dsh-risk-guard

DeepSeek Harness agents execute hundreds of tool calls every day, yet a single misstep—deleting the filesystem root, reformatting a disk, or force-pushing to main—can cause damage that nothing can undo. dsh-risk-guard was built to close that gap: it silently audits the entire agent session without altering sandbox policies or popping up dialogs, and pulls the fuse right before an irreversible catastrophe executes.

Three pillars define its capability. First, zero-cost deterministic risk scoring—no LLM calls, no extra API spend. It assigns explainable scores to destructive deletes, credential reads, writes outside the workspace, dependency installs, network egress, and heavy builds, with consecutive high-risk calls accumulating a bonus. Second, a precise insurance fuse that blocks only the worst outcomes: protected-path deletion, disk wipe or format, force-push to main/master or a configured protected remote, and network commands such as curl, wget, or ssh that reference credential files. Everything else runs untouched. Third, an operation bill—invoke /risk-guard at any time to review scores, risk tags, redacted details, and every fuse denial for the current session, a single turn, or full history. All records are stored as local JSONL; secret-shaped values are redacted before they hit disk. The plugin makes zero network calls and ships zero telemetry.

It fits any developer or team that has DeepSeek Harness in its pipeline and wants to keep the agent autonomous while still being able to trace what happened and stop disasters before they land. If you run unattended agent environments or need to answer a compliance question with exactly what the agent did, this plugin slots in without touching your existing sandbox setup.