@techs/dsh-decision

@techs/dsh-decision

English | 简体中文

The decision layer for DeepSeek Harness (dsh): host-neutral typed judgments (binary, categorical, and ordinal), a versioned deterministic policy, and four dsh integration points: tool guardrail (tools/pre-execute), model routing (agent/request), result judging (tools/post-execute), and machine approval (approval/request).

As a Cordis plugin, it exposes ctx.decision for judgment provider registration:

export const inject = ["decision"];

export function apply(ctx: Context): void {
  ctx.effect(() => ctx.decision.registerProvider(provider));
}

Providers are separate packages; @techs/dsh-decision-jev is the first implementation. Any model that can provide structured probability judgments can be integrated through a JudgmentProvider adapter, even if it uses a different API protocol. The adapter declares its supported question types, registers with ctx.decision, and is selected by decision.config.provider. A new model needs its own threshold evaluation; Jev's calibration does not transfer automatically.

This package includes no provider, route table, or credentials. Its ./kernel export has no Cordis or Schemastery runtime dependency and is shared with Pi. The default shadow mode observes without changing behavior. Outbound state is redacted for recognizable secret patterns, and each judgment writes a sanitized audit record. enforce is experimental.

Install it with the Jev adapter into an existing dsh profile (requires a Vercel AI Gateway key):

export AI_GATEWAY_API_KEY=your_gateway_key
dsh plugin --profile web add @techs/dsh-decision @techs/dsh-decision-jev
dsh web

Each package declares a dsh.bundle, so installation adds both configuration layers. The defaults observe tool risk and retain dsh's native permission mode.

• Design and v2 plan (Chinese)
• Integration and verification (Chinese)
• Threshold evaluation (Chinese)
• Repository README