Kaola Project Runner

KPR 是 KPM 软件开发能力的镜像，通过迭代升级跟进，并继续被维护。

Let one agent work through another agent’s CLI.

One controlling agent can coordinate multiple projects at once. With Kaola-Delegator, each project gets its own delegated Host running one Project Runner Agent; that Runner assigns the project’s authorized workers to separate issues. Each worker can use Kaola Workflow to autonomously advance its issue, verify the result, and report back. Project records and worktrees stay within that project’s canonical repository; workers and authorization are not pooled across projects.

This repository provides:

• twelve self-contained worker Agent Skills — Claude Code, Codex CLI, Cursor CLI, Devin CLI, Grok CLI, Kimi CLI, OpenCode, ZCode, Droid CLI, dsh (DeepSeek Harness), Antigravity, and Hermes Agent. A controlling agent starts a session in a Git repository, sends instructions, reads replies and runtime evidence, and stops that exact owned session. Communication uses structured ACP (Agent Client Protocol) only.
• the main control-plane Skill Project Runner (kaola-project-runner): recover authorization, plan, dispatch, heartbeat, accept before finalize, own close-out.
• the external Kaola-Delegator Skill (kaola-delegator) for hands-off delegation from an outer Agent.

Pair a chosen entry with Kaola Workflow when that work needs a recoverable path from issue to verified delivery.

Contents

1. Choose an entry — Delegator / Project Runner / Platform Runner / Workflow Next
2. Kaola Workflow integration — normal path, orchestrator binding, evidence-backed exception
3. Install and begin
4. Select workers — runtimes, model selection, worker classes (Expert / Worker / Elite), preset catalog (all 12 runtimes, 28 presets), platform notes
5. Authorization — authorization by class, choosing a worker, seat binding and switching, authorization before a new Host
6. What you specify and what agents own
7. Daily use — session commands, steering, watching ACP sessions
8. For agents
9. Further documentation · Development · License and use

Choose an entry

Pick the most direct entry for how much you want to control. Do not force every task through every layer. Each layer finishes its own job and does not repeat the next.