dsh-cctui
Maintenance: Activeagentforce314/dsh-cctui
Claude Code style TUI for Deepseek-Harness
Install
dsh has no central install command — add this plugin’s entry (documented in its README below) to your profile or patch config, then restart.
How installs work2
stars
0
forks
TypeScript
Language
NOASSERTION
License
2026-08-18
Created
2026-08-19
Last push
README
dsh-ccTUI
Claude Code style TUI for Deepseek-Harness — 🐳 ported from the clawcodex ui-tui and
packaged as a dsh bundle plugin (npm package dsh-cctui).
The full Claude-Code-style look, feel, and interactions — streaming markdown transcript, tool trail
with diff cards, approval and question prompts, plan review, session switcher and resume,
slash commands, model picker, context bar — running in-process against harness services
(ctx.agents, session/event, the approval waterfall, ctx.userQuestions, ctx.commands).
██████╗ ███████╗██╗ ██╗ ██████╗ ██████╗████████╗██╗ ██╗██╗
██╔══██╗██╔════╝██║ ██║ ██╔════╝██╔════╝╚══██╔══╝██║ ██║██║
██║ ██║███████╗███████║ █████╗ ██║ ██║ ██║ ██║ ██║██║
██║ ██║╚════██║██╔══██║ ╚════╝ ██║ ██║ ██║ ██║ ██║██║
██████╔╝███████║██║ ██║ ╚██████╗╚██████╗ ██║ ╚██████╔╝██║
╚═════╝ ╚══════╝╚═╝ ╚═╝ ╚═════╝ ╚═════╝ ╚═╝ ╚═════╝ ╚═╝
🐳 Claude Code style TUI for Deepseek-Harness
╭─── dsh-ccTUI v0.3.0 ─────────────────────────────────────────────────────────────────────────╮
│ │
│ Welcome back, you │ ▾ Available Tools │
│ │ harness: bash, create_goal, edit, …+22 │
│ : ' : │ │
│ ___'_______ │ ▸ Available Skills (0) │
│ /' o \--. │ │
│ | ___\ │ ────────────────────────────────────────────── │
│ \____________/ \/ │ 25 tools · 0 skills · /help for commands │
│ \__/ \__/ │ │
│ │ │
│ ──────────── │ │
│ Model deepseek-v4-flash │ │
│ Path ~/workspace/dsh-ccTUI │ │
│ Perms Ask for approval · /permissions │ │
│ Session cctui-56b37bee-41fd-4feb-b270-5988… │ │
│ │
╰──────────────────────────────────────────────────────────────────────────────────────────────╯
Install
Requires Node ≥ 22.19, npm, pnpm, and the dsh CLI (npm install -g @deepseek-ai/dsh).
git clone https://github.com/agentforce314/dsh-ccTUI.git
cd dsh-ccTUI
./install.sh # builds, then: dsh plugin --profile dsh-cctui add "$PWD"
dsh --profile dsh-cctui # launch (or ./bin/dsh-cctui.js)
Model/provider configuration comes from your dsh profile (agent-default-model settings or
an - id: cctui config override in the profile's cordis.patch.yml: provider, model,
cwd, sessionId).
API key
The TUI stores no key; the harness resolves one per request. Write it to the managed store
at $DSH_HOME/.credentials.yaml (default ~/.dsh/.credentials.yaml) — a YAML mapping of
credential reference to value and nothing else:
DEEPSEEK_API_KEY: sk-…
chmod 600 that file and chmod 700 ~/.dsh: dsh-credentials-local refuses to read a
document carrying any group or other permission bit, and fails at boot naming the repair.
The document is watched, so a key stored while the TUI is running takes effect on the next
request — first run is "browse models, store the key, prompt again", no restart in between.
Until one resolves, the route stays registered and /model stays browsable; it is the
request that fails, with MISSING_CREDENTIAL naming every entry point it looked at.
The launching environment beats the store (DEEPSEEK_API_KEY=… dsh --profile dsh-cctui) and
is deliberately read-only from inside; the store in turn beats <cwd>/.env and
~/.dsh/.env. Keep the key itself out of cordis.patch.yml — adapter config carries only
apiKeyEnv, the reference to resolve. One key covers both the conversation and the
web_search tool.
Mode 0600 stops other OS users, not the model: the harness never hands it the document's
path and never loads the value into the environment, but bash and the filesystem tools run
as you and the shipped policy confines writes rather than reads.
Highlights
- Conversation loop: streamed deltas render live; reasoning behind Ctrl+O/Ctrl+R density toggles; busy verbs and spinners; Esc interrupts (Ctrl+C never kills the app).
- Tools: Claude-style
⏺ Tool(args)/⎿ resulttrail, each result summarised the way the original does it —Read 3 lines,Found 37 files,Received 1.2KB (200 OK),Did 1 search in 2s— with a failed call carrying the tool's own message and a running one its own clock; write/edit diffs render as structured diff cards; sandbox-escalation approvals pop the approval box (1approve /2+Enter deny); todo lists pin under the busy line. - Subagents: a delegation reads
⏺ Subagent(Review the diff)/⎿ Done (2 tool uses · 1.2k tokens · 11s), and its children stream into/agents— goal, live tool calls, tokens, spawn tree. - Sessions:
/sessions(Ctrl+X) lists live and persisted sessions;/resume <id>replays a persisted transcript, tool results and all;/new,/title,/rename. - Commands: every harness
ctx.commandsentry (e.g./plan,/goal) appears in the completion menu and dispatches through the harness;/modelopens the picker backed by the llm catalog;/effort,/context,/usage,/help,/status. - Modes: Shift+Tab cycles default → plan → bypassPermissions (mapped onto the harness plan-mode controller and approval policy).
Configuration
Environment knobs use the DSH_CCTUI_ prefix — e.g. DSH_CCTUI_INLINE=0 (alternate screen
instead of inline scrollback), DSH_CCTUI_THEME=light|dark, DSH_CCTUI_HOME (data dir,
default ~/.dsh-cctui), DSH_CCTUI_FPS=1. CLAUDE_CODE_SCROLL_SPEED is still honored as a
migration fallback for people coming from Claude Code.
Development
npm install
npm run typecheck && npm test # 149 files / ~1900 tests
npm run e2e # PTY e2e against a real dsh boot + scripted LLM
npm run e2e:install # the real `dsh plugin add` install path
Tool rows are the fiddliest surface to eyeball, so they have their own harness:
python3 scripts/tool-gallery.py [tool…] [--expand] drives a real dsh boot through one
harness tool per scenario (via the scripted test/e2e/probe-llm.mjs) and prints the trail it
rendered — the thing to diff against the same call made to Claude Code.
Versioning
Each shipped change increments the patch digit by one; it runs to 99 before rolling over into the minor digit:
0.2.0 → 0.2.1 → … → 0.2.13 → … → 0.2.99 → 0.3.0 → 0.3.1 → …
Write patch numbers without leading zeros (0.3.1, not 0.3.01 — the latter is not valid
semver and npm rejects it). Bump package.json in the change's own commit, then tag
v<version> on main after it merges.
Architecture and porting details: docs/ARCHITECTURE.md,
docs/PLAN.md, docs/PORTING-NOTES.md. The adapter
boundary (src/harness/ is the only place that may import @deepseek-ai/*) is CI-gated.
Provenance & license
MIT. Substantial portions are ported from the MIT-licensed clawcodex ui-tui (including its
forked Ink renderer) with integration patterns from the MIT-licensed dsh-TUI project — see
NOTICE.md.
More in CLI & Terminal
deepseek-reasonix
by esengine
DeepSeek-native AI coding agent for your terminal. Engineered around prefix-cache stability — leave it running.
dashi-taskboard
by chuspeeism
Task board for AI coding agents, available as a CLI or plugin for DeepSeek Harness, Claude Code and Codex.
mnemon
by mnemon-dev
LLM-supervised persistent memory for AI agents — graph-based recall, cross-session knowledge, single binary. Works with DeepSeek Harness, Claude Code, OpenClaw, and any agent runtime.
sivtr
by ariestar
A unified agent memory workspace for human and agent
