keli-wen/cospan

Cospan treats "model × companion tools" as the scheduling unit: by writing /cospan-codex in the prompt along with the target model, reasoning intensity, and timeout, the main agent organizes the context, launches a background session, waits for results non-blockingly, and merges them upon return. The README demonstrates use cases such as having Gemini dispatch two GPT models to review different modules in parallel, having Codex drive a browser to test the checkout flow, and delegating repository scanning and race condition checks to Gemini and DeepSeek V4.1 Flash to save quota for flagship models.

Cospan treats "model × companion tools" as the scheduling unit: by writing /cospan-codex in the prompt along with the target model, reasoning intensity, and timeout, the main agent organizes the context, launches a background session, waits for results non-blockingly, and merges them upon return. The README demonstrates use cases such as having Gemini dispatch two GPT models to review different modules in parallel, having Codex drive a browser to test the checkout flow, and delegating repository scanning and race condition checks to Gemini and DeepSeek V4.1 Flash to save quota for flagship models.

Why it matters The main agent asynchronously delegates tasks like reviews and repository scanning to other models and tools, avoiding being blocked; when installed on Harness, it also provides a unified sidebar to view all background tasks.

Stars 21

Reasons

  • MIT licensed, created on 15 11, evolved from the author's previous agy-staff project
  • Provides doctor --check to verify installed command-line tools and parameter conventions without consuming tokens

Usage

  • Requires Node.js 20 or higher and the command-line tools to connect (agy, claude, codex, dsh); in Harness, run dsh plugin --profile desktop add github:keli-wen/cospan, and in Claude Code, run claude plugin marketplace add keli-wen/cospan before installing

Risks

  • Each tool assigned work is billed or consumes quota according to its own account, and parallel dispatching accumulates costs
  • The README recommends letting the agent download and follow installation instructions itself; if you are uneasy, you can install manually

Open

GitHub

Explore the latest AI Radar updates