AMI Claude Mods
Suite of production-grade modular extensions, reactive terminal UI panels, and function hooks for Claude Code CLI integrated with the Amiga IA ecosystem, optimizing context consumption, safe exits, multi-profile guards, and Cleanroom TDD isolation.
The Problem
As autonomous coding agents tackle large-scale workflows in Claude Code CLI, developers encounter key operational hurdles: rapid context window depletion from repetitive mechanical file searches, orphaned background Docker containers or dangling processes upon exit, accidental commits across mismatched Git identities or mutations on production Kubernetes clusters, and zero visibility into active hook configurations.
The Solution
A suite of six decoupled, fully tested modular extensions (mods) and function hooks integrated with the Amiga IA ecosystem, operating within Claude Code's sandboxed ESM runtime and verified via 'claude-code/testing'. Features an interactive hook and settings inspector panel (hooks-viewer), a mechanical read circuit breaker enforcing subagent delegation (undelegated-reads), a pre-exit resource checklist (close-checklist), an inter-agent messaging inbox (child-sessions), Double-Blind Cleanroom isolation for bias-free TDD (cleanroom-guard), and configurable multi-profile Git/Kubernetes safeguards (context-guard).
Key Learnings
- Host-Gated Isolation over Node.js Runtime: Plugins operate in an isolated ESM container without arbitrary require() or direct fs access; all I/O flows through the EngineInterface ($) handle, guaranteeing sandboxed safety and uniform cross-platform behavior across Windows, macOS, and Linux.
- Top-Level Function Scope for Engine Reference ($): Claude Code's static plugin validator enforces that the $ handle cannot be passed to nested or anonymous functions within the register closure, requiring pure module top-level helper functions.
- Cross-Platform Asynchronous Environment Resolution: Avoiding silent runtime fallbacks when querying OS-specific environment variables (USERPROFILE vs HOME) by resolving them concurrently with Promise.all.
- Double-Blind Cleanroom Synergy with Amiga IA: Decoupling test authors from code implementers to eliminate confirmation bias during agentic TDD workflows and ensuring clean integration with the ami-orchestrate-cleanroom skill.
- Mandatory Declarative State Registration in TypeScript Contracts: Any reactive atom state key must be formally registered in the PluginState interface of types/index.d.ts to prevent untyped collisions during engine compilation and validation.
Production-Grade Extensions and Safety Guards for Claude Code
AMI Claude Mods (ami-claude-mods) is a curated suite of modular extensions and function hooks built specifically for the Claude Code CLI engine (Early Access 2.1+) and integrated with the Amiga IA agent ecosystem (amiga-ia). Engineered for developers and engineering teams leveraging AI agents in intensive daily coding workflows, the suite resolves key operational, context, and security pain points when running agentic loops in the terminal:
- Engine Inspection & Diagnostics (
hooks-viewer): Reactive side pane and interactive/hooks-panelcommand visualizing active hooks, registered mods, and configuration layers (user,project,local) in real time, bringing complete observability to the execution environment. - Context Efficiency & Delegation (
undelegated-reads): Status bar gauge and circuit breaker that detects bursts of consecutive mechanical searches (Read,Grep,Glob) on the primary thread and prompts delegation to exploration subagents, preserving the main reasoning context window. - Pre-Exit Resource Audit (
close-checklist): Exit guardrail triggered via/checklistor session termination (session.end). It audits uncommitted Git changes, running Docker containers/volumes, and lingering child processes to prevent local environment bloat. - Inter-Agent Messaging Inbox (
child-sessions): Non-blocking banner rendered above the prompt that tracks and prioritizes blocking questions or notifications from concurrent background subagents, complete with quick action buttons to reply, view, or dismiss. - Double-Blind Cleanroom Isolation (
cleanroom-guard): TDD guardrail that strictly blocks test-generating subagents from reading implementation source code (and vice versa), ensuring unit tests remain completely unbiased and immune to confirmation bias. - Multi-Profile Git & Kubernetes Safeguards (
context-guard): Declarative JSON-driven sentinel that verifies active GitHub CLI accounts against working repository paths and blocks un-scopedkubectlmutations on production Kubernetes clusters without explicit context flags.
Event-Driven Architecture & Reactive Terminal UI
All components in the suite remain fully decoupled and execute within Claude Code’s sandboxed ECMAScript runtime:
- Zero Core Binary Modifications: Extends agent capabilities by subscribing to the engine’s native event bus (
session.start,tool.call,command.run,session.end) via the officialEngineInterface($). - Reactive Terminal Surfaces (Ink/React): User interfaces (side panes, toasts, status bar badges, and prompt banners) leverage declarative reactive primitives (
Box,Text,Button) resolved dynamically at runtime. - Typed State Management: Inter-turn state persistence is managed through strongly typed reactive atoms, ensuring clean memory isolation between loaded mods and deterministic pre-validation checks.