Andy-Xie-1145/dsh-auto-review-ww

Sandbox-first auto review: in-sandbox work runs unreviewed, and only an escalation past the workspace pays for one LLM review that fails closed.

A different trade from the shipped experimental auto review, and the README tabulates it: that one runs on a danger-full-access base and reviews every request, while this one keeps real workspace-write confinement and reviews only escalations, so in-sandbox work costs zero review tokens. Each escalation sends a five-section snapshot to the session's own provider — a fixed review policy, the session's working directory, project instruction texts, a filtered role-tagged history of instructions, checkpoints and executed tool calls with their arguments, and the exact pending call with its full arguments, schema and approval reason — and gets back a closed JSON answer of risk, decision and reason. The policy asks one question, whether this call should run now given what is visible in the session, and denies on a mismatch between claimed purpose and real effects, on unrelated effects, and on anything unknown.

Agent Capabilities ★ 0
View on GitHub ↗

Install

dsh plugin --profile web add github:Andy-Xie-1145/dsh-auto-review-ww

The README's web-profile line; the GUI Plugins → Add plugin route accepts either the package name or the repository URL. Then fully restart Harness and pick the Auto review preset — the client only decorates the official auto value. The README stresses that the shipped @deepseek-ai/dsh-experimental-auto-review row must stay enabled: auto is a reserved preset id that only the official row provides at runtime, and sessions whose last permission/preset event is auto cannot cold-open without it.

Compatibility

Built against DSH 0.2.0-rc.2 and designed to coexist with the official auto-review row, which owns the reserved auto state while this package owns the auto-review preset and its reviewer. It works with any model route, because the reviewer rides the session's own provider and model. Failure is closed on purpose: if the reviewer itself fails, the call is denied and the reason is published back to the requesting agent; if the plugin crashes, the preset stays selectable and degrades to ordinary human approval rather than silently allowing.

Details

Alternatives

sperictao/dsh-auto-review-jev · Andy8647/dsh-auto-approval · Jiao-XXX/dsh-auto-approve

More plugins in Agent Capabilities

Browse more in Agent Capabilities

Guides for Agent Capabilities plugins