dsh-tui-plugin

A standalone terminal UI bundle for the DeepSeek Harness: an interactive in-process host surface that runs on top of the official dsh base bundle — session list, streaming conversation with light markdown, tool cards, approvals, questions, background jobs, subagents, goals, plan mode, model selection (with reasoning-effort cycling), settings (file editing), skills, and @ file/skill references.

It is a standard Cordis bundle plugin (cordis.patch.yml + tui-runner/tui-startup function plugins). It mounts no Host, HTTP server, or browser — it drives the same ctx host services the official Web surface proxies.

Requirements

• Node ^22.19 or >=24
• The official DeepSeek Harness CLI installed, so the dsh-base bundle and the cmdline/exit services exist.

Install & run

This is a bundle plugin for the official dsh CLI: install it into a profile's bundle stack, then boot that profile. It layers the terminal runner over the official dsh-base bundle.

npm i -g @deepseek-ai/dsh
npm i -g dsh-tui-plugin

# one-time: create/use a profile and add this bundle to it
dsh plugin --profile tui add dsh-tui-plugin
# (the first `add` creates the `tui` profile; it also installs dsh-base)

# every launch: boot the profile
dsh --profile tui

dsh --patch <path> takes a patch file path, not a package name. Bundle plugins are loaded by listing the package name in the profile's dsh.profile.bundles (which dsh plugin --profile tui add dsh-tui-plugin does for you) and installing it with pnpm in the profile directory.

From there the TUI opens in raw mode: 1–8 views, / command palette, @ mentions, mouse selection (copies on release), wheel + PgUp/PgDn scrolling, /lang for 中文.

The status line shows the model route with its reasoning effort (provider/model · high), accumulated token usage (↑in ↓out plus cache traffic), and while a turn runs, its elapsed time and output rate (12.3s 45/s). Resuming an interrupted session clears the stuck running state and shows a "session resumed — send a message to continue" notice instead.

Session lifecycle & compaction

• The sessions view marks heavy sessions (large accumulated prompt-side context, ◆) so you can spot the expensive ones at a glance.
• Resuming a heavy session (≥ 40k prompt tokens) offers /compact first: continuing would re-send the whole context on every request, missing provider prompt caches and billing full uncached input each time. Press y to compact, n/Esc to continue as-is. /compact is also always available from the command palette.

Build

npm i && npm run build   # emits lib/ via tsc

Test

npm test                 # unit suites (key decoder, screen, views, fold, popups,
                         # app state machine, runner wiring, i18n)

tests/real-composition.spec.ts boots the shipped base + tui patches through the real Loader; it needs the official base patch on disk:

DSH_BASE_PATCH=/path/to/deepseek-harness/packages/bundle/base/cordis.patch.yml npm test

(The official @deepseek-ai/dsh package may ship its base patch; point DSH_BASE_PATCH at it when it does.)

Publish checklist

1. Use the name as-is — dsh-tui-plugin is free on the npm registry (verified), and every reference (package.json, cordis.patch.yml, src/) already matches it. Only if you want a personal scope (e.g. @your-scope/dsh-tui-plugin) do you rename — and then also update the two dsh-tui-plugin name: entries in cordis.patch.yml (the patch rows load this package by its own name).
2. Verify dependency versions against the registry (npm view @deepseek-ai/dsh-llm version etc.) — the ranges here match the day this was scaffolded; the harness moves fast.
3. npm publish (with --provenance when your registry supports it).
4. Add the dsh-plugin GitHub topic on your repository.

Upgrading

The bundle pins @deepseek-ai/dsh-* service APIs that are stable, but official releases occasionally adjust the base patch rows. When a new @deepseek-ai/dsh lands, bump the ranges and re-run the composition test against the new base.

License

MIT — derived from the DeepSeek Harness project (MIT).