940842546/dsh-permissions

English | 中文

Claude Code-style permission rules engine for DeepSeek Harness (dsh). A dual-face Cordis plugin: a host engine on the tools/pre-execute waterfall plus a visual editor in Settings → 权限 (Permissions).

Highlights

• Four rule tiers with strict precedence: hard > deny > ask > allow.
• hard outranks full access: hard rules keep blocking even when the session is on the full-access preset (approval policy never); ask rules follow the session policy and auto-pass under full access.
• Scopes: global rules apply everywhere; per-workspace rules merge on top (deny always wins on conflict).
• Wildcard matching for file tools (read/write/edit/glob/grep/read_image):
  • write(*.pem) — path ends with .pem
  • write(*secret*) — path contains secret
  • write(.ssh) — path segment .ssh anywhere (case-insensitive, \// normalized)
  • write(C:\users\*) — absolute-path prefix
  • bare write — every invocation
• Persistence: rules live in the dsh-permissions settings namespace and survive restarts (<harness home>/settings.yaml).
• Model transparency: active rules are injected into the system prompt ([active-permission-rules]).
• Visual editor: staged (draft) editing — changes apply only after Save & Apply, with one-click presets (protect sensitive dirs / key files / dangerous commands).

Screenshots

Top: scope & rule builder	Panels / tester / decision log	Staged saving

Rule syntax

Rule	Meaning
pwsh	every call of that tool
pwsh(npm run)	first argument starts with npm run
write(*.pem)	file path ends with .pem
write(*secret*)	file path contains secret
write(.ssh)	path segment .ssh anywhere
pwsh(*)	every call (explicit)

Non-file tools match the raw first argument by prefix. grep additionally matches its path argument.

Install

Option A — official installer (recommended):

dsh plugin add dsh-permissions

Option B — manual patch row: append the insert list from this repo's cordis.patch.yml to your profile patch (~/.dsh/cordis.patch.yml or ~/.dsh/profiles/<profile>/cordis.patch.yml), after installing the package where the profile resolves it (~/.dsh/node_modules/dsh-permissions):

- insert:
    - id: permissions
      name: dsh-permissions

Then restart the app. The Settings → 权限 page appears automatically; the engine starts with safe defaults (16 hard rules protecting .ssh / .aws / .gnupg / AppData / *.pem / *.key / *.env / *.htpasswd, plus deny: pwsh(rm -rf *)).

Permissions page

• Engine toggle, three one-click preset cards, a point-and-click rule builder (action × tool × match mode × value → live preview), and four colored rule panels.
• All edits are staged: nothing affects the agent until you click Save & Apply; Discard restores the last saved state.

Security notes

• The engine only narrows the session's existing sandbox/approval posture: allow skips this plugin's own ask but never bypasses the DSH sandbox or tools.guard guards.
• Denials surface to the model as Error: 权限规则拒绝… (hard: 硬规则拒绝（高于 full access，不可豁免）…), so the agent can route around blocked calls.
• The settings route (GET/POST /api/dperm/rules) is the namespace owner's own endpoint — the DSH api-proxy's settings allowlist intentionally does not expose third-party namespaces.

Development / publishing

See PUBLISH.md for the release checklist and the pitfalls we hit (client package exports must include ./package.json; bundle id must equal the package name; no unbounded method refs into React's useSyncExternalStore).

License

MIT