dsh-plugin-vet
A trust-layer plugin for deepseek-harness (DSH) that covers a download → scan → audit → score → decide → runtime-watch pipeline: a deterministic static scan (11 rules), an LLM-driven audit protocol, a two-part scorecard, an optional runtime guard (T1 sentinel plus T2 fs and child_process hooks) and optional honeypot decoys. Monitor-and-alert only — vet never blocks or kills on its own in the default configuration.
Installation
dsh plugin --profile <profile> add @jieai/dsh-plugin-vetOverview
A trust-layer plugin for deepseek-harness (DSH) that covers a download → scan → audit → score → decide → runtime-watch pipeline: a deterministic static scan (11 rules), an LLM-driven audit protocol, a two-part scorecard, an optional runtime guard (T1 sentinel plus T2 fs and child_process hooks) and optional honeypot decoys. Monitor-and-alert only — vet never blocks or kills on its own in the default configuration.
- Deterministic static scan with 11 configurable rules (R1-R20) produces a deterministic, unforgeable verdict.
- Agent-side audit of sensitive points and quality issues follows the vet-audit-protocol skill.
- A final two-part scorecard is handed to a human or model to decide disposition.
- Optional runtime guard adds a T1 sentinel (memory/fd/child-process monitoring) plus T2 fs, child_process and network hooks.
- Optional honeypot lures plant fake credential files; T2 reports touches (read/write/delete) as a separate honeypot alarm class.
- Default mode is report: vet never auto-uninstalls, never kills processes, never rewrites configs and blocks nothing.
- Safety tiers standard, hardened and paranoid; deny mode and the N7 confirmation block are explicit opt-ins.
- Includes OSV exact-version queries, an official-package content-hash baseline and runtime contract snapshots.