v0.4.2

Official MCP Registry metadata for the local stdio adapter. The published npm package declares io.github.apixly-ai/jev-filter, matching the active Registry record.

• Install and configure the server using the README. Inference requires your private TypeSafe API key.
• Native npm bundles cover macOS arm64/x64 and Linux arm64/x64; the Python wheel and source archive are also attached.
• All five public npm packages match these release tarballs. A fresh macOS arm64 installation passed doctor and real MCP initialize/tools/list checks without API keys or inference. The release CI also passed.

What's Changed

• Give Jev speed a visible, verifiable place on the homepage by @JIA-ss in #26
• Make Jev Filter documentation discoverable and useful for AI search by @JIA-ss in #27
• Fix live README discovery metadata and verify the documentation site by @JIA-ss in #29
• Connect maintained skills to bounded Jev workflows by @JIA-ss in #28
• Fix the agent quickstart search description by @JIA-ss in #30
• Add a standard text sitemap fallback by @JIA-ss in #31
• chore(deps): bump the actions group across 1 directory with 5 updates by @dependabot[bot] in #21
• docs: add M8ven Trust Index badge by @JIA-ss in #33
• docs: simplify repository homepage layout by @JIA-ss in #34
• chore: prepare MCP metadata and manual Registry publishing by @JIA-ss in #35

New Contributors

• @dependabot[bot] made their first contribution in #21

Full Changelog: v0.4.1...v0.4.2

v0.4.1 — Fast selection, traceable evidence

Jev Filter gives your agent the relevant evidence while keeping every original recoverable. This release pairs paid live measurements with a browser fix grounded in observed scroll-container state.

Measured advantages

• 91.6–97.2% less returned tool context. All 24 expected ID sets matched in 24 real agent runs across three synthetic scenarios and two primary models; filtered completion was 12/12, raw completion 11/12.
• 2.58× Jev batch speed. Automatic batch-parallel filtering handled 96 records and two predicates in 1.246 s, versus 3.215 s for single-record parallel calls. Input tokens fell 56.5%; all 12 runs across four arms were exact. This is a Jev-stage comparison.
• More relevant code found. Opt-in recall plus native filtering recovered 10/12 labeled matches versus 4/12 with exact candidates, with zero false positives in this fixture. It returns more evidence and still misses cases without lexical overlap.
• 13/13 paid interface checks passed. Python CLI, freshly installed npm 0.4.0 JavaScript/native, official MCP, local page extraction and typed surveys use the same evidence contract. These earlier interface measurements retain their original package version.

Same-output screening response

Additional paid validation of the released 0.4.1 runtime returned the exact same 48 IDs from 96 synthetic records in all nine runs. Median execution-path time was 1.313 s for Jev API, 13.421 s for Luna through signed-in Codex CLI, and 14.543 s for Astra through the same CLI. CLI timing includes startup and the primary-agent turn; native inference latency is not isolated, and Jev has no primary continuation in this test. Eight simple bilingual templates repeated twelve times and three repeats per arm make this a screening smoke sample.

Per-run speed evidence and the incremental usage receipt preserve exact outputs, timing, context bytes and actual token counts. Jev used more input tokens than either primary path here; this result establishes neither input-token savings nor a whole-agent speedup.

Grounded browser progress

Observed scroll position, remaining distance, container identity and clipping now reach the planner and independent verifier. The pinned source A/B improved nested-scroll completion from 0/3 to 3/3 while retaining the strict 0.55 probability / 0.10 margin policy. Six ordinary goal runs still require review; completing more steps adds calls and time. Freshness, origin and irreversible-action gates remain in place.

A separate benchmark-only primary-model recovery reference completed 6/6 synthetic goals, including detecting one premature Jev DONE. This reference is not an automatic fallback shipped in the runtime.

Evidence and limits

Methods, source data and reproducible commands include all negative experiments and unknown usage. Five of six whole-agent cells were slower by 0.1–18.4%; cold API-equivalent cost ranged from 23.2% lower to 1.1% higher. Subscription billing is unknown. Small synthetic fixtures do not establish universal speed, accuracy or invoice savings.

The bilingual homepage now leads with measured context and batch speed, supported by charts and inspectable evidence.

Install and verify

npm install -g @apixly/jev-filter@0.4.1
jev-filter doctor

Published files include Python distributions and macOS/Linux native npm bundles. All eight checksum-listed artifacts have verified GitHub build provenance; all five npm registry integrities match the Release files. A fresh macOS arm64 registry install passed 11 offline CLI/JavaScript/official MCP checks. Other native platforms passed CI; their package bytes and provenance were verified separately.

Publication verification receipt · Merged change · Full changelog

Jev Filter 0.4 — Give your agent the signal

Give your agent the signal

Jev Filter 0.4 brings clearer decisions, richer browser evidence and direct JavaScript/TypeScript and MCP integration. Large tool outputs become typed decisions with source-level proof; your agent can focus on the selected evidence and the cases that need reasoning.

A stronger decision core

• Review ambiguity before action. Per-question confidence, probability and margin policies preserve the evidence and explain what needs another look.
• Keep accounting honest. Hosted errors retain sanitized provider codes, retry attempts and unknown usage, including verifier and text-helper failures.
• Verify the current state. Unique checked/pressed/selected/value checks, richer observed controls and guarded nested scrolling support independent outcome checks. Stale observations cannot prove completion.
• Inspect behavior and change. Bounded opt-in code recall and complete tracked Git before/after evidence retain source identities, modes and freshness.
• Connect the same core. Typed JavaScript client, Python API, native CLI and four read-only MCP tools preserve partial packets and receipts.
• Evaluate your policy. Offline grouped holdout diagnostics and optional official System One Adapter benchmarks make tradeoffs inspectable.

Evidence you can inspect

The release passed the protected PR checks across Python 3.10–3.14, Linux/macOS/Windows Node, four native targets, real local browser fixtures and platform desktop checks. Local validation: 302 Python tests, 85.31% coverage, 11 Node tests, actual CLI/JS/MCP E2E and installed native tarballs with no system Python/rg.

Fresh fixed-source offline program A/B met 13/13 expected decision/accounting cases versus 3/13 on the baseline; false fixture actions fell from 3 to 0. Browser contract cases went from 3/15 to 15/15. Extra retrieval improves fixture recall while reducing precision and increasing context/time; safer decisions increase review. These tests make zero semantic-model calls and do not establish new model accuracy, open-web success or invoice savings.

Inputs, raw results and reproducible commands · Integration guide · 中文文档

Install

npm install -g @apixly/jev-filter@0.4.0
jev-filter doctor

Registry publication follows the signed GitHub release workflow; verify the separate Publish npm run before assuming an upload is installed. Python users can install the attached wheel or the v0.4.0 Git source. Native npm packages support macOS/Linux; Windows uses WSL or native Python.

Filtering uncertainty is opt-in. Hosted defaults require probability 0.55 and top-two margin 0.10 and return needs_review; zero thresholds reproduce legacy behavior. MCP supports connection-scoped 2024-11-05 through 2025-11-25 revisions. New-tab following and per-request in-flight MCP cancellation remain outside this release.

Changelog · Merged implementation · Compare v0.3.0…v0.4.0

v0.3.0

What's Changed

• feat: hosted execution for browser, desktop and survey (0.3.0) by @JIA-ss in #22
• docs: README usage, recorded-run replay and real-world audit (0.3.0) by @JIA-ss in #23

Full Changelog: v0.2.4...v0.3.0

v0.2.4

What's Changed

• fix: wait up to ~15 minutes for npm to process each published package by @JIA-ss in #19
• fix: read stdin and write output as UTF-8 regardless of console codepage (0.2.4) by @JIA-ss in #20

Full Changelog: v0.2.3...v0.2.4

v0.2.3

What's Changed

• fix: run the Python package natively on Windows by @JIA-ss in #17
• chore: release 0.2.3 and correct the native Windows install instructions by @JIA-ss in #18

Full Changelog: v0.2.2...v0.2.3

v0.2.2

What's Changed

• feat: auto-open dashboard and tolerate npm processing delays (v0.2.2) by @JIA-ss in #16

Full Changelog: v0.2.1...v0.2.2

v0.2.1

What's Changed

• fix: automate verified npm publication after GitHub releases by @JIA-ss in #12
• fix: use absolute tarball paths for npm publication by @JIA-ss in #13
• ci: remove duplicate checks and complete dependency auditing by @JIA-ss in #14
• fix: dashboard port collision recovery (v0.2.1) by @JIA-ss in #15

Full Changelog: v0.2.0...v0.2.1

v0.2.0 — Local token and USD dashboard

Token and USD statistics, with a local dashboard

Jev Filter now has an opt-in numeric-only usage ledger and a read-only localhost dashboard. Inspect estimated input-token reduction, input dollar value, Jev cost and net value; filter by model, UTC date and counting method; export filtered JSON or self-contained HTML.

• Local byte estimates and tiktoken counts; explicitly enabled OpenAI/Anthropic counting adapters.
• Per-call pricing snapshots, unknown usage and negative values remain visible.
• No source text or credentials in the ledger; no change to semantic selection, authorization or result caching.
• Counts compare normalized candidate JSON with the full returned packet once. They are not whole-agent or subscription invoice savings. Typed batches record costs without claiming an unsupported reduction.

Published on npm: npm install -g @apixly/jev-filter@0.2.0. All five package versions have registry integrity matching these release tarballs. GitHub Trusted Publisher bindings are configured for the maintained release workflow. Configure recording with jev-filter stats configure, then open jev-filter stats dashboard.

English setup and measurement guide · 中文配置与统计口径

Verification

102 local tests passed, 88.56% coverage; desktop/mobile dashboard filtering, export and reconciliation; Python 3.10–3.14 CI; native tarball installation on macOS/Linux x64/ARM64 without system Python/ripgrep; security, documentation and public-content checks. Counter APIs were tested with mocked providers; no new billable inference benchmark was run for this release.

Assets include Python wheel/sdist, main npm package, four native platform packages, dependency SBOM and SHA256 checksums. GitHub build provenance is attached to the release assets.

Full changelog: v0.1.0...v0.2.0

v0.1.0

What's Changed

• Release Jev Filter v0.1.0 with native npm packages and reproducible evidence by @JIA-ss in #8

New Contributors

• @JIA-ss made their first contribution in #8

Full Changelog: https://github.com/apixly-ai/jev-filter/commits/v0.1.0