dsh-cursor-codex

Connect DeepSeek Harness (dsh) to the Cursor editor and the OpenAI Codex CLI — as an ACP agent, an MCP server, or a headless worker.

Quick start

# 1. Install dsh (or use `npx @deepseek-ai/dsh` everywhere)
npm install -g @deepseek-ai/dsh

# 2a. ACP (Zed / JetBrains / any ACP client)
dsh plugin --profile acp add @jeremy9682/dsh-acp
dsh --profile acp
#     (before the npm publish lands: `dsh plugin --profile acp add file:/path/to/dsh-cursor-codex/acp`)

# 2b. MCP (Cursor / Codex)
#   Cursor:  merge templates/cursor/mcp.json into ~/.cursor/mcp.json
#   Codex:   cp templates/codex/dsh.config.toml ~/.codex/dsh.config.toml
#            codex -p dsh "your task"

# 2c. One-shot headless (any shell)
dsh --profile headless "<complete self-contained task>"

Both 2b and 2c ship skills: install skills/cursor-delegate-to-dsh and skills/codex-delegate-to-dsh into the respective editor's skills directory so the agent knows when to delegate.

Why three channels?

Cursor and Codex are ACP agents, not ACP clients — neither can load a third-party ACP agent today. The full picture, verified against official docs, is in docs/integration-guide.md.

Verified

• ACP profile: initialize → session/new → session/prompt → streamed answer → end_turn (tested against dsh 0.1.0-rc.6).
• MCP server: initialize → tools/list → dsh_health → dsh_delegate real task (tested against dsh 0.1.0-rc.6).

License

MIT