A CLI that makes AI coders prove their work before shipping

What it does

Zeroshot is a JavaScript CLI that coordinates multiple AI agents to implement, review, and verify code changes from GitHub issues, Markdown files, or plain text prompts. It classifies tasks by complexity, dispatches a planner and worker in isolated environments—local directories, git worktrees, or Docker containers—and loops the results through independent validators until they approve or produce actionable rejections. All state is persisted to SQLite, so interrupted runs can resume without losing context.

The interesting bit

The validators are deliberately blind: they never see the worker’s context or code history, which prevents the model from rationalizing its own mistakes. The conductor scales the agent swarm to the task—trivial jobs get one agent, standard tasks get a planner plus worker and three to five validators—so you aren’t burning API tokens on a full committee to fix a typo.

Key highlights

• Supports Claude Code, OpenAI Codex, Gemini CLI, and OpenCode as backend providers, switchable per run.
• Integrates with GitHub, GitLab, Jira, and Azure DevOps issues, auto-detecting the provider from git remotes.
• Runs in isolation modes ranging from none to Docker, with crash-safe SQLite state recovery.
• Enforces cost ceilings to prevent runaway API spend during long iteration loops.
• Can open pull requests and auto-merge on approval with the --ship flag.

Caveats

• Windows support is explicitly deferred while the team hardens reliability; Linux and macOS only for now.
• It shells out to provider CLIs rather than calling APIs directly, so you need the relevant vendor tools installed and authenticated.
• The README warns that poorly defined tasks (“make the app faster,” “improve the codebase”) will fail because validators require clear acceptance criteria to verify completion.

Verdict

Worth a look if you are a senior engineer or platform team trying to automate issue backlogs and PR workflows without sacrificing correctness. Skip it if you want exploratory refactoring or vibe-driven prototyping where “done” is a feeling.