dsh-jev

Semantic tool routing and typed System One decisions for DeepSeek Harness (DSH) powered by TypeSafe Jev.

Modeled after pi-typesafe, this plugin integrates TypeSafe Jev into DeepSeek Harness with Cordis service injection, DSH tool definitions, slash commands (/typesafe and /jev), threshold calibration, spend caps, and pre-turn lifecycle hooks.

---

Why Jev in Coding Agents?

Most coding agents spend expensive, slow frontier LLM reasoning tokens on "small mechanical decisions":

• Which category or module does this bug belong to?
• Is this issue blocking user workflows?
• Is the modification risk low, medium, or high?
• Does this PR require human review or can it continue automatically?

TypeSafe Jev provides a fast, structured judgment model: you provide a state and typed questions, and it returns calibrated probabilities, option distributions, and rubric scores in well under a second for a fraction of a cent ($0.042 per million input tokens; output tokens are free).

1. Free the main model from mechanical triage: Let DeepSeek / frontier models focus on deep reasoning, code generation, and complex debugging.
2. Deterministic thresholds in code: Jev outputs real probability distributions instead of prose so your workflow can branch with concrete cutoffs:

   if (riskProbability > 0.85) {
     requireHumanReview();
   } else {
     continueAutomatically();
   }
   

3. Independent batched questions: Evaluate Choice, Score, and Noul questions concurrently without question contamination.

---

Features

• Typed Judgments (typesafe_evaluate / jev_evaluate tools): Run fast, calibrated System One decisions directly from any LLM turn in DSH using choice (categorical selection), noul (0-1 truth probability), and score (rubric scale).
• Skill Discovery (jev_find_skill tool): Semantically matches and suggests the most relevant specialized agent skills (SKILL.md) for any task without cluttering prompt context.
• Tool Discovery (jev_find_tools tool): Semantically evaluates and identifies relevant registered tools in DeepSeek Harness for a user task.
• Human Slash Commands (/typesafe & /jev):
  • /typesafe status — Displays TypeSafe configuration, auth state, session usage, daily spend caps, and registered tool/skill counts.
  • /typesafe login [key] — Validates and saves API key to ~/.pi/agent/pi-typesafe/auth.json (with owner-only permissions).
  • /typesafe logout — Clears stored API key and resets auth state.
  • /typesafe enable — Enables TypeSafe agent evaluation tool calls for this session.
  • /typesafe disable — Disables TypeSafe agent evaluation tool calls.
  • /typesafe test [prompt] — Run live connectivity test or evaluate prompt against Jev (defaults to built-in bug triage sample).
  • /typesafe playground [json] — Run direct JSON state & questions without polluting agent context.
  • /typesafe calibrate — Historical sample threshold calibration toolkit (AUC, precision, recall sweep).
  • /typesafe skills [query] — Semantically search and rank available skills directly from chat.
  • /typesafe tools [query] — Semantically search and rank available tools directly from chat.
  • /typesafe auto [on|off] — Toggle automatic per-prompt skill suggestions.
  • /typesafe help — Display help message.
• Spend & Cost Tracking:
  • Session request limit (default 20 attempts).
  • Daily request, token, and USD caps (PI_TYPESAFE_MAX_REQUESTS_PER_DAY, PI_TYPESAFE_MAX_INPUT_TOKENS_PER_DAY, PI_TYPESAFE_MAX_USD_PER_DAY).
  • Cross-process 31-day persisted usage ledger at ~/.pi/agent/pi-typesafe/usage.json.
• Threshold Calibration Toolkit: Calibrate optimal decision thresholds from labeled historical samples with Mann-Whitney rank AUC, precision, recall, and optimal F1 picking (calibrate, replay, formatCalibration).
• Post-Run Gate CLI (dsh-jev-gate / jev-gate): Standalone binary for CI/CD pipelines, subagents, or verification gates. Checks git diff, files, or stdin against natural language acceptance criteria using Jev probability.
• Graceful Fallback & Fail-Open: Fails open to local keyword heuristic shortlists when Jev is unconfigured or offline.

---

Installation in DeepSeek Harness

In your DSH profile directory (e.g. ~/.dsh/profiles/web):

# Install via GitHub or link local workspace
pnpm add file:C:/Users/lldois/workspace/dsh-jev

In ~/.dsh/profiles/web/package.json, add "dsh-jev" to dsh.profile.bundles:

{
  "dsh": {
    "profile": {
      "bundles": [
        "@deepseek-ai/dsh-base",
        "@deepseek-ai/dsh-web-app",
        "dsh-jev"
      ]
    }
  }
}

---

Setup & Credentials

You can configure your TypeSafe API key via:

1. Slash Command:

   /typesafe login your_api_key_here
   

   Saved securely to ~/.pi/agent/pi-typesafe/auth.json.
2. Environment variable:

   export TYPESAFE_API_KEY=your_key_here
   export PI_TYPESAFE_ENABLED=1
   

3. Secret files:
   • ~/.pi/agent/pi-typesafe/auth.json
   • ~/.dsh/secrets/typesafe_api_key
   • ~/.pi/agent/secrets/typesafe_api_key

Verify your setup by running:

/typesafe status

---

Tools

1. typesafe_evaluate / jev_evaluate

Used by the model or code to get structured decisions, classifications, triage, and scoring.

{
  "state": {
    "title": "升级后无法登录",
    "body": "输入密码后一直回到登录页"
  },
  "questions": {
    "area": {
      "type": "choice",
      "instructions": "这个问题属于哪个模块？",
      "criteria": {
        "auth": "登录与身份验证",
        "ui": "界面与布局",
        "other": "都不符合"
      }
    },
    "blocking": {
      "type": "noul",
      "instructions": "这个问题是否阻止用户继续使用产品？"
    },
    "severity": {
      "type": "score",
      "instructions": "评估这个问题的严重程度：",
      "criteria": [
        "仅影响外观",
        "存在可用绕过方案",
        "阻止核心流程"
      ]
    }
  }
}

---

Running Tests

npm test

---

License

MIT © lldois