jev-browser 0.1.2

jev-browser 0.1.2

Complex-case hardening and full-quality demo recordings.

Controller and perception

• Goal-aware candidate ordering: elements whose name or href mention goal/input keywords are listed first, so long pages (Wikipedia: 500+ links) no longer truncate the relevant link away. Keywords that match most elements are ignored.
• Dropdowns report their current option, so a value already set is not selected again.
• "Probably done" is not done: a model stop at 70–85 % goal probability first spends one more step on an untried action; only ≥ 85 % (or no alternatives) ends the run.
• The observation hash includes the viewport: a scroll that reveals something counts as a change; scrolling at the end of a page does not.
• A page change counts as progress even when the backend reports a verification error (e.g. Enter navigated before the fill could be verified).
• go_back never strands a run on about:blank (Chrome history and ego alike).
• ego: clicks intercepted by an image/overlay fall back to a forced click, then a DOM click.

Demos (recorded in ego lite, attached below as 1080p MP4)

• Wikipedia multi-hop: search, pick the result, follow links across three articles — 6 steps, $0.0029, 21 s.
• Sign in with a secret password, then a contact form with a dropdown — 10 steps, $0.0013, 18 s.
• Three identical "Start free trial" buttons, the goal names the Team plan — 3 steps, $0.0003.
• GitHub repository navigation to docs/ATOMIC_PLANNING.md — 3 steps.

scripts/make-demo.mjs composes these from any run made with --step-screenshots <dir>.

Verified

Unit + e2e on macOS with live Jev (ego lite and headless Chrome): 55 passed, 0 failed, 5 skipped (Safari opt-in). CI on Ubuntu: unit + mock e2e.

jev-browser 0.1.1

jev-browser 0.1.1

Demo recordings and better run introspection.

• run --step-screenshots <dir> (MCP: step_screenshots_dir): save the page exactly as Jev saw it before every step, plus final.png.
• Journals record targets: id → element description for the top click / type / select candidates, so a journal reads without the page.
• scripts/make-demo-gif.mjs: compose a side-by-side GIF/MP4 from any run (left: the page, right: Jev's goal_done / blocker / action / target judgment and the executed action).
• README demos recorded in ego lite: navigating the NanoJev GitHub repository to docs/ATOMIC_PLANNING.md; choosing the Team plan among three identical "Start free trial" buttons (probability 1.00).
• CI: unit + mock e2e on Ubuntu with headless Chrome. The live suite is meant to run locally; no API key is needed or stored in the cloud.
• Docs: full CLI reference, "what leaves your machine" data-flow section, Windows --copy note, all environment variables.

Install / upgrade: git pull in your clone (the installed skills are symlinks), or download jev-browser-skill-0.1.1.zip for a manual copy.

jev-browser 0.1.0

jev-browser 0.1.0

First release of jev-browser: browser use / computer use for coding agents where
TypeSafe Jev supplies calibrated per-step judgments and a code controller
owns memory, budgets and termination (the split introduced by
NanoJev).

Highlights

• Backends: ego lite (default; reuses your logged-in sessions, hands the tab to you on blockers, resumes with --space-id), Chrome via CDP (launch or attach, headless), Safari via safaridriver.
• Interfaces: CLI (run, observe, judge, pick, doctor, config, install, mcp) and a dependency-free MCP stdio server (jev_browse, jev_observe, jev_judge, jev_pick, jev_doctor, jev_config).
• Agents: one command installs into Claude Code, Codex, Cursor, the skills.sh directory and Claude Desktop (MCP), with backups, --dry-run and --uninstall.
• Safety: typed values are selected from inputs, never generated; secrets never reach the model or the journal; step / cost / time budgets; local JSONL journals.
• Zero dependencies: Node 22+ and TYPESAFE_API_KEY.

Verified

• Unit suite + e2e scenarios per backend against a fixture site (navigation, search + typing, login with a secret, choosing among identical buttons, cart, form + dropdown, scroll-to-reveal, blocker hand-off, limits), CLI round-trip, ego hand-off → resume.
• macOS with live Jev: 54 passed / 0 failed / 5 skipped (Safari opt-in), three consecutive identical runs; ~$0.02 per full suite.
• GitHub Actions on ubuntu-latest with the mock judge and headless Chrome.
• Real sites via ego lite: TypeSafe docs → Choice page (2 steps, $0.0005); GitHub → docs/ATOMIC_PLANNING.md (4 steps, $0.0016).

Install

git clone https://github.com/ChenYCL/jev-browser-skill.git && cd jev-browser-skill
export TYPESAFE_API_KEY=...
node skills/jev-browser/bin/jev-browser.mjs doctor
node skills/jev-browser/bin/jev-browser.mjs install

Or: claude plugin marketplace add ChenYCL/jev-browser-skill && claude plugin install jev-browser@jev-browser-skill,
or npx skills add ChenYCL/jev-browser-skill --skill jev-browser.
The attached jev-browser-skill-0.1.0.zip is the bare skill directory for manual installs.

Known limitations

Same-origin iframes and new tabs are not followed; canvas / drag-and-drop / uploads are not handled; Safari is implemented to the WebDriver spec but not yet exercised live; goals should be in English for best accuracy.

Full changelog: CHANGELOG.md