecture audit.

System-of-systems architecture setup and audit across multiple projects. Use when initializing .architecture-portfolio or reviewing duplicated capabilities, project boundary conflicts, technology-stack sprawl, shared infrastructure, cross-project data flows, hidden coupling, deployment and domain complexity, reusable components, or a one-year technical governance roadmap. Requires an explicit portfolio registry and does not replace per-project audits.

Evidence-backed, persistent architecture audit and profile setup for one repository. Use when initializing .architecture, onboarding to a codebase, reviewing architecture health, preparing a refactor, investigating structural debt, or assessing module boundaries, ownership, contracts, resilience, security, observability, tests, deployment, and over-design. Automatically initializes missing project governance unless the user explicitly requests read-only work. Produces candidate findings for independent verification, not confirmed conclusions, fixes, or remediation plans.