socai-io/jev-social

Problem The repository supports a no-clone GitHub install, but the public npm name jev-social is still unclaimed: the registry endpoint returned HTTP 404 on 2026-09-29. Publishing the verified CLI would shorten the first run to npx --yes jev-social@<version> and make the project discoverable through npm without changing the local-browser or read-only execution boundary. The package metadata is already prepared (name, version, bin, explicit files, Node 22+, and the public npm registry). The repository-side release path is now merged in #114; npm-side package ownership and trusted-publisher binding remain. Current state Repository workflow merged in #114 at 193213e. The GitHub npm environment has a one-minute wait and only the v* tag policy. Exact-main Node tests, plugin-security scan, and CodeQL succeeded. Public npm package exists and the initial package claim is complete. npm trusted-publisher binding, provenance, and clean-install verification are complete. Merging #114 did no

Problem Captured TikTok or Instagram evidence can render as a native <video controls> element in both the result card and the detail dialog. The current production renderer creates that element without an accessible name, so a screen-reader user may hear only a generic “video” control and cannot tell which captured record it belongs to. The image preview remains decorative because the adjacent card or dialog heading already names the record. This issue is specifically about naming interactive video controls. Suggested scope In public/app.js, give every rendered video preview a concise accessible name derived from the same evidence title used by its card or detail dialog. Pass that context into renderMediaPreview instead of querying nearby DOM text after rendering. Keep media-candidate selection, preview fallback behavior, controls, posters, and evidence data unchanged. Add a focused regression that exercises the production rendering helper or production source and fails if the con

Problem The complete evidence table is wrapped in .table-scroll with overflow: auto. On narrow viewports, keyboard-only users cannot move focus into that scroll container, so columns outside the visible area are difficult or impossible to reach without a pointer. Proposed change Make the existing table scroll container a named, keyboard-focusable region and give its focused state a visible high-contrast outline. Keep the table markup, headings, and visible copy unchanged. Suggested scope: public/index.html public/styles.css one focused regression under test/ Acceptance criteria The .table-scroll element can receive keyboard focus without JavaScript and exposes a region role. The region's accessible name is derived from the existing evidence-table heading and description; do not add duplicate visible text. :focus-visible produces a clearly visible outline using the existing palette and does not permanently change layout. The table retains its existing aria-labelledby relationship and

Problem Earlier prototype copy used an illustrative “about 30 minutes → about 50 seconds” comparison. The current product is materially different: Jev now selects every operation, total time depends on live browser state and platform response, and the repository deliberately avoids claiming a universal speedup. The README still needs a useful timing comparison, but it must come from a reproducible protocol that separates model latency, browser execution, media work, and evidence coverage. Proposed scope Define three fixed research tasks: one each for Instagram, TikTok, and LinkedIn-capable builds. Pin the Jev model, socai version/commit, result limit, max steps, browser profile mode, region, and warm/cold conditions. Record per-step Jev latency, socai/browser latency, total wall time, records captured, comments read, media downloaded, stop reason, and failure rate. Run enough repetitions to report median and p95 rather than a single demo take. Publish raw, privacy-safe benchmark

Goal Make Jev Social a reliable, inspectable reference implementation for typed Jev decisions driving real socai browser evidence—not a polished shell around uncertain platform state. Roadmap Phase 1 — Know what can run #18 — Detect socai version and per-platform capability before routing. #19 — Report browser/profile readiness separately from CLI installation. #14 — Document safe Chrome-session and platform-login diagnostics. Keep unsupported platforms out of Jev’s action space. This phase is complete on main. #43 consumes socai v0.6.0's observational status contract, separates browser connection and platform login state from CLI capability, and keeps the browser-facing payload strictly allowlisted. Phase 2 — Never lose observed evidence #20 — Checkpoint each streamed decision and evidence batch atomically. Restore interrupted runs without raw stdout or filesystem paths. #22 — Make the four-card visual summary prefer genuinely previewable media. Phase 3