Skip to content

Orbit v0.1.0 — Liftoff

Latest
Latest

Choose a tag to compare

View all tags
@zheng-weihao zheng-weihao released this 06 Jul 15:26
v0.1.0
b133b15
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

🚀 Orbit v0.1.0 — Liftoff

Source code as agent knowledge — multi-repo Git workspaces where agents read, code, and ship.

Orbit manages multi-repo Git workspaces for AI coding agents. Instead of embeddings, vector stores, or stitching together scattered docs, agents work directly against full source worktrees — grepping API signatures, tracing call chains, then editing, committing, and pushing, all from one directory. Every task gets its own isolated workspace, so multiple agents can run in parallel without stepping on each other.

This is our first public release. The core workflow is stable, covered by 177 tests across 20 files, and already in daily use. We're shipping now because the foundation is solid enough to build on — together.

✨ Highlights

1. One source of truth — read and write unified

No RAG. No stale training data. Agents verify against real source, then edit, commit, and push from that same source. One mechanism for both reading and writing.

2. Agent self-service with progressive loading

Agents decide what they need — no human setup required:

Level 0  orbit repos        → Name + one-line brief (~50 tokens/repo)
Level 1  orbit info <repo>  → Repo roles + key entry points — when to add, where to start (~200 tokens)
Level 2  orbit add <repo>   → Full source — grep, read, trace

3. Parallel workspace isolation

Each workspace is an independent combination of multi-repo worktrees. Multiple agents, multiple tasks, zero conflicts. Polyrepo teams don't have to migrate to a monorepo — every repo keeps its own Git history and CI.

4. A knowledge feedback loop

Agents capture discoveries with orbit jot (cheap, ~20 tokens) and fold them into persistent repo memos via orbit memo. The next session's agent picks up those memos at Level 1 — no re-exploration needed.

5. First-class agent integrations

Ships with Claude Code and Qoder plugins — SessionStart hooks for deterministic workspace detection, PreToolUse hooks that auto-approve safe commands, and full skill definitions. Zero-config for supported agents.

📦 What's Included

Core Commands

  • clone — Add repos to the shared pool (supports fork push-URLs)
  • repos — Browse the pool roster (--json, plus memoBehind staleness)
  • info — Repo memo (roles + entry points) and auto-fetch, with two-layer staleness detection
  • memo — Read and write repo knowledge (--scaffold to generate a template, --refresh to rebuild the index)
  • sync — Fast-forward pool repos to upstream (force reset, branch switching)

Workspace Lifecycle

  • new — Create a workspace with a goal (--exec to launch an agent immediately, --name for custom naming)
  • add — Pull a repo into the workspace as a full worktree (--ref to check out a tag or branch)
  • switch — Branch operations within a workspace (-c to create, or switch to an existing branch)
  • jot — Lightweight discovery queue (push/pop) for capturing knowledge in flight
  • done — Mark work complete and record PR URLs
  • prune — Reclaim done workspaces (--dry-run, --verify, --older, with branch protection)

Workspace Reactivation

Setting a goal on a done workspace clears its completion record and PR history, reactivating it for reuse — no need to spin up a new workspace for follow-up work.

Status & Context

  • status — Workspace state at a glance
  • goal — Read, set, or clear the workspace goal (pipe-friendly)
  • context — A complete context dump in one call (--prime for cold starts, --json, or single-key queries)

Configuration & Diagnostics

  • config — Project-level settings (e.g. agent.recommend)
  • doctor — Environment health check (git ≥ 2.20, bash ≥ 3.2, optional deps, project structure)
  • completion — Shell completion for zsh and bash

Agent Integrations

  • Claude Code plugin (skill + SessionStart hook + PreToolUse hook)
  • Qoder plugin (skill + SessionStart hook + PreToolUse hook)
  • Unified install.sh with --claude / --qoder (alias --qodercli) / --zsh / --bash flags

⚡ Quick Start

# Install the runtime and an agent plugin
./install.sh --claude          # or --qoder

# Add repos to the pool
orbit clone git@github.com:org/backend.git
orbit clone git@github.com:org/frontend.git

# Set your recommended launch command (one-time)
orbit config agent.recommend 'claude "orbit start"'

# Create a workspace and launch
orbit new "Update the API definition and adjust frontend calls"
cd task-01 && claude "orbit start"

Or bootstrap without cloning the repo first:

curl -sL https://raw.githubusercontent.com/orbcli/orbit/main/install.sh | bash

⚠️ Known Limitations

  • Convention-based isolation — Agents are kept inside workspace boundaries by skill-level conventions, not filesystem sandboxing. This is fine for the single-owner model, but it isn't hardened for untrusted or adversarial multi-agent setups.
  • No color output yet — All CLI output is plain text; TTY-aware coloring is planned.
  • IDE worktree recognition — A few edge cases around the .git worktree-pointer file (VS Code + GitLens, JetBrains) are unconfirmed. Qoder and VS Code are confirmed working.
  • Skill-only agents lack deterministic detection — Without the plugin hooks, agents fall back to the orbit start trigger phrase at session start.
  • No export / import yet — Sharing team knowledge currently requires access to the same filesystem; a portable export is on the roadmap.
  • Dependencies — Requires git ≥ 2.20 and bash ≥ 3.2. Optional: jq (JSON parsing) and gh (GitHub CLI features).

🙏 Get Involved

This is day one. The foundation is in place — directory model, branch strategy, metadata contract, knowledge system — and it's built so that whatever comes next never forces the base to be reworked.

If you work with AI agents across multiple repos and you're tired of fragmented context, take Orbit for a spin:

More to come. Godspeed. 🛰️

Assets 2
Loading