A360

a360 — architecture plugin for a360ai

A Claude Code plugin that teaches Claude how the a360ai monorepo (avvocato360ai — a legal practice management platform) is structured, and enforces its architecture so generated code follows the conventions instead of drifting.

This plugin is exclusive to a360ai. It installs at user scope but stays completely inert in any other repository (see Scoping below). It is a sibling of the generic nja plugin, but a360ai uses its own in-repo base classes (not the @carlonicora/* libraries) and keeps its architecture docs in the repo — so the rules and skills differ.

Skills

Skill	Use when
a360-architecture	Before editing/creating any TS file under apps/api/src/features, apps/web/src/features, or packages/shared/src. Routing table → the layer reference doc (entity, serialiser, DTO, repository, service, controller, migration, model, interface, frontend service). The authority.
a360-writing-plans	Writing an implementation plan. Wraps superpowers:writing-plans with a360ai compliance: inline citations to references/*, a self-audit, skill-wins-over-plan dispatch, and an audit step in the final verification task.
a360-verify	Auditing uncommitted changes against the architecture rules — before committing, before handing back, or after implementing a feature. Read-only: reports violations with evidence, does not fix.
a360-handoff	Compact the current conversation into a handoff document for a fresh agent.
a360-merge-worktree	Validate (lint/build/test) and merge an a360ai git worktree back into dev, then clean up.

The architecture reference docs (mirrored from a360ai's docs/architecture/) are bundled under a360-architecture/references/, so the skill is self-contained.

Enforcement (hooks + linter)

Three layers, all jq-dependent, all harness-side (zero model-context cost), all inert outside a360ai:

1. PreToolUse reminder (hooks/remind-architecture.sh) — before each Edit/Write/MultiEdit, if the file matches the routing table it points at the exact reference doc. Soft; never blocks.
2. a360-lint (scripts/a360-lint.sh) — deterministic checker for the greppable anti-patterns in references/anti-patterns.md: fetch() in the frontend, raw result.records, manual SKIP/LIMIT, hardcoded node labels / endpoint strings, hand-written Cypher with no security filter (cypherService.default()/securityService.userHasAccess()), .nullable()/ z.union()/z.discriminatedUnion() in LLM output schemas, controllers importing repositories, @IsString() on date DTOs, and more. Two tiers (BLOCKING / WARN). Suppress a false positive with a comment containing a360-lint-ignore.
3. Stop gate (hooks/architecture-gate.sh) — runs a360-lint over the uncommitted diff when Claude tries to finish; if any BLOCKING violation remains, it blocks completion and feeds the list back. Grep-fast — no build/test.

Scoping. A repo counts as a360ai only if it has docs/architecture/AI-QUICK-REFERENCE.md or its root package.json is named avvocato360ai. Everywhere else, every hook no-ops. (a360-lint run manually always checks whatever you pass it.)

Install

/plugin marketplace add carlonicora/a360
/plugin install a360@a360

Layout

a360/
├── .claude-plugin/marketplace.json
└── a360/
    ├── .claude-plugin/plugin.json
    ├── hooks/
    │   ├── hooks.json                 # PreToolUse + Stop
    │   ├── remind-architecture.sh
    │   └── architecture-gate.sh
    ├── scripts/
    │   ├── a360-lint.sh               # deterministic anti-pattern checker
    │   └── a360-detect.sh             # "is this a360ai?" guard
    └── skills/
        ├── a360-architecture/         # routing table + bundled references/
        ├── a360-writing-plans/
        ├── a360-verify/
        ├── a360-handoff/
        └── a360-merge-worktree/

License

MIT