mac-proxy-sync

Wreckit takes a text backlog and moves each item through codebase research, planning, story-by-story implementation, and pull request creation for you to review. State and artifacts stay in Git-trackable JSON and Markdown files, making runs inspectable and resumable. It requires an AI agent or API access and can consume substantial tokens; an optional Sprite-based sandbox isolates riskier execution.

Use Codex Autoresearch when a repository task has a numeric outcome, such as fewer test failures, better coverage, lower latency, or a smaller binary. The skill confirms the target, editable scope, metric command, regression guard, and run mode, then lets Codex try one focused change per iteration. Improvements are committed and retained; failed or regressive trials are reverted through Git, with an auditable event history. It can run interactively or in the background, but each experiment requires a clean named Git branch and sufficient access to create and revert commits.

paper2code takes an arXiv URL, audits which implementation choices are specified or missing, and generates a project whose code and configuration point back to paper sections and equations. It focuses on the paper’s core contribution, does not download datasets or build full training infrastructure, and does not claim the resulting implementation is correct.

Agent Sandbox Skill gives Claude Code, Codex, and Gemini CLI managed E2B environments for engineering work that should not touch the local filesystem or production systems. Agents can install dependencies, run untrusted code, scaffold and build full-stack applications, expose a hosted port, and perform Playwright-based visual checks. Commands cover small sandbox tasks as well as plan, build, host, and test workflows, with lifecycle tools that preserve a sandbox across turns. An E2B API key is required, and compute and hosted previews depend on the E2B service.