Preface¶
If you need to use DeepSeek Harness (DSH) models, sessions, commands, approvals, tools, and subagents in a terminal while wanting an interaction style closer to Claude Code, the following introduces cogine-ai/dsh-claude-tui.
It is an unofficial DSH TUI client, described as:
Unofficial, high-fidelity Claude Code-style TUI for DeepSeek Harness, verified against real PTY captures
First, install it locally, then enter the terminal interface. After that, you can continue working with DSH’s real models and sessions.
What It Is¶
dsh-claude-tui is maintained by cogine-ai, under the MIT license, with the current package version being 0.1.5.
It solves a specific problem: placing DSH’s runtime capabilities into a terminal interface, allowing users to complete model selection, session resumption, input editing, approval confirmation, tool observation, and subagent management in the command line.
Note: This is an independent community project with no affiliation, endorsement, or sponsorship from Anthropic or DeepSeek. The term “Claude Code” in the project is used to identify a fixed-version interaction target and does not contain Anthropic source code.
Core Features¶
Terminal Interface¶
The TUI provides the following interface elements:
- Claude-shaped welcome panel
- prompt
- menus
- transcript
- status rows
- approvals
- questions
- Agent states
The welcome panel displays actual runtime information:
- TUI version
- Harness version
- bundled or system runtime
- DSH Home
- tool mode
DSH Runtime Capabilities¶
This TUI reads and displays DSH-side capabilities rather than hardcoding models, effort, credentials, or approval behaviors into the interface.
Verified capabilities include:
- DSH-owned models
- durable Sessions
- commands
- approval policies
- tools
- structured questions
- subagents
- Provider/model catalog
- advertised effort levels
- saved defaults
- masked API-key entry
- credential-source visibility
Input and Execution Information¶
The input side supports:
- Multiline editing
- image paste
- submit or steer
- cancellation
- history search
- slash completion
- bounded
@file mentions
The execution side can view:
- Reasoning and tool activity
- compact/expanded results
- cache hit rate
- token totals
- TTFT
- throughput
- turn outcome
For Session and Agent workflows, it supports:
- New or resumed Sessions
- graceful flush
- foreground/background subagents
- an active-agent roster
Installation and Setup¶
The runtime requires Node.js 22.19+ or 24+.
One-time launch:
npx --yes --legacy-peer-deps dsh-claude-tui
legacy-peer-deps is a temporary npm installation bypass used to handle the dense upstream peer graph in rc2.
To pin to this release version:
npx --yes --legacy-peer-deps dsh-claude-tui@0.1.5
For repeated use, you can install globally:
npm install --global --legacy-peer-deps dsh-claude-tui@0.1.5
Then run:
dshtui
Typical Usage¶
First, start the TUI:
npx --yes --legacy-peer-deps dsh-claude-tui
For real model requests, you need to provide credentials for the selected DSH provider.
Check or enter credentials:
/provider
Switch DSH-exposed models and effort levels:
/model
Alternatively, use keyboard shortcuts:
Option+P
Alt+P
Resume a session:
dshtui --resume
Resume a specific session:
dshtui --resume <session-id>
Toggle plan mode:
Shift+Tab
This shortcut executes DSH’s /plan or /plan off command.
Paste an image:
Ctrl+V
It reads an image from the desktop clipboard and adds Claude’s [Image #N] composer marker.
The image submission path validates and persists bytes, writing to rc2’s:
ctx.attachments
And uses the attachment-aware:
execute(agent, line, images, signal)
Use Cases and Considerations¶
Suitable for the following scenarios:
- Needing to place DSH’s models, sessions, tools, and approval workflows into a terminal
- Wanting familiarity with Claude Code-style input, viewing, and execution feedback
- Needing to resume sessions, view subagents, and handle approvals in the command line
- Hoping to complete daily agent work without leaving the DSH runtime
Before use, note the following:
- This is an independent community project, not an official Anthropic or DeepSeek project.
- It runs with the current DSH process permissions. Before installation, you should check the source code, license, and dependency lock.
- The license is MIT.
- Real model requests require credentials for the selected DSH provider.
- This version pins the bundled runtime to DeepSeek Harness
0.1.1-rc.2, with external runtime requirements of>=0.1.1-rc.2 <0.1.2. - Do not run different Harness versions concurrently against the same
$DSH_HOME. - Unless the DeepSeek Harness documentation specifies a supported migration path, do not downgrade the Home.
Conclusion¶
The value of dsh-claude-tui lies in placing DSH’s models, sessions, commands, approvals, tools, and subagents into a terminal workflow closer to Claude Code. It is not an official client nor a simulation of Claude’s cloud service, but an unofficial TUI for the DSH runtime.
- GitHub: https://github.com/cogine-ai/dsh-claude-tui
- Plugin Directory Page: https://www.skillhub.cn/plugins/cogine-ai/dsh-claude-tui