[timrmcclain]

Summary

Fixes

• Quota colors reflect real availability. A window with healthy remaining % is no longer shown green when the account is blocked by a different, exhausted window (e.g. 5-hour at 100% while the 7-day cap is at 0%).
• Stuck "Querying quota..." fix. pruneQuotaCache compared only key counts, so resetting an orphaned loading entry to idle was silently dropped whenever no accounts were added or removed.
• Kimi quota. A brand-new, zero-usage account legitimately returns {}; that is now treated as "no data yet" instead of an "unrecognized format" error. Payloads with real but wrong-shaped content still error. The empty reply also records its check time, so the card no longer says "not checked yet" next to "no usage limits reported".
• Proxy survives the GUI. The core process is no longer tied to the GUI's lifetime (Windows kill-on-close job object and Linux PR_SET_PDEATHSIG removed; normal exit records state instead of killing the core). Closing, crashing, or updating the GUI no longer drops active proxy connections. The existing adopt_existing_core_processes path reattaches on next launch; explicit Stop/Restart still stop the core.
• desktop_mapping_error row fix. Error matching now skips rows that has_mapping() excludes, so a config error is no longer attributed to an inert row that shares a source model.
• App Version row. getVersion() was always denied without core:app:allow-version, so the row stayed on "Detecting" unless the update check supplied a value.

Features

• Claude Desktop model picker. "Pick from catalog" adds models from already-discovered accounts in bulk instead of one typed row at a time.
• "Just reset" indicator. A quota window whose remaining % climbs between two successful fetches is flagged as just reset (Accounts panel, dashboard cards and Quota Lookup).
• Proxy page. Proxy controls and connection details move from Overview to their own page; Overview keeps a one-line status that links there. The page stays reachable while the proxy is stopped so it can be started from it.
• Easier plugins.
  • Advanced settings → Extensions shows a short guide with an "Open Plugins" button; the raw native-plugin form sits behind "Show advanced settings".
  • The plugin store has a "Recommended for you" section matched to the connected account types, with one plain sentence per plugin and a third-party / proxy-restart warning. Only entries from the official store listing are recommended.
• AI plugin finder. In the plugin store, the user describes a goal in a sentence. The GUI sends that plus a trimmed public catalog as one chat request through the local proxy (new Tauri command ask_plugin_finder), using the GUI's own client key on the Rust side so it never reaches the webview. A small, fast model is picked automatically.
  • The reply is treated as untrusted: only ids present in the loaded store are shown (official listing preferred, at most three). Each match shows why it fits, what changes, and what to watch out for.
  • Installing still goes through the normal install dialog.

Merge with current main

Upstream main is merged in (19 conflicted files). Where both sides built the same surface:

• Kept both where they coexist (Home overview cards alongside the account dashboard; upstream's AgentClientList).
• Upstream wins where its refactor supersedes the branch's inline version (usage token composition; events view split into UsageEventsView).
• Fork-only dashboard components now use the renamed /credentials endpoints (they still called /auth-files, so the dashboard showed no accounts after the merge).
• Quota Lookup row markup was replaced by the merge; the blocked and just-reset row states were restored on top of upstream's version.
• xAI quota checks keep the branch's behavior of not sending a paid inference request as a fallback when billing returns empty.

Note for reviewers

This branch comes from a personal fork and also carries fork-specific customization alongside the fixes above:

• branding
• a custom account dashboard
• personal docs and preview images
• a personal build version string, 0.3.12+personal-desk.26

Happy to split the fixes into a smaller PR if that's preferred. The plugin recommendations are a hand-curated list of store ids; that list may be better kept fork-only.

Test plan

• tsc --noEmit: clean
• cargo test: 764 passed, 0 failed, 8 ignored
• bun test: 780 passed, 4 failed. The same 4 also fail on unmodified upstream main (authFileUsageSummary, operationFeedback, textContrast, uiLocalization). In one full run a fifth, timing-sensitive quota test (keeps both secondary pages consistent with Overview for resetDue) failed; it passed 3/3 when run alone.
• Built and installed locally: GUI swapped while the proxy kept the same PID, port 8317 stayed up, and live client connections continued through the swap.
• Plugin guide, recommendations and AI finder checked in the browser-mock dev build.
• Finder request path exercised against a live local proxy: model auto-selected, 200 response.
• Visual review of the merged Home, Agents, Usage, and Version pages
• AI finder end to end inside the installed desktop app: a live search through the local proxy returned matches named by display name.

🤖 Generated with Claude Code