jevify
Runtime model calls often hide small structured decisions inside expensive or hard-to-validate generation paths.
Why this exists
Everyone is talking about performance gains and token savings from moving decisions off general-purpose LLMs. The harder question is what that looks like in practice, in your app: which calls, which lines, what changes, and how you would see the gain rather than take it on faith.
jevify is built to make that concrete, in three steps you can inspect:
- See where it applies. The audit points at real call-sites (
path#line), shows how each response is consumed, and separates the bounded decisions — a queue, a score, a yes/no — from the generation that should stay generation. A clean audit is a valid answer.
- Picture the change. Each candidate gets a before/after sketch: the primitive, the TypeSafe pattern, the closest official cookbook, and what your app would need to adapt.
- Measure it on your own traffic.
/jevify migrate runs the new path in shadow beside the current one and logs both: cost per decision, latency p50/p95, agreement, labeled disagreements, fallback rate. The gain becomes a number from your data, or the migration stops.
The pieces share one method: this skill is the canonical source, the MCP service at jevify.lovable.app runs the same audit remotely and keeps reports under your account, the open-source local MCP server runs the same engine offline on your machine, and the Lovable skill brings it into Lovable projects. Every effect stays a hypothesis until shadow data says otherwise.
Before and after
// Before: a model call returns a label; the app only uses the label.
const result = await generateObject({
schema: z.object({ queue: z.enum(["sales", "support", "other"]) }),
prompt: `Route this ticket: ${ticket.message}`,
});
return dispatch(result.object.queue);
// After /jevify migrate: JEV beside the current path, behind a flag. Shadow never blocks or changes the answer.
const MODE = env.JEVIFY_TRIAGE_MODE ?? "shadow"; // off | shadow | on
const MIN_CONFIDENCE = 0.8; // placeholder until tuned on shadow data
if (MODE === "on") {
const jev = await decideQueue(ticket).catch(() => null); // error or timeout → current path, never "no"
if (jev && jev.answer !== "insufficient_context" && jev.confidence >= MIN_CONFIDENCE) {
return dispatch(jev.answer);
}
}
const current = await currentModelRoute(ticket); // the existing generateObject call, unchanged
if (MODE === "shadow") void decideQueue(ticket).then((jev) => logShadow({ current, jev })).catch(() => {});
return dispatch(current);
decideQueue sends a Choice question over { ticket }; its request shape comes from the current TypeSafe docs at migration time.
What the audit produces
From the sample report over this repository's evals/ fixtures (local mode, --wide):
16 runtime call-sites in 13 files · 10 JEV candidates (1 inside a composite) · 2 generation tasks retained
1 deterministic · 1 embedding search · 1 human review · 1 unknown · 1 consolidation group · 1 wide opportunity
evals/fixtures/ticket-router.ts#3 | JEV_CANDIDATE | Choice | LOW | dispatch() uses only the label
evals/fixtures/announcement.ts#6 | GENERATION_REQUIRED | — | — | publish(r.text)
evals/fixtures/ban-account.ts#6 | HUMAN_REVIEW | — | HIGH | policy requires a moderator
Consolidation: consolidation.ts #6 #7 #8 → one request with Choice + Score + Noul over { ticket }
How that was produced and what it does not prove: evals/results-local-v1.3.md.
Install
- Agent skills:
npx skills add lucioamor/jevify
- Claude Code:
claude plugin marketplace add lucioamor/jevify then claude plugin install jevify@jevify (untested until the v1.3 branch is published)
- Lovable: import
https://github.com/lucioamor/lovable-skill-jevify under Settings → Skills → Add → Import from GitHub
- Local MCP server (offline, no account):
claude mcp add jevify -- npx -y @nxlv-ai/jevify (source)
See INSTALL.md for MCP connection details.
Commands
/jevify audits and reports without editing source; add --wide for non-call semantic-code opportunities./jevify migrate <finding> plans and, after explicit approval, migrates one candidate in shadow mode.
Installed as a Claude Code plugin, the command may be namespaced as /jevify:jevify.
How classification works
In local mode, the agent reads call-sites, prompts, response parsing, and consumers, then writes jevify-report.md when it has file access or reports in Lovable chat. Local mode sends no code to the jevify service, although the agent itself may use remote processing.
In MCP mode, the jevify service classifies a narrow source window and stores its service report under the authenticated account. That is triage: the agent must inspect the consumer for every JEV_CANDIDATE and DETERMINISTIC_CODE, preserve the service result, and record confirmation or disagreement. The service also exposes an optional decision proxy at /api/public/jevify/decision and a harness; the audit and migrate flows do not call JEV.
Use with TypeSafe's official skill
Install the official TypeSafe agent skill linked from llms.txt for detailed question design. /jevify identifies where a migration may fit and governs how to validate it safely; it does not reproduce TypeSafe's skill.
Each candidate is routed, after classification, to a primitive, one of TypeSafe's patterns, and the closest official cookbook, with a line on what the app must adapt. The skill carries only the selection signals; the recipes stay in TypeSafe's live docs, and their measured results are never presented as expected effects for your app.
Honest limits
- Static inspection can miss dynamic calls and cannot prove production behavior.
- MCP classifications see limited context and require consumer verification.
- Agreement with the current model does not prove correctness; label divergence samples.
- Availability, pricing, data terms, API fields, SDKs, and model names must be checked before production use.
- HIGH-risk decisions remain shadow-only until an authorized human approves a separately defined action policy.
License and authorship
Created by Lucio Amorim, Lovable Partner. Dual-licensed by content type:
- Skills, documentation, report templates and other prose: CC BY 4.0. Retain attribution and indicate changes when redistributing.
- Code (
scripts/, launch/, evals/, .claude-plugin/, .mcp.json and any other source files): Apache License 2.0.
