dsh-jev-verify
Jev (TypeSafe System One decision model) for DeepSeek Harness: fast decision tools (choice/score/noul, ~300ms parallel), auto-guard (deterministic + Jev risk/loop), in-chat board (jev_overview). Boot-stable by design: every registration is quarantined; a failure degrades one feature, never the profile. 适合需要毫秒级快速判断、风险防范和稳定启动的用户。
설치
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:xienda/dsh-jev-verifyUsage
Ask the agent to use jev_decision for any fast judgment. Example prompt:
Use
jev_decisionwith state = "" and these questions:department(choice, criteria billing/technical/sales),is_urgent(noul),severity(score, 3 levels). Report answers + confidence.
Or request jev_verify at any time to confirm the endpoint is healthy, and
jev_guard_status to see how the auto-guard is behaving.
Direct tool call shape
{
"state": "I was double-charged for my subscription and I want a refund.",
"questions": [
{ "name": "department", "type": "choice",
"instructions": "Which team should handle this?",
"criteria": { "billing": "Payment or subscription issues",
"technical": "Bugs or integration problems",
"sales": "Pricing or account questions" } },
{ "name": "is_urgent", "type": "noul",
"instructions": "The message conveys urgency or time-sensitivity" }
]
}
Returns per-question answers (choice/score/noul + confidence + probabilities),
usage, latencyMs and estimatedCostUs.
Configuration
| Key | Default | Meaning |
|---|---|---|
apiKeyEnv | TYPESAFE_API_KEY | credential-ref / env var for the key |
apiKey | — | literal key override (secret) |
baseURL | https://api.typesafe.ai/v1 | API base |
model | jev-latest | model (pin e.g. jev-1.13.0) |
timeoutMs | 15000 | per-call timeout |
maxQuestionsPerCall | 25 | question cap per call |
verifyEnabled | true | register jev_verify |
autoGuard.enabled | false | master switch for auto-guard hooks |
autoGuard.safetyCheck | true | pre-execute risk check for guarded tools |
autoGuard.loopCheck | true | semantic stall detection |
autoGuard.tools | bash/pwsh/run_code/terminal | tools the safety check applies to |
autoGuard.denyThreshold | 0.85 | Jev noul ≥ threshold ⇒ deny / advise |
autoGuard.maxJevCallsPerSession | 50 | guard Jev budget per session |
autoGuard.loopConsecutive | 3 | same-tool calls before loop check |
autoGuard.loopCooldownMs | 60000 | cooldown after a stall verdict |
autoGuard.loopMinChars | 200 | min result length for loop check |
autoGuard.statusTool | true | register jev_guard_status |
autoGuard.determinismFirst | true | run the free deterministic check before Jev |
dashboard.enabled | false | serve the optional standalone board page |
dashboard.basePath | /jev | path of that board page |