Jev Ultrafast asks which browser control to use next. Fast Jev Compaction asks whether an earlier tool result still belongs in the conversation. In both, the application code uses the answer to decide what happens next.
These ten GitHub projects are listed on Jevfast.com, with source references and a "What Jev decides" chart for each. They cover different uses with decisions that can be traced in code. The order does not rank their stars, quality or measured performance.
Reading the charts
Each chart puts the application's instruction and input beside the typed answer. Follow the source link to check how the code uses that answer.
Ultrafast and Fast Jev Compaction show recorded samples. The other eight use illustrative answers. Their captions and chart labels identify which kind of evidence you are looking at.
| Project | Decision pattern |
|---|---|
| Jev Ultrafast | Choose a browser operation and its target |
| Fast Jev Compaction | Keep useful context without rewriting it |
| Jevgrep | Judge whether code is evidence |
| LiteLLM | Choose a model tier before generating an answer |
| Hermes Jev Skills | Combine several judgments about one situation |
| OpenMuse Generative UI | Choose a response format and score its candidates |
| Jsort | Build a ranking from pairwise judgments |
| JEV Shield | Turn command judgments into permission outcomes |
| fx | Review a pending action for malicious activity |
| Jev Trader | Ask a bounded market-direction question |
1. Jev Ultrafast: browser actions
GitHub repository · Jevfast decision view
Ultrafast sends the current page, numbered controls, recent actions and the user's goal to Jev. One request asks for an operation and a compatible target. When the app needs to enter text, a separate small language model supplies the field value.
The target belongs to the observation that produced it. Before acting, the code must check that the page still matches.
Recorded sample: CLICK has probability 0.77; the separate target question gives the ticket-type control 0.88. Neither number measures whole-task success, and this sample is not a general benchmark.
Pinned source for the request.
2. Fast Jev Compaction: retaining tool history
GitHub repository · Jevfast decision view
For each eligible tool call, this plugin asks two Noul questions: should the call stay, and should its full result stay? Code compares the answers with a threshold and keeps both, shortens the result, or removes the pair.
This lets it keep the record of an action while dropping a bulky output. The call and its result can have different retention rules.
Recorded trial on a synthetic demo conversation: two NO answers concern an earlier edit after the task moves to a changelog entry. The chart shows two of fourteen questions.
Pinned source for the request.
3. Jevgrep: finding code evidence
GitHub repository · Jevfast decision view
Jevgrep prepares candidate code blocks with their paths and declaration context. It asks Boolean questions about relevance and whether a block belongs to the scope named in the query.
That scope check matters when a file contains familiar words but concerns a different API. The questions let the caller distinguish a relevant block from one about the wrong component before sending more context to a coding agent.
Illustrative request: the two YES selections show relevance and scope answers. They are not a captured search result.
Pinned source for the request.
4. LiteLLM: routing to a model tier
GitHub repository · Jevfast decision view
When its Auto Router uses the Jev classifier, LiteLLM sends the prompt and configured tier descriptions in a Choice request. The selected tier determines which configured model pool receives the request.
The routing candidates need to match the deployment. Jev chooses where completion generation happens; the chosen model generates the answer.
Illustrative tier choice. A configuration can disable NON_REASONING or replace the tier IDs. The displayed selection was not measured.
Pinned source for the request.
5. Hermes Jev Skills: several questions about one state
GitHub repository · Jevfast decision view
Hermes Jev Skills packages decisions for routing, memory, compaction and other agent features. Its documented CLI example describes a failed deployment and asks Noul, Choice and Score questions about that same state.
The questions cover whether work is blocked, whether to retry or escalate, and how severe the situation is. Each keeps its own answer type. The example demonstrates the schema; its selected answers should not be copied into a production policy.
Illustrative answers to the documented CLI request, not an observed deployment decision.
Pinned source for the request.
6. OpenMuse Generative UI: presentation and candidate fit
GitHub repository · Jevfast decision view
OpenMuse's Jev adapter receives the user's message along with the agent's summary, context, candidate options and allowed controls. It asks for a presentation Choice and a fit Score for each candidate.
The adapter treats presentation and candidate quality separately. It validates the response and can keep the agent's original control when confidence is insufficient.
Repository test fixture: comparison cards versus prose, with candidate fit scored from 0 to 3. This is not a live user session.
Pinned source for the request.
7. Jsort: ranking texts in pairs
GitHub repository · Jevfast decision view
Jsort puts two texts in state as A and B, then asks a Noul question: does A rank above B under the user's criterion? The returned pairwise probabilities feed a fitted ranking scale.
The comparisons produce a ranking without asking for one score per item. The criterion still needs to be clear enough to apply consistently across pairs.
Illustrative request. The YES selection shows the answer format, not an observed ranking result.
Pinned source for the request.
8. JEV Shield: reviewing command permissions
GitHub repository · Jevfast decision view
For ambiguous commands, JEV Shield supplies command context and asks for an allow/ask/block Choice. A Noul question also asks about irreversible harm. Local enforcement code interprets the answers.
That code needs a policy for conflicting or unavailable answers. Jev's review is one part of enforcement; it cannot replace an operating-system sandbox.
Illustrative ASK and NO answers. They do not certify the safety of an actual command.
Pinned source for the request.
9. fx: checking a pending action for malicious activity
GitHub repository · Jevfast decision view
fx has an optional Jev permission reviewer. It sends the active review policy, context and exact pending tool call, then asks a Choice question with clear and caution answers.
The question concerns concrete malicious activity or prompt injection. Its scope is narrower than an assessment of every destructive or operational risk, so a clear answer should not be read as approval of every possible consequence.
Illustrative clear answer to the pending-action review.
Pinned source for the request.
10. Jev Trader: a market-direction question
GitHub repository · Jevfast decision view
Jev Trader sends compact market information, including order-book and recent-trade context, with a configured block horizon. Jev selects buy or sell through a Choice question about the expected direction.
The prediction, execution and measured outcome need to stay separate. A constrained answer is easy to inspect, but the format cannot establish whether the forecast is correct.
Illustrative buy answer and source-defined request fields. The chart contains no captured market values, trading result or evidence of profitability.
Pinned source for the request.
Checking an integration
Across these projects, the app supplies evidence and candidate choices, Jev returns a typed judgment, and code applies it. The questions range from action and destination choices to relevance, retention and scores.
Before adopting one of these integrations, read its input fields and candidate set, then check answer validation and fallback behavior. Keep the distinction between a recorded response and an illustrative answer in view as you read the source.
Disclosure: Favo Yang maintains Jevfast independently of TypeSafe AI and the listed project creators. This article was prepared with AI assistance and reviewed by the author. The charts come from Jevfast's explanations of the source code; the original projects belong to their authors. No performance, security or trading benchmark was run for this article.
Top comments (0)