下一代AI程式碼工具24小時內獲得78顆GitHub星標

Navigation Menu

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

To see all available qualifiers, see our documentation.

Autonomous engineering teams for Claude Code. Point Zeroshot at a GitHub issue, walk away, and return to verified, production-ready PRs.

License

Uh oh!

There was an error while loading. Please reload this page.

covibes/zeroshot

Folders and files

Latest commit

History

Repository files navigation

zeroshot CLI

2024 was the year of LLMs. 2025 was the year of agents. 2026 is the year of agent clusters.

Autonomous engineering teams for Claude Code.

Install

Platforms: Linux, macOS

Requires: Node 18+, Claude Code CLI, GitHub CLI

You know the problem. Your AI agent:

AI is extremely capable. But not when one agent does everything in one session.

Context degrades. Attention drifts. Shortcuts get taken.

Zeroshot fixes this with multiple isolated agents that check each other's work. The validator didn't write the code, so it can't lie about tests. Fail? Fix and retry until it works.

Point at a GitHub issue, walk away, come back to working code.

Demo

Sped up 100x — 90 minutes, 5 iterations until validators approved

The full fix cycle. Initial implementation passed basic tests but validators caught edge cases: race conditions in concurrent updates, ABA problem not fully handled, retry backoff timing issues. Each rejection triggered fixes until all 48 tests passed with 91%+ coverage.

A single agent would say "done!" after the first implementation. Here, the adversarial tester actually runs concurrent requests, times the retry backoff, and verifies conflict detection works under load.

This is what production-grade looks like. Not "tests pass" — validators reject until it actually works. 5 iterations, each one fixing real bugs the previous attempt missed.

When to Use Zeroshot

Zeroshot requires well-defined tasks with clear acceptance criteria.

Known unknowns (implementation details unclear) → Zeroshot handles this. The planner figures it out.

Unknown unknowns (don't know what you'll discover) → Use single-agent Claude Code for exploration first, then come back with a well-defined task.

Long-running batch tasks → Zeroshot excels here. Run overnight with -d (daemon mode):

Crash recovery (zeroshot resume) means multi-hour tasks survive interruptions.

Rule of thumb: If you can't describe what "done" looks like, zeroshot's validators can't verify it.

Commands

Q: Why Claude-only (for now)?

Claude Code is the most capable agentic coding tool available. We wrap it directly - same tools, same reliability, no custom implementations to break.

Multi-model support (Codex CLI, Gemini CLI) is planned - see #19.

Q: Why do single-agent coding sessions get sloppy?

Three failure modes compound when one agent does everything in one session:

Zeroshot fixes this with isolated agents where validators check work they didn't write - no self-grading, no shortcuts.

Q: Can I customize the team?

Yes, see CLAUDE.md. But most people never need to.

Q: Why is it called "zeroshot"?

In machine learning, "zero-shot" means solving tasks the model has never seen before - using only the task description, no prior examples needed.

Same idea here: give zeroshot a well-defined task, get back a result. No examples. No iterative feedback. No hand-holding.

The multi-agent architecture handles planning, implementation, and validation internally. You provide a clear problem statement. Zeroshot handles the rest.

How It Works

Zeroshot is a multi-agent coordination framework with smart defaults.

Zero Config

The conductor classifies your task (complexity × type) and picks the right workflow:

Model Selection by Complexity

Set model ceiling: zeroshot settings set maxModel sonnet (prevents opus)

Zeroshot is message-driven - define any agent topology:

Coordination primitives:

The easiest way to create a custom cluster: just ask Claude Code.

Example prompt:

Claude Code will read existing templates, create valid JSON config, and iterate until it works.

Built-in validation catches failures before running:

See CLAUDE.md for cluster config schema and examples.

Crash Recovery

Everything saves to SQLite. If your 2-hour run crashes at 1:59:

Isolation Modes

Git Worktree (Default for --pr/--ship)

Lightweight isolation using git worktree. Creates a separate working directory with its own branch. Fast (<1s setup), no Docker required. Auto-enabled with --pr and --ship.

Docker Container

Full isolation in a fresh container. Your workspace stays untouched. Good for risky experiments or parallel agents.

When to Use Which

Default mode: Agents are instructed to only modify files (no git commit/push). You review and commit yourself.

When using --docker, zeroshot mounts credential directories so Claude can access tools like AWS, Azure, kubectl.

Default mounts: gh, git, ssh (GitHub CLI, git config, SSH keys)

Available presets: gh, git, ssh, aws, azure, kube, terraform, gcloud

Custom mounts (mix presets with explicit paths):

Container home: Presets use $HOME placeholder. Default: /root. Override with:

Env var passthrough: Presets auto-pass related env vars (e.g., aws → AWS_REGION, AWS_PROFILE). Add custom:

More

Contributing

See CONTRIBUTING.md for development setup and guidelines.

Please read our Code of Conduct before participating.

For security issues, see SECURITY.md.

MIT — Covibes

Built on Claude Code by Anthropic.

About

Autonomous engineering teams for Claude Code. Point Zeroshot at a GitHub issue, walk away, and return to verified, production-ready PRs.

Topics

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

There was an error while loading. Please reload this page.

Stars

Watchers

Forks

Releases

  21

Packages

  0

Uh oh!

There was an error while loading. Please reload this page.

Contributors

  4

Languages

Footer

Footer navigation