User repository on GitHub
Pi extension suite powered by Jev: selective context compaction and model routing
- Forks
- 2
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Pi Jev — iefnaf
Two pi extensions driven by Jev judgments: one replaces summary-based context compaction with keep, drop and shorten calls, the other scores each prompt for difficulty and switches pi's model before the turn runs.
- Stars
- 10
- Language
- TypeScript
- License
- MIT
- Published by
- Community
- Author
- iefnaf
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Pi Jev — y0usaf.
About this project
Selective context compaction and per-turn model routing for pi, powered by Jev. pi-jev uses Jev's typed judgments to decide which tool outputs still matter and how demanding a user request is. It retains useful text instead of generating a new compaction summary, and can switch models before each turn. Both features are optional; if Jev is unavailable, pi continues with its normal compaction or the current model.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- Compaction asks Jev per tool call whether to keep it, whether to keep its full result, and how stale it is, then renders retained content verbatim.
- Routing scores only the current user prompt on a three-level difficulty rubric, and switches model only at confidence 0.6 or above with a target configured.
- Every failure path falls back to pi's own behaviour: Jev errors, aborted requests, empty spans or reductions below minReduction.
- Settings resolve environment, then .pi/jev.json, then the global agent file, then defaults; API keys are environment-only and never written to config.
Watch out
Both features are optional and pi keeps its normal behaviour whenever Jev is unavailable, which is the right default but means you should verify the hook actually fired. Requirements are pi installed, an API key for one Jev transport, and routing targets configured and authenticated inside pi; routing stays inactive until at least one target is set. Verbatim retention applies to retained text only: images become [image] placeholders and assistant thinking blocks are omitted during conversion. A middle-band prompt keeps the current model rather than resetting to a default, including a model an earlier turn selected. If both keys are present TypeSafe wins unless JEVC_PROVIDER says otherwise, and settings writes are global by default, including disabled, unless you pass -l for project scope. The recorded 32-message to 16-message compaction example is explicitly illustrative, not a latency or reduction guarantee. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
One recorded run converted a span from 32 messages and 39,379 characters to 16 messages and 9,553 characters in 757 ms, using a single Jev request and 652 output tokens, and the README calls that illustrative rather than a guarantee. Requests are batched to fit your configured state and request budgets, so a compaction can consume several Jev requests. Unlike a summarizer, it keeps the retained tool text rather than rewriting it, which is the real behavioural difference.
The default is conservative: missing or malformed per-call answers retain the affected call. Long results you keep can be shortened to the first truncateHeadChars with an explicit truncation marker, while a confident staleness judgment can rescue a borderline result. If Jev fails, the request aborts, the span is empty, or the estimated reduction is below minReduction, pi generates its normal summary instead.
The thresholds are deliberately narrow. The cheap target needs difficulty at or below 0.5 and confidence at or above 0.6 with routing.cheap set; the strong target needs difficulty at or above 1.5. Anything in the middle band, at low confidence, unanswered, or with no target configured keeps the current model, and that includes a model an earlier turn selected rather than resetting to a default. Invalid model references and missing provider auth also keep the current model.
You need a key for one transport, TypeSafe or OpenRouter, auto-detected from the environment with TypeSafe winning when both are present unless JEVC_PROVIDER=openrouter. To bypass both features, run /jev set disabled true or launch pi with JEVC_DISABLED=1. Configuration changes are read on the next hook, so you do not need to restart the session, and API keys are never written to config files.
pi install npm:@alexlikevibe/pi-jev, or install from the GitHub URL or a local checkout, or load a single extension with pi -e for development. Use /reload if you install into a running session. To bypass both features for one process, launch with JEVC_DISABLED=1, or run /jev set disabled true and set it back with /jev set disabled false. Settings writes are global by default, including disabled; append -l for project scope.
Highest priority first: environment variables, then the project file .pi/jev.json, then the global ~/.pi/agent/jev.json, then built-in defaults. API keys are environment-only and never written into configuration files. Hooks reload configuration on every event, so changes apply to the next turn or compaction without restarting pi, and environment overrides keep winning over settings changed through the menu.
No. Verbatim retention applies to retained text, not to every original message field. Images become [image] placeholders and assistant thinking blocks are omitted during conversion, while tool inputs are serialized into the transcript. There is also one boundary exception: when pi cuts inside a turn and the span ends on a tool result, that call is never dropped outright, so its metadata and a bounded head of the result are kept.
More Jev resources like this
13GitHub stars
Pi extension that classifies each task with Jev over OpenRouter and routes it to a Pareto-optimal model balancing measured quality, cost, and latency under a role policy.
Jev Agents & MCP#community#jevbest#catalog-import
10GitHub stars
Node router that sends each query to the cheapest capable LLM tier by asking Jev to classify demand against your tier descriptions, with no LLM call spent on routing.
Jev Agents & MCP#community#jevbest#catalog-import
Community
rajdhakad9826/jev-routerTypeScript
rajdhakad9826/jev-routerTypeScript5GitHub stars
A Claude Code function hook that sends AskUserQuestion dialogs to Jev and draws a probability per option above the dialog, so you still choose while seeing how confident the model is.
Jev Agents & MCP#community#jevbest#catalog-import
Community
BeLazy167/typesafe-modTypeScript
BeLazy167/typesafe-modTypeScript3GitHub stars
Model router that asks Jev eleven questions in one call and lets a YAML policy turn the answers into a tier. Every request starts on the fast tier and only escalates on evidence such as multi-step reasoning, hard difficulty, high stakes or safety, and every decision is logged without the message text.
Jev Agents & MCP#community#jevbest#catalog-import
Community
iamvatsalpatel/tiershiftTypeScript
iamvatsalpatel/tiershiftTypeScript2GitHub stars
Model router that never names a model: Jev classifies the task family, then a cost-per-task Pareto walk picks the cheapest frontier point clearing its quality floor.
Jev Agents & MCP#community#jevbest#catalog-import
Community
tylerjharden/ailerixTypeScript
tylerjharden/ailerixTypeScript6GitHub stars
A per-prompt capability router for coding agents: a local BM25 shortlist of installed skills, MCP servers, subagents and commands, then one Jev request injects at most one suggestion plus two runner-ups.
Jev Agents & MCP#community#jevbest#catalog-import
Community
bestagentkits/jev-skillfulTypeScript
bestagentkits/jev-skillfulTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.