dsh-multi-folder

English | 中文

Secondary working directories for a DeepSeek Harness project — edit a source repo, a test repo, and a docs repo side by side without leaving the primary workspace.

A DeepSeek Harness plugin bundle that gives one project (workspace) a set of secondary working directories:

• The agent's core cwd and every other core attribute keep pointing at the primary workspace.
• Under Workspace Write mode the agent gains the same read / write / edit / execute permissions on the configured secondary directories as on the primary workspace — enforced by re-rooting the session's own sandbox policy, so every mode keeps its semantics (read-only still denies, workspace-write allows, danger-full-access allows).
• The directory list is injected into the system prompt and re-rendered per session assembly.
• Configuration changes notify the agent through a non-interrupting message queue — delivered at the next message boundary (user send or tool-call end), and only when the directory set actually changed.
• Configurable before the session starts: the session-creation page (new-session screen) offers a Multi-folder entry that reads and edits the same per-workspace configuration through a sessionless remote API (multiFolder/* endpoints) — no session id required.
• No new tools. Everything is a framework-level change (tool-pipeline interception) plus a UI-level change (a session-scoped header entry).

Requirements

• Node.js >= 20
• A DSH profile composed from @deepseek-ai/dsh-base + @deepseek-ai/dsh-web-app

Install

Link this repository into a DSH profile:

dsh plugin --profile web add dsh-multi-folder

Then restart the DSH backend (host composition loads at process start) and refresh the browser page (the client bundle is served no-cache).

Usage

A Multi-folder button appears in the session header, and a second entry appears on the session-creation page: a chip row directly above the composer card (the same band the git-branch chip uses), aligned with the workspace/preset chips of the new-session screen. Clicking it opens the panel as a popover anchored to the chip. Exactly one session-creation entry is ever shown — the plugin registers three candidate seats and elects the best available one (upstream conversation.hero.workspaceExtras chip > conversation.input.dock row > fixed bottom-right launcher for shells declaring neither). The panel lets you:

Equivalent slash command for the user:

/multi-folder list
/multi-folder add "D:\path\to\repo"
/multi-folder remove "D:\path\to\repo"
/multi-folder set "D:\a" "D:\b"

The agent needs nothing extra: read / glob / grep work everywhere, and write / edit / pwsh / bash are intercepted and re-rooted automatically when the target path (or workdir) falls inside a configured secondary directory.

Permission model

Each confined command runs under exactly ONE writable root — the workspace root the call is re-rooted to (the Windows ACL runner grants a single workspace write SID per process tree). Consequences:

• A command whose cwd stays the primary workspace cannot create files inside a secondary directory. git -C <secondary> commit, cd <secondary> inside a script, git clone <url> <secondary>, or absolute-path writes all fail with an OS-level Permission denied (e.g. fatal: Unable to create '.../.git/index.lock': Permission denied).
• Symmetrically, a command re-rooted to a secondary directory cannot write to the primary workspace (or another secondary directory) in the same invocation.
• Rule for file-creating commands: set workdir to the directory the command writes into. For git, run the command from inside the repository (pass workdir pointing at it) instead of using git -C from the primary workspace.
• Reads are unrestricted and need no workdir.

When a shell run ends in such a denial and references a configured secondary directory, the plugin attaches a short diagnostic hint to the tool result explaining the workdir fix.

How it works

• Interception — a listener on the tools/execute around-dispatch waterfall short-circuits write / edit / pwsh / bash calls whose resolved path (or workdir) lands inside a configured secondary directory, and executes them with the session's standing sandbox policy re-rooted to that directory ({ ...standingPolicy, workspaceRoot: secondaryDir }). The mode itself is untouched, which is what gives every sandbox mode its identical primary-workspace semantics for free. Paths are canonicalized through fs.resolve + processPath before matching, so .., symlinks, and case differences behave correctly.
• Prompt injection — one ordered systemPrompt section with a text provider evaluated per assembly, rendering only for sessions whose workspace has configured directories.
• Notifications — a pending notice armed by the command handler (only on actual change) is consumed at the next boundary by either the agent/pre-step waterfall (prepend into the entering message batch) or the tools/post-execute waterfall (attach as additionalContexts), whichever fires first — the framework's native plugin-sourced notice context.
• Configuration & security boundary — per-workspace config lives in a host-owned store outside every agent sandbox root (<DSH_HOME>/storages/multi-folder/<workspace-key>.json). Direct write/edit attempts against the config file are rejected with an explicit message — the agent can never self-grant directories; configuration is user-managed by design. See SECURITY.md.
• Sessionless remote API — a multiFolder namespace registered through ctx.typert.register (hand-written src-json descriptors) plus a plain-object service provided as multiFolder. Its list/add/remove/set methods are keyed by workspace path and share one validated core with the /multi-folder command, so the creation page can configure directories before any session exists.
• Client — a hand-maintained factory bundle (window.__ModuleLoader__.load), no build toolchain required. The panel drives the host through two channels: the Remote BFF (ctx.remote.commands.execute) for sessions, and the shared /api RPC channel (ctx.connection.rpc.call) for the sessionless endpoints.

Project layout

Development

No build step: the host half is plain ESM and lib/client.js is a hand-maintained factory bundle in the DSH client-modules format. Tests run with Node directly:

node test/smoke-host.mjs    # host apply smoke test + remote API behavior
node test/intercept.mjs     # interception / command / notification behavior
node test/smoke-client.mjs  # client bundle + panel flows (React shim)

Before modifying lib/client.js, see docs/design.md for the bundle contract.

Documentation

• docs/design.md — architecture and security model
• docs/upstream-hero-slot.md — the upstream conversation.hero.workspaceExtras slot change (B1) and its plugin-side consumption

Contributing

See CONTRIBUTING.md. Issues and pull requests are welcome.

License

MIT