Guardrails heat 1 · Quiet★ 0heat incomplete (no X)

dsh-decision-guard

A MIT DeepSeek Harness plugin (package dsh-decision-guard) that uses hosted TypeSafe Jev for sensitive-command verdicts, per-session tool schema restriction, skill-catalog pruning, and glob relevance ranking—all off by default with fail-open native fallbacks.

What it is

Gate risky shell and shrink DSH context with one Jev connection.

What it does: Sensitive-command verdicts plus tool/skill/glob pruning via hosted Jev.

Why it belongs in the showcase: Distinct DSH-native multi-gate on TypeSafe, beyond dsh-jev / dsh-jev-mcp wrappers.

Caveats: Opt-in features and provider reachability.

Evidence

Public MIT repo, created 2026-11-03, 0 stars as of 2026-11-05. README documents TypeSafe API key setup, api.typesafe.ai CONNECT tunnel for CN HTTP 451, 43 sensitive-command rules, tools.restrict + tool_search, skillLimit defaults, and rename from dsh-sensitive-guard at 0.2.0.

Caveats

  • DeepSeek Harness desktop plugin — not a generic Claude/Codex router
  • All four features are opt-in; mis-restrict needs tool_search recovery
  • CN users may need the built-in proxy tunnel to reach api.typesafe.ai

Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.