Changelog

​

Resume a session, OpenAI fixes

• Resume a session. Open any saved session and click Resume to keep recording into it later. The transcript, answers, notes, Ask chat and session id all carry over, and Assist is handed what it knew from the earlier sitting, so a two-day interview stays one session and the model answers day two with day one in context. Duration counts recorded time only, the session shows how many sittings it has, and the transcript marks where each resume happened. To make this work, Assist’s compacted session memory is now stored on the session row (local SQLite, deleted with the session) instead of only in memory.
• OpenAI models work again. Every OpenAI request failed with 400 Unsupported parameter: 'max_tokens'. Raven now sends max_completion_tokens, which GPT-5.x requires.
• An OpenAI key can be saved next to an Anthropic key. Settings validated the two keys in separate calls and the second was rate-limited, leaving “Please wait a moment and try again” and a spinner that never stopped. Both keys now validate in one call.
• Effort picker no longer offers Max on GPT-5.6 models. The API rejects it. If you had selected it, Raven falls back to Low.

​

macOS in-app self-update fix

• macOS updates now install. Earlier builds could download an update, show Restart & update, relaunch, and still be on the old version. Bundled libraries shipped read-only, so the macOS updater could not clear the download’s quarantine attribute and aborted the install. Libraries are now writable, with a build guard so it cannot regress. If you were stuck on 2.4.1 or 2.4.2, updating to 2.4.3 applies cleanly. See in-app updates.
• Updater errors are now logged instead of silenced, so any future update problem is diagnosable.

​

Stealth dock fix and key validation

• Undetectable no longer shows a Dock icon. In stealth mode the macOS Dock icon could intermittently reappear when the dashboard was shown; it now stays hidden.
• API key validation no longer hangs. Entering an OpenAI, Deepgram, AssemblyAI, or Anthropic key and clicking Save & Validate or Test Connection could spin forever if the network stalled. Validation now times out with a clear “Could not reach …” message, so the key can be saved once reachable.

​

Streaming Ask, saved chats, and macOS self-update

• Ask streams now. Answers type out token-by-token in both per-meeting Ask and Ask across all meetings - no more waiting for the whole reply. The input sits at the bottom like a normal chat.
• Ask chats are saved. Each meeting keeps its own Ask conversation, restored when you reopen the session. Ask across all meetings now supports multiple chats - start a new one, switch, rename, or delete - each auto-titled from your first question. A turn is never lost if you navigate away mid-answer; generation finishes in the background and the full answer is saved.
• Microsoft Teams 1:1 calls are now detected for auto-start, not just scheduled meetings.
• macOS updates in-app, like Windows. Now that builds are notarized, click Update now to download and Restart & update to install - no more grabbing a disk image by hand. (Existing 2.4.0 Macs get 2.4.1 through the installer one last time; in-app self-update takes over from 2.4.1 onward.)

​

Meeting intelligence, after the call

• Ask your meetings. A new Ask tab on every session answers questions from that call’s transcript, and a global Ask across all meetings searches your whole history with cited sources. It’s conversational with memory - follow-ups build on earlier answers, and older turns are summarized (never dropped) so long chats keep their context.
• Action items. Every session now surfaces a structured action-items checklist - task, owner, and deadline when they were spoken.
• Follow-up email draft. One click turns a session into a ready-to-send recap email. Drafts are saved with the session.
• Talk ratio. See your You-vs-Them speaking share live in the overlay and on each session.
• Export. Save any session recap as Markdown or PDF, with optional transcript.
• Auto-start on meeting detection. Raven can notice a Zoom, Google Meet, Teams, or Webex meeting and offer to start - set it to prompt, start automatically, or off in Settings. No bot ever joins your call.
• Smarter notes. Your mode’s note-template instructions now guide the summary, so notes follow the structure you set.

• Onboarding now asks for permissions before API keys, and saves your wizard step and keys - Quit & Reopen (e.g. after the macOS Screen Recording restart) resumes where you left off instead of wiping keys.
• The Mac update modal no longer offers a version you already have (same-version GitHub feed no longer shows “Raven X is ready”).
• Send Feedback (user menu and About) opens a form - no GitHub account required. Docs point to docs.useraven.ai.

• Fixed sessions that stayed Untitled with a blank summary after a call. Notes now stream at the model’s max output so reasoning can’t eat the whole reply, and missing notes are retried on boot.
• Notes generation stays on the fast, cheap models (Haiku / Luna); live assist is unchanged with the full model catalog and effort settings.

• Packaged the tray-icon assets into the Windows and macOS installers so the system-tray icon shows up reliably.

• First open-source, BYOK desktop release for Windows and macOS. Publishing it created the GitHub update feed (latest.yml / latest-mac.yml) so installed copies update cleanly.

• Reliability: graceful handling of client disconnects during AI streaming, plus internal telemetry and stability fixes.

• Hotfix for a real-time sync endpoint that could return errors on certain requests.

• Refreshed the Anthropic (Claude) and OpenAI model lineups to the current generation.
• Made the interview coaching mode more general-purpose, and hardened the app under load testing.

• Windows installers are now Authenticode-signed via the release pipeline, reducing SmartScreen friction.
• Dependency and security cleanup.

• Added anonymous usage telemetry with clear in-app links to the Privacy Policy and Terms. First release with real users.
• Observability and reliability improvements.

• Restored automatic GitHub Release creation in the publish pipeline.

• Fixed a fatal crash when the app booted twice on macOS.
• Launched the documentation site at docs.useraven.ai.

• Enabled macOS notarization for smoother Gatekeeper installs, with update-UX polish.

• macOS updates now quit and install correctly, with immediate download feedback.

• Per-account database isolation for safer multi-user and multi-environment use.

• Production-grade update experience: user-initiated downloads, an update button that reflects the real update state, and a macOS auto-update (zip) target.

• Reproducible native echo-cancellation (AEC) build pipeline with versioned artifacts.
• Cache-first profile loading, consolidated IPC error handling, and wired-up session summaries.

• Packaging and version bump.

• First build: a live transcript tab in the overlay with auto-scroll, a six-step onboarding with an overlay tour, a redesigned Settings screen, and broad security hardening plus a large automated test suite.