Our sandbox completed local commands, not a live support session
Our run on October 5, 2026, at commit c469e5b recorded a successful Agent Assist build in 13 seconds. The attached harness records installation separately at 22 seconds and the test command at 7 seconds. These are wall-clock command durations in our sandbox, not model-response times. The build tail shows TypeScript and Vite completing; it does not establish that a browser session, hosted request or deployment worked.
Our harness records command outcomes, exit codes, timeout flags, elapsed time and parsed test summaries. The stored parser summary reports 14 passed, 0 failed and 0 skipped tests. The raw tail names engine.test.ts and questions.test.ts, then repeats the final totals. We have not reconciled complete logs or verified exhaustive unique coverage. The successful commands provide evidence for evaluating this subproject locally, without establishing what every test covers or whether other demos work.
The MrKeyoor testing methodology explains how command records inform reviews. The supplied methodology page was fetched on October 31, 2026. It provides provenance context, but does not prove that its current procedure governed our October 5 run. This review uses the stored run and adds no new execution.
Agent Assist selects replies while code enforces policy
Agent Assist uses Jev to select handwritten replies. The Agent Assist README describes a support console where each incoming customer message triggers one request containing questions about intent, churn risk, frustration, flags and macro selection. Refund eligibility remains a TypeScript calculation based on plan and tenure. This division gives a developer a concrete boundary to inspect: the model judges language, and application code decides how to act on its output.
The confidence threshold is part of that boundary. At confidence 0.6 or above, the selected macro is auto-filled; below it, the interface presents candidates for the agent to choose. A none result means the agent writes a reply. The documented hands-free path can auto-send an auto-filled reply 1.5 seconds later. Before adapting that behavior, decide which actions require confirmation and test cases where a plausible suggestion would violate policy.
The Jev Experiments introduction describes separate applications with their own documentation. Start with Agent Assist if support workflows are your question. Our sandbox evidence applies to that subdirectory and supplies no build or runtime result for the other applications.
Live evaluation requires credentials and approved conversation data
The documented live request sends the last six conversation messages, customer plan, tenure, prior-ticket count and macro summaries to TypeSafe. According to the README, the API key stays in the local server and never reaches the browser bundle. Keeping the key server-side does not keep conversation text local. Evaluate this path with data you are permitted to send to the service; use mock mode for interface work that does not require model judgments.
The README reports a 93 ms median panel refresh from its own scripted live run and describes hand-checked judgments on seeded conversations. Those are upstream observations, not our measurements. Its four-second language-model baseline is explicitly simulated. Our run therefore establishes neither that latency nor a measured speed advantage over another deployed copilot. A useful next evaluation should record answer quality separately from panel-update time, using your account, network and cases where mistakes carry a cost.
The recorded audit outputs conflict
The stored audit summary records zero advisories. Separately, the raw installation output says "2 critical severity vulnerabilities". Neither output explains the disagreement, and the installation message does not identify affected packages or establish exploitability. Both remain part of the record. A zero summary cannot support a clean security conclusion when the supplied raw output conflicts with it.
Scan coverage is unknown because the record has no scannedPackages field. The stored dependency count measures footprint and cannot fill that gap. Obtain a current audit report identifying the dependency tree and affected packages before adopting code. That would be a separate check; this correction does not claim that one occurred. The stored license signal is also null, which leaves licensing unresolved rather than establishing a legal conclusion.
Alternatives address different parts of the application
AI SDK provides a unified API for multiple model providers. Its repository documents text generation, structured output, agents and UI integration. Evaluate it when generated responses or provider choice are central to the application. It is an application toolkit that could complement a decision-oriented design; the supplied evidence does not establish Jev integration or a drop-in replacement for Agent Assist.
NanoJev makes its model and dataset publicly downloadable without signing in. Its repository also supplies local inference instructions for a CUDA environment and reports game evaluations. Evaluate it when inspecting or operating a decision model is the goal. This is a separate model evaluation: the supplied game results do not establish support-triage accuracy or compatibility with Agent Assist.
Ratings apply to the evidence available here
The five-point rubric is: 1 for demonstrated obstacles that prevent the intended evaluation, 2 for a limited reference requiring substantial adoption work, 3 for adequate evidence with material gaps, 4 for strong evidence within the stated scope, and 5 for broader corroboration across that scope. Unknown dimensions receive no numerical rating. These are editorial assessments, not measured product benchmarks.
Setup ease is 4 for the measured subproject because installation and build succeeded, with live setup still requiring credentials and validation. Documentation is 4 for the supplied Agent Assist instructions covering requests, policy rules, confidence handling, mock mode and startup. Community remains unassessed because the supplied primary extracts do not establish maintainer responsiveness or sustained contribution activity. Maturity is 2 for adoption as a support application: local commands succeeded, but live workflow validation, audit reconciliation and licensing remain unresolved.

