dsh-code

A DeepSeek Harness terminal coding agent for developers who prefer a focused TUI workflow.

English · 简体中文

[!IMPORTANT] dsh-code is an independent community project, not an official DeepSeek distribution. DeepSeek Harness is also in developer preview, so compatibility-breaking upstream changes may occur between pinned baselines.

Why dsh-code?

DeepSeek Harness provides an official Web UI and a plugin-first agent runtime. dsh-code is for developers who prefer to stay in the terminal: it packages the same DSH agent semantics in a compact, keyboard-driven interface that works naturally beside shells, editors, Git, and remote development environments.

The product draws on the interaction ideas of Pi and uses @earendil-works/pi-tui for terminal rendering. It does not fork or replace the agent core. The Agent Loop, sessions, model adapters, tools, sandbox, permissions, MCP, Skills, Plan/Todo, and sub-agents remain owned by the pinned DSH runtime.

In short:

DeepSeek Harness agent runtime + Pi-inspired terminal UX + pi-tui renderer

Requirements

Linux, macOS Intel/Rosetta, Windows ARM, and standalone installations without Node.js are not supported in the first release.

Installation

npm

npm install -g @tsingwill/dsh-code

Install the current release candidate instead of the stable channel:

npm install -g @tsingwill/dsh-code@next

Verify the installation:

dsh-code --version
dsh-code --help

The scoped npm package is @tsingwill/dsh-code; the installed command remains the shorter dsh-code.

Build from source

git clone --recurse-submodules https://github.com/guoxiucai/dsh-code.git
cd dsh-code
corepack enable
corepack prepare pnpm@11.7.0 --activate
pnpm install --frozen-lockfile
pnpm run build:lib
pnpm run build
node lib/bin.js

Preview

Highlights

• Terminal-native workflow — streaming Markdown, five-line collapsed reasoning and verbose tool bodies, always-visible line-numbered file diffs, selectable/copyable results, themed paste markers, shell blocks, and a bottom-pinned composer.
• Standard and PTC modes — Standard makes direct tool calls; PTC composes multi-step work in one TypeScript program. PTC child calls use native tool rows, with long bodies collapsed and file diffs always visible.
• DeepSeek Harness semantics — uses DSH's public session/events and services; there is no second agent loop, session store, permission engine, or tool registry.
• Model setup in the TUI — configure DeepSeek, OpenAI, or an OpenAI-compatible endpoint through an inline, reversible wizard.
• Safe project startup — canonical-path trust records and read-only, workspace-write, or danger-full-access permission presets.
• Persistent sessions — continue the latest session, search/resume/delete history, inspect session statistics, fork a completed turn, and compact context.
• Agent visibility and decisions — dedicated Plan/Todo states, queued-user message feedback, tool progress, retry and compaction indicators, one-shot approval bars, structured questions, plan review, and a clickable active sub-agent indicator with cancel/remove controls.
• Fast terminal controls — slash-command completion, fuzzy @ file and folder completion, direct ! shell mode, inline selectors, and keyboard-first navigation.
• Independent installation — stores product data under ~/.dsh-code, keeps a separately installed dsh command untouched, and supports explicit updates.
• Adaptive visuals — a DeepSeek-blue palette tuned independently for dark and light terminal backgrounds.

Architecture

dsh-code is deliberately a thin terminal host over a fixed DSH baseline:

flowchart TB
  User["Terminal user"] --> CLI["dsh-code launcher"]
  CLI --> TUI["Terminal host<br/>Pi-inspired UX + pi-tui"]
  TUI --> API["Public DSH services<br/>session/event + AgentHandle"]
  API --> DSH["@deepseek-ai/dsh-base<br/>Standard / PTC Agent Preset"]
  DSH --> Runtime["Agent Loop · Sessions · Models · Tools<br/>Sandbox · Permissions · MCP · Skills<br/>Plan/Todo · Sub-agents"]

The launcher owns only product concerns: command parsing, ~/.dsh-code home isolation, project trust, session selection, profile initialization, updates, and delegation to the upstream DSH executable. The TUI renders structured events and sends input back through the public AgentHandle API. The TUI exposes only the upstream standard (Standard) and code (PTC) Agent Presets. New sessions default to Standard and may select PTC at startup or via /mode before the first turn. The mode locks when that turn starts; resume always reconstructs the preset recorded in the event log, so an existing tool history never changes schema underneath itself.

See the accepted architecture decisions in docs/adr/ and the exact upstream revision in UPSTREAM_BASELINE.md.

Quick start

cd /path/to/your/project
dsh-code

On the first launch for a project:

1. Review the canonical project path and choose a permission preset.
2. If ~/.dsh-code/.credentials.yaml has no saved API credential, dsh-code automatically opens the inline provider setup.
3. Select a provider, save the first API token and default model, then send a task in the editor. Use /config later to add or change providers.

For the official DeepSeek API, /config asks for the API key and default model. For an OpenAI-compatible service, the wizard keeps five explicit values:

1. provider route ID;
2. base URL;
3. credential environment-variable name (pre-filled from the route ID);
4. API key;
5. model ID.

The wizard uses DeepSeek-compatible examples, supports Esc to return to the previous step, and writes values only after the final step succeeds. Credentials are stored owner-only in ~/.dsh-code/.credentials.yaml.

Usage

Command line

Interactive commands

Additional commands supplied by the pinned DSH profile remain discoverable through / autocomplete.

Essential keys

Approvals and structured questions

When a tool requests sandbox escalation or a hook returns ask, dsh-code pins a one-shot Allow once / Reject bar above the composer. The choice applies only to that request; Esc cancels it and does not create a persistent grant.

The DSH ask_user_question tool and plan-mode review use the same bottom-pinned interaction area. Single-choice questions submit immediately; multi-choice questions use Space to toggle entries and Enter on Continue to submit. Select Type an answer… and press Enter to edit that same row in place; no panel content shifts. While editing, Esc returns to the preceding menu. Long plan Markdown is bounded to six lines and scrolls with PgUp / PgDn. Parallel requests are queued so one prompt cannot replace another.

Sessions, configuration, and isolation

By default all dsh-code state lives under ~/.dsh-code:

~/.dsh-code/
├── .credentials.yaml       # owner-only provider credentials
├── profiles/dsh-code/      # fixed DSH profile + terminal host patch
├── projects/               # canonical-path trust records
└── sessions/               # persisted sessions grouped by project

Set DSH_CODE_HOME to use a different root. On launch, dsh-code sets the delegated DSH_HOME to this isolated directory and disables DSH telemetry. It does not import or overwrite standalone DSH settings, credentials, sessions, plugins, or MCP configuration, so a globally installed upstream dsh binary remains independent. The bundled DSH skill registry does perform read-only discovery from compatible skill roots: project .dsh/.agents/.codex/.claude, dsh-code user ~/.dsh-code/skills, and user ~/.dsh/.agents/.codex/.claude skill directories. dsh-code does not install, delete, copy, or update those skills; /skills reports the winning upstream registry entry and its source in a single-level picker. Space enables or disables the selected entry and Enter places an enabled, user-invocable skill in the editor. Its enable/disable switch is a dsh-code-only overlay stored at ~/.dsh-code/skill-preferences.json; it never edits the source SKILL.md or changes another product's skill state. A disabled skill is hidden from both the model catalog and user slash invocation in dsh-code.

dsh-code owns separate MCP configuration at ~/.dsh-code/mcp.json (user scope) and .dsh-code/mcp.json (trusted-project scope). /mcp shows only those owned rows by default; Import from other agents… performs an on-demand, read-only scan of standalone DSH, OpenAI Codex, and Claude Code. The imported snapshot is then managed independently and never follows or modifies its source. Space enables or disables a row, project scope overrides a same-named user row, and add/import/edit/remove operations hot-reconcile the public upstream MCP client inside the current process. A green ● connected means at least one tool is registered; connecting, disabled, overridden, error, and not-connected states have distinct indicators. Imported environment/header values may contain credentials, so both files are private (0600 where supported) and the project file is gitignored. Legacy project MCP rows in .dsh-code/cordis.patch.yml are migrated once while unrelated project plugin rows remain intact. Stdio server stderr is isolated from the alternate-screen UI and written to rotating private logs under ~/.dsh-code/logs/mcp/<server>.stderr.log.

Updating

Updates are explicit; dsh-code does not silently update itself:

dsh-code update --check
dsh-code update
dsh-code update --channel next
dsh-code update --version 0.1.1

The update command is supported for npm-global installations. Source checkouts should be updated with Git and rebuilt with the same tool that installed them.

Development and verification

pnpm run typecheck
pnpm test
pnpm run build

The repository pins DeepSeek Harness as the deepseek-harness/ git submodule. Product code stays at the repository root; upstream changes belong in a dedicated baseline update or should be contributed to DSH first.

Release design, platform compilation, candidate verification, and the update strategy are documented in docs/NPM_RELEASE.md.

Contributing and security

• Read CONTRIBUTING.md before opening a pull request.
• Use GitHub Issues for public bug reports and feature requests.
• Report vulnerabilities privately as described in SECURITY.md.
• Never attach unredacted API keys, session logs, credentials, or crash logs.

Relationship and attribution

dsh-code is a downstream, independent community project. It is not affiliated with or endorsed by DeepSeek AI or the Pi maintainers.

• Agent runtime: DeepSeek Harness
• TUI renderer and interaction inspiration: Pi
• Product distribution and terminal host: this repository

The dsh-code terminal-whale logo adapts the official DeepSeek whale silhouette with a terminal window and prompt. The DeepSeek name and official whale artwork belong to their respective owners; see NOTICE for complete attribution.

License

MIT © 2026 guoxiucai. Third-party notices are listed in THIRD_PARTY_NOTICES.md.