dsh-tianshu-tui — DeepSeek Harness coding terminal

Chinese | English

dsh-tianshu-tui is an interactive terminal TUI panel plugin for the official [DeepSeek Harness]. The rendering core is a self-developed, minimalist ANSI engine, delivering a lightweight, ultra-smooth rendering experience. (Evolved from the author's [Tianshu-harness] https://github.com/huiliyi37/Tianshu-harness. The UI is a pure presentation layer: all agent state comes from the session event stream — streaming Markdown / tool cards, 16+ themes, slash commands and pickers, input history with local preference persistence, and LSP diagnostics. On top of that, engineering-layer enhancements were made, such as image & vision bridging, intelligent code retrieval, memory & cross-session recall, etc.)

Docs

Doc	Description
Getting started	Installation, launch and FAQ
Interaction manual	Full table of shortcuts and commands
Configuration	Assembly options, environment variables and runtime configuration
Architecture	Layering, data flow and design decisions
Themes	16 built-in themes and customization
Plugin ecosystem	Companion plugins and extension points
VS Code	Using in VS Code
ADAPTER.md	TUI ↔ harness boundary contract
Contributing guide	PR conventions and validation matrix
Development notes	Structure, build and release

Installation

This package is not a standalone program. You must first have the official CLI @deepseek-ai/dsh (npm latest, currently 0.1.5-rc.1; requires ≥ 0.1.5-rc.1, peer dependency aligned). Just npm i this package alone and it will not run.

One-click install (recommended): The repo ships a cross-platform script that auto-detects Node/pnpm and, via pnpm, installs the official CLI + assembles this plugin and launches it (for China networks, defaults to the npmmirror registry):

# macOS / Linux（bash）
bash <(curl -fsSL https://raw.githubusercontent.com/huiliyi37/dsh-tianshu-tui/main/scripts/install-tui.sh)
# 只安装不启动：
bash <(curl -fsSL https://raw.githubusercontent.com/huiliyi37/dsh-tianshu-tui/main/scripts/install-tui.sh) --no-launch

# Windows（PowerShell）
powershell -ExecutionPolicy Bypass -Command "irm https://raw.githubusercontent.com/huiliyi37/dsh-tianshu-tui/main/scripts/install-tui.ps1 | iex"
# 只安装不启动：
powershell -ExecutionPolicy Bypass -File scripts\install-tui.ps1 -NoLaunch   # 克隆仓库后本地跑

1. Prepare your environment

• Node.js ^22.19 || >=24
• pnpm (dsh plugin forwards to it; if you don't have it, corepack enable is enough — Node ships with corepack)

⚠ npm 11 OOM gotcha: the official CLI @deepseek-ai/dsh has a large dependency tree (60+ sub-packages); npm 11 (bundled with Node 24) will hit a JavaScript heap out of memory on install (hangs for several minutes then OOMs, reproducible in practice). Use pnpm instead (commands below). If you were using npx -y @deepseek-ai/dsh and it hangs / reports a heap OOM, just switch to pnpm.

Do not type dsh directly. If an old dsh already exists on PATH (e.g. ~/.local/bin/dsh, where dsh --version is below 0.1.0-rc.8), it will fall back to the local staging dir and produce ERR_FS_EISDIR / Path is a directory .../@deepseek-ai/dsh. Always use the pnpm dlx command below.

2. Install this plugin into the tui profile

pnpm dlx @deepseek-ai/dsh plugin --profile tui add @huiliyi37/dsh-tianshu-tui

pnpm may warn about missing peers — you can ignore it: the peer is provided by the official dsh host, so there's no need to install it separately. Without pnpm, you can also run npx -y pnpm dlx @deepseek-ai/dsh plugin --profile tui add @huiliyi37/dsh-tianshu-tui.

After installing from npm, each launch checks against npm latest: if a new version exists, it's written into the profile and a restart is prompted to take effect. You can also run /update in the TUI to manually check (check-only, no install; it prints the update command). To skip the online check, set DSH_TUI_SKIP_UPDATE=1. github: / link: installs are not rewritten to npm packages.

You can also install from Git: pnpm dlx @deepseek-ai/dsh plugin --profile tui add github:huiliyi37/dsh-tianshu-tui (the repo already contains lib/index.js, no need to build).

3. Launch

pnpm dlx @deepseek-ai/dsh --profile tui

Seeing the welcome page with the dsh-tianshu-tui brand means success. Ctrl+Q or /exit to quit.

If the official CLI is installed globally (pnpm add -g @deepseek-ai/dsh) and dsh --version is not lower than 0.1.0-rc.8, you can simply replace the pnpm dlx @deepseek-ai/dsh above with dsh.

4. Agent presets (/preset)

The command is /preset (not /presets). This package's bundle targets the official web version: it turns off the host's agent faces and mounts @deepseek-ai/dsh-agent-presets (dependency pinned to 0.1.5-rc.1; npm latest is still stuck on the outdated 0.0.1-rc.1). Running plugin add for this package installs the roster in one shot; new sessions mount in setup, and /preset switches among the official shipped faces (standard / PTC / minimal / creative) — it does not stack on top of dsh-base tools.

Usage:

• /preset — lists the capabilities and tool set of each preset, * marks the current one; the footer / welcome top bar also shows the current short name (standard / PTC / minimal / creative). After you've chatted, it also appends the wire: tool face of the most recent request.
• /preset <id> — switches to the specified preset (only in a blank session: run /session new first, then switch; ptc/creative are aliases of code/cordis).

If npx still reports ERR_FS_EISDIR, it's a conflict between the old install fallback in ~/.dsh/profiles/node_modules and the official CLI. Switch to a clean directory and launch again:

DSH_HOME=/tmp/dsh-tianshu pnpm dlx @deepseek-ai/dsh plugin --profile tui add @huiliyi37/dsh-tianshu-tui
DSH_HOME=/tmp/dsh-tianshu pnpm dlx @deepseek-ai/dsh --profile tui

Do not run tsdown on this package in the DeepSeek Harness workspace root: it would write the unpublished @deepseek-ai/dsh-root into the bundle, and loading will always fail.

For the re-ask image capability, also assemble the companion package vision-ask/ in the same repo. The LSP model tool face (lsp_goto_definition / lsp_find_references / lsp_diagnostics) is now bundled in this package (companion plugin @huiliyi37/dsh-lsp, auto-inserted via the bundle patch) — install one TUI package and you get both the display bridge and the model tool face, sharing the same LSP server set (no double spawn). The TUI bridge's diagnostics source detection order: built-in companion plugin lsp service (getDiagnostics shape) → official ctx.lsp seam (deepseek-harness's dsh-lsp, adapted via query(getDiagnostics)) → built-in bridge fallback. ⚠ Users who previously installed the old community version (github:omdsh-dev/dsh-lsp) should plugin remove the old version before upgrading — assembling both old and new simultaneously will register the same model tool name twice.

Update notes

Current npm latest: @huiliyi37/dsh-tianshu-tui@1.0.0-rc.1 (GitHub Release).

1.0.0-rc.1 (2026-09-14): First candidate release of 1.0.0; version number enters the 1.0 line. Collects four convergences since 0.1.2-rc.31 — retry path body loss fix (after a successful LLM retry, the body of that step was previously being dropped wholesale: each host attempt is an independent attempt, so step-level dedup was a systematic mistake); failed attempts now get a ⟳ 未完成的尝试 marker prepended before their body, so it isn't read as a continuous answer with the successful body; assistant stream semantics consolidated into a single maintenance point in adapter/assistant-stream (previously implemented in four places, requiring four-way coordination for the same host behavior change); removed two dead states with no consumers (SessionManager snapshot layer, TranscriptView.streaming aggregate). ⚠ Requires host ≥ 0.1.5-rc.1; node ≥ 24.4.

0.1.2-rc.31 (2026-09-14): Fixes #58 — on a 0.1.5 host, the assistant body in a normal turn was not being rendered (attempt events only appear on error/interrupt paths; normal-turn body arrives as an embedded stream with assistant/message; now falls back to rendering at the message boundary, with dedup guarded by the same-step streaming delta). ⚠ Requires host ≥ 0.1.5-rc.1; also ensure node ≥ 24.4 (the 0.1.5 host CLI entry depends on import.meta.main, so older node silently produces no output).

0.1.2-rc.30 (2026-09-12): Host upgraded to 0.1.5 (official latest switched to 0.1.5-rc.1) — batched rewrite of streaming events assistant/chunk → assistant/attempt, new sessionPersistence contract (snapshot list + open/read handles), legacy library time rendering hardening; real-device pty e2e all green. ⚠ Requires host ≥ 0.1.5-rc.1.

0.1.2-rc.29 (2026-09-09): Host upgraded to 0.1.2-rc.1 (official latest is now stable, fixing #56 startup crash) — Session.events removed, replaced with snapshotEvents(); userQuestions now uses a waterfall answerer (global registration); CallId→ToolCallId; preset code→ptc; fork lineage now uses isSeeded + inheritedEventCount; bundle patch adds subagent-model-selection-settings; legacy hosts (≤0.1.1-rc.2) fail loudly at startup with upgrade/rollback guidance. ⚠ This upgrade requires host ≥ 0.1.2-rc.1.

0.1.2-rc.28 (2026-08-29): Responds to #55 vim improvements — mode-aware cursor shape (NORMAL inverted block / insert bar), two-phase history search input (n/N allowed in the edit phase, Enter then jumps, search target explicitly annotated), search-hit substring highlighting (including /scroll); also auto-refill of the input line on delivery failure + a consistency guard for the README key table.

0.1.2-rc.27 (2026-08-29): Flows back two Tianshu items — actionable at error moments (beyond error-at-bottom + guidance, the most recently delivered message is auto-refilled into the input line, ↩ indicates "may not have been fully processed", just edit and resend; successful turns clear the bottom material, drafts not preemptively written); plan-review decision card visual layering (dim decision divider + approve ❯ / success primary-action highlight, rendering unchanged when no theme is passed).

0.1.2-rc.23 (2026-08-27): LSP trio aligned with the upstream 0.6.0 official seam line (single lsp tool with four operations + local provider defaulting to tsserver); diagnostics source capability-gated to prevent the /lsp panel from degrading; host peer aligned to ^0.1.1-rc.2.

0.1.2-rc.22 (2026-08-27): First bundled version of the LSP model tool face (companion plugin auto-mounted; install one TUI package to get display bridge + model tool face; please remove the old community version first).

0.1.2-rc.21 (2026-08-27): Three community feedback fixes — input box cursor inversion (#50), preset defaulting to standard (#48), root-causing and reporting malformed SSE payload to the official side (#49).

For the full version history see CHANGELOG.md; in the TUI, /changelog to view (/changelog all for everything).

Highlights

• A full session workspace in the terminal — real-time rendering, append-only scrolling transcript, session recovery on launch, /fork to explore branches, /rewind to roll back (session truncate + optional file rollback), /export to export the Markdown transcript, mid-turn steering (/steer / Ctrl+T).
• End-to-end images — clipboard paste (Ctrl+V / terminal menu paste), inline rendering via terminal graphic protocol (kitty / iTerm2), delivery via the harness attachment service, letting vision-capable models actually see them — when the main model can't read images, it automatically turns them into descriptions via a separate vision model (vision bridge).
• A complete input surface — grok-style slash dropdown (fuzzy prefix match, MRU ordering, ghost preview), @-path Tab completion and @mention expansion, bracketed paste, optional vim key bindings, external editor (Ctrl+E), history search (Ctrl+F/Ctrl+R) — Ctrl+. brings up the full key table at any time.
• In-terminal interaction surface — structured question panel (numeric-key selection, plan-review feedback mode), pending approval cards with inline diff preview, mode cycling (Shift+Tab: normal → plan → always-approve), command palette, plus real-time status / config / skills / tasks / delegation tree / workflow panels.
• Reasoning visualization — the think channel flows as a live header line, collapses into compact lines in the scroll region (✻ 思考 (3.2s) · 12 行), Ctrl+O expands in place (par with competitors: collapsed by default).
• Personalized harness integration — /doctor terminal diagnostics, /memory project memory browser, /btw side-questions to the background agent, /model + /effort hot switching (immediately effective in the current session).
• Auditable by construction — the TUI itself registers no prompts, tools, or context faces; user input becomes ordinary log messages, and all rendered state is derived from session events.
• Evolves in lockstep with the harness — co-developed with the harness-side capabilities on top of the 2026-08-09 baseline snapshot (250+ commits): image/vision pipeline, DeepSeek Spark model engineering, session persistence and file snapshots, memory, verification gates and failure routing, code intelligence, git tools. See the next section.

Capabilities co-evolving with the harness (since the 2026-08-09 baseline)

The terminal UI evolved from Tianshu-Tui (Apache-2.0; per-file provenance in SOURCE-MAP.md). This bundle was then developed in sync with harness-side work on top of the DeepSeek Harness baseline snapshot snapshots/20260809T140917Z — 250+ commits from 2026-08-10 to 2026-08-13. The following capabilities live in the host harness (a separate package, not distributed with this bundle); the TUI is their primary interaction surface:

• Image pipeline & visual bridge — image ContentBlock added to the merge-extensible content vocabulary; dsh-llm-deepseek serializes user image blocks into OpenAI-style image_url content parts — user images are now end-to-end reachable on the wire (clipboard → input line → session → model request). Models declare vision capability via supportsVision (LlmModelInfo + llm-deepseek catalog). dsh-vision-bridge covers text-only primaries: at agent/pre-step it describes image attachments via a separate vision model (visionAutoBridge automatically picks the first vision-capable model when no provider/model is specified; fallback model + data URL validation; the prompt automatically selects between general-structure and OCR-level precise transcription based on UI/error keywords); the description is injected as a plugin-source user message — Model-visible ⟺ logged; bridge failures degrade to a visible hint, never failing the entire turn.
• DeepSeek Spark aliases — The official API has no spark model, and this host does not register a deepseek-spark provider. /model spark-flash / spark-pro map to the registered deepseek-official route, with wire models deepseek-v4-flash / deepseek-v4-pro respectively.
• Session persistence & file snapshots — Session.truncate rolls back the event log and resets derived state; the persistence backend adds deleteFrom and a truncate coordinator, surviving rollback across reloads; dsh-fs-snapshot ports FileHistory (trackEdit / rewindToBoundary), taking a snapshot before a write tool executes. TUI entry point: /rewind (session truncation + optional file rollback).
• Memory — dsh-memory (MemoryService + Markdown file backend, non-git fallback) and tool-memory (memory_save / memory_search + memory summary injection) provide cross-session recall. TUI entry points: /memory, /remember.
• Verification gates & failure routing — dsh-evidence-gate enforces RED-first verification: obligation state machine, edit/verification counters, TDD gate (enforce mode), probe suggestions + cooldown, L2 final-review gate; native integration with str_replace_editor and headless-agent assembly. dsh-agent-router predicts step failures based on turn history and routes work — including verification subagent dispatching and per-profile tool restrictions — with real-turn e2e coverage.
• Code intelligence & retrieval — dsh-semantic-index (BM25 + salience/RRF/vector fusion, incremental updates) exposed via the semantic_search tool; dsh-meridian code index (node:sqlite schema, tree-sitter parsers for TypeScript/Python/Go, graph/impact/flow queries, behavior signals, background backfill) exposed via repo_graph and <codebase-index> summaries; dsh-pheromone file-level pheromones + atomic JSON persistence, surfaced via file_info and the focus semantics of the read tool.
• Git service & tools — dsh-git service seam (GitLocal CLI provider, service-as-plugin) + dsh-tool-git single git tool for the model (operation discriminant: status / diff / log / commit), assembled into the base bundle.

Features

Session management

Capability	Description
/session new|list|switch	Create, list, switch sessions; on recovery, the full transcript is replayed through the same rendering bridge
Recovery panel	On launch, writes the list of recoverable sessions into the scroll region
/fork [directive] · /branch	Fork the current session (history copied to a new child session), optionally with a starting directive
/rewind	Roll back to a specified message — session truncation and/or file rollback to the pre-boundary snapshot
/export	Export the current session transcript as a Markdown file
/clear	Clear the current session's scroll region view

Input surface

• Slash command menu — Type / to open a dropdown menu: fuzzy prefix matching, ↑↓ / PageUp / PageDown to select, Tab to accept, Enter to submit, MRU ordering, argument placeholder ghost and input line ghost preview.
• Clipboard and image paste — Ctrl+V reads clipboard images (falls back to text); terminal menu paste detects images; paste paths that look like images are loaded as attachments; Alt+W / vim yank copies the selection to the system clipboard via OSC52.
• Image submission — Attached images display a 📎 N images marker; on submission they render inline below the user bubble and reach the model via the attachment service; the bubble carries a vision hint (forwarded / bridged via vision model / not sent). Oversized images are adaptively compressed before sending: long edge capped at 1568px (PNG preserves transparency), progressive JPEG 0.82 → 0.55 → 1024px + 0.55 until below the provider limit, always downscaling never upscaling.
• Vim editing key bindings (#51) — Esc enters NORMAL, key table aligned with Claude Code: h j k l / w e b W B E / 0 $ ^ / gg G / f F t T ; , navigation and find replay, d c y × motion + numeric prefix, text objects iw aw iW aW, line-level dd cc yy Y and p P paste, x X D C s S r o O J u .; v/V visual selection includes cursor-end characters. Multi-line drafts keep column movement with j/k; Chinese consecutive segments form words without per-character jumping. Runtime /vim toggle, /vim default sets the startup default. Two-key insert sequence → Esc (vimInsertRemaps, e.g. {"jj":"esc"} written to prefs; 1-second window to prevent false triggers). Cursor shape differs by mode: NORMAL inverted block, insert bar (#55). NORMAL / opens session history search (two-phase: input filters in real time, after Enter confirmation n/N jumps — the search term may contain n/N).
• Editing — external editor (Ctrl+E), Tab file completion, @mention expansion, input history, multi-line input, bracketed paste (multi-line / long text pastes into the input line as a whole block, not submitted line by line); the input line is drawn as a complete rounded-rectangle frame.
• Mid-run queue (par with Claude Code queue) — messages submitted while the agent is running enter a local queue above the input track (instant echo, not sent directly); on turn end they're auto-delivered in order; on interrupt, no delivery (leaving room for you to pull back with ↑); empty input ↑ pulls the queue head back to the input line; switching sessions discards the queue and echoes the count. Mid-turn immediate correction still goes through /steer / Ctrl+T.
• Image re-ask — the companion plugin @deepseek-ai/dsh-vision-ask in the same repo registers sent images and answers targeted model questions via ask_image (see vision-ask).

Rendering & Projection

• Conversation flow — markdown rendering, tool-family coloring + per-tool timing, parallel tool calls collapsed into groups.
• Live tool card resolution — settled tool results are rendered as scrolling-region cards per the harness presenter's intent: diff results render as structured red/green file diffs (shared with the approval preview), terminal results include a command title + cwd + exit/signal badges, and the rest collapse into text cards.
• Reasoning channel — real-time shimmer header line while thinking, per-segment collapsed scroll row at end of segment, Ctrl+O to expand full text in the live area.
• Fluency collapsing — repetitive routine tool traffic collapses under the quiet policy; compact mode (/density) keeps only the header line.
• Turn status — braille spinner + phase text status line, workflow run summary, delegation tree, task pane, config/skills panel as live-region panels; when a turn ends (not interrupted) and there are tool calls, a dimmed summary line is emitted (turn N · 读X 改Y · 耗时).
• Subagent runs — one live spinner line per run; terminal states land as ✓/✗/◌ entries in the scrolling region.
• Window chrome — welcome page (brand header, friendly short session id, environment check line), top bar (cwd + git branch + model), bottom three-line region: input line (bottom border colored by mode) → footer (mode badge + shortcut hints) → metrics line (model / token usage / cache hit rate).
• Themes — built-in palette + custom:<name>; automatic terminal detection with 16-color fallback; respects NO_COLOR; when a custom theme loads, contrast warnings are issued against the declared background (WCAG < 3.0, non-blocking).

Interaction panels

• Structured questions — numeric-key selection, Esc to cancel, overlap protection; plan-review feedback mode (f to enter, Enter to submit Keep planning + custom feedback).
• Approval cards — y/N/Ctrl+C to settle pending approvals; inline diff preview when the tool can produce one; blind-approval hint when diff is not viewable; requests from non-current sessions are delegated to the next listener.
• Mode cycling — Shift+Tab cycles normal → plan → always-approve; plan state drives the footer badge, always-approve is a session-level local state (reset on switch/exit).
• Real-time panels — /status (goal/todos/plan projection snapshot; subagent domain in /subagents), /config (terminal notifications (system notification + BEL beep on completion, beep also reachable over SSH sessions) / compact rendering + host settings / permissions / credentials), /skills browse (↑↓ for details), /tasks pane, /subagents delegation tree, /workflow runs. Inspection panels are mutually exclusive, Esc to close. When host services are not assembled, the corresponding sections collapse; /config's terminal section is independent of the host. Other panels echo a ⚠ warning when the backing service is missing (no silent blanks).
• Command palette (Ctrl+P) / key table (Ctrl+.) / history search (Ctrl+F/Ctrl+R) overlay.

Model and vision

• /model — view and switch models (default + current session hot-switch); spark-flash / spark-pro aliases map to deepseek-official + official wire id deepseek-v4-flash / deepseek-v4-pro. /model <provider/model|alias> [off|high|max] sets the reasoning level within the same command.
• /effort — set the reasoning level (off / high / max; auto reverts to the model default), current session hot-switch.
• Vision bridge — image reading capability is declared per model (supportsVision, auto-refreshed via the llm catalog) and drives bubble hints; when the main model can't read images, the auto-selected vision model generates image descriptions before submission (one-shot path; see known limitations). Bridge availability sources: the assembler passes vision.bridgeEnabled, or the host vision bridge plugin provides the visionBridge service (the TUI auto-detects by service existence before submitting images); if neither is present, images are not sent and a warning is shown.
• Vision copilot — after assembling the companion plugin @deepseek-ai/dsh-vision-ask, each sent image is registered with a short id (img_1 …), and the model can repeatedly ask via ask_image — targeted questions, different angles, unlimited times; repeated same-image same-angle questions hit the per-image description cache. Details and configuration in the vision-ask README.
• /mcp — lists connected MCP servers and tool counts; tools <name> views a server's tool inventory.

Commands

Command	Action
/session new|list|switch	Session management (list/selector grouped by today/yesterday/this week/older)
/fork [directive] · /branch	Fork the current session, optionally with a starting directive
/rewind	Two-phase rollback (message list → granularity)
/export [path]	Export transcript as Markdown
/scroll	Paginated viewer: browse full transcript (scroll / live search / n·N jump / g·G first/last)
/clear	Clear scroll region view
/compact	Compact session context
/steer <text>	Mid-turn steering (correct direction without interruption)
/model [target] [effort]	View/switch model (aliases: spark-flash, spark-pro)
/effort off|high|max|auto	Set reasoning level (hot switch)
/theme [name]	Switch theme
/density	Toggle compact tool card rendering
/vim [on|off|default]	Toggle vi/vim editing keybindings; default writes the startup default
/lsp	Toggle LSP diagnostics panel (auto-fetches diagnostics for files the agent touches; diagnostic badges on tool cards)
lsp_goto_definition · lsp_find_references · lsp_diagnostics	LSP model tool surface (companion plugin lsp/ registered; goto definition / find references / file diagnostics)
/status	Toggle status panel (goal/todos/plan projection + session summary section)
/config [notify [on|off]]	Toggle settings panel (empty input n notifications, d density). No-arg toggles the panel
/skills	Toggle skill browser panel
/tasks	Task pane (background tasks)
/goal	Goal management (create / pause / resume / complete / block)
/subagents	Delegation tree panel
/workflow	Workflow run panel
/btw <question>	Ask a side question to the background agent
/remember <text>	Save a memory entry
/memory	Memory browser (list / filter / delete / preview)
/doctor	Terminal diagnostics + fix guidance
/mcp [tools <name>]	List MCP servers; view a server's tools

Shortcuts

Key	Action
Enter	Send
Shift+Enter	Newline (or \+Enter for continuation)
Ctrl+N	New session
Ctrl+S	Restore most recent session
Ctrl+Q	Quit (same as /exit)
Ctrl+P	Command palette
Ctrl+.	Keybinding reference overlay
Ctrl+F / Ctrl+R	History search (type to filter; after Enter confirm, n/N next, p/P previous)
Ctrl+O	Expand/collapse the most recent reasoning block
Ctrl+E	Open input line in $EDITOR (configurable via editorKey)
Ctrl+T	Mid-turn steering
Ctrl+Enter	Preempt: interrupt the current turn and immediately send the draft (requires terminal support for the kitty keyboard enhanced protocol; RIVET_KITTY_KEYBOARD=1 forces it on)
Ctrl+C	Interrupt the in-flight turn (double-tap on idle empty input to quit)
Ctrl+V	Paste clipboard image (falls back to clipboard text when no image)
Alt+W	Copy selection to system clipboard (OSC52)
Shift+Tab	Cycle modes: normal → plan → always-approve
Tab	@-path completion; accept selected slash menu item
Ctrl+U	Delete to beginning of line
↑/↓	Input history (acts as selection when slash menu is open; with queued messages, ↑ on empty input retrieves the head of the queue)
PageUp/PageDown	Slash menu page up/down
Esc	Close menu/overlay/inspection panel; cancel pending question; double-tap when idle to rewind
t	Approval card: remember this tool (auto-approve this tool within the session; other tools still require per-card approval)
a	Approval card: allow within session (always-approve + settle current request)

Assembly

The bundle patch inserts the tui-runner plugin on top of dsh-base:

- id: tui-runner
  name: '@huiliyi37/dsh-tianshu-tui'

TuiRunnerConfig (all optional): stdin/stdout (stream injection, defaults to process streams), initialSessionId, editorKey (defaults to ctrl_e; ctrl+o is reserved for reasoning expansion), vimEnabled (defaults to false), vision (supportsVision / bridgeEnabled / bridgeSource; when not provided, supportsVision is auto-refreshed via the llm catalog and bridgeEnabled is auto-detected by the host visionBridge service's presence — the visual bridge plugin should provide this service when assembling), workflowHistoryLimit (defaults to 50), lsp (enabled / timeoutMs; enabled by default with a 2000ms per-fetch timeout — local language service bridge: when the agent touches a file, an LSP server is lazily started by extension (typescript available by default via npx; pyright/gopls/rust-analyzer/clangd/jdtls detected on PATH) to fetch diagnostics, displayed on tool card badges and the /lsp panel; diagnostics go only into the TUI local display cache, not written to session events, and no model surface is registered).

Service dependencies: sessions/agents/agentDefaultModel are required (mandatory injection); goals/subagents/memory/compact/tasks/skills/sessionProjections/workflowEngine/planMode are optional — when not assembled, the related commands and panels fail loudly with "unavailable" reports, never silently swallowing, and never blocking TUI startup.

Verification

npm test

Model Experience

None — the TUI renders logged session events and forwards regular user input; it registers no prompts, tools, or context surfaces.

KV Cache Impact

No direct impact; user input submitted via the TUI becomes a regular log message, and its request impact belongs to the session and loop packages.

Known Limitations & TODOs

• Image re-asking requires a companion plugin — The ask_image tool and session image registry live in @deepseek-ai/dsh-vision-ask (a separate package in the same repo); the TUI bundle itself does not ship them. Without the companion plugin, already-sent images cannot be re-asked, and repeated descriptions at the same angle will invoke the vision model again; the visual bridge still covers the one-time submission description path.
• LSP bridge runtime limitations — The model tool surface (goto/find/diagnostics) is bundled (@huiliyi37/dsh-lsp); deep integrations like Tianshu edit-diff's diagnostics-narrowing remain future work. When server initialization exceeds the timeout (default 2s), diagnostics are silently absent; the next file touch re-fetches; large repos keep tsserver in resident memory (lazy start mitigates, no idle reclamation); switching sessions does not restart the server (rootUri reuses the first session's cwd).
• app.ts monolith (~3.2k lines) — The pending state machine is already controllerized (question/approval), but rendering composition and key arbitration remain in app.ts; the C4 split plan (pure-function panel sections) is ongoing.
• Projection layer partially wired — Of the four pure fold models, turn-summary (turn/end summary line) and summary-state (/status session summary section, still functional when the host projection bus is absent) are wired; activity-status/activity-store are intentionally left unwired (statusline is a self-contained projection, no benefit to replacing; activity-store has no consumer yet). Current state is documented in docs/projection-layer.md.

License & Provenance

Apache-2.0. The terminal rendering engine evolved from Tianshu-Tui (Apache-2.0); per-file provenance and modification notices are in SOURCE-MAP.md and NOTICE.

Friend Links

Project	Description
dsh-web-ui	DSH Web UI plugin and skin collection
dshfind	DeepSeek Harness Chinese learning and sharing community
deepseek-harness-ux	Long tasks without screen flooding: key progress clearly visible, auto-collapses on completion
dsh-TUI	Claude Code-style full-screen interactive terminal plugin
DSH-better-sidebar	Full sidebar workbench: third-party Tabs, file/terminal/Git/subagents
DSH Desktop	DeepSeek Harness community desktop app (Electron, Windows x64 / macOS Apple Silicon installers), download and run, no command-line environment setup needed. Built-in local Harness host and plugin system, supports iOS/Android remote task dispatching and agent progress tracking.
dsh-whale-report	Turns session, token, cost, tool call, and risk anomalies into readable Agent reports