| name | jev-decision |
|---|---|
| description | Selectively assess bounded evidence relevance, classifications, rubric scores, or verification gaps using the installed Jev advisory MCP tools or managed CLI. Use only when a separate assessment can improve a consequential decision or reduce a large saved evidence artifact before context ingestion. Routine commands and simple tasks need no Jev call. |
Use your usual model for reasoning, coding and straightforward work. Jev is optional advice. Native permissions and executable tests remain authoritative.
Prefer the installed Jev MCP tools when available. Check jev_status if readiness is unclear. For CLI-only clients, use the absolute managed launcher supplied by your installed copy of this skill. Never install from a mutable checkout or place a credential in a command, setting or prompt.
Send only the smallest relevant, sanitized evidence. Batch related typed questions against one complete evidence window. Preserve descriptive rubric levels, failures, exit codes, test summaries, sources and original artifacts. For a large saved log, use jev_read_evidence before loading it into model context; approved roots and content limits apply.
Treat status unavailable or offline as no advice. Continue normal reasoning and verification. Do not retry repeatedly after budget exhaustion or authentication failure. Confidence and absent usage can be unknown. A high assessment never authorizes a command or proves a task is complete.
Automatic pruning stays off until independently labeled development and held-out validation demonstrates useful savings while preserving required evidence. Keep uncertain content. Do not claim percentage savings, billing savings or improved correctness without measurements.
Use jev_guard_command only for nontrivial risk triage and jev_verify_completion only to identify evidence gaps. Never use either as a permission gate or completion certificate. Never mutate memory or benchmark grades solely from Jev output.
If a Jev guard hook asks for confirmation or blocks a shell command, tell the user what was flagged and let them decide. Never rephrase, split or obfuscate a command to get past the guard.