dsh-plugin-auto-review

When a sandboxed tool call in DeepSeek Harness hits an approval gate, the pipeline stalls waiting for a human prompt. Delegated subagents cannot surface an unattended dialog at all, silently breaking automation loops. dsh-plugin-auto-review embeds two provider-grade review policies—Codex Guardian and Grok escalation—directly into DSH. The underlying model is supplied through any compatible DSH LLM adapter, so you are never locked to a vendor SDK.

Two preset policies ship out of the box: Codex Guardian evaluates compliance for sandboxed shell operations, while Grok escalation handles calls that warrant elevated scrutiny. Select a reviewer globally or per session in the Web UI, or wire in a custom route that points to any compatible model. Session-level selections are inherited by child sessions, so the approval chain never fragments across a delegation hierarchy.

Safety is the first-class concern. The model fires only when an exact tool call reaches a genuine approval boundary; ordinary allow-listed paths incur zero model cost. Denials, timeouts, interruptions, or malformed outputs all fail closed, falling back to DSH native manual flow. The plugin never bypasses sandbox permissions—retries pass through the same approval gate—and review transcripts are bounded with injected plugin context excluded from user authorization anchors. This plugin suits teams running unattended automation pipelines, using Codex or Grok models, or orchestrating large-scale tool calls within DSH.