Jev Decision Guide

---
name: jev-decision-guide
description: Set up or upgrade Qualixar Jev Decision Layer in a workspace, or use bounded task, tool, skill, recipe, review, and context decisions; automatic prompt guidance is separately consented and selective.
---

# Qualixar Jev Decision Layer

This independent Qualixar integration adds typed decisions alongside your coding model. TypeSafe Jev is the hosted route; Laya is an optional local route on an Apple-Silicon Mac, installed with `jev laya-install` and verified before it can be chosen. The model advises; the agent and native host retain the task, tools and permissions.

1. Check `jev_auto_status` for the current workspace before a live decision. If it is not enrolled, or the user asks to change the saved scope, call `jev_setup` with that exact workspace path. It opens the private two-step browser wizard; only the user enters a provider key and confirms scope there. Never ask for a key in chat or create consent on the user's behalf. A reviewed upgrade keeps the existing Keychain key and daily budget identity when the provider is unchanged.
   If automatic Jev prompt guidance was enabled in the setup review, an eligible coding prompt may already include a receipt-bearing advisory file or skill suggestion. Use that suggestion only if relevant; do not make a duplicate routing call merely to confirm it. No prompt is sent automatically to a hosted provider without that separate opt-in.
2. Every tool that sends text needs `workspace_path` (absolute) and `data_classification`: the value the grant accepts, as the session guidance or `jev_auto_status` states it. Under a Jev + Laya grant, pass `internal-minimized` for ordinary text and `restricted` for private or client content; a restricted decision is made by Laya on this Mac and never sent to Jev. Never pass a wider value than the grant allows.
3. For a closed task, tool or skill shortlist, call `jev_route` with 2–12 candidates, each exactly `{"id", "description"}` (id: a letter first, then letters, digits, `_` or `-`; no dots; never `unknown`, which the tool adds itself). Treat `unknown` or a provider error as a handoff to ordinary agent judgment, not a reason to guess or retry indefinitely. The result never authorizes execution.
4. Use `jev_recipe_catalog` to show available use cases. `jev_recipe_try` accepts exactly one recipe's fields and returns experimental advice; the hand-authored fixture examples and uncalibrated thresholds are not model-quality proof. A recipe with fixed options chooses only among its own options; for the user's own list, use `jev_route`.
5. Use `jev_verify` to check an extraction against its source (branch on `trustworthy`, never on an empty suspect list) and `jev_rerank` to check whether retrieved passages answer a question (each memory needs `content`; obey `should_abstain`).
6. Use `jev_review_diff` only for an explicitly supplied, screened unified diff. Run the relevant tests and independent code review afterward; Jev's risk and focus answers cannot approve a change.
7. Use `jev_reduce` when a bounded extractive view is useful, and `jev_recall` for exact omitted text. The original source remains authoritative. A shorter view or decision count is not a measured host token, cost or time saving.

Keep all normal host instructions, the user's other tools and memory, any existing router, and host permission checks intact. If the local broker or provider is unavailable, continue the user's authorized work without silently substituting a fixture. Do not send secrets or whole private repositories to a hosted provider. Restricted non-secret text may go to a hosted provider only when the exact workspace policy selects Jev maximum; under Jev + Laya or Laya only, restricted text stays on the Mac. Never infer either permission from this skill alone.