dsh-baton

Claude Code and Codex each hand a task to DeepSeek Harness, and its subagents build this rainy-night street in 1 h 41 min (40 s, sound on):

https://github.com/user-attachments/assets/72e2fdff-275e-48bc-8e5a-39ffb025a0a8

See it run

One Codex chat starts a task, queues a follow-up, steers the running turn, and waits for all three:

Codex and Claude take turns in the same Harness session, and every message is labelled with who sent it:

Quick start

Requires DeepSeek Harness ≥ 0.2.1-rc.2 (signed in) and Claude Code or Codex. Nothing else: no Node.js, no terminal. The steps are the same on macOS and Windows.

1. Add the plugin in Harness: Plugins (插件) → Add plugin (添加插件) → type dsh-baton-host → Install (安装) → Enable now (立即启用).

Switched on, the plugin registers dsh-baton for Claude Code and Codex, for every folder, and Harness says so.

Codex only, once: open Codex Settings → Hooks and Trust the dsh-baton hook (from 0.6.1). It lets Codex end its turn after handing a task over and be woken when the run ends, so waiting uses no model requests. Until you trust it, Codex waits with dsh_wait as before, and the first hand-over reminds you.

2. Check: in a new Claude Code or Codex conversation, ask it to call dsh_list_workspaces. It lists your Harness workspaces.

That is all. If Harness says it could not register, run npx -y dsh-baton-mcp setup in a terminal (this needs Node.js ≥ 22.15); it does the same registration. Updating, removing and Windows details are further down.

How it works

How each client waits (nothing to configure):

• Claude Code waits for a run in the background (the wait_command that dsh_send returns) and is woken once, when the run ends.
• Codex with the dsh-baton Stop hook trusted (from 0.6.1): after dsh_send it ends its turn; the hook waits for the run outside the model and, when the run ends or asks something, continues the same conversation with the result. One model request per run instead of one every 270 s.
• Codex without the hook waits inside dsh_wait, up to 270 s per call, just under its 300 s tool timeout. An unfinished wait returns only the state, so Codex calls again without summarising. Codex gets no wait_command: its terminal cannot wake it, so it would poll.
• Codex asks you to approve MCP tool calls: approve the dsh-baton tools. For unattended codex exec runs, add --approve-for-me.

How a hand-over is checked (from 0.6.0)

You don't have to take DeepSeek's word for it. Before the run, dsh-baton notes what is in the folder. When the run ends it shows exactly which files were created, changed or deleted, and runs your project's own tests itself (npm test, or whatever the client passes as checks). A second, read-only DeepSeek agent then goes through the request point by point and ends with a table: requirement, met or not, and where (file:line). Claude or GPT reads just that, sends anything missing back in one message, and opens code only where a row says "not met". Unless cwd names another folder, the new Harness session starts in the folder you're already working in.

Token usage

In short: hand over the big, long jobs. A feature that spans many files, a large refactor, a whole small app: that is where it saves the most. Claude Code's own bill drops by 65–81%, and Codex gets 2 to 5 times as many tasks out of a ChatGPT Plus 5-hour window (5 times on a medium task with GPT-6 Astra).

The honest part: for small tasks, Codex on GPT-6.1 Sol is cheap enough that doing it alone costs less in total than handing it over, and at medium size it comes out about even. The hand-over pays off when you want Claude or GPT to plan and review, a cheaper model to do the typing, and you'd rather not keep hitting the 5-hour limit.

Who this is for: most of us pay for Claude or ChatGPT with a plan (Plus, Pro, Max 5x…), while DeepSeek is pay-as-you-go from a balance you top up, so every hand-over is money spent outside your plan. If your plan's limits already cover how you work, you don't need this: keep it simple and keep the money. If you're like me and keep running long, messy jobs with round after round of changes, and keep hitting the limit, a little DeepSeek spend gets you a lot more out of the plan you already pay for.

Cost per task at list price, DeepSeek V4.1 Flash 1017 at its peak price:

Share of the ChatGPT Plus 5-hour window per task: Sol small 1.9% → 0.8%, Sol medium 4.3% → 1.0%, Astra small 7% → 3.5%, Astra medium 22% → 4%.

• Two tasks, asked the way a user would. The small one takes Claude Code or Codex about 7–8 minutes on its own; the medium one is a small app with a server and a page. Long jobs beyond that were not measured. The requests are in Chinese, linked as written:
  • Small: a command-line task manager in Node, no dependencies, split into modules, with tests. Request · acceptance check
  • Medium: a task manager web app, Node server plus page, REST API, stats, tests and a README. Request · acceptance check
• The only difference in the hand-over runs is one line added at the end: 交给 DeepSeek Harness 去做。 ("Hand it to DeepSeek Harness.")
• Same model and effort in both arms, three runs per cell (Astra: one run each, two for the small hand-over), and every run judged by the same acceptance check, which the model never sees. All 12 hand-over runs passed it the first time.
• Prices are list prices. DeepSeek is counted at its peak price, the most conservative case; the real DeepSeek bill came to about 58% of that.
• Every run, the raw data and the scripts: run log and full report.

Tools