@yadsh/dsh-decision-tools

Agent tools for typed decisions through configured System One-compatible servers, including local Jeff. No dependency on Jev Compaction or its session hooks.

Tools

Tool	Use
dsh_decision_capabilities	List configured backend ids, declared kinds, credential readiness and limits; no inference
dsh_decision_ask	Send a batch of arbitrary noul, choice and score questions
dsh_decision_classify	Exclusive classification, or one yes/no question per label with multilabel: true
dsh_decision_check	Evaluate a named checklist with yes/no/uncertain results
dsh_decision_score	Rate text or JSON context against ordered levels
dsh_decision_rank	Score and rank candidate texts for a goal, retaining all candidates and stable tie order

noul returns the source score and yes, no or uncertain. Default yes/no thresholds are 0.7/0.3. choice returns the winning label, the complete distribution, backend confidence and an uncertainty flag. score preserves the provider's ordinal expectation on 0..levels.length-1, confidence, probabilities and the rubric. It is not converted into a made-up percentage. Choice/score confidence below 0.5 is marked uncertain by default.

These scores are model outputs, not guarantees of factual correctness. In Jeff, confidence describes distribution peakedness; displayed probabilities can be temperature-scaled while the ordinal score uses an untempered distribution. Do not recompute score from displayed probabilities. Questions can affect each other when Jeff runs with JEFF_ISOLATE=none; multiple-label questions are not guaranteed statistically independent. Ranking may likewise depend on the other candidates in the request. Validate task-specific quality before treating these scores as action gates.

Configuration

The host entry namespace is dsh-decision-tools. A host-only plugin has no custom browser bundle or settings card; its Schemastery Config is the native configuration surface. Example for a self-hosted backend:

- id: dsh-decision-tools
  config:
    defaultBackend: local
    backends:
      - id: local
        baseUrl: http://scorer.example:8000
        model: jev-latest
        apiKeyEnv: DECISION_API_KEY
        kinds: [noul, choice, score]
    maxStateTokens: 2400
    maxRequestTokens: 2600
    maxStateChars: 9000
    timeoutMs: 45000
    maxConcurrency: 1
    maxQueue: 8

Only the operator configures URLs, models, credential-variable names and supported kinds. A tool may select an existing backend id, never supply a URL or credential. The default backend points at http://localhost:8000 and reads JEFF_API_KEY; in a container configure the actual service address. An empty apiKeyEnv intentionally disables the Authorization header. No other provider is selected automatically after a failure.

All config values are validated: 1–8 active calls, 0–64 queued calls, 100–120000 ms deadline, 1–128 questions, 2–64 labels, and 1 KiB–1 MiB response cap. Defaults are one active call, eight waiting calls, eight questions, sixteen labels, 45 seconds, and 256 KiB response bytes. Token estimates are heuristics, not the backend tokenizer or billable usage. The state is never silently truncated: oversize state/questions are refused with STATE_LIMIT or REQUEST_LIMIT, so the agent must choose an explicit excerpt or split the call. The concurrency gate covers this plugin's tools; other plugins using the same inference server have their own concurrency controls.

Examples

Choose an approach for a stated constraint:

{
  "state": "Need to inspect an imported report while preserving original files.",
  "instructions": "Which approach best matches the requirement?",
  "labels": [
    { "id": "copy", "description": "Copy the input and inspect the copy" },
    { "id": "overwrite", "description": "Edit the original file directly" }
  ]
}

Pass that to dsh_decision_classify. For dsh_decision_score, provide state, instructions and levels, e.g. ["irrelevant", "partly relevant", "directly relevant"]. Levels can also be { "what": "high", "examples": ["blocks the task"] } objects, preserving the native score contract.

For dsh_decision_rank:

{
  "goal": "Inspect report data while preserving original files",
  "levels": ["poor fit", "partial fit", "good fit"],
  "candidates": [
    { "id": "copy", "text": "Copy the report, then inspect the copy" },
    { "id": "edit", "text": "Rewrite the original report" }
  ]
}

The result includes a descending ranking and the underlying typed answers. Equal scores keep input order. No chosen candidate is executed by this plugin.

For dsh_decision_ask, pass state and a questions array:

{
  "state": { "message": "Report export crashes and blocks the task" },
  "questions": [
    {
      "id": "defect",
      "type": "noul",
      "instructions": "Does this describe a defect?"
    },
    {
      "id": "urgency",
      "type": "score",
      "instructions": "Rate urgency",
      "criteria": ["low", "medium", "high"]
    }
  ]
}

For dsh_decision_check, pass state and criteria entries with id, instructions and optional yes/no criterion descriptions. All question, label and candidate ids must be unique simple identifiers starting with a letter.

Security model

Calls send only explicit supplied context to an operator-configured endpoint; they do not read the conversation, files or network sources automatically. Bearer keys are read from environment variables per call and never returned. Redirects are rejected. No request/response body is logged. Errors contain a stable category and actionable summary without raw backend bodies. Canceling a tool or disposing the plugin cancels active and waiting calls; the wall deadline includes queue wait and response streaming. Response bytes are capped before JSON parsing, duplicate JSON keys and malformed typed answers are rejected, and returned labels must exactly match the request. Configuration-declared capabilities do not certify a backend or its quality.

Deployment and validation

Install using dsh plugin --profile <profile> add @yadsh/dsh-decision-tools. On a preset-based deployment, grant the six exact tool names above to the intended agent role. Installing the bundle only registers global tools; it does not bypass role filters or prove discovery in an existing chat. Start a new chat and verify its actual tool catalog after applying the preset.

Run pnpm --filter @yadsh/dsh-decision-tools run check and pnpm tarball:verify:packages plugins/dsh-decision-tools. Unit checks include all question kinds, malformed responses, full-request/state limits, queue cancellation, deadlines, disposal, multilabel classification and stable ranking. Live tests must separately verify the local provider, all three question kinds, and the agent-visible tools on the target host. A compatible wire format is not a quality or calibration guarantee.

Requirements and license

DeepSeek Harness >=0.1.7-rc.2 <0.2.0; Node.js ^22.19.0 or >=24.0.0. MIT.