JEV FOR REVIEW ATTENTION
Which parts of a code change deserve a closer look? Paste a small unified diff, screen bounded risk signals, then ask Jev to point to one original hunk only when you choose to locate evidence.
Open the review workbench ↓Start with synthetic examples or edit the goal, constraints, diff and related test material. Parsing and scope checks happen locally before a request. The original added and deleted lines remain visible even when Jev finds no clear location.
Paste a conventional text diff with diff --git, ---/+++, and @@ hunk headers. Binary, renamed-path, and unparseable diffs are rejected before any request.
1 locatable hunk; both serialized request States fit the 8,000-character limit.
One explicit browser request, with four or five Boolean/Noul questions. Provider failover may cause another underlying attempt. No automatic evidence-location call follows.
Screening returns risk probabilities, not written comments. Afterward, choose one dimension to locate evidence with a separate request.
Old and new columns are repository file line numbers from the hunk header. A deleted line has no new-file number; an added line has no old-file number. Textarea line numbers are never used.
src/paginate.ts@@ -1,4 +1,4 @@ export function paginate(items: string[], page: number, size: number) {- const start = (page - 1) * size;+ const start = page * size; return items.slice(start, start + size); }Jev returns a Boolean/Noul probability for correctness, security, reliability and compatibility risk, plus test gap only when related test material is supplied. These are attention signals, not findings or written explanations.
For the dimension you select, Jev makes one Choice among the parsed hunk IDs, no_match, and insufficient_context. The workbench shows every option probability and separate provider confidence before it highlights any original text.
The threshold moves Priority review and Suggested check bands on existing probabilities. Higher thresholds flag fewer dimensions, but do not make a change safer. A low signal never proves there is no bug.
This page accepts a supplied, bounded diff. In a real repository, a reviewer still needs the base and head revisions, full callers and contracts, compiler or linter diagnostics, relevant tests actually run, provenance of the patch, and someone to inspect the highlighted lines. An unprovided test excerpt is unknown evidence, not proof the project lacks tests.
Dev Agrawal's Jev Review discovers local changes and runs a longer staged workflow: screening, file profiles, evidence selection, mechanism classification, severity, and reviewer routing in a local dashboard. jev-mcp exposes agent-callable review and completion-gate tools. This browser workbench does not read your repository, run tests, verify completion claims, or install either integration.
The parser accepts conventional JavaScript/TypeScript text diffs with up to six valid hunks. It rejects binary or unsupported formats and any input whose full serialized State exceeds 8,000 characters; it never silently cuts away part of a patch. It does not fetch GitHub, scan a repository, execute submitted code, write comments, approve a PR, merge, or repair a defect. A chosen hunk is a review prompt, not proof of a vulnerability or an exhaustive search.
Your goal, constraints, diff and supplied tests go through this site's server to its configured TypeSafe Direct or Vercel Gateway provider. Omit secrets and unrelated proprietary code. Data handling depends on the configured provider path; do not assume that code remains in the browser or that every request has ZDR. See the access and data-handling guide and API guide.
TypeSafe documents Noul probabilities, Choice distributions, and confidence as distinct outputs. The review threshold and the location-confidence cutoff are this site's application policy, not TypeSafe's recommended code-review settings. This new scenario has not received real-model semantic validation here.