Automated approval review for DeepSeek Harness. Inspired by Codex 0.147's --approve-for-me (Guardian auto-review).

English | 中文

approval/request hook: in review mode, read-only tools (read, grep, glob, ssh_list, inspection tools, ...) are auto-approved; in auto mode every request is approved — except dangerous commands, which are always denied (fail-closed security floor).
tools/pre-execute hook: dangerous shell commands (rm -rf /, mkfs, dd of=/dev/sdX, fork bombs, chmod -R 777 /, curl|sh, shutdown, ...) are hard-blocked before dispatch, so they never reach the approval prompt.
approval_policy_status tool: report the current mode and activity.
RPC (approve-for-me/status, approve-for-me/set-mode) for client halves and other plugins.

Mode	Read-only tools	Other tools	Dangerous commands
`off`	ask human	ask human	ask human
`review` (default)	auto-approve	ask human	auto-deny
`auto`	auto-approve	auto-approve	auto-deny

dsh plugin --profile web add github:DamonKoy/dsh-plugins#path:/packages/dsh-approve-for-me

Restart dsh web. Default mode is review.

~/.dsh/dsh-approve-for-me.json:

{ "mode": "auto" }

Valid modes: off | review | auto. You can also switch at runtime via the approve-for-me/set-mode RPC (memory-only, resets on restart).

Auto-approval never bypasses the dangerous-command denylist.
Denials are fail-closed: a throwing listener falls through to the human answerer, never to silent approval.
The policy engine (lib/policy.js) is unit-tested; run node --test test/ to verify.

MIT

dsh-approve-for-me

같은 패키지 이름의 다른 저장소