dsh-plugin-tui

Usage

dsh --profile tui                       # start at the prompt
dsh --profile tui "explain this file"   # send a first message, then stay interactive
dsh --profile tui -r                    # resume the latest tui session here
dsh --profile tui -s                # resume a specific session
dsh --profile tui -w ./scratch          # new session in another directory
dsh --profile tui -l                    # list this profile's sessions here
dsh --profile tui -p danger-full-access # full access from launch (no sandbox, no prompts)
dsh --profile tui -p workspace-write -y # sandboxed, but approvals auto-allowed

Inside the REPL:

Input	Meaning
/exit, /quit, Ctrl-D	leave (flushes the session first)
/session	print the session id
/sessions	pick a session interactively (↑/↓ move, Enter switches, Esc cancels); /sessions list prints the static list
/resume	switch to another session in place (index from /sessions list, or an id prefix)
/history [n]	reveal n earlier exchanges of the current session (switching in shows only the most recent ones)
/new	start a fresh session, releasing the current one
/title [text]	show the session title, or rename it (a rename pins the title)
/model [id]	pick a model interactively (↑/↓, Enter, Esc) or hot-switch by name (id or provider/id) — takes effect from the next step
/effort [id]	pick reasoning effort interactively, or set by name (/effort default restores the provider default)
/help	built-ins plus every harness-registered command
Ctrl-C while a turn runs	cancel the turn, stay in the REPL
Ctrl-C twice while idle	leave
y / N at an approval prompt	allow / reject one gated tool call
/approvals [ask|auto]	auto allows every request with a dim audit line (the sandbox still applies); ask restores prompts. dsh --profile tui -y starts in auto
a number (or free text) at a ⁇ question	pick an option / answer in your own words; multi-select takes comma-separated numbers

A status bar pinned to the terminal's bottom row shows the live model and effort, the approvals mode, the session title, and the working directory — it refreshes as those change and survives resizes; the transcript scrolls above it. Thinking is collapsed by default — the spinner area shows a live three-line window over the reasoning stream, and ctrl+o toggles the full dim stream. /model and /effort choices are remembered per profile (tui-model.json beside the profile; the machine-wide default is never touched) — /model default clears the preference. Tool calls always render as one task-summary line (⏺ Group counts (+14 lines)); the full command or script goes to the same per-turn detail stream ctrl+o reveals. The toggle only ever covers the current (or just-finished) exchange — starting a new one clears it. Resuming or switching to a session replays its recent exchanges statically (never re-running anything); /history pages further back, and your terminal's own scrollback covers the rest. The prompt appears in well under a second — MCP servers connect in the background and their tools join late. While a turn runs, a braille spinner covers quiet stretches; token deltas are micro-batched (~16ms) so heavy streams stay smooth. Answer text renders as markdown, a line at a time — bold headings, colored bullets and inline code, green fenced code blocks. Tool calls render through each tool's own presenter — ⏺ $ ls -la — List files, ⏺ Write foo.txt — with the ⏺ colored by the call's category (read, edit, execute, …), falling back to a compact raw-arguments line for tools without one. Todo snapshots draw as a ☑/◐/☐ checklist, each exchange ends with a dim ↳ 1.2k in · 89 out accounting line, and prompt history persists across runs (up-arrow recalls earlier inputs).

Ctrl-C works as a keystroke because raw-mode readline owns the terminal; the launcher's own SIGINT handler still answers an external kill -INT with a full teardown, which is the behavior you want from outside the TUI.