dsh-llm-approve-for-me

In DeepSeek Harness, every sandbox permission escalation—running a Shell command, writing a file—pops up a manual approval prompt that interrupts your coding flow. dsh-llm-approve-for-me delegates that decision to the LLM you configure in DSH: once the approve-for-me preset is active, a built-in tool-less reviewer role makes a single fast call per request and returns allow, deny, or ask, letting low-risk operations pass automatically while surfacing only the requests that genuinely need a human.

The review path is deliberately minimal: no Agent or Session is created, the sub-agent catalog stays clean, no history context accumulates, and there are no command whitelists, regexes, or dangerous-operation rules. The reviewer must emit strict JSON; on timeout, model error, or malformed output the request falls back to native human approval with the failure reason logged—never a silent default-allow. A session-top panel shows the last 100 approval records (isolated per session) and a visual settings form where you can tweak the reviewer model, timeout, and max tokens online, effective on the next review without a restart.

Best for developers who lean heavily on DSH sandbox tools, want to cut down manual approval interruptions, and still require a fail-safe that routes back to a human when the LLM cannot decide. It is not a security product and does not replace least-privilege, backups, or isolation; it simply hands the per-request go/no-go call to the LLM you trust, so you can focus your attention on the moments that truly need a human eye.