DeepSeek Harness for VS Code

Your coding agent, with every change in view.
Bring DeepSeek Harness (DSH) into VS Code: work with your code, review native diffs, and follow each task with built-in Trace and usage insights.

English | 简体中文

Install for VS Code · Open VSX · Download VSIX · Changelog

An independent community project. Issues welcome.

For JetBrains IDEs (IDEA, PyCharm, etc.), please see dsh-intellij-integration.

Highlights

Why DSH?

• See what changed. Review tool edits in VS Code's native side-by-side diff, even outside a Git repository.
• Decide before execution. Approval cards show commands and target files, with proposed diffs for supported file writes.
• Start with context. Bring files, selections, Git diffs, or paused debugger state into a task without copying everything by hand.
• Pick up where you left off. Resume persistent sessions and follow tools, subagents, Todos, and token usage in the Activity panel.
• Smarter agent decisions, no extra install. Turn on the built-in Laya/Jev integration to catch loops, trim noisy output, and check completion claims against evidence.

Quick start

Requires VS Code 1.106.0 or later and a configured DSH model provider with credentials.

1. Install the extension from the Marketplace or Open VSX links above, or search for harcochen.dsh-vsc-integration in Extensions.
2. Open chat. Open and trust your project folder, then run DSH: Open Chat from the Command Palette. The extension automatically starts or connects to a Runtime; by default, it attempts a managed Runtime download when no usable environment is available.
3. Set up your provider. Run DSH: Configure API Key for DeepSeek credentials. For other providers, use DSH: Open dsh Web UI in Browser. Select or register a DSH Workspace, then choose a model.
4. Give it a task. Type @ to reference a file, or right-click a selection for DSH actions. Follow the task, respond to approval requests, and open diffs from tool cards to review the result.
5. (Optional) Turn on Laya/Jev. Enable dsh.jev.enabled and the features you want, run DSH: Configure Jev API Key, then accept the Runtime restart. See Laya / Jev System One decisions.

A DSH Workspace groups sessions in Harness and can be associated with a project path. When using the same Runtime, you can continue sessions created in the Web UI.

Try it on real work

Features

Native diff for every edit, no Git required

After a write/edit tool call, open the target file to see VS Code's native side-by-side diff. The before-image is reconstructed by replaying hunks backwards from the Session log, so it also works in non-Git repositories and Git-ignored files.

Preview before approval

The approval card shows the actual command line, working directory, and target files that will be written. For supported file-writing tools, open a native diff of the proposed change before approving it.

A write whose target file still has unsaved editor changes is not released: approval is refused, the card names the files and stays pending, and you can save or revert them and approve again.

Laya / Jev System One decisions, built in

Jev is TypeSafe's System One decision model. It does not generate text. It reads an agent state and returns calibrated, typed answers in milliseconds. Laya is an Apache-2.0 open-weights alternative that serves the same POST /v1/systemone request and response shape. The extension bundles the dsh-jev-integration Runtime plugin (shared with the JetBrains plugin) and mounts it into every Runtime it starts. You do not install a separate plugin or edit your DSH profile.

Use Jev: enable dsh.jev.enabled plus the features you want, run DSH: Configure Jev API Key, and restart the Runtime when prompted. The key is encrypted in VS Code SecretStorage.

Use Laya: point dsh.jev.baseUrl at an HTTPS endpoint that serves the Laya System One API, set dsh.jev.model to the model name that server expects, and configure an API key. If your server ignores the key, any placeholder value works. The extension accepts only https:// endpoints and falls back to the TypeSafe default for anything else.

Everything is off by default. Enabled features may send the samples they need to the configured endpoint. Existing or externally managed Runtimes are never modified.

Sessions, account, and schedules (dsh 0.1.7-rc.2)

• DSH: Manage Sessions pins sessions and restores archived ones. DSH: Archive Session asks whether to stop a running task before it archives the session. DSH Workspaces also work without an open VS Code folder.
• DSH: Manage DeepSeek Account signs in through the browser, signs out, and shows the account profile and balance.
• The Schedule dock lists reminders across sessions. You can edit daily, weekly, or cron rules, view delivery history, and delete reminders.

Slash commands enumerated live from the Runtime

The slash menu dynamically fetches commands registered by the Runtime for the current session (/plan, /compact, /goal, etc.) and merges them with the extension's own IDE commands.

Editor and Git context

• Right-click the current file, selection, or Git diff to explain, fix, review, or generate documentation.
• Right-click Ask about resource in Explorer to ask about a file or folder.
• The @ menu autocompletes project files and previous Sessions.
• DSH: Capture AppShot (macOS only) captures a window screenshot and inserts it into the conversation as a draft.

Sessions, Trace, and Activity at a glance

The sidebar provides a native conversation-outline TreeView. Trace, token usage, Todo lists, and subagents are gathered in the Activity panel. The UI supports VS Code's dark and light themes.

DSH: Open Chat in Editor Tab mirrors the same conversation in an editor tab, so the chat can sit next to the file you are editing. Both surfaces show one session and one stream — switching between them does not restart or fork anything.

Autonomous debugging (off by default)

With dsh.autonomousDebugging enabled, the extension exposes a loopback MCP endpoint inside this window and the agent can drive the VS Code debugger: debug_start launches a launch configuration that already exists in the workspace, debug_breakpoint adds, removes and lists breakpoints, debug_control continues, steps and waits for the next pause, and debug_context reads the paused stack, variables and source. Variables whose names look like secrets are replaced with [redacted by dsh-ide] before they leave the window. The endpoint binds 127.0.0.1 only, checks the Host header and a per-launch token, and never lets the model invent a launch configuration. It applies to a Runtime this window starts; switching the setting needs a Runtime restart, which the extension offers when you change it.

Credentials and balance

The bottom bar shows your current balance, including peak and off-peak pricing. Low balances are highlighted clearly.

FAQ

Do I need to install DSH manually? Usually no. The extension looks for a usable local environment and attempts to download a managed Runtime when needed. The first download requires network access; dsh.installWhenMissing controls automatic installation.

How is Laya/Jev integrated? The extension package carries the IDE-neutral dsh-jev-integration Runtime package and mounts it only into a Runtime started by this extension; existing or externally managed Runtimes are not modified. Laya is supported through the same System One contract by changing dsh.jev.baseUrl (HTTPS only). Jev is disabled by default. dsh.jev.enabled is the master switch; individual switches enable loop guard, result shaping, completion evidence checks, tool pruning, skill routing, decision tools and local deterministic safety checks. The endpoint and model are configurable, while numeric thresholds use built-in defaults. Configure the API key with DSH: Configure Jev API Key; it is encrypted in VS Code SecretStorage, with TYPESAFE_API_KEY and $HOME/.dsh/.env available as fallbacks. Enabled features may send their required samples to TypeSafe System One. Changes require a Runtime restart. This build exposes the shared Runtime subset of upstream dsh-jev, not its full agent-loop, dashboard, browser, or mobile bundle.

Can I connect to an existing Runtime? Yes. Set dsh.serverUrl to your running dsh web address and set dsh.serverToken to its launch token when the token is not already in the URL. This extension accepts valid SemVer versions at or above dsh 0.1.5-rc.1, including newer prereleases and stable versions. The default download and approved upgrade target is 0.1.7-rc.2; a compatible local installation is reused without downgrading. The extension reads session history through the public Remote page/follow APIs. Session log storage and migration are owned by the Runtime.

The Remote source audit targets upstream tag dsh-v0.1.7-rc.2 (477b4f420553e8a52c2fbccc464d7561b239c443); see the RPC adaptation report. The carrier handles multipart binary unary responses and client-to-host stream frames. The IDE exposes Account management, Schedule Activity Dock, and default Workspace initialization against the updated RC.2 contracts. Remote Workspace Files, Jobs, plugin-manager, permission-preset, and terminal surfaces remain candidates and are not exposed in the IDE.

The default dsh.command: "auto" probes dsh --version on PATH, then in the npm global prefix. A compatible local CLI is used directly. An incompatible CLI gets an upgrade prompt before any plugin download: it shows the current version, target and installation path. Approval upgrades a verified older npm global installation to dsh.runtimeVersion, then probes that same CLI again. Declining or closing the prompt uses pinned pnpm, then npx, then the managed CNB Runtime; missing CLIs also use this fallback. Upgrade failure offers fallback or cancellation. Unknown versions and older installations outside the active npm prefix get manual guidance. Diagnostics never prompt or install. Explicit local paths follow the same upgrade flow; explicit pnpm/npx keeps package-manager startup. If you previously saved dsh.command: "pnpm", reset it or select auto to enable local-first discovery.

Default app arguments are web --no-open; pnpm/npx gets its required prefix automatically when no argument override is saved. Existing package-manager argument overrides are preserved, and auto mode strips their package prefix when selecting a local CLI. Shared Runtime discovery still runs before choosing a new launcher, so fallback reuses a healthy Runtime instead of starting a second one.

If a standalone Runtime asset is unavailable from the configured download source, use a compatible local CLI, the pinned pnpm/npx fallback, or an existing instance. After compilation, node scripts/verify-runtime-discovery.mjs checks selection and actual startup arguments in an isolated POSIX CLI environment without downloads or model requests.

Does DSH support multi-root workspaces? DSH supports multiple independent Workspaces, but each Session has one working directory (cwd). A VS Code multi-root workspace is therefore represented by the first workspace folder for Runtime startup; use separate DSH Workspaces or Sessions when roots need different working directories.

Does DSH automatically identify secrets or personal information? No. Context is based on files, selections, and attachments that you explicitly choose; DSH reports size/truncation but does not send workspace content to an additional secret/PII classifier.

What if startup fails? Run DSH: Diagnose Environment, then DSH: Show dsh Runtime Logs from the Command Palette. Include your extension version, OS, and redacted error details when opening an issue.

Does it support Chinese? Yes. Commands, chat, Activity, and Trace follow VS Code's display language, with English and Simplified Chinese available.

Architecture and runtime

The extension connects to the Runtime through RC Remote RPC, using HTTP calls and a multiplexed WebSocket for live session updates.

Multiple VS Code windows discover each other through Runtime advertisements, then fall back to port 3080 and any configured dsh.serverPort. Every candidate is health-checked before use: authentication and a successful session/list call establish a usable connection, and advertised versions below the minimum are excluded. External services remain externally owned and are not stopped on disconnect. If authentication credentials are missing, set dsh.serverUrl to the full launch URL, including its token.

Advertisements are discovery metadata only. No advertisement grants or denies permission to start, so a missing, stale, or unreadable one can never block startup — the worst case is one failed health probe followed by this window launching its own Runtime.

Each window publishes exactly one file, <ownerId>.json, under dsh-runtime-advertisements-<user> in the OS temporary directory, carrying its endpoint, launch URL, version, PIDs, and composition hash. A window writes only its own file and never reclaims another's; legacy dsh-runtime.lock files are still read as hints, never written or removed. Readers take the sixteen most recent entries, so an abandoned file cannot crowd out live ones.

Advertisement lifecycle:

• Only a ready endpoint is published. A launch that fails before producing a URL leaves nothing behind.
• Startup briefly coordinates through a loopback mutex — at most 250 ms waiting and 500 ms held — and rechecks for a shared Runtime before spawning. Losing that race never blocks a launch; it just means this window starts its own Runtime.
• Explicit stop, dispose, and failed launches withdraw the advertisement.
• An unexpected launcher exit drops ownership but keeps the advertisement while its endpoint still answers, because package-manager wrappers routinely exit while the Runtime they started keeps serving. Only an explicitly refused loopback connection withdraws it; a timeout or an ambiguous host keeps the record.
• Startup is pinned to --host 127.0.0.1 and, without dsh.serverPort, to an OS-assigned port. A pinned port that loses a bind race retries once on an OS-assigned port.
• Shutdown stops owned process trees before withdrawing the advertisement. POSIX uses separate process groups; Windows uses scoped taskkill /T while the root identity is known.

After npm run compile, run node scripts/verify-runtime-discovery.mjs and node scripts/verify-runtime-shutdown.mjs to check launcher/port selection, advertisements, upgrade confirmation, and shutdown using isolated temporary directories, child processes, and loopback listeners.

graph TD
    A[VS Code Extension Host] <-->|RC Remote RPC| B[Standalone Harness Runtime]
    A <-->|Typed Full-State Bridge| C[React Webview UI]
    B <-->|CNB Distribution| D[Managed Local Engine]
    A <-->|Process Lock| E[Multi-Window Shared Runtime]

Configuration

Search dsh in VS Code settings for the full list.

Other ways to install

From GitHub Releases — download the .vsix from Releases and run Extensions: Install from VSIX.... Pre-release builds go to Open VSX flagged as pre-release and to GitHub Releases; on Open VSX only users who switched that extension to its pre-release version receive them, and they never reach the VS Code Marketplace, which does not accept SemVer pre-release version numbers. From 0.8.0 on, stable releases use even minor versions (0.8.x) and pre-release builds use the next odd minor (0.9.x), so a stable release never supersedes a newer pre-release.

Build from source:

git submodule update --init --recursive
npm install
npm run check
npm run package

Then install the generated .vsix via Extensions: Install from VSIX....

Extension API

Other VS Code extensions can hook into the API DSH exports.

const registration = api.registerConversationNavigation([
    { seq: 42, label: "Review the PPO implementation", detail: "Training config" },
]);
context.subscriptions.push(registration);

const dsh = vscode.extensions.getExtension<import("dsh-vsc-integration").DshExtensionApi>(
    "harcochen.dsh-vsc-integration",
);
const api = await dsh?.activate();
context.subscriptions.push(
    api?.registerAgentStatusPresentation({ label: "🐋 Diving" }),
);

Development and testing

npm install
npm run check      # TypeScript check (host + webview)
npm test           # Release gate: webview check + compile + test suite
npm run compile    # Build to dist/
npm run package    # Compile + vsce package
npm run release    # Test + version bump + CHANGELOG archive + tag

The checked-in Remote smoke runner uses historical 0.1.5-rc.2 V3 fixtures. It does not verify the 0.1.7-rc.1 multipart, uplink, or archive changes or the 0.1.7-rc.2 business RPC changes; see the current RPC adaptation report for that audit boundary.

To run the historical Remote integration smoke against a compatible installed launcher:

npm run compile
node scripts/verify-remote-runtime.mjs --launcher /absolute/path/to/dsh

This smoke run uses a temporary DSH home/workspace and a loopback model stub. It does not use your sessions or external model credentials. The runner requires Node.js >=22.15.0 with node:zlib Zstandard support (zstdCompressSync; Node 23 users need >=23.8.0).

To verify the managed Runtime release logic:

node scripts/verify-managed-runtime.mjs              # remote contract only
node scripts/verify-managed-runtime.mjs --full       # install and smoke-test

More information

• Changelog
• Product TODO
• Third-party notices

Acknowledgments

Thanks to dsh-reasoning-effort for the chibi runner sprite reference. The conversation outline takes inspiration from the dsh-milestone project.

License

MIT