cue-agent

Cue

Turn GitHub issues into tested, reviewed draft pull requests — using headless coding agents.

Cue (pronounced kyoo, like “cue the lights”) is a CLI that drives coding agents through a deterministic issue pipeline:

flowchart LR
    A["Issue<br/>agent:ready"] --> B["Triage Plan"]
    B --> C["You Approve<br/>agent:approved"]
    C --> D["Dev & Test Gate"]
    D --> E["Code Review"]
    E --> F["Draft PR"]
    F --> G["You Merge"]
    F -. "agent:revise" .-> E

Instead of babysitting interactive agent sessions in your terminal, Cue runs coding agents through a deterministic pipeline where GitHub acts as the state store: labels drive the state machine, issue comments carry implementation plans, and draft PRs deliver the finished code.

Documentation & Guides: hamedniroomand.github.io/cue

Why Cue?

• Asynchronous & Non-Blocking: Assign an issue with a label and move on. No watching streaming tokens in a terminal.
• Human Control on Irreversible Steps: You review and approve the implementation plan before code is written, and you merge the final PR. Agents never touch main or force-push.
• Deterministic Quality Gates: Real test and lint commands (bun test, npm test, pytest, etc.) validate changes — pass/fail is never left to LLM self-evaluation.
• Isolated Worktrees & Scrubbed Secrets: Agent edits happen in dedicated git worktrees outside your project root. GitHub tokens (GH_TOKEN) are completely scrubbed from agent subprocesses.
• Multi-Engine Flexibility: Works with OpenAI Codex (codex), Claude Code (claude), and Google Antigravity (antigravity).

Quick Install

curl -fsSL https://raw.githubusercontent.com/hamedniroomand/cue/main/install.sh | bash

Or, if you already have Bun ≥ 1.1:

bun install -g cue-agent

Requires authenticated gh and codex CLIs (or Claude / Antigravity). Then, inside any project repository:

cue init

Label an issue agent:ready and run cue process. Full walkthrough: Getting started.

Documentation

• Why Cue — how Cue differs from the built-in orchestration of agent CLIs
• Getting started — prerequisites, installation, and first issue
• Pipeline and labels — the state machine and human/Cue collaboration loop
• Commands — init, process, run, cleanup, status, ui, upgrade
• Configuration — .cue/config.json schema and adapters
• Dashboard — local web UI for transcripts, costs, and issue tracking
• Contributing — setup and architecture for developing Cue

License

MIT