Jev Browser

The project clearly identifies itself as unofficial, names TypeSafe System One as an external decision service, and supplies consistent author, repository, and MIT license attribution, justifying full source-attribution credit. CI has read-only contents permission and releases use OIDC instead of a long-lived npm token. Irreversible actions are routed to needs_confirmation and require explicit allow_irreversible. However, that classification is probabilistic, browser_act can perform actions directly, and arbitrary-site automation has broad effects, so least privilege, confirmation, and external-effects controls are not complete. The README explains that page descriptions go to Jev and that the calling LLM normally does not receive snapshots, but it does not fully specify transmitted fields, retention, or service privacy terms. API keys use environment variables and private benchmark data is git-ignored, yet there is no documented redaction, storage, logging, or cleanup policy for credentials, page text, or persistent profiles. Dependencies are few and CI uses npm ci, but the supplied evidence contains no lockfile, vulnerability scanning, update automation, or commit-pinned Actions. No general undo, transaction, or recovery mechanism is documented, so rollback scores zero.

The README, package metadata, and workflows agree on package identity, entry points, Node requirements, test commands, and release procedure, providing strong static self-consistency. Dependency and Chromium installation are documented and CI covers Node 20 and 22. The core decision path nevertheless depends on the external TypeSafe API, network access, and a user key, with no documented offline substitute, retry policy, or degradation behavior during service outages. Failure messaging is unusually detailed: likely_done, needs_login, needs_confirmation, error, stuck, ambiguous, and blocked are defined along with the caller's expected response, supporting full credit for failure_messages.

The material covers MCP, library, and CLI users and names concrete scenarios including forms, dynamic pages, iframes, shadow DOM, uploads, and non-English interfaces. Capability boundaries are explicit for ordered sub-goals, open-ended objectives, and comparisons across many values, with check, snapshot, and direct-action takeover routes, justifying full credit. Step-writing guidance and typed tool parameters improve trigger precision, but selection and irreversibility decisions remain probabilistic and require caller handling of low-confidence outcomes, so trigger_precision is not complete. Node requirements, Chromium setup, environment variables, persistent profiles, headed/headless behavior, and generic MCP client configuration provide thorough environment-fit evidence.

The README is well organized into quick start, source setup, tool and status references, library and CLI examples, repository layout, and release instructions. Installation commands and network/key prerequisites are concrete. Examples cover the principal interfaces, limitations include practical workarounds, and MIT metadata matches the complete LICENSE, supporting full scores for those criteria. Naming is consistent across the package, binaries, and documentation, but the project is only at version 0.1.1 and makes no compatibility or stability commitment, limiting naming_stability to 2. A version number, tag-based release workflow, and generated release notes exist, but no changelog or upgrade guidance is supplied, so versioning_changelog is thin. The author, issue tracker, and release repository are identified, but there is no maintenance policy, support horizon, response ownership, or co-maintainer information. Unknown publisher identity is not itself penalized; the deduction concerns incomplete maintenance responsibility.

Tool results expose actionable statuses, URLs, action histories, scores, candidates, and page text, and the documentation explains how callers should continue, supporting full output-usability credit. The split between LLM planning and compact Jev decisions offers clear marginal value over repeatedly sending full page snapshots. However, accuracy, latency, irreversible-action, and token-saving claims are repository-authored benchmark statements, with some expressly labeled estimates, so marginal_value remains at 2. The README gives approximate call latency, normal call counts, and end-to-end timing, but omits TypeSafe API pricing, quotas, failure overhead, and ongoing operating cost, leaving the cost-benefit case incomplete.

The README points results to RESULTS.md and identifies benchmark commands, task organization, and ground-truth checks, giving major claims a traceable path. The supplied evidence does not include RESULTS.md, test sources, a lockfile, or raw benchmark artifacts, so the 40/42 result, zero false-done claim, and token totals cannot be checked item by item; claim_traceability therefore stops at 2. README, package.json, CI, and release configuration corroborate installation, entry-point, testing, and publishing claims, while effectiveness and guard-performance claims remain largely single-source, limiting cross-source corroboration. The documentation explicitly labels the context comparison as an estimate and distinguishes observed results, limitations, and caller responsibilities without claiming independent validation, supporting full fact-inference separation.