Show HN: Codex Swarm – Local ChatGPT swarm for coding with Git-tracked agents (github.com)

Codex Swarm is an open-source framework that turns your local IDE + OpenAI Codex plugin into a predictable, git-tracked multi‑agent workflow. Instead of freeform chat, it provides an opinionated orchestrator that interprets a user goal, drafts a plan, and delegates ordered steps to JSON-defined agents (PLANNER, CODER, REVIEWER, DOCS, etc.) stored in .AGENTS/*.json. The canonical backlog lives in tasks.json; a small CLI (python scripts/tasks.py) regenerates a human-readable tasks.md. Every atomic task maps to exactly one commit (emoji‑prefixed, task-ID referenced) so progress and provenance are auditable. There’s no separate runner or daemon—the Codex plugin is the runtime, all file operations happen locally, and network calls only run when initiated from your IDE. Setup requires the OpenAI Codex plugin, Git, and Python 3.10+. For AI/ML practitioners this is significant because it moves AI-assisted coding from ephemeral chat breadcrumbs to repeatable, human-in-the-loop workflows with guardrails: approvals, role permissions, commit-linked task metadata, and dynamically loadable agent prompts. That structure helps teams enforce review, documentation, and compliance steps while keeping the model’s scope limited to repository content visible to the plugin. It’s extensible (create domain agents via JSON templates), plugin-agnostic across IDEs that support Codex, and useful for building research, QA, or docs specialists—though it’s explicitly not an autonomous remote executor and depends on the Codex plugin runtime.