dsh-tui

An open-source terminal front door for DeepSeek Harness (dsh).

@tomowang/dsh-tui is an out-of-tree mode bundle: it stacks on @deepseek-ai/dsh-base exactly like the shipped dsh-web-app and dsh-headless bundles do, but drives the agent from your terminal instead of a browser. The package is both a Cordis plugin (terminal input and presentation) and a dsh bundle (dsh.bundle.patch in package.json points at cordis.patch.yml); everything else — model adapters, tools, session persistence, sandbox and approval policy — stays in dsh-base and remains patchable underneath.

How it works

• The TUI renders only from the durable session log: it replays agent.session.snapshotEvents() on startup and follows session/event live, so --resume shows the exact history the log carries — the harness's "model-visible ⟺ logged" invariant does the heavy lifting.
• The interface runs full-screen in the terminal's alternate screen buffer, with an application-owned transcript viewport (scroll with the mouse wheel/trackpad, PageUp/PageDown) that auto-follows new output until you scroll up.
• Line input maps to the agent inbox: agent.followup() while idle, agent.steer() while a turn is running, Ctrl+C cancels the running turn.
• tui-startup parses this app's flags (everything after the launcher's own) through dsh-cmdline and publishes them as an ordinary Cordis service; the runner row reads them via the bundle patch, mirroring dsh-headless.
• Both stdin and stdout must be real TTYs; the plugin fails loud instead of degrading, so pipes keep using dsh --profile headless.

Features

• Status bar — session id, active LLM provider/model, current agent preset, live run state with spinner, queued-message count, and logged event count.
• Stats line — turn/step counts, LLM/tool wall time, TTFT and decode tok/s, cache-hit %, billed tokens, and a compact context-usage summary; sections hide themselves until there's data.
• /model provider management — switch the active model by picking from a provider's catalog, and add, edit, or delete custom LLM providers (route, base URL, API key, model discovery) without leaving the terminal.
• Agent presets — start a fresh session on a given preset with --agent-preset, or browse and switch presets from /presets (fixed once the session's first turn has run).
• Session inspectors — /trajectory for a turn/step event ledger with a detail view and filtering, /context for a context-window usage breakdown, /plugins for the loaded Cordis plugin tree and fiber state.
• Collapsed tool calls with a live spinner — a running tool call shows as a single spinner line in the prompt area; once its result lands, it settles into one collapsed ✓/✖ transcript line rather than an inline multi-line card.
• Tool Cards overlay — /tools or Ctrl+O opens a scrollable browser over the session's tool calls/results, each shown expanded to its full presentation by default (Enter/Space collapses a card back to its title).
• Reasoning display — a model's reasoning/thinking content never floods the screen: an animated ✦ thinking line stands in for it while still streaming, and it collapses to a one-line ✦ think · … summary ahead of the visible answer once settled in the transcript; the full text is always available via /trajectory.
• Markdown rendering — assistant text wit

…