AI-Whisper: Terminal-Native Claude + Codex Dual-Agent Collaboration Orchestration System

AI-Whisper is an innovative terminal-first AI programming collaboration platform integrating Claude and Codex into an autonomous 'implement-review' pair. It uses a 'baton handoff' mechanism to ensure only one agent has execution rights at any time, enabling a closed-loop workflow without continuous human supervision. Key features include structured serial collaboration, real terminal sessions, recoverable execution, and human-in-the-loop escalation.

Single AI coding assistants often lack code review capabilities, leading to potential quality issues. AI-Whisper addresses this by pairing two leading coding agents (Claude and Codex) to form an autonomous collaboration team. Unlike swarm architectures with concurrent agent actions (which cause chaos), it uses a strict serial model for predictable and debuggable workflows.

Baton Handoff: Only one agent holds execution rights at any time. Terminal Native: Agents run in real terminal sessions (e.g., whisper collab mount claude in one terminal, Codex in another). Workflow Steps:

1. Role assignment (implementer/reviewer).
2. Implementer executes steps.
3. LLM evaluator checks deliverables.
4. Review loop (if not up to standard, reviewer feedback guides iteration).
5. Recoverable state (persists workflow for resume).

Tech Stack: TypeScript (92.5%), JavaScript (3.8%), Shell (3.7%). Structure: Monorepo with packages/cli (command line), core (logic), skills (agent skills). Dependencies: Node.js 22+, Claude Code CLI, Codex CLI, LLM evaluator (with credentials), optional tmux. Skills: Install via whisper skill install (supports workflow validation, start, status reporting).

Autonomous Mode Risks: Agents have full access (read/write files, execute commands) in mounted workspaces (using --dangerously-skip-permissions etc.). Recommendations: Use in trusted code directories, monitor via dashboard, review results before release. Escalation: System hands control back to humans on issues like exhausted iterations, blocked agents, low confidence—this is a designed feature, not a failure.

Suitable For: Heavy AI coding users wanting mutual agent checks; terminal-first developers; long structured workflows (spec→plan→implement→review); multi-agent review needs. Not Suitable For: One-off quick coding; invisible background automation; AI coding beginners needing guidance. Workflow Modes: Spec-driven (using docs/spec.md) and open-loop (ralph-loop for open goals).

Installation: npm install -g ai-whisper OR pnpm install + build from repo. Configuration: Install skills (whisper skill install), set up LLM evaluator credentials. Run: Mount agents in two terminals, start workflow (e.g., Run spec-driven-development using docs/spec.md), monitor via whisper collab dashboard. Contribution: Apache 2.0 license; accepts DCO-signed contributions (use git commit -s).

AI-Whisper pioneers a new direction for AI coding tools—moving from single agents to structured multi-agent collaboration, from instant responses to recoverable workflows, and from full autonomy to supervised autonomy. Its key innovations (baton handoff, human-in-the-loop, terminal native) offer a reference architecture for teams looking to upgrade AI assistants from code completers to development partners. For Claude/Codex users, it can boost automation of complex tasks while ensuring code quality.