amux-dev

Amux

Research, planning, and code review skills for Claude Code with dual-engine cross-validation (Claude + Codex).

What's Included

Skills (3)

• research — Deep research with 20+ sources and confidence tracking, powered by agent teams with Codex cross-validation
• planning — Pre-implementation planning that researches approaches using Context7, Serper, GitHub MCPs, and optionally btca for source-level codebase research, with dual-engine evaluation via Codex MCP
• code-review-pipeline — Multi-reviewer code review using agent teams (architecture, implementation, tech practices, tests, UI), each cross-validated with Codex

Agents (6)

Pipeline reviewers (5): architecture, implementation, tech practices, tests, UI — dispatched by the code-review-pipeline skill based on file types. Each performs Claude analysis and calls the native codex MCP tool for cross-validation.

Standalone reviewer (1): code — reviews completed work against the original plan and coding standards. Invoked via /review-code.

Commands (8)

• /research — Start a deep research session
• /code-review-pipeline — Run the full review pipeline
• /review-architecture, /review-code, /review-implementation, /review-tech-practices, /review-tests, /review-ui — Individual review commands

Hooks

• session-start — Announces available skills and detects Codex CLI presence (note: CLI presence does not guarantee MCP usability)
• task-loop-hook — Generic task loop that blocks exit while any skill's task-loop.json has complete: false. Used by research (and available for future long-running skills).
• pre-commit-quality-gate — Runs quality checks before commits

Dual-Engine Architecture

Each reviewer agent independently:

1. Performs Claude-based domain review
2. Calls the native codex MCP tool with cwd set to the repo root
3. Validates the Codex response — empty, non-JSON, or MCP error-text responses are treated as Codex-unavailable
4. Merges findings with classification (AGREE/CHALLENGE/COMPLEMENT) only if Codex returned valid JSON
5. Returns unified JSON with engine tags and cross-validation status

Cross-validated findings (flagged by both engines) receive a confidence boost — these are the highest-signal issues. Reviewers gracefully degrade to Claude-only when Codex is unavailable or returns unusable output.

Prerequisites

• Claude Code with plugin support
• Codex CLI (optional, for dual-engine mode): npm i -g @openai/codex
• Codex MCP server is declared as an MCP dependency (uses codex mcp-server — requires Codex CLI installed)
• btca (optional, for source-level codebase research in planning): bun add -g btca then claude mcp add --transport stdio btca-local -- bunx btca mcp

Installation

Claude Code (via Plugin Marketplace)

/plugin marketplace add guyathomas/amux-marketplace

/plugin install amux@amux-marketplace

Verify

/help
# Should list /amux:research, /amux:code-review-pipeline, etc.

License

MIT — see LICENSE