dsh-tui-pro

dsh-tui-pro

Community-maintained interactive terminal UI plugin for DeepSeek Harness.

Release status

Version 1.7.1 is the current release. The current source targets 1.8.0; its verification status is recorded in TESTING.md.

Do not use the existing v1.0.0 GitHub tag; it predates the repaired source and bundle metadata. Install from npm or a verified GitHub Release.

The 1.8.0 source reserves /exit for whole-TUI shutdown. /quit and empty-input Delete close the current project session, preserve its history, and return to the fixed assistant. The assistant can read user/assistant dialogue and image placeholders from another Active project session without exposing reasoning or tool traffic. See REPAIR_PLAN.md for the implementation record.

Intended installation

Install the released package with:

dsh plugin --profile tui add @lk251066/dsh-tui

The package itself owns the TUI plugin and its cordis.patch.yml profile layer. There is no separate bundle package.

Intended features

• Alternate-screen terminal interface with a full outer frame
• Persistent right-side active sessions grouped by workspace, with compact activity and status
• Internal transcript paging with fixed input and status areas
• /sessions main-area browser across complete history, with direct activate, remove, and open actions
• Double Escape checkpoint browser for returning the current conversation to an earlier turn, restoring its prompt for editing
• Enter steering, Tab next-turn queueing, empty Up recall, and Esc queue recovery while the agent runs
• Alt+Left and Alt+Right, /switch, and sidebar-click active-session switching
• /quit and empty-input Delete session closing, with complete history preserved
• Assistant-only paged reading of visible user/assistant dialogue from Active project sessions
• Direct transcript drag selection with automatic clipboard delivery
• Fixed bottom selectors for questions, approvals, models, reasoning effort, skills, details, themes, permissions, renaming, and goals
• Alt+V clipboard image attachment for vision-capable models
• Per-session /memory on|off, enabled by default for the assistant and disabled by default for project sessions
• A fixed personal assistant entry with workspace-session tools
• /new for the active project and /new <path> for another project
• Turn-level You and Assistant headings, syntax highlighting, diff rendering, and Markdown rendering
• Thinking blocks are hidden by default; /details ... reasoning on expands settled reasoning
• Tool cards support hidden, collapsed, and expanded views; diff cards show colored additions and deletions
• Approval dialogs with risk confirmation

New untitled conversations show the large welcome view. Existing conversations use the compact dsh v{version} — {title} header, while cwd, branch, model, and context stay in the bottom status line.

Launching in a directory resumes its first manually ordered active project session. If that directory has no active session, dsh creates one and adds it to the workspace. To close a project session, open it and run /quit, or press Delete while the input is empty. A running turn is cancelled, the session leaves Active, the assistant opens, and history remains available through /sessions. The assistant cannot be closed.

The TUI commands are /help, /model, /effort, /clear, /details, /theme, /rename, /fork, /status, /context, /agents, /jobs, /settings, /memory, /skill, /export, /sessions, /switch, /new, /assistant, /exit, and /quit. /skill opens the searchable bottom selector; /skill <name> [instructions] loads one user-invocable skill directly. /exit cancels the current turn and exits the TUI; /quit closes the current project session and returns to the assistant. /help, /status, /context, /agents, /jobs, /settings, and /sessions use the scrollable chat main area while the frame and sidebar remain fixed. dsh-base may additionally provide /feedback, /goal, /compact, /permission, and /plan. /queue, /palette, /reload, /fleet, and /memories are not part of this bundle.

Press Alt+V to read an image from the system clipboard and add an [Image #N] placeholder to the active session draft. Submission succeeds only when the selected model explicitly advertises image input; otherwise the draft is retained so a vision-capable model can be selected. PNG, JPEG, WebP, and GIF are supported. Windows uses PowerShell, macOS requires pngpaste, and Linux uses wl-paste or xclip.

Development

See DEVELOPMENT.md for local setup and TESTING.md for the verification sequence.

License

MIT