Skip to main content

​
crewship decisions

Evaluate explicitly supplied text using Jev through TypeSafe or OpenRouter. This local pilot needs no running Crewship server. It returns suggestions and never assigns work, executes actions, changes permissions, or deletes memory. Results support --format json, yaml, and ndjson; quiet suppresses results. The default view prints the complete JSON document. A --dry-run always prints the provider request as JSON for inspection or reuse, regardless of format. Errors follow the CLI’s global error formatting. Set TYPESAFE_API_KEY or OPENROUTER_API_KEY securely in the environment. Credentials are not read from the server vault. Calls are billed directly to that key and are not recorded in server Paymaster accounting.

​
Shared flags

The input/request limit is 96 KiB and the question limit is 64. These local bounds do not guarantee compliance with the provider’s token limits. There are no automatic retries; errors exit nonzero without printing upstream bodies.

​
evaluate

crewship decisions evaluate accepts a JSON object with state, questions, and an optional model. Instructions must be strings. Questions support noul, choice (2–255 options), and score (2–10 text levels).
Example request.json:
Remove --dry-run to send the supplied state and questions. The response includes model version, answers, probabilities where applicable, token usage, and elapsed latency. Provider-reported cost may be absent, meaning unknown. The shared --threshold is accepted but only affects triage.

​
triage

crewship decisions triage accepts plain text. One request asks about the primary subsystem, reproduction evidence, explicit production impact, and requested destructive actions.
The result preserves all answers and reports needs_review and its reason. Unknown categories, low confidence, or production/destructive signals request review. Even needs_review=false is only an advisory category suggestion, never permission to execute the request. The threshold is not calibrated on production Crewship data or Czech input.

​
rerank

crewship decisions rerank accepts a query and 1–30 already-authorized candidates. Each candidate needs a unique ID and nonempty text. The caller owns access control; the command does not fetch memory or validate tenant scope.
Example candidates.json:
The four-level relevance rubric evaluates all candidates in one call. Output retains every candidate ID and original position; ties preserve input order. Scores are rubric values, not BM25 scores or calibrated relevance probabilities. On API failure, the command exits nonzero: callers should retain original retrieval order. The shared --threshold does not affect reranking.

​
Pilot validation

This recovery includes the local CLI and its bounded provider client. Provider responses and failures are exercised with isolated HTTP fixtures; dry-run recipes are exercised through the CLI. Hosted inference and confidence calibration have not been verified. See the provider’s OpenRouter Jev tutorial for the hosted decision interface. Routine decision routing and the earlier local SemIf/MLX experiments remain outside this recovered CLI. This command supplies no execution authority.