Skip to content

Releases

Cue 1.33.3 — stronger reasoning and stable opening answers

Choose a tag to compare

@balabommablock-cpu balabommablock-cpu released this 27 Oct 07:31

Cue now reasons from relevant role/company context without treating a rehearsal answer bank as the boundary of its knowledge. Mixed Markdown packs separate career evidence, opportunity reference and draft answers. Personal advice can propose new choices while historical claims remain grounded.

In cue mode, a follow-up keeps the old card until a complete opening line is ready, then appends finished cue lines from the same generation. Hidden preparation is confirmed against the final question and surrounding speech; changed inputs discard it. Thinking text remains excluded.

Validation: 524 automated checks; six live Sol Fast quality cases; a packaged-server/real-ASR/real-model rapid-follow-up replay; source/runtime secret scans; blank-profile and package/signature checks. First useful openings were measured in seconds, not reliably below one second. See docs/VALIDATION.md for current native checks and limits.

Apple silicon Mac preview, ad-hoc signed and not Apple notarized. No Windows installer. Existing users retain their own local account, context and downloaded models. Fresh packages contain no biography or saved account.

Final installed-native combined capture passed with Mac speaker playback and the Brio microphone. The rapid spoken case retained its added constraints and the final recommendation appeared automatically without thinking text. Shared-room background conversation can still trigger an unwanted suggestion; this is not perfect speaker identification or a sub-second guarantee.

Cue 1.33.2 — conversation context and interruptions

Choose a tag to compare

@balabommablock-cpu balabommablock-cpu released this 27 Oct 02:31

Cue now keeps actual spoken conversation separate from earlier AI suggestions. Paused case setup, candidate answers, clarifications, changing constraints and interruptions stay available to follow-ups. The previous answer stays visible until a replacement is complete, and thinking text stays out of the reader.

Improved cross-input echo handling, including spoken numbers versus digits, without dropping independent overlapping questions. Preserves the classic green overlay and existing local account/context.

Validation: 492 automated checks, real speech/model software replay, native room-microphone conversation replay, dual-input hardware checks, source/runtime secret scans, and package/signature verification. See docs/VALIDATION.md for test boundaries and observed timings. This does not guarantee sub-second replies or perfect speaker attribution.

Apple silicon Mac preview. Ad-hoc signed, not Apple notarized. No Windows installer. Use assistance where participants or rules permit it.

Final installed-app dual-input replay also passed: both physical sources registered the same spoken case, and the history contained one assembled question/answer with both requested parts. Recording was stopped after verification.

Cue 1.33.1 — stable, answer-only cues

Pre-release

Choose a tag to compare

@balabommablock-cpu balabommablock-cpu released this 26 Oct 19:41

Cue 1.33.1 improves answer speed and stability for permitted conversations.

  • Keeps the current answer until a follow-up or correction is complete; the first answer still streams.
  • Ignores common acknowledgements and candidate speech instead of restarting the answer.
  • Removes thinking/progress commentary from the reader and filters model reasoning events.
  • Uses shorter cue answers and preserves recent numbers for short what-if follow-ups.
  • Fixes timing metadata that could discard an otherwise valid Codex answer.
  • Keeps saved context, accounts and local model caches when updating.

469 automated checks passed. Six live GPT-6.1 Sol Fast text-provider scenarios returned first text in 1.4–4.6 seconds and complete answers in 2.7–6.8 seconds after preparation. The installed Mac app passed a typed numerical case, its follow-up and the recorded-voice test. These are observations, not guarantees. Updated-app live computer audio may require renewed macOS permission. See docs/VALIDATION.md for test scope.

Mac preview: Apple silicon, macOS 14.2+, ad-hoc signed and not notarized. Windows is not supplied. Fresh installs contain no previous user's context or credentials. DMG and ZIP checksums are included.

Cue 1.33.0 — easier setup and Mac installer

Choose a tag to compare

@balabommablock-cpu balabommablock-cpu released this 21 Oct 05:07

Cue 1.33.0 makes the path from download to a first answer easier.

  • A Mac disk image with Cue, an Applications shortcut and short installation instructions. ZIP remains available.
  • Simpler setup: the current AI comes first, with alternatives under Change AI; unfinished setup resumes at the saved step.
  • Try a typed question or an included sample voice with the microphone off. Choose whether the real conversation will happen on the Mac, a phone or in the room.
  • Download cancellation works when closing setup, pressing Escape or cancelling explicitly. Late results cannot switch the selected AI or report false readiness. Model changes cancel unfinished loads.
  • Smaller-model retry and alternate AI choices appear after a local setup failure. Loading and cancellation messages stay consistent.
  • The installed version is read from the app, and text-only sessions say Ask a question.
  • The classic translucent green reading interface, user context, account choices and credentials remain intact.

Validation: 447 tests passed on Mac; Linux verification and selected Windows core checks passed. Real clean-profile cancellation, cached local inference and sample-voice transcription/answering were checked in the built UI. The installed 1.33.0 app also passed a Qwen3 8B/Whisper sample-voice test with recording off, confirmed its Settings version and restarted cleanly; its displayed 3.3 seconds to first text is one observation, not a speed guarantee. The disk image was mounted read-only to verify its contents; the final image checksum and app signature passed. Source and packaged runtime scans found no credentials.

Mac preview: Apple silicon and macOS 14.2+. The app is ad-hoc signed and not Apple notarized; macOS may block it on a fresh device. Developer ID signing/notarization and recipient-device physical audio checks remain release work. No Windows installer, universal capture invisibility or fixed response-time guarantee is included. Successful paid OpenAI/Claude calls still need a valid user API key and balance.

Fresh downloads contain no developer account or personal context. Existing users keep their own local data. Model weights download separately.

Live audio follow-up: the installed 1.33.0 app answered a spoken test using the physical microphone, Mac loopback, and both sources enabled. Both meters returned to Off after Stop. The test used Mac speaker playback and a room microphone; it does not certify overlapping independent speakers or every meeting provider. Stale macOS permission entries had to be refreshed before the live check.

First-use guide, including audio and installation help.

Cue 1.32.0 — verified one-click local AI

Choose a tag to compare

@balabommablock-cpu balabommablock-cpu released this 21 Oct 04:37

Cue 1.32.0 keeps the classic translucent green overlay and adds a verified one-click local AI setup.

  • Download Qwen3 1.7B, 4B or 8B plus English Whisper inside Your AI → On this Mac. No Terminal or separate model app.
  • Setup generates a real answer before reporting ready. Empty or cancelled tests cannot leave a false-ready state.
  • Safer default for Macs with less than 16 GB RAM; cached model files are reused.
  • OpenAI and Claude API keys with account model lists and exact model IDs remain available, alongside eligible ChatGPT accounts.
  • Hardware recommendations, local AI limits, a suggested announcement and screenshots are in docs/.
  • 437 automated tests passed on Mac. Linux verification and Windows portable-core CI passed. Real clean-profile Qwen3 1.7B download, inference and cached reload were checked on an Apple M4 Max with 36 GB RAM. The installed native app also loaded Qwen3 8B, passed its self-test and answered a typed question.

Mac preview: Apple silicon, macOS 14.2+. Ad-hoc signed, not Apple notarized. No Windows installer is included; Windows core checks do not validate native audio, graphics inference or sharing. Qwen3 4B, fresh-recipient physical audio and successful paid-API calls still require hardware/account-specific checks. This release makes no universal invisibility or latency promise.

The app starts with a blank user profile. Upgrades preserve the user’s own local data. Credentials, developer context and owner analytics secrets are excluded. Model weights download separately under their own licences.

Cue 1.31.0 — OpenAI, Claude and the classic green overlay

Choose a tag to compare

@balabommablock-cpu balabommablock-cpu released this 21 Oct 04:09

Cue 1.31.0 keeps the classic translucent green reading overlay and adds OpenAI and Claude API connections.

  • Connect your own OpenAI or Claude Console API key; keys are encrypted on your Mac and never returned by credential-status calls.
  • Load models available to your account, or enter an exact compatible text-model ID, including a fine-tuned model.
  • OpenAI Responses with older Chat Completions compatibility; Claude Messages streaming.
  • Optional Claude workspace selection. Clear errors for invalid keys, unavailable models, quota limits and interrupted answers.
  • Existing Cue users keep their context, account and cache. Early Community-only installations retain their existing data folder.
  • ChatGPT sign-in, local AI, voice following, reading presets and Share View remain available.

430 automated Mac tests passed. Provider HTTP/SSE flows were tested with synthetic fixtures; no paid OpenAI/Claude API credentials were used for those checks. Native key entry and invalid-key handling were checked in the installed app. This is not a model accuracy or latency guarantee.

Apple silicon Mac, macOS 14.2+. Model weights download separately. Ad-hoc signed; not Apple notarized. Share View has limited capture boundaries and does not promise universal invisibility.

Cue Community 1.30.1 — Mac preview

Pre-release

Choose a tag to compare

@balabommablock-cpu balabommablock-cpu released this 21 Oct 03:25

Cue Community is an MIT-licensed Mac preview with blank personal context and no shared account.

  • Add and review your own CV or notes.
  • Connect your own eligible ChatGPT account, or download a local model.
  • Choose Compact, Comfortable or Large text. Export/import display and answer preferences without exporting accounts or context.
  • Separate Community app data and launcher identity.
  • 404 automated tests passed. Source and bundled-runtime secret scans passed, with a narrow exception for an upstream tokenizer code expression. No reported npm vulnerabilities at publication.

Download the Apple silicon ZIP, extract it and move Cue.app to Applications. Requires macOS 14.2+. Models download separately. The SHA-256 file lets you verify the download.

Preview limits: the app is ad-hoc signed, not Apple notarized, so macOS may block opening it. This is not a Windows build. Live speech/model behavior needs checking on the recipient's Mac. There is no universal invisibility or sub-second latency guarantee.

Share View filters Cue from its own presentation source. Other recording and screen-sharing paths can capture the overlay. Read the screen-sharing boundaries and validation report in the repository before using it.

Cue Community 1.30.0 — superseded preview

Choose a tag to compare

@balabommablock-cpu balabommablock-cpu released this 21 Oct 03:19

This initial preview has been superseded by Cue Community 1.30.1.

Use 1.30.1 for the corrected public-registry lockfile and clean-machine build checks. The same preview limits apply: Apple silicon only, ad-hoc signed, not notarized, and no universal invisibility guarantee.