This release brings major improvements to voice conversations, search, agents, Home, and CLI workflows, plus many reliability fixes across everyday chat use.
- GPT-6 Astra support: Berd now uses a newer Goose backend with proper GPT-6 Astra support, including its larger context and output limits.
- Voice engines: Voice Conversation now supports Apple Speech Recognition, Siri voices, OpenAI speech-to-text, and OpenAI text-to-speech, with clearer setup in Voice settings.
- Faster spoken responses: Berd can start speaking assistant replies while they stream instead of waiting for the full response to finish.
- Better voice turn-taking: Interruptions, false positives, mute boundaries, and active user speech are handled more reliably so Berd does not talk over you or resume stale audio.
- Background voice calls: Voice conversations can continue when you switch sessions or apps, with floating controls for mute, hang up, and returning to the active chat.
- Voice setup fixes: Starting a voice call now works more reliably after installing models, granting microphone access, changing settings, or starting from a new chat.
- Voice quality improvements: Siri audio is normalized to avoid garbled playback, Pocket playback quality is improved, and Pocket speed changes can apply during active speech.
- AirPods mute support: On supported macOS devices, AirPods and Beats mute controls now toggle Berd’s voice microphone state.
- Chat search:
Cmd+Kand Session History search now cover the full session set, not just chats already loaded in the app. - Model picker recency: Recently used models now appear near the top of the model picker and in the compact recommended view.
- Databricks model names: Unity Catalog models show cleaner, readable names in the chat composer while preserving the full model ID behind the scenes.
- Chat tables: Markdown tables now scroll vertically when needed and wrap long cell contents like URLs and file paths.
- Queued sends: Messages sent while a chat is still starting are no longer silently stranded, and retry behavior is more reliable.
- Voice-submitted messages: Chats now scroll to voice-submitted messages immediately, before the assistant responds.
- Artifact viewer: Open documents and images now stay in sync with files on disk, including external edits, shell writes, deletions, and recovery after temporary read failures.
- Sidebar recovery: “View all chats” now appears whenever more sessions exist, giving users a path to Session History even when the sidebar has not loaded those chats.
- Agent avatars: Custom avatars and generated gloopies now appear consistently across chat surfaces, delegation activity, active-agent summaries, pickers, mentions, and messages.
- Avatar library: Custom gloopies are now reusable across agents and appear first in the Gloopies collection.
- Agent creation: Newly saved agents open directly to their profile, descriptions validate more naturally, and the profile avatar affordance now opens the avatar collection gallery.
- Agent import: Agents can now be imported from ZIP files, and native drag-and-drop imports are more reliable.
- Agent cards: Shared agent cards have a cleaner layout and avoid obsolete metadata.
- Home canvas: Brand-new installs now start with a refined Home layout, and users can add text labels to organize their canvas.
- Home reliability: Berdy’s Home avatar stays visible when clicked, and onboarding now enters Berd directly after the welcome page.
- Settings: About has been merged into System settings, with updates and app details in one place.
- Composer focus: Closing model and project pickers now returns focus to the composer so you can keep typing.
- Mention suggestions:
Escapeand outside clicks now dismiss mention suggestions reliably without reopening for the same token. - Skills: The Skill Builder now warns before discarding unsaved edits.
- Skills toolbar: Skills search, import actions, and wide-screen alignment have been cleaned up.
- Public builds: Consumer builds no longer expose Block-internal Skill Discovery or send users to internal sign-in flows.
- Feedback: Eligible builds can show response ratings and sampled session feedback prompts, with surveys staying visible after transcript remounts.
- Windows app launch: Clicking Berd while it is already running on Windows now focuses the existing window instead of opening a duplicate instance.
- Berd Help: The bundled help guidance now checks whether
berdctl feedbackis available before recommending it. - berdctl folders:
berdctlcan now attach and detach project folders after a project already exists, and folder commands better handle~versus expanded home paths. - berdctl errors: CLI session and project commands now show more useful backend error details instead of only generic messages.
- bb apps: The bundled Apps CLI gained commands to list, inspect, debug, check readiness, roll back, delete, and manage access for deployed apps.
- Buzz Handoff skill: Berd now publishes a Buzz Handoff skill for bringing Buzz context into a private agent conversation and sharing an approved reply.
- Monitoring tools: Berd now bundles native monitoring and orchestration skills so agents can run longer-lived monitored workflows and safely retry delivered updates.
- Experimental — Remote SSH sessions: You can opt into running sessions on an SSH host while keeping the Berd UI local, with reconnect support and in-composer SSH environment setup.
- Experimental — OpenAI Realtime voice: A new Realtime voice mode combines a fast spoken interface with the full Berd coding agent behind it.
- Experimental — Pull request tracking: Berd can show related pull requests in the Changes rail and an in-app Pull Requests popover from the top bar.
- Experimental — Chat on canvas: Pinned chats can expand into live, interactive Home canvas cards.
- Experimental — Prompt pins: You can pin reusable prompts to Home and run them with one click.
Full Changelog: v0.6.2...2b87b50
Source commit: f7ce2748cc465f23f4210f892b3cae83d778f93e
The Windows NSIS installer and Linux packages lack platform-native code signatures; their updater archives remain minisign-authenticated.