dsh-fs-allowlist

English | 中文

DSH plugin: approval-free writes into whitelisted directories. Let the write/edit file tools modify configured whitelist directories (e.g. your Obsidian vault) directly — no more "sandbox denial → escalation retry → manual approval" — plus a Settings → Plugins → Directory Allowlist GUI and automatic approval for bash escalations that touch whitelisted paths. Behavior everywhere else stays exactly as before.

Features

Layer	Mechanism	Covers
write/edit tools	Wraps the ctx.fs.checkedTarget fence; whitelisted paths pass early	Approval-free vault reads/writes
bash commands	tools/pre-execute indexes the command text; approval/request auto-answers allowed-once when the command or escalation reason hits a whitelisted path (recursive check)	Approval-free bash writes (command text first; misses fall back to manual approval)
Settings GUI	settings.plugins.tab section + server-side HTTP management routes	Add/remove directories, toggle bash auto-approve visually

Configuration is hot-reloaded: GUI edits apply instantly; manual edits to the config file are picked up within 3 seconds. No DSH restart needed.

How it works

write/edit layer

Every write/edit tool call passes through the ctx.fs.checkedTarget() fence mounted by dsh-fs-sandbox (outside the workspace → FS_SANDBOX_DENIED → model retries with escalation → approval prompt). This plugin wraps that fence:

• If the canonically-resolved target lands inside a whitelisted root → pass it through early (same fresh-target semantics as the fence);
• Otherwise (miss / read-only mode / resolution failure) → delegate to the original fence untouched.

File semantics — atomic writes, version guards, read-before-write checks, diff events — live in the parent dsh-fs-local provider below the fence and are not touched. Whitelist matching uses the same realpath (deepest existing ancestor) semantics as the fence, so symlink redirection cannot escape the allowlist.

bash layer (approval auto-answer)

bash runs under the Seatbelt process sandbox, whose profile cannot be extended from a plugin. Instead two listeners cooperate: tools/pre-execute records each bash call's command text by callId, and the approval/request answerer looks it up and matches the command text plus the escalation reason against the allowlist — path-like tokens are extracted, ~ is expanded to home, then recursively checked under whitelisted roots (a parent entry covers all its children). The command text almost always contains the target path, so approval no longer depends on the model writing it in the justification. Misses always fall back to the human prompt (fail-safe). Toggle it off any time in the settings page.

Install

dsh plugin --profile web add /path/to/dsh-fs-allowlist

or from GitHub after release:

dsh plugin --profile web add "github:wzn16/dsh-fs-allowlist#main"

Installs into $DSH_HOME/profiles/web; restart DSH Desktop to take effect (refreshing the browser is not enough). The profile links the source directory, so editing the source only needs a restart — no reinstall.

Configuration

$DSH_HOME/fs-allowlist.json (default: ~/Library/Application Support/dsh-desktop/harness/fs-allowlist.json on macOS; the resolver also understands $DSH_HOME, Windows %APPDATA%, Linux XDG paths, and the open-source ~/.dsh layout):

{
  "extraWritableRoots": [
    "/Users/you/ObsidianVault"
  ],
  "bashAutoApprove": true
}

• Prefer the Settings → Plugins → Directory Allowlist page — no hand editing;
• Missing file / invalid JSON / empty list → empty allowlist, the plugin is fully transparent (= uninstalled behavior): one-step rollback;
• bashAutoApprove defaults to true when absent.

Behavior matrix

Scenario	Behavior
write/edit inside the workspace	approval-free (native, unchanged)
write/edit inside a whitelisted root	approval-free
write/edit anywhere else	deny → escalate → approval (unchanged)
read-only mode writing a whitelisted root	still denied (stricter knob wins)
bash touching a whitelisted root (command or reason matches)	auto-approved, no prompt
any other bash escalation (no match)	normal approval prompt

Known limitations

1. The bash auto-answer matches the command text plus the justification against the allowlist (recursive path check + substring fallback) — still a heuristic: if neither mentions a whitelisted path, a prompt appears (fail-safe); a match approves that one call.
2. checkedTarget is an internal seam of the shipped bundle: if a DSH upgrade refactors it, the plugin degrades loudly to "warn + transparent" (equivalent to today's behavior) without breaking sessions; adapt per release.
3. A whitelist entry means the agent may write there without asking — keep the list deliberately narrow.

Rollback

dsh plugin --profile web remove dsh-fs-allowlist (or empty fs-allowlist.json) + restart DSH.

License

MIT