ninetyhe-90/dsh-decision-review

[!NOTE] This bundle installs with mode: off and makes no calls. Enabling it requires a compatible POST /v1/decision service plus an explicit model and baseUrl.

Why this plugin

Asking one model to check its own answer usually repeats its own reasoning. A second opinion is useful only when the reviewer sees the question, the primary choice, the evidence, and the constraints without joining the original reasoning or holding tool permissions.

The primary model must commit to primaryChoice before review starts. The suggestion and its confidence return to the primary model, which still decides whether to adopt, keep, or ask the user.

flowchart LR
    A[Primary model picks primaryChoice] --> B[__decision_review tool]
    B --> C[POST /v1/decision]
    C --> D[choice + confidence]
    D --> E{confidence}
    E -->|>= minConfidence| F[ok: adoptable]
    E -->|>= abstainBelow| G[not_adoptable]
    E -->|< abstainBelow| H[abstained]
    F --> I[Primary model decides]
    G --> I
    H --> I

Capabilities

Capability	Behaviour
Three modes	off registers nothing; shadow logs only; advisory returns a suggestion
No decision hijack	Review returns advice and confidence; the primary model still chooses
Structured output	status, recommendation.choice, confidence, reasonCode
Confidence tiers	Adoptable, keep-but-not-recommended, or abstained
No retries	A failed review returns immediately; the primary judgement stands
Redacted logs	Point, status, reason code, model, HTTP status, request id, input fingerprint, latency
Permission isolation	Permissions, approvals, writes, network targets, and allowlists never reach the reviewer

Protocol

POST {baseUrl}/v1/decision
Authorization: Bearer {DECISION_API_KEY}

The body has exactly three top-level fields: { model, state, questions }.

Field	Source
state.question	Tool input question
state.primaryChoice	Tool input primaryChoice
state.evidence	Tool input evidence
state.constraints	Tool input constraints
questions.recommendation.criteria	Tool input options
Response	answers.recommendation.choice and confidence

probabilities are informational only. baseUrl is the gateway root; an address already ending in /v1 or the endpoint path is rejected at startup.

Install

dsh plugin --profile <name> add dsh-decision-review
dsh plugin --profile <name> add /path/to/dsh-decision-review
dsh plugin --profile <name> add github:ninetyhe-90/dsh-decision-review

Configuration

Field	Type	Default	Description
mode	off | shadow | advisory	off	Run mode
model	string	none	Required reviewer model name
baseUrl	string	empty	Reviewer gateway root, without /v1
apiKeyEnv	string	DECISION_API_KEY	Environment variable holding the credential
allowedDecisionPoints	string[]	[]	Decision point names allowed to call review
timeoutMs	number	5000	Per-review timeout
minConfidence	number	0.75	Adoptable threshold
abstainBelow	number	0.5	Abstain threshold
promptSection	boolean	true	Inject the prompt section
promptSectionOrder	number	30	Prompt section order
systemExtra	string	empty	Extra domain constraints

- insert:
    - id: decision-review
      name: dsh-decision-review
      config:
        mode: advisory
        model: your-decision-model
        baseUrl: https://your-gateway

Failure codes

reasonCode	Meaning
timeout / cancelled	Local timeout or parent request cancelled
upstream_timeout	Upstream timeout reported by the gateway (HTTP 504)
credential_unavailable	Credential environment variable missing
authentication_failed	Credential missing or rejected (HTTP 401)
model_not_authorized	Credential cannot use the model (HTTP 403)
model_unsupported	Gateway does not know the model
invalid_request	Request body rejected (HTTP 400)
route_unavailable	HTTP 404; check baseUrl
rate_limited / quota_exceeded	Upstream limit or exhausted budget
context_too_large	HTTP 413
invalid_response / response_too_large	Unparsable or oversized response
provider_error	Any other network or server failure

Security boundaries

• The reviewer receives no tools, files, network access, or approval rights.
• Privilege escalation, user confirmation, writes, network targets, and allowlists stay outside review.
• Review output is advice; it cannot bypass the primary model or host confirmation.
• Logs omit request bodies and raw errors.

Compatibility

• Node.js 20.11+
• DeepSeek Harness 0.2.x (ctx.tools, ctx.logger; optional ctx.systemPrompt)
• @deepseek-ai/schemastery 3.18.x

Layout

src/decision-core.mjs    config validation, input normalisation, prompt section, reason codes
src/decision-model.mjs   tool registration, review request, result mapping
test/                    config, protocol, and tool behaviour
cordis.patch.yml         bundle layer, mode: off

Development

npm install --include=dev
npm test
npm run check

Roadmap

• Three modes and confidence tiers
• Structured failure codes and redacted logs
• Configurable decision point allowlist
• Bundle packaging
• npm release
• Minimal reference reviewer service

License

MIT