[Plugin] weiwen-law-dsh — white-box causal guardrail: every tool call adjudicated before execution along a causal logic chain

[Shaky77]

weiwen-law-dsh (Weiwen's Law / 唯稳律) — a white-box causal guardrail plugin for dsh.

What it does

Every tool call is adjudicated before execution along a deterministic causal logic chain:

• Hard red lines: destructive filesystem operations, credential-file reads, and forced pushes are blocked outright.
• Fail-stop escalation: repeated boundary violations escalate to a stop instead of slipping through.
• Volition boundary: prompts that try to pry into the agent's own mind/volition are refused.
• Fault containment: a faulting link is cut and must complete a verify-fix-reenter loop before retry — no blind re-runs.
• Risk verdicts: executions that cannot be audited up front (e.g. curl | bash) get a low/mid/high risk verdict instead of a blind allow.

Deterministic rule engine — runs fully local, zero API cost, no model calls on the deny/allow path. Every decision stays auditable after the fact.

Install

dsh plugin add "github:Shaky77/weiwen-law-dsh"

Extras

• 6 white-box self-check tools exposing the engine's live state.
• 36 multi-model trial runs archived in-repo; 118 tests passing.
• Declares dsh.bundle.patch; one-command installable.

P.S. A printed lightweight version of the framework was delivered to your Beijing office via EMS back in June — this plugin is the runnable, live form of that same material.

Author: Shaky77 (夏祺). License: AGPL-3.0.
Repo: https://github.com/Shaky77/weiwen-law-dsh

[Rehanguards]

Adjudicating every tool call before execution is the right approach to prevent agent prompt injections. Pre-execution checks need to run deterministically on payload buffers before shell or API invocation to guarantee safety. Would love to know how you're structuring the causal chain logic to keep latency low during nested tool calls!