dsh-typesafe
TypeSafe Jev (System One decision model) for DeepSeek Harness: typed decisions in three question primitives, confidence-gated routing, a cost meter, and an automatic prompt-injection guard over untrusted tool results. 适合需要高可靠性类型化决策、成本监控和安全防护的Agent工作流。
インストール
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:979569650/dsh-typesafeドキュメント
README 全文を読む ↗Configuration reference
Every key is defaulted, so id + name alone is a working install. Edit them in
the settings card, or override the row in a profile cordis.patch.yml.
| Key | Default | Meaning |
|---|---|---|
enabled | true | Master switch. Off removes the prompt section and the guard. |
apiKeyEnv | TYPESAFE_API_KEY | Credential reference to resolve. |
baseURL | https://api.typesafe.ai | API origin. |
model | jev-latest | Alias or pinned version id. Pin a version if you tuned thresholds against it. |
timeoutMs | 30000 | Per-request timeout. |
maxStateChars | 60000 | Rejects oversized calls before they are billed. |
inputPricePerMTok | 0.042 | Cost reporting only. |
promptEnabled | true | Whether to tell the agent when to use the tools. |
promptOrder | 700 | Prompt section placement, in the behavioural-rule band (TEAM_POLICY 600 … PTC_ONLY 800). |
guardEnabled | true | Screen untrusted tool results. |
guardTools | web_fetch, web_search, read_page, read_url, browse, scrape, mcp__*__fetch, mcp__*__fetch_content, *fetch_content | Which tool results to screen. Globs, so one entry covers a family. Only tools that bring outside text in belong here. |
guardWarnThreshold | 0.5 | Injection probability at which a result is flagged. |
guardMinChars | 200 | Skip shorter results. |
guardMaxChars | 6000 | Chunk size for screening. Every chunk is screened, so this trades request count against request size — it is not an excerpt budget. |