Jelee0145/dsh-auto-review

README excerpt

DSH Auto-Review English 简体中文 Codex-style automated approval review for DeepSeek Harness (DSH). Instead of a human prompt or a blanket denial, sandbox-escalation requests are decided by an isolated reviewer subagent that judges two independent questions — how risky is this exact operation? and did the user actually authorize this effect? — and returns a structured allow/deny verdict. Status: Community Beta (v0.1.0-beta.1). This is not an official DeepSeek AI product or plugin. It is a community project distributed separately from DSH. Expect rough edges; see Known Limitations. How it works (brief) A DSH agent that needs to write outside its sandbox retries the operation once with sandbox permissions + a justification . DSH routes that escalation through the approval seam; Auto-Review claims the ask and hands the reviewer subagent: - Exact operation — the full validated tool arguments. - Trusted context — the user's genuine messages (authorization evidence only; the agent's justification is marked untrusted). - One-invocation semantics — the grant applies to this exact call and nothing else. The reviewer returns { verdict, riskLevel, userAuthorization, reason } . A critical risk is d…

Category