dsh-decide

dsh-decide

English | 简体中文

Multi-provider decision models for DeepSeek Harness (DSH): one pair of agent tools over TypeSafe Jev, TokenRhythm NeoHorse-Jev, OpenAI Decisions, and local System One servers.

Decision models are not chat models. They take a body of text (state) plus named typed questions and return probabilities / labels / scores — no prose, no streaming, no messages. That is why dsh-decide adds tools and a settings page instead of a model in the model picker.

Providers

id	Service	Dialect	Default model	Key
typesafe	TypeSafe Jev	System One	jev-latest	TYPESAFE_API_KEY
tokenrhythm	基元律动 NeoHorse-Jev	System One	NeoHorse-Jev-4B	TOKENRHYTHM_API_KEY
openai	OpenAI Decisions (beta)	OpenAI Decisions	gpt-6-luna	OPENAI_API_KEY
local	Local System One server (Laya via Unsloth, self-hosted models)	System One	laya	optional / keyless

Any other endpoint can be added as a custom provider with either dialect. The first enabled provider whose key resolves is used by default; set defaultProvider to pin one. Each provider card links to its website with the aichen referral tag appended.

Install

Development (this checkout; replace the path with your clone):

dsh plugin --profile desktop add link:F:\path\to\dsh-decide

Then restart DSH (host plugins are not hot-reloaded). In this repository the link is already installed; dsh plugin adds both the dependency and the bundle entry.

Published install is dsh plugin --profile desktop add dsh-decide once the package is on npm.

Settings

Settings → Decisions has four tabs:

• Providers — per-provider enable/disable, model, base URL, path, key status, save, test, list models, set default, and a form for custom providers.
• Examples — copyable prompts and the exact tool request, so users can start without reading docs.
• Behavior — confidence thresholds, request limits, cache, audit, and the optional execution guard.
• Usage — request/token/cost totals per provider for today and the last 7 days, plus recent decisions.

The page follows the DSH language setting. English and Chinese are complete; the other shipped DSH languages (Japanese, Spanish, French, German, Russian, Portuguese) translate the main interface and fall back to English for long help text. Contributions for the remaining strings are welcome.

Keys are written through the DSH credentials service; the page only ever reports whether a key is set and where it comes from. The state you judge is sent to the selected provider — do not put secrets in it.

Tools

decide

{
  "state": "I was charged twice for my March subscription.",
  "questions": [
    { "name": "billing", "type": "yesno", "instructions": "Is this about billing?" },
    {
      "name": "team",
      "type": "choice",
      "instructions": "Which team should handle this?",
      "choices": [{ "label": "billing" }, { "label": "technical" }]
    },
    {
      "name": "urgency",
      "type": "score",
      "instructions": "How urgent is this?",
      "levels": [{ "label": "not urgent" }, { "label": "this week" }, { "label": "today" }]
    }
  ]
}

Returns one answer per question with probability / choice / score, confidence where the provider supplies it, and uncertain from the configured thresholds. Set includeProbabilities: true to get the full distribution. provider overrides the default; state must be a string (serialize JSON first).

decide_gate

Asks one yes/no question — by default “could this action irreversibly delete, overwrite, or expose real data?” — and maps the probability to allow, review, or deny. Use it before destructive actions. The action text is sent to the provider; failures are reported as errors, never as approval.

decide_rerank (beta)

Scores a list of passages against a query and returns them sorted. Useful before stuffing retrieved context, when keyword search returned too much:

{
  "query": "How do I rotate the API key?",
  "documents": [
    { "id": "doc-1", "text": "..." },
    { "id": "doc-2", "text": "..." }
  ]
}

decide_judge (beta)

Judges a candidate answer or tool result against explicit yes/no checks and returns PASS / FAIL with one probability per check. Useful as a review gate or LLM-as-a-judge:

{
  "task": "Answer the user's question about refunds",
  "output": "...",
  "checks": ["cites a source", "answers the question", "contains no invented policy"]
}

Experimental: loop triage

When enabled in Behavior → Experimental, the plugin listens on agent/pre-step, asks the decision model whether the new message is quick, deep, or tool, and appends a short hint to the turn. It is off by default, costs one decision call per turn, and its behavior may change between releases.

How to use in DSH

1. Open Settings → Decisions → Providers, pick a provider, press Test, then Use as default.
2. In chat, describe the decision and the options: “Use decide to route this ticket to billing / technical / sales and score its urgency.” No skill is installed; the agent picks tools from their descriptions, and naming decide just makes the choice deterministic.
3. Before a risky shell command, say “call decide_gate first and continue only if it says allow.”
4. For RAG or review flows, ask for decide_rerank / decide_judge.
5. Optionally enable the execution guard or experimental loop triage in Behavior (both off by default; do not combine the guard with dsh-jev-guard).

The full walkthrough, example prompts, and troubleshooting table: docs/usage.md. 中文说明：docs/usage.zh.md.

Configure without the page

- id: decide
  name: dsh-decide
  config:
    defaultProvider: tokenrhythm
    maxQuestions: 8
    timeoutMs: 20000
    maxRetries: 2
    thresholds:
      choice: 0.7
      score: 0.7
      yesnoLow: 0.35
      yesnoHigh: 0.65
    providers:
      tokenrhythm:
        enabled: true
        dialect: systemone
        baseURL: https://tokenrhythm.studio/v1
        path: /systemone
        apiKeyEnv: TOKENRHYTHM_API_KEY
        model: NeoHorse-Jev-4B

Precedence: profile config > ~/.dsh/decide/settings.json (written by the page) > defaults. Keys are never stored here.

Limits, confidence, cost

• Each provider declares its own limits (question count, choice/score counts, state size, body size). They are enforced locally before a request, with an actionable error.
• Do not reuse thresholds across providers. Jev, Laya and OpenAI compute confidence differently; the defaults threshold on the top probability for choice and on noul/predicate probability for yesno. Per-provider threshold overrides are supported.
• Cost is estimated from usage.input_tokens and a per-provider price (TypeSafe: $0.042/M input, output free). TokenRhythm and OpenAI prices are left unknown until configured.
• Timeouts and network failures are not retried; only 429/5xx/529 responses are (the API has no documented idempotency key).

Privacy

• state is sent to the selected provider. Local providers keep it on the machine.
• Audit records (~/.dsh/decide/decisions.jsonl) store a hash and character count by default; storeState: full is opt-in.
• API keys live in the DSH credential store or the environment, never in settings or logs.

Development

npm test

The suite is offline: adapters, dialect normalization, limits, thresholds, retries, audit, routes, and a fake Cordis host. Node 20+; no runtime dependencies.

Known gaps

• The tools/pre-execute guard is implemented but off by default (guard.enabled), and only checks bash/pwsh. Do not enable it together with dsh-jev-guard.
• decide_rerank and decide_judge are beta; loop triage (experimental.triage) is experimental and may change between releases.
• Images / box questions are not sent yet.
• Native Laya /predict (not System One) is not adapted; use the Unsloth System One endpoint or a proxy.
• OpenAI Decisions is beta; the native path is configurable (baseURL + path) if it moves.

About

Maintained by 老陈聊AI (Xiaohongshu). Provider website links carry the aichen referral tag so the author can track partnerships; the tag is a single constant in lib/providers.js.

License

MIT