dsh-safe-guard

Deny destructive shell commands at the dsh tools/pre-execute waterfall before they run: rm -rf /-class deletions, git push --force (without --force-with-lease), raw block-device writes, mkfs, Windows drive-root wipes/formats, chmod -R 777 /, fork bombs, plus configurable extra patterns.

Second development of the official permission-gate shape (docs/cookbook/extension-cookbook.md § "A hook plugin") in the deepseek-harness repo, with the pure-rules/injection-point split recommended by dsh-handbook ch.4.

Install

dsh plugin --profile web add link:/plugins/dsh-safe-guard

The dsh.bundle manifest appends this package to the profile's bundle list; its patch mounts one host row (id: safe-guard).

Config

- id: safe-guard
  name: dsh-safe-guard
  config:
    extraDenyPatterns:
      - 'curl[^|;&]*\|\s*sh'

Patterns are case-insensitive regex over the whole command text; an invalid pattern fails at load.

Hot-loaded user rules ($DSH_HOME/safe-guard.json)

{ "denyPatterns": ["curl[^|;&]*\\|\\s*sh"], "askPatterns": ["docker\\s+system\\s+prune"] }

Edited from the DSH Launcher Control Deck → Safety rules tab and re-read every 1.5 s. denyPatterns deny outright; askPatterns return { kind: 'ask' }, so the dsh web UI asks for approval before the command runs. Invalid entries in this file are skipped (the guard never goes down because of a typo); the built-in rules and extraDenyPatterns are evaluated first and are unaffected.

Guarded surfaces

bash.command, pwsh.command, terminal_send.text. All other tools delegate untouched. Denials return { kind: 'deny', reason } and never call next(); every allow delegates, so downstream policy (sandbox, permission, plan mode) still runs.

Model Experience

None, as this package registers no tool, prompt section, or injected context; a denial surfaces to the model only through the standard denied tool result produced by the tools pipeline.

KV Cache effect

Independent: the guard adds no request content and preserves any already-reusable prefix.