Jev Browser v0.5.1 — compatible System One endpoints

Compatible local / self-hosted System One endpoints

v0.5.1 keeps the existing TypeSafe/Jev wire contract and allows a custom baseURL / JEV_BASE_URL to run without a hosted Jev API key.

• Hosted default behavior is unchanged: JEV_API_KEY / TYPESAFE_API_KEY is still required.
• A custom endpoint may be keyless and must implement POST /v1/systemone with the existing state/questions -> model/answers/usage envelope.
• Browser candidate validation, confidence validation, policy hooks, cancellation, and mutation rules are unchanged.
• No provider-specific Python/model runtime or router is added to Jev Browser.

Laya experiment

Laya 0.2.1 was connected through a local System One shim without any Jev Browser wire-format changes, confirming protocol compatibility.

The current public Laya checkpoints are not recommended as the default general browser backend based on these controlled known fixtures: English Laya passed 4/9 semantic regression tests and 2/5 compact English browser tasks; multilingual passed 2/9 semantic tests and 2/5 compact Japanese browser tasks. This is a small compatibility experiment, not a general model benchmark.

On this host with CPU inference after model load, English Laya median forward latency was about 119.7 ms for 1 question, 373.3 ms for 10 questions, and 1707.7 ms for 50 questions. This supports the decision-frontier idea but does not imply equivalent browser accuracy.

Executed verification

• Tested head: 031ec77c2473744da5197129f15afc858ac39cfc
• Release commit: 2ca31b23d463c3bad43f28130344b02867a73e91 (identical Git tree 8262e9bd5b697352e3796784546449b612dae9e6)
• Local regression suite: 273/273 passed
• Fresh installed-tarball SDK / persistent CLI / MCP checks: passed, including semantic paths
• Runtime dependency audit: zero reported vulnerabilities
• Exact-head PR CI: 6/6 — https://github.com/tontoko/jev-browser/actions/runs/35445926642
• Post-merge main CI: 6/6 — https://github.com/tontoko/jev-browser/actions/runs/35473837353
• Tarball SHA-256: a76095d916626ddf0296ad9fc3b2385c0edf87eeb4bf2f7b7de58414c0c6f8eb

Attached verification files record the release evidence. Backend confidence/accuracy is distribution-specific; callers should calibrate non-Jev endpoints independently.

Install

npm install --save-dev https://github.com/tontoko/jev-browser/releases/download/v0.5.1/tontoko-jev-browser-0.5.1.tgz
npx playwright install chromium

Hosted Jev uses its API key. A compatible local endpoint can instead be selected with JEV_BASE_URL=http://127.0.0.1:8765.

Jev Browser v0.5.0 — grounded semantic verification

Grounded semantic verification

v0.5.0 adds an explicit semantic verification layer without weakening deterministic Playwright assertions.

• locateSemantic() binds a natural-language description to a real observed target/evidence source; it never returns model-generated selector code.
• compareSemantic(), compareSemanticBatch(), and assertSemantic() distinguish passed / failed / inconclusive and retain the grounded evidence used.
• minConfidence defaults to 0.8. minSourceConfidence defaults to the effective comparison threshold but can be tuned separately. Both are decision-score thresholds, not probabilities of correctness.
• Low source confidence can short-circuit before the comparison frontier. Low comparison confidence and insufficient evidence never pass.
• Exact grounded equality is deterministic and skips the semantic comparison request.
• Independent questions share Jev decision frontiers; transport chunks at the same dependency level do not inflate serialDecisionDepth.
• Semantic SDK, CLI, and MCP paths share one implementation and are read-only.
• Usage reports requests, questions, serial decision depth, tokens, provider time, observation time, and local verification time.

Caller verification remains authoritative

A release-gate real-Jev run found a case where the model declared a goal complete before a caller-supplied deterministic until became true. The runtime now records that rejected completion state and asks Jev for grounded progress. A repeated done without progress becomes condition-unmet; no mutation is invented or replayed. Completion that becomes true while a decision is in flight remains valid.

Executed verification

Release tree: faebb81d2fdc9b4bc1fcef00357b280c02472009

• Tested head: 9594d2919ba0dd783b0e4712fb6121c730a142a4
• Release commit: 92a318b1f4215f064e9dd663172fb08534d2abdc (identical Git tree)
• Local regression suite: 270/270
• Real Jev core + modern-control suite: 47/47
• Real Jev semantic suite: 9/9
• Premature-completion reproduction: 5/5 consecutive real-Jev passes
• Exact-head PR CI: 6/6 — https://github.com/tontoko/jev-browser/actions/runs/35412849539
• Post-merge main CI: 6/6 — https://github.com/tontoko/jev-browser/actions/runs/35413249402
• Fresh installed-tarball SDK / persistent CLI / MCP checks passed, including semantic paths
• Runtime dependency audit reported zero vulnerabilities at release preparation
• Tarball SHA-256: f87bbeb98f033b18d08174a3b4f384652b326e59e9db937de7e866e8c724e99f

Attached VERIFICATION.md, machine-readable verification.json, raw real-provider TAP logs, and SHA256SUMS contain the evidence.

The semantic fixtures are controlled regression/calibration evidence, not an unseen-site accuracy benchmark. Jev confidence is not a calibrated probability. Deterministic Playwright/native assertions remain the stronger oracle whenever exact application truth is available. Multi-commit checkpoint/resume is not part of v0.5.0.

Install

npm install --save-dev https://github.com/tontoko/jev-browser/releases/download/v0.5.0/tontoko-jev-browser-0.5.0.tgz
npx playwright install chromium

Native operations require no AI key. Jev-backed operations use JEV_API_KEY / TYPESAFE_API_KEY.

Jev Browser v0.4.0 — modern controls without candidate explosion

Modern controls, still grounded

v0.4.0 extends the one-instruction SDK / CLI / MCP workflow to large native select inventories, standards-associated ARIA comboboxes, and crowded pages without adding a second planner model, model-authored selectors/code, or site-specific scripts.

• Exact named select values remain local. Long prose-driven lists are resolved in bounded partitions only after that control is chosen.
• Native select state includes index + label + value semantics, so replacing an option at the same index is detected as drift.
• ARIA comboboxes use only the popup declared by aria-controls / aria-owns, require one unique visible enabled exact option, use the normal Playwright/policy lane, and verify selection after the click.
• Truncated whole pages can select one real semantic region for action and a separate result region for readback. Explicit caller scope is never widened.
• Opaque-name fixtures no longer reveal caller JSON paths to make the evaluation easier.

Executed verification

• 236/236 local regression tests on tested source 3b1fb9c4ab59eb1ddd2133bb078fa08fee01c54d.
• 26/26 real-Jev modern-control cases on that source: 24 create/save combobox tasks across four implementations × six distinct layouts, plus 1000-option and distant-option cases. All 24 create/save cases submitted exactly once.
• 6/6 cross-platform PR CI jobs on the exact tested commit: https://github.com/tontoko/jev-browser/actions/runs/35347810823
• 6/6 post-merge main CI jobs on the identical release tree: https://github.com/tontoko/jev-browser/actions/runs/35348464871
• Installed-tarball SDK, persistent CLI and MCP checks passed.
• Runtime dependency audit reported zero vulnerabilities.

The squash merge 65b6aee01f332574aca1bd6dff3b75dcc307c5bb has the identical Git tree as the tested source. Attached VERIFICATION.md, verification.json, raw live-modern.tap and SHA256SUMS contain the reproducible evidence.

These synthetic fixtures are regression evidence, not an unseen-site accuracy benchmark or a measured comparison against another browser agent. UI readback is not proof of database durability/exactly-once execution. Arbitrary widget conventions, Canvas-only reasoning, and multi-commit/resumable business workflows remain outside v0.4.

Install

npm install --save-dev https://github.com/tontoko/jev-browser/releases/download/v0.4.0/tontoko-jev-browser-0.4.0.tgz
npx playwright install chromium

Native operations require no AI key; Jev-backed operations use JEV_API_KEY / TYPESAFE_API_KEY.

Jev Browser v0.3.0 — one instruction to verified creation

One instruction + data, through SDK / CLI / MCP

The shared goal runtime opens a native form, batches independent input bindings, executes browser writes serially, handles onward steps and ordinary save confirmations, and checks a fresh result record. Nested student/guardian data, native selects and multiselects, booleans, confirmation fields, shadow/iframe forms and asynchronous validation are covered. Existing Playwright-native tools remain available.

Result metadata distinguishes applied inputs, saved fields read back, unobserved fields, caller assertions and uncertain effects. A model-only completion opinion is not a verified save. An unknown submission is not replayed.

Structured extraction now batches independent record frontiers without mixing each record's candidate evidence: the deterministic 20-row regression uses 2 logical requests instead of 21.

Verified release

• 206/206 local regression tests.
• 21/21 real-Jev integration cases at the final source.
• 100/100 controlled real-Jev creation tasks (20 fixture families x 5 variants), with actual saved bodies and exactly one submission independently checked. This is not an unseen-site success rate.
• Installed tarball checks for native and one-goal operations through SDK, persistent CLI and MCP.
• All six public platform/browser CI jobs: https://github.com/tontoko/jev-browser/actions/runs/35287408952
• PR #3 merged. Release commit 8a70925 has the same Git tree as the tested source 7bd9363.

See attached VERIFICATION.md, machine-readable verification.json and raw final-source TAP logs. SHA256SUMS covers the distribution and evidence files. Earlier pre-fix live runs failed and were corrected; the final 21-case count is not a claim of three consecutive final-source runs.

Install

npm install --save-dev https://github.com/tontoko/jev-browser/releases/download/v0.3.0/tontoko-jev-browser-0.3.0.tgz
npx playwright install chromium
npx jev-browser --help

Set JEV_API_KEY in your secret environment for natural-language decisions. Native operations do not require an AI key. The package is distributed here; this release is not an npm registry publication.

See docs/goal-runtime.md for contracts and limitations. UI readback is not a database-durability/exactly-once guarantee. Object-array creation, arbitrary multi-commit business workflows and autonomous Canvas reasoning are outside the claimed automatic path.

Jev Browser v0.2.0

Jev Browser 0.2.0

A shared Playwright/Jev core with a persistent CLI, an MCP server, and a typed SDK. Native operations and deterministic assertions need no model key; grounded natural-language operations use the official Jev API.

Included

• 40 shared MCP/CLI commands: navigation, snapshot refs, form controls, tabs/frames/dialogs, uploads/downloads, screenshots/PDF, traces, storage/cookies, diagnostics and assertions.
• Authenticated named CLI sessions across independent processes, plus JSONL sessions and one-shot commands.
• SDK act/observe/extract/run and agent().execute(), existing Playwright Page ownership and native assertions.
• Nested objects and arrays of DOM records, source evidence and per-row field isolation; explicit input bindings and verbatim quoted text.
• Chromium, Firefox and WebKit launch options; persistent profiles, CDP and Playwright WebSocket connections.
• English/Japanese setup, migration/security/API documentation and a CLI agent skill.

Verification on this exact release commit

Commit: 53d02dfdbbbfa21a42adb3d2fc310e4fe8375cab. Hosted CI passed all six jobs, including installed-package SDK/CLI/MCP checks in every job.

Environment	Passed	Failed	Platform-specific skips
macos-latest / Node 24 / chromium	133	0	0
ubuntu-24.04 / Node 22 / chromium	133	0	0
ubuntu-24.04 / Node 24 / chromium	133	0	0
ubuntu-24.04 / Node 24 / firefox	133	0	0
ubuntu-24.04 / Node 24 / webkit	133	0	0
windows-latest / Node 24 / chromium	131	0	2

The two Windows skips are POSIX symlink fixtures; the portable upload-root authorization test still runs on Windows. The explicit real-Jev suite passed 16 cases three consecutive times (48/48), with model jev-1.13.0, on synthetic pages only. Gitleaks history scan and npm dependency audit reported zero findings. These results do not establish arbitrary-site accuracy or comprehensive prompt-injection resistance. Independent Codex review was unavailable because its account had reached the usage limit; no independent approval is claimed.

Install

Download the attached tarball, verify it against SHA256SUMS, then:

npm install --save-dev ./tontoko-jev-browser-0.2.0.tgz
npx playwright install chromium
npx jev-browser --help

Node.js 22.15 or newer. The tarball includes compiled JavaScript and type declarations. No npm registry publication is implied.

Compatibility and security

This is a functional alternative to Playwright MCP/CLI and Stagehand workflows, not an identical API or command-line re-export. Use the migration guide. The standalone AI layer is DOM-based; outer clients can use screenshots and coordinate mouse tools, but autonomous Canvas/image interpretation and generated prose are not claimed. Hosted browser infrastructure and arbitrary Node-side MCP code execution are not included. The Jev hosted service is separate from this Apache-2.0 code release.

Keep session tokens, authentication state, traces and page data private. Model confidence is neither permission nor an assertion; only deterministic facts establish verified completion. Browser mutations are never automatically retried.