Repository navigation
Releases: TED-H-com/multi-ai-chat-desktop
Release list
v1.9.6
What's Changed
- docs: point Latest downloads at v1.9.5 by @teddashh in #111
- docs: hand off the v1.9.5 cycle by @teddashh in #112
- fix(adapter): recognize ChatGPT ProseMirror composer by @teddashh in #114
- fix: handle ChatGPT Chat/Work layout and signed-out Grok by @teddashh in #118
- docs: prepare the v1.9.6 release by @teddashh in #120
Full Changelog: v1.9.5...v1.9.6
v1.9.5
What's Changed
- docs: point Latest downloads at v1.9.4 by @teddashh in #107
- fix(workflow): fail a step when its provider page is replaced by @teddashh in #108
- fix(engine): yield before every strategy's first composer write by @teddashh in #109
- docs: prepare the v1.9.5 release by @teddashh in #110
Full Changelog: v1.9.4...v1.9.5
v1.9.4: ChatGPT replies reach the app again
v1.9.4 — ChatGPT replies reach the app again
Provider-compatibility patch / Provider 相容性補丁
Multi-AI Chat Desktop v1.9.4 recovers ChatGPT response capture when the user-turn anchor detaches in the middle of a turn. A reply that ChatGPT finishes on screen now reaches the app instead of leaving the step silent until the 600-second host timeout.
Multi-AI Chat Desktop v1.9.4 會在 ChatGPT 的 user turn 錨點中途離開 DOM 時救回回覆擷取。ChatGPT 畫面上已經回完的內容,不再一路沉默到 600 秒的 host 步驟逾時。
A Windows zh-TW v1.9.3 Roundtable run showed the failure: Meta finished two speeches, then ChatGPT answered on screen and the app received nothing. v1.9.3 latched the anchor onto ChatGPT's optimistic user bubble; ChatGPT replaced that bubble with a collapsed Show more turn, so the latched node left the document and every live reply failed the follows-the-anchor test. With no chunk, the completion check was never armed, and the turn-completion timeout — which needs captured text before it can fire — stayed disabled.
一份 Windows zh-TW 的 v1.9.3 Roundtable 紀錄重現了這個失敗:Meta 完成兩輪發言後,ChatGPT 在畫面上回了,app 卻什麼都沒收到。v1.9.3 把錨點鎖在 ChatGPT 的樂觀 user 泡泡上;ChatGPT 之後用折疊的 Show more turn 取代它,被鎖住的節點離開 DOM,所有實際回覆都過不了「必須排在錨點之後」這道檢查。沒有 chunk 就不會啟動完成判斷,而必須先有已擷取文字才會生效的回合逾時也跟著失效。
v1.9.4 keeps the adopted anchor latched after its node detaches, treats a disconnected anchor as absent rather than as a hard stop, and permits an anchorless read only once generation has been observed with this wait's pre-send baseline already snapshotted. Assistant turns that existed before the send are still excluded.
v1.9.4 在錨點節點脫離後仍保留已採用的錨點,把斷線的錨點當成「沒有錨點」而不是直接停止讀取,並且只有在這次等待已經拍下送出前基準、而且確實觀察到生成之後,才允許無錨點讀取。送出前就存在的 assistant turn 仍然被排除。
Verification / 驗證
- CI on #105: frontend (734 tests), three-platform Clippy, and CodeQL passed; the adapters job was skipped because no adapter changed.
- #105 CI:frontend(734 項)、三平台 Clippy、CodeQL 通過;本次沒有動 adapter,adapters job 略過。
- Three new regressions cover a reply read after the latched anchor detaches, a completion when no anchor ever resolves, and a negative case that a pre-send assistant message is never emitted through the fallback.
- 新增三項回歸測試:錨點脫離後仍讀到回覆、完全沒有錨點時仍能完成、以及送出前就存在的 assistant 訊息不會經由 fallback 送出。
- Meta AI authenticated send, receive, and completion were confirmed live on Windows zh-TW with
v1.9.3(two Roundtable speeches, 909 and 1,333 characters). Meta stop, new-session reset, and profile persistence remain unverified. - Meta AI 已登入狀態的送出、接收與完成,已在 Windows zh-TW 的
v1.9.3實機確認(兩輪 Roundtable 發言,909 與 1,333 字)。Meta 的停止、新對話重置與 profile 持久化仍未驗證。 - The ChatGPT live retest has not been run on
v1.9.4. The fix is covered by focused tests only. - ChatGPT 的實機重測尚未在
v1.9.4上執行。 這個修復目前只有針對性測試覆蓋。 - Known gap: a reply that ends with no thinking detector, no strong activity signal, and no resolvable anchor can still stay silent until the host step timeout.
- 已知缺口:如果一則回覆結束時既沒有思考偵測、也沒有強活動訊號、又無法解析錨點,仍會沉默到 host 步驟逾時。
Downloads / 下載
- Windows x64 installer and portable zip
- Apple Silicon macOS DMG
- Linux x86_64 AppImage
SHA-256:
83f41da9b7ce3bcdce09be0f4805355b1f2c194962790420b41aa72064f6015c Multi-AI.Chat.Desktop_1.9.4_x64-setup.exe
133f66dbb923d251750e227927a7821a8166d82b3e88cecd76ce501aebe570ff Multi-AI-Chat-Desktop-1.9.4-windows-portable.zip
9c984912c8d1d9f8f0f87daab4bcc0df256a0e2bdb28a76ab426e7a910e3b52a Multi-AI.Chat.Desktop_1.9.4_aarch64.dmg
183e2045ce33185d0066bc776c1063ea6187112fb710c027dc3892c3092e19d7 Multi-AI.Chat.Desktop_1.9.4_amd64.AppImage
Full changelog: v1.9.3...v1.9.4
v1.9.3: Meta login is recognized
v1.9.3 — Meta login is recognized
Provider-compatibility patch / Provider 相容性補丁
Multi-AI Chat Desktop v1.9.3 reports Meta as logged in when its composer is visible, editable, and not inert. A leftover login button or inert prehydration field no longer hides a usable editor.
Multi-AI Chat Desktop v1.9.3 會在 Meta 輸入框可見、可編輯、且不是 inert 時,把狀態標成已登入。殘留的登入按鈕或 inert 預載欄位不再蓋過可用的編輯器。
A Windows v1.9.2 debug bundle showed Facebook/Instagram login returning to the pane while Meta stayed logged_out. The hydrated composer is also recognized as a textarea or contenteditable textbox (adapterVersion 3).
Windows v1.9.2 debug bundle 顯示 Facebook/Instagram 登入已回到窗格,但 Meta 仍是 logged_out。本版也認得 hydration 後的 textarea 與 contenteditable 文字框(adapterVersion 3)。
A disabled, readonly, or inert composer still stays logged out. This does not claim authenticated send/receive.
停用、唯讀或 inert 的輸入框仍是登出。本版沒有宣稱已完成登入後的送收。
Verification / 驗證
- CI on #102: adapters, frontend (731 tests), three-platform Clippy, and CodeQL passed.
- #102 CI:adapters、frontend(731 項)、三平台 Clippy、CodeQL 通過。
Downloads / 下載
- Windows x64 installer and portable zip
- Apple Silicon macOS DMG
- Linux x86_64 AppImage
SHA-256:
6242a6ded3a10364698eb3eeb099713eb2b026a65d193a3b76b77ecabcd0ad3b Multi-AI.Chat.Desktop_1.9.3_x64-setup.exe
916464a6499746576c2b7eacfc6a1bc76c80e639af5688d4064ee019e152dc3b Multi-AI-Chat-Desktop-1.9.3-windows-portable.zip
eefb2d2c1e3801c38bbcb332d63d569fc526f266e1a9a683ece2d8aec1954b60 Multi-AI.Chat.Desktop_1.9.3_aarch64.dmg
773a312de505c85906b328e0b2979669f0148db2645142d87666f06c364794e9 Multi-AI.Chat.Desktop_1.9.3_amd64.AppImage
Full changelog: v1.9.2...v1.9.3
v1.9.2: Meta Facebook/Instagram login stays in-pane
v1.9.2 — Meta Facebook/Instagram login stays in-pane
Provider-compatibility patch / Provider 相容性補丁
Multi-AI Chat Desktop v1.9.2 keeps Meta Log in on Facebook or Instagram inside the app. A login in the system browser cannot return to the isolated Meta profile.
Multi-AI Chat Desktop v1.9.2 讓 Meta 的 Facebook/Instagram 登入留在 app 視窗內。在系統瀏覽器登入無法回到隔離的 Meta profile。
A Windows v1.9.0 debug bundle showed nav-blocked www.facebook.com, after which Meta stayed logged_out through restart. Bundled Meta SSO now includes Facebook and Instagram HTTPS hosts (adapterVersion 2). Those hosts stay out of the Meta app surface, so the engine is not injected on identity pages.
Windows v1.9.0 debug bundle 顯示 nav-blocked www.facebook.com,之後 Meta 一直是 logged_out,重開也一樣。本版把 Facebook/Instagram HTTPS 登入網域收進 Meta SSO(adapterVersion 2)。這些網域不屬於 Meta app surface,引擎不會打進身份頁。
Also includes the v1.9.1 Grok Heavy resume-timer patch.
本版同時包含 v1.9.1 的 Grok Heavy 恢復生成計時器修正。
How to sign in / 如何登入
- Complete Facebook or Instagram sign-in in the Meta pane or the in-app login window.
- 請在 Meta 窗格或 app 內登入視窗 完成 Facebook/Instagram 登入。
- Do not use the system browser that previously jumped out; that session cannot come back.
- 不要用先前跳出的系統瀏覽器登入;那個 session 回不來。
Verification / 驗證
- CI on #100: adapters, frontend, three-platform Clippy, CodeQL passed.
- #100 CI:adapters、frontend、三平台 Clippy、CodeQL 通過。
- Authenticated Meta send/receive was not re-run for this tag. Please confirm login reaches
logged_inand survives restart on a real Windows build. - 本 tag 沒有重做已登入的 Meta 送收。 請用 Windows 實機確認登入後為
logged_in,且重開仍在。
Downloads / 下載
- Windows x64 installer and portable zip
- Apple Silicon macOS DMG
- Linux x86_64 AppImage
SHA-256:
a852d8f550c224b1375f0b84cb0cd0a06c9f1dfba0a6c307ea2ffe38d80edab7 Multi-AI.Chat.Desktop_1.9.2_x64-setup.exe
be9128c68b4a15ca7591843eb89e212b06d73cdb94d418b147708654fc1a7a7f Multi-AI-Chat-Desktop-1.9.2-windows-portable.zip
c76bc577853bd4ecb642b14892dbddc0c00a7fe1b7144f6e577e4e5b53ee0e24 Multi-AI.Chat.Desktop_1.9.2_aarch64.dmg
b57fdd4a4f9611a9a745e3147e958738e50027f7ceee0c81dd2bd13138d455ae Multi-AI.Chat.Desktop_1.9.2_amd64.AppImage
Full changelog: v1.9.1...v1.9.2
v1.9.0: Optional Meta AI standby
v1.9.0 — Optional experimental Meta AI standby
Stable release / 正式版本
Multi-AI Chat Desktop v1.9.0 ships five provider choices with exactly four active. ChatGPT, Claude, Gemini, and Grok remain the default. Meta AI is an experimental optional standby in Settings.
Multi-AI Chat Desktop v1.9.0 提供五家可選、同時只啟用四家。ChatGPT、Claude、Gemini、Grok 仍是預設;Meta AI 是 Settings 裡實驗性的可選備用。
Optional Meta AI / 實驗性 Meta AI
- Settings can swap Meta AI with exactly one default provider. Local profiles stay; roles and free-mode targets repair to the active four. No workflow grows a fifth seat.
- Settings 可用 Meta AI 替換恰好一家預設 provider。本機 profile 會保留;角色與自由分送 targets 會修到目前啟用的四家。Workflow 不會變成五席。
- Meta AI has its own portrait. Event log and debug bundle label it Meta AI without storing prompts or replies.
- Meta AI 有獨立 portrait。Event log 與 debug bundle 會標成 Meta AI,不儲存 prompt 或回覆。
- Guest or email/mobile access depends on what Meta's site currently offers. Facebook and Instagram sign-in are not embedded.
- 訪客或 Email/手機登入取決於 Meta 網站目前提供的方式。Facebook 與 Instagram 登入不支援內嵌。
Recovery and preflight / 恢復與預檢
- Login, reload, report, and replay errors can be retried. Duplicate and stale clicks are ignored.
- Login、reload、report、replay 失敗可以重試;重複或過期點擊會被忽略。
- Workflows preflight the current four even if a standby still looks Ready. Debate seats cover Meta when it is active.
- 即使備用仍顯示 Ready,workflow 也只檢查目前四家。Meta 啟用時 Debate 席位會涵蓋它。
- Standby errors lead to Settings. Saving a lineup is protected from overlapping autosave or Close actions; malformed replay providers are rejected instead of replaced.
- 備用狀態提供「設定」入口;切換儲存不受 autosave/關閉交錯覆寫,無效的 replay provider 會明確拒絕。
- Meta readiness and submission require a visible, editable composer, including a fresh check before sending.
- Meta 就緒與送出前都會確認輸入框可見且可編輯。
Verification / 驗證
- Local
pnpm verify: 724 Vitest tests across 62 files, 22 agent tests, TypeScript, ESLint, injected build, and adapter validation passed. Frontend production build passed. - 本機
pnpm verify通過:62 個檔案/724 項 Vitest、22 項 agent tests、型別/lint/injected build/adapter 驗證;前端正式建置通過。 - Automated coverage includes Meta replacing each default provider in Debate, Consult, Coding, Roundtable, and Brainstorm, plus snapshot/replay and standby recovery.
- 自動測試涵蓋 Meta 替換原四家各席位的 Debate、Consult、Coding、Roundtable、Brainstorm,以及 snapshot/replay 與備用狀態恢復。
- Three-platform release CI passed: 724 frontend tests, 22 agent tests, and 91 Rust tests on each platform; macOS ad-hoc signature verification passed. All four downloaded artifacts matched GitHub SHA-256 digests; the portable ZIP passed CRC and content checks.
- 三平台 CI、macOS 簽章、四個下載檔的 SHA-256 及 portable ZIP 完整性均通過。
- No new authenticated Meta or VM smoke was run for v1.9.0.
- v1.9.0 沒有新增已登入的 Meta 或 VM smoke。
- Prior evidence is unchanged: Windows packaged launch; an earlier Apple Silicon report that opened the app and logged into ChatGPT, Claude, and Gemini while Grok remained on Cloudflare; Linux CI-only packaging. Earlier no-login standby checks exist; they are not authenticated Meta send/receive.
- 先前證據不變:Windows packaged launch;較早的 Apple Silicon 回報能開啟 app 並登入 ChatGPT、Claude、Gemini,但 Grok 卡在 Cloudflare;Linux 僅 CI packaging。先前有未登入的 standby 檢查,那不是已登入的 Meta 送收。
Downloads / 下載
- Windows x64 installer and portable zip
- Apple Silicon macOS DMG
- Linux x86_64 AppImage
Notes / 注意事項
- Windows artifacts remain unsigned and may trigger SmartScreen. macOS is ad-hoc signed, not notarized. Linux remains CI-packaged without a new maintainer real-device launch report.
- Windows 產物仍未簽章,可能觸發 SmartScreen。macOS 為 ad-hoc 簽章、未 notarize。Linux 仍由 CI 封裝,沒有新的 maintainer 實機啟動回報。
- A real-account ChatGPT↔Grok slow handoff, the Grok Cloudflare challenge path, and a new Apple Silicon launch/provider-login smoke were not manually repeated.
- 真實帳號的 ChatGPT↔Grok 慢速接力、Grok Cloudflare challenge,以及新的 Apple Silicon 啟動與 provider 登入 smoke 都沒有人工重做。
- Meta AI does not bypass login, regional availability, rate limits, or later provider-side gates.
- Meta AI 不會繞過登入、區域、用量或後續 provider 端限制。
Full changelog: v1.8.9...v1.9.0
v1.8.9
v1.8.9
This release fixes the two failures found in the Windows v1.8.8 live smoke:
- Grok returns the original embedded pane to grok.com after the exact denied/external auth.x.ai first-login handoff, without expanding provider permissions.
- ChatGPT recognizes the current submit control and collapsed long prompts, does not accept an optimistic bubble while the draft is still idle, and treats the observed plain Pro thinking phase as active so Astra is not interrupted.
- Ordinary successful-but-empty bridge pulls no longer become a fake [Error: bridge degraded]; the workflow owns the exact 600-second inactivity timeout.
Recommended Windows checks
- Close the older app, install or extract v1.8.9, and confirm all previously signed-in providers reach Ready without restarting the app.
- Run a long Grok Heavy -> ChatGPT Astra -> Grok Heavy serial workflow without clicking ChatGPT Send manually. Confirm the long handoff is consumed once, Pro thinking remains active, no later input interrupts Astra, and no local bridge degraded error appears.
- If a disposable Windows profile or test account is available, complete the first login for all four providers and confirm Grok leaves Completing sign-in and becomes Ready without restarting the app. Do not delete the main profile to perform this check.
Windows artifact verification
- Installer SHA-256: BEA221D18A0A47D1B7AE22B79AF8C534BF18A5C0F1AEF39C75307F30650C9170
- Portable SHA-256: 0BC2F332273CDCEA2DDA32B4CEB97533DAECD0FB36B9ACF5B5D24B6513FD4F11
- Installer FileVersion/ProductVersion: 1.8.9
- Portable executable FileVersion/ProductVersion: 1.8.9
What changed
- Fix Grok first-login recovery and ChatGPT Astra handoffs in #94
Full changelog: v1.8.8...v1.8.9
v1.8.7: Reliable provider handoffs and safe workflow failures
v1.8.7 — Reliable provider handoffs and safe workflow failures
Maintenance update / 維護更新
Multi-AI Chat Desktop v1.8.7 makes slow structured handoffs into ChatGPT complete or fail promptly, prevents remounted history from completing the wrong step, and restores safe terminal behavior for unattended provider errors.
Multi-AI Chat Desktop v1.8.7 改善慢速 structured handoff 傳入 ChatGPT 時的完成與失敗判定,避免重新掛載的歷史回應誤完成目前步驟,並恢復無人操作 provider 錯誤的安全終止行為。
Highlights / 主要更新
- ChatGPT send confirmation is now tied to the matching newly rendered user turn. A silent delivery failure reaches a bounded error path in about 10–19 seconds instead of consuming the 600-second response timeout; exact long Unicode/Markdown prompts and composer remounts are covered by regression tests (#88).
- ChatGPT 送出確認現在綁定相符、剛渲染的使用者訊息。靜默送出失敗會在約 10–19 秒內進入有界錯誤路徑,不再耗盡 600 秒回應逾時;精確長篇 Unicode/Markdown prompt 與 composer remount 均有回歸測試(#88)。
- Response capture is anchored after that matching user turn, so a remounted historical answer cannot complete a retry or a new workflow. Cancel cleanup fully settles before a replacement run starts (#88).
- 回應擷取固定在該筆相符使用者訊息之後,因此重新掛載的歷史答案無法誤完成重試或新 workflow;取消清理也會在替代流程開始前完整 settle(#88)。
- Debate, Consult, Coding, and Roundtable again stop terminally on unattended provider errors, with fail-fast cleanup for an unfinished Consult sibling. Brainstorm keeps explicit Retry, Skip, and Cancel recovery; request IDs, FIFO ownership, stale-action rejection, and run isolation remain intact (#89).
- Debate、Consult、Coding、Roundtable 遇到無人操作的 provider 錯誤時會再次安全終止,Consult 未完成的並行 sibling 也會立即清理。Brainstorm 仍保留明確的「重試/略過/取消」,並保留 request ID、FIFO ownership、過期操作拒絕與 run isolation(#89)。
- Provider errors are no longer mislabeled as timeouts in diagnostics or recovery dialogs; English, Traditional Chinese, Japanese, and German copy is included. Independent heartbeat coalescing also keeps useful diagnostics in the bounded event buffer (#88, #89).
- Diagnostics 與恢復對話框不再把 provider 錯誤誤標成逾時,並包含英/繁中/日/德文案;獨立 heartbeat coalescing 也能在有限 event buffer 中保留有用診斷(#88、#89)。
- The multilingual official site and guides are refreshed, and the patched Vitest 4.1.11 line resolves GHSA-82fw-gwwq-j7x9 with zero known pnpm audit findings (#85, #86).
- 多語官方網站與指南已更新;Vitest 4.1.11 修補線解決 GHSA-82fw-gwwq-j7x9,pnpm audit 回報零筆已知問題(#85、#86)。
Validation / 驗證
- 522 frontend tests across 54 files and 22 Agent-ready tests passed.
- 54 個檔案共 522 個 frontend tests,以及 22 個 Agent-ready tests 全部通過。
- 81 local Windows Rust tests passed. Cross-platform Rust tests and warnings-denied Clippy passed on Windows, macOS, and Linux; CodeQL passed for Actions, JavaScript/TypeScript, and Rust.
- 本機 Windows 的 81 個 Rust tests 全部通過;Windows、macOS、Linux 的 Rust tests 與 warnings-denied Clippy 均通過;Actions、JavaScript/TypeScript 與 Rust CodeQL 亦通過。
- All four downloaded assets matched GitHub's byte sizes and SHA-256 digests. The macOS DMG passed CI signature verification. The Windows portable executable reports ProductVersion and FileVersion 1.8.7, includes the
PORTABLEmarker and README, stayed running for a 12-second isolated launch smoke, and was then closed. - 四個下載產物的檔案大小與 SHA-256 均符合 GitHub 記錄。macOS DMG 通過 CI 簽章驗證。Windows portable 執行檔的 ProductVersion 與 FileVersion 均為 1.8.7,包含
PORTABLE標記與 README,在 12 秒隔離啟動測試中持續執行後關閉。
Downloads / 下載
- Windows x64 installer and portable zip / Windows x64 安裝檔與 portable zip
- Apple Silicon macOS DMG / Apple Silicon macOS DMG
- Linux x86_64 AppImage / Linux x86_64 AppImage
Known limits / 已知限制
- Windows artifacts are unsigned and may trigger SmartScreen.
- Windows 產物尚未簽章,可能觸發 SmartScreen。
- The macOS package is ad-hoc signed but not notarized. A new Apple Silicon launch and provider-login smoke is still pending.
- macOS package 使用 ad-hoc 簽章但尚未 notarize;新版 Apple Silicon 啟動與 provider 登入 smoke 仍待人工驗證。
- A live real-account ChatGPT↔Grok slow handoff and the Grok Cloudflare challenge path have not yet been manually verified for this release.
- 本版尚未以真實帳號人工驗證 ChatGPT↔Grok 慢速接力與 Grok Cloudflare challenge 流程。
- Linux packaging is CI-verified only; no new real-device check has been performed.
- Linux 封裝目前只有 CI 驗證,尚未進行新的實機檢查。
Full changelog / 完整變更: v1.8.6...v1.8.7
v1.8.6: Brainstorm recovery, login compatibility, and dependency security
v1.8.6 — Brainstorm recovery, login compatibility, and dependency security
Maintenance update / 維護更新
Multi-AI Chat Desktop v1.8.6 lets long Brainstorm runs recover safely from provider limits, restores ChatGPT login-page detection, and clears four high-severity development-dependency alerts.
Multi-AI Chat Desktop v1.8.6 讓長時間腦力激盪在 provider 用量受限時安全恢復、修復 ChatGPT 登入頁偵測,並清除四筆高風險開發期依賴警示。
Highlights / 主要更新
- Brainstorm now pauses on usage limits, bridge degradation/timeouts, and structured provider errors with explicit Retry, Skip, and Cancel choices. Skip records a safe placeholder and continues to the next seat without feeding raw error text into later prompts (#80, #81).
- 腦力激盪遇到用量限制、bridge degraded/逾時或結構化 provider 錯誤時,現在會暫停並提供明確的「重試/略過/取消」。選擇「略過」後會記錄安全佔位內容並繼續下一席,不會把原始錯誤文字傳給後續提示(#80、#81)。
- Brainstorm snapshots advance to graph v4 so older v3 runs cannot silently adopt the new recovery semantics. Retry cancellation also reactivates and cleans up the reserved turn correctly (#81).
- 腦力激盪 snapshot 升為 graph v4,舊版 v3 執行不會靜默套用新的恢復語意;重試後取消也會正確重新啟用並清理保留的 turn(#81)。
- The bundled ChatGPT adapter is now v7 and recognizes the redesigned
/auth/loginform without expanding URL or permission scope (#78). - 內建 ChatGPT adapter 更新至 v7,可辨識新版
/auth/login表單,且沒有擴大 URL 或權限範圍(#78)。 - The development-only
fast-uridependency moves from 3.1.5 to patched 3.1.7, resolving all four open High Dependabot alerts; both production and full dependency audits report zero known vulnerabilities (#82). - 僅供開發使用的
fast-uri由 3.1.5 更新至已修補的 3.1.7,解決全部四筆 High Dependabot alerts;production 與完整依賴 audit 均回報零已知漏洞(#82)。
Contributors / 貢獻者
Thank you @Rumi-3653 for contributing the ChatGPT adapter fix in #78, and thank you @ufgeorge for reporting the Brainstorm interruption in #80.
感謝 @Rumi-3653 在 #78 提交 ChatGPT adapter 修正,也感謝 @ufgeorge 在 #80 回報腦力激盪中斷問題。
Validation / 驗證
- 488 frontend tests across 54 files and 22 Agent-ready tests passed.
- 54 個檔案共 488 個 frontend tests,以及 22 個 Agent-ready tests 全部通過。
- Cross-platform Rust tests and warnings-denied Clippy passed on Windows, macOS, and Linux; CodeQL passed for Actions, JavaScript/TypeScript, and Rust.
- Windows、macOS、Linux 的 Rust tests 與 warnings-denied Clippy 均通過;Actions、JavaScript/TypeScript 與 Rust CodeQL 亦通過。
- All four downloaded release assets matched GitHub's byte sizes and SHA-256 digests. The macOS DMG passed CI signature verification; the Windows portable executable reported version 1.8.6, included its
PORTABLEmarker, stayed running for a 12-second isolated launch smoke, and closed cleanly. - 四個下載產物的檔案大小與 SHA-256 均符合 GitHub 記錄。macOS DMG 通過 CI 簽章驗證;Windows portable 執行檔回報版本 1.8.6、包含
PORTABLE標記,在隔離啟動測試中持續執行 12 秒並正常關閉。 pnpm audit --prodand the fullpnpm auditboth report zero known vulnerabilities; GitHub reports zero open Dependabot alerts after #82.pnpm audit --prod與完整pnpm audit均回報零已知漏洞;#82 合併後,GitHub 顯示零筆開放中的 Dependabot alerts。
Downloads / 下載
- Windows x64 installer and portable zip / Windows x64 安裝檔與 portable zip
- Apple Silicon macOS DMG / Apple Silicon macOS DMG
- Linux x86_64 AppImage / Linux x86_64 AppImage
Known limits / 已知限制
- Windows artifacts are unsigned and may trigger SmartScreen.
- Windows 產物尚未簽章,可能觸發 SmartScreen。
- The macOS package is ad-hoc signed but not notarized. Apple Silicon launch and login still require manual verification.
- macOS package 使用 ad-hoc 簽章,但尚未 notarize;Apple Silicon 啟動與登入仍待人工驗證。
- Live provider login—including the Grok challenge path—and a live 48-turn Brainstorm recovery run have not been manually verified for this release. ChatGPT v7 has focused live-DOM selector evidence, not a full in-app logged-out workflow.
- 本版尚未實機驗證 provider 登入(包含 Grok challenge)與完整 48-turn 腦力激盪恢復流程。ChatGPT v7 有聚焦的 live-DOM selector 證據,但不是完整 app 內 logged-out workflow。
- Linux packaging is CI-verified only; no new real-device check has been performed.
- Linux 封裝目前只有 CI 驗證,尚未進行新的實機檢查。
Full changelog / 完整變更: v1.8.5...v1.8.6
v1.8.5: Reliable recovery, faithful transcripts, and UI polish
v1.8.5 — Reliable recovery, faithful transcripts, and UI polish
Maintenance update / 維護更新
Multi-AI Chat Desktop v1.8.5 improves stuck-provider recovery, transcript fidelity, responsive layout, and interface copy, while refreshing vulnerable build dependencies.
Multi-AI Chat Desktop v1.8.5 改善 provider 卡住時的恢復流程、對話記錄完整度、響應式版面與介面文案,並更新有安全風險的建置相依套件。
Highlights / 主要更新
- Grok recovery is more conservative and bounded: it considers reload only after an explicitly logged-in session remains stuck beyond the watchdog, while provider security challenges remain passive and cannot be bypassed (#65).
- Grok 恢復流程更保守且有明確界限:只有已確認登入的 session 持續卡住超過 watchdog 時才考慮 reload;provider 安全驗證仍保持被動,且不會被繞過(#65)。
- Transcript extraction better preserves code-block line breaks and rendered inline/display mathematics; transcript scrolling and replay interactions are also easier to use (#68, #70, #71).
- 對話擷取能更完整保留 code block 換行與 inline/display 數學式,並改善 transcript 捲動與 replay 操作體驗(#68、#70、#71)。
- Toolbars wrap better on narrow windows, preset details use a consistent layout, and stale UI copy is corrected while preserving the preset naming contract (#66, #72, #75).
- 窄視窗下的 toolbar 換行更自然、preset 詳情版面更加一致,也修正過時文案並維持既有 preset 命名契約(#66、#72、#75)。
- Vulnerable JavaScript build dependencies were refreshed;
pnpm audit --audit-level=lowreports zero known vulnerabilities (#76). - 已更新有風險的 JavaScript 建置相依套件;
pnpm audit --audit-level=low目前回報零個已知漏洞(#76)。
Contributors / 貢獻者
Thank you @DaveTseng2019 for contributing #65, #66, #68, #70, #71, #72, and #75.
感謝 @DaveTseng2019 提交 #65、#66、#68、#70、#71、#72 與 #75。
Validation / 驗證
- 485 frontend tests across 54 files, 22 Agent-ready tests, and 81 Rust tests passed.
- 54 個檔案共 485 個 frontend tests、22 個 Agent-ready tests,以及 81 個 Rust tests 全部通過。
- Cross-platform CI and warnings-denied Clippy passed on Windows, macOS, and Linux; CodeQL also passed.
- Windows、macOS、Linux 的跨平台 CI 與 warnings-denied Clippy 均通過,CodeQL 亦通過。
- All four downloaded release assets matched GitHub's SHA-256 digests. The macOS DMG passed CI signature verification, and the Windows portable build launched successfully in an isolated smoke test.
- 四個下載產物的 SHA-256 均與 GitHub digest 相符;macOS DMG 通過 CI 簽章驗證,Windows portable build 也在隔離環境中成功啟動。
Downloads / 下載
- Windows x64 installer and portable zip / Windows x64 安裝檔與 portable zip
- Apple Silicon macOS DMG / Apple Silicon macOS DMG
- Linux x86_64 AppImage / Linux x86_64 AppImage
Known limits / 已知限制
- Windows artifacts are unsigned and may trigger SmartScreen.
- Windows 產物尚未簽章,可能觸發 SmartScreen。
- The macOS package is ad-hoc signed but not notarized. Apple Silicon launch and login still require manual verification.
- macOS package 使用 ad-hoc 簽章,但尚未 notarize;Apple Silicon 啟動與登入仍待人工驗證。
- Live provider login—including the Grok challenge path—has not been manually verified for this release.
- 本版尚未實機驗證 provider 登入流程,包含 Grok challenge 路徑。
- Linux packaging is CI-verified only; no new real-device check has been performed.
- Linux 封裝目前只有 CI 驗證,尚未進行新的實機檢查。
Full changelog / 完整變更: v1.8.4...v1.8.5