Proposal for a top level ui toggle button to switch between chat and agent mode.

[fossFriend]

To the devs at DeepSeek, thank you for a very versatile model. DeepSeek has been one of my main LLM's for the past year.

I recently delved in agentic, so for the past year mostly used the web interface for coding. There is a strange benefit from working with the LLM in chat based ui, and that is token savings, the co-authoring with the LLM, and the general tightness of keeping the LLM on track.

Recently I tried out agentic and although I admit I am very new to it, one session chewed up 1.5M tokens using pi connected to DeepSeek v4 Flash, and most of that time I felt I was not having a meaningful say in what the agent was doing. I blew my token budget on a free tier api which I really regret. This would not have happened if I was doing this through in chat mode.

This is where a tradeoff would be very welcome, by giving DeepSeek-Harness two working modes that can be triggered by the user, namely, agent mode and chat mode. There are many times I am working as product developer, debugger and tester, while the LLM is working as coder. This is a very good work load for me because I get to keep the overall picture in mind and to steer the LLM onto granulated issues like fixing one bug at a time, adding one feature at a time, knowing when to revert, etc. Many times I am able to identify bugs far more efficiently than the LLM and immediately propose a tweak to test. This is where chat mode really shines without blowing a token budget because the LLM is on autopilot.

Having a chat/agent toggle would solve both usage cases, and give a very cool feature that I dont see in any harness in the market right now. I would love to see this in DeepSeek-Harness first. This would cater to the full spectrum of users from chat based coders to agentic based coders, and make your harness even more versatile.

[JinkaiLiu]

Thanks for articulating this so clearly. This workflow resonated with me, so I built a small community plugin that implements it:

dsh-autonomy

It adds an always-visible Chat | Agent switch on a dedicated row above the DeepSeek Harness Web composer. It works inside the same session, without changing the model, preset, history, sandbox, or permission policy.

• Chat removes tool schemas from the model request and guards tool execution.
• Agent restores the normal DSH tool loop and policies.
• The selected mode is stored per session and survives reloads and full DSH restarts.
• Switching modes does not disturb an unsent draft.

Install from npm:

npx @deepseek-ai/dsh plugin --profile web add dsh-autonomy
npx @deepseek-ai/dsh web

npm: https://www.npmjs.com/package/dsh-autonomy

Version 0.1.0 has been tested with DeepSeek Harness 0.1.0-rc.6, including clean registry installation, same-session switching, session isolation, and cold-restart recovery.

One boundary is intentional: switching to Chat cannot undo a tool action that has already started; the existing Stop control is still the right way to cancel an active turn.

I would really appreciate it if you tried it and shared whether it matches the workflow you had in mind.