Everything, newest first.

Evaluations Workbench: score repos 1–5 against a custom weighted rubric (e), a weighted-average badge on every card, sort by eval score, an evaluated-only filter, and a one-command auto-decide from Vee’s fit suggestion.

Fit delta tracking: re-scan any repo and a ↑ green / ↓ amber badge shows exactly how the verdict moved. Triage % pill, hover preview with decision and note, r to re-scan the focused card, and three smart palette filters.

Keyboard-first decisions: a d popover with Vee’s suggestion, bulk assign in select mode, triage progress chips, search-term highlighting, decision dates, a “recently decided” sort, and a compact density toggle.

Event-delegated library grid (O(1) wiring, zero listener leaks), a hover preview flyout with ELI5 and strengths, inline clickable capability tags, and an extended j/k/n/c/p/Esc keyboard suite.

AI head-to-head compare for library cards, per-repo personal notes, full-text search across ELI5 summaries, one-click Markdown export, and a natural-language filter — type a question with ? and get a ranked list back.

Persistent Ask history across sessions, multi-turn conversations with full prior context, library pinning, URL-paste analysis without navigating, and a fit-distribution bar in the stats.

A Tech Radar that organises your library by Adopt / Trial / Hold / Reject, decision badges in the header and verdict, five new keyboard shortcuts, and Markdown radar export.

Every tab deep-linkable via URL hash, a clickable repo name, f for a fresh scan from anywhere, star counts on cards, batch-scan notifications, per-repo tab memory, and a decision filter.

Workflow friction removed: a board popover from the analysis, a diff callout on the verdict, a paste-URL error screen, Copy-for-Slack, a fit emoji in the tab title, and a refresh-stale batch shortcut.

Ask This Repo — grounded Q&A using the cached analysis with five turns of history, no extra AI call for context, and an a shortcut that jumps straight to it.

A full evaluation workbench around each scan.

• Docs Quality — an A–F grade across six README/docs dimensions.
• Decision Log — Adopt / Trial / Hold / Reject with a note, indexed in IndexedDB.
• Compare mode — a side-by-side capability diff for any two library repos.
• Scaffold Export — a CLAUDE.md-ready integration template as a .md download.

Collections, the lens mascot, and a tactile motion pass — all behind reduced-motion.

• Boards — group the repos you’re weighing; membership travels in backup.
• Vee — one theme-aware SVG lens: wide-open on a strong fit, narrowed on risky.
• Actionable errors — a bad key links to Settings; a transient failure offers Retry.

Removed the Claude subscription sign-in (Anthropic prohibits third-party use of those tokens) and shipped the clean Console-API-key path. Grok, OpenRouter, and ChatGPT sign-ins are unaffected.

Connect OpenAI without pasting a key — the same OAuth login the Codex CLI uses — completing the three big coding-CLI logins: Claude, Grok, and ChatGPT.

A dozen-plus providers through one data-driven registry — OpenAI, DeepSeek, Groq, NVIDIA NIM, Kimi, Zhipu GLM, Qwen, MiniMax, Azure, a custom endpoint, and local Ollama with no key at all.

A Select toggle reveals checkboxes across the library — pick any number and remove in one confirmed action; Esc backs out.

Five new themes — Nord, Gruvbox, Rosé Pine, Catppuccin Latte, Solarized Light — for thirteen in all, with a proper fix for light mode and status colours kept readable via color-mix.

The foundation: portable export/import of your whole library, hardening against README prompt-injection, BM25 search, self-healing retries, and full keyboard and reduced-motion accessibility.