lukethecat/dsh-plugin-warroom-garak

README 摘要

dsh-plugin-warroom-garak A DeepSeek Harness plugin that adds one model-facing tool, garak scan — an authorized garak baseline red-team sweep against a target LLM endpoint, with a built-in authorization gate, a budget cap, and an auto-written evidence report . Part of the war-room approach: garak owns the probes (the "weapons"); this plugin is the orchestrator that runs them against an authorized target, keeps the run in the dsh session log (auditable/replayable), and produces a compliance-ready evidence artifact. v0 / developer preview. Built against @deepseek-ai/dsh- 0.1.0-rc.x and modeled on the official dsh-tool-web plugin. Two seams are version-sensitive and may need a small fix on first build in your environment — both are commented in the source: 1. defineTool parameter/output schema DSL ( src/index.ts ) 2. garak report JSONL field names / hit detection ( src/garak.ts ) What it does 1. Refuses to run unless authorization.authorized === true with a non-empty scope . 2. Caps attempts per probe at maxGenerations (budget / circuit-breaker). 3. Spawns garak against your REST target, parses the JSONL report into per-probe hit rates. 4. Writes a one-page Markdown evidence report (ta…

分类