v0.9.8 — Attached images readable by any model

Fixed

• Attached images now work on every model — attaching an image only passed its file path with a "read it" note, so a text-only brain (DeepSeek, GLM, …) replied that it couldn't read the image. The daemon now transcribes each attached image to text (visual description + verbatim OCR) with a vision model — Gemini Flash first, OpenAI gpt-4o-mini fallback — and inlines that into the prompt, so any brain can read it. The original file still rides along for natively-multimodal brains to read directly. (Needs a Gemini or OpenAI key in ⚙ CONNECT; falls back to the old behaviour without one.)

Update

bagidea update

🤖 Generated with Claude Code

v0.9.7 — Agent models in the roster, orb polish

Added

• Each agent's model is shown in the roster — the agents panel now shows a "🧠 <model>" line under every agent (e.g. deepseek-v4-pro, kimi-for-coding, glm-4.6; Claude for the default brain), with the full provider/model on hover. The CEO — your stand-in, not an AI agent — shows none.

Fixed

• Orb edge looks smooth — the circular clip sat exactly on the orb's glowing rim, so its hard edge cut the glow against the colourful wallpaper and looked jagged. The orb art is now inset a few pixels, leaving a thin transparent halo so the clip falls on empty space instead of the glow.
• No caption chrome behind the orb on click — despite being undecorated, the orb window still carried a system menu + min/max styles, so clicking it flashed a white caption bar and a system icon / window buttons in the corners. Those styles are dropped and the non-client area removed, so nothing draws behind the orb (without disturbing the transparent compositing).

Update

bagidea update

🤖 Generated with Claude Code

v0.9.6 — Orb click-through

Fixed

• Orb no longer blocks the desktop around it — the orb's window is wider than the visible circle (Windows pads it to a minimum width) with transparent margins, so anything beneath them — e.g. desktop icons — couldn't be clicked, and the orb looked off-centre. The window is now clipped to a circle centred on the visible orb (sized from the real client rect, re-applied on DPI/monitor changes): the margins are clipped away and click straight through to the desktop, the orb sits dead-centre, and a stray title-bar sliver on click is gone too.

Update

bagidea update

🤖 Generated with Claude Code

v0.9.5 — Per-model context windows, Kimi Code, orb polish

Added

• Kimi Code provider — the Kimi Code coding plan (kimi.com/code) is a separate service from the general Kimi · Moonshot API: its own sk-kimi-… keys, its own Anthropic-compatible endpoint (https://api.kimi.com/coding), and a single model (kimi-for-coding, 256k). Now a one-click built-in provider — pick Kimi Code, paste the key, Connect. Previously such a key failed against the Moonshot endpoint with a confusing 401.

Improved

• Context window is now per-model and auto-detected — the usage meter and the compaction point used one coarse number per provider, so models were badly mis-sized (DeepSeek showed 128k and compacted at ~115k despite a real 1M window). Each model now resolves its own window from a researched table (Claude 4.6/4.8 1M, DeepSeek V4 1M, Gemini 2.5 1M, GPT-4.1 1M, GLM-4.6 200k, Qwen3-Coder 1M, Kimi K2 256k, Grok, Llama, Mistral, …) and, where a provider advertises it on its model list (OpenRouter, Groq, Together, …), the live value wins automatically. The compaction budget is derived from that window (~80%), so big-context models run far longer before summarizing. Still overridable per provider via providerConfig.contextWindow / contextBudget.
• Gender-aware agents — an agent now self-identifies by the gender of its assigned voice (♀/♂), so the voice you hear and the words match (Thai ครับ/ผม vs ค่ะ/ฉัน) in both chat and realtime calls.

Fixed

• Voice hotkey (Right Ctrl) no longer wedges — a missed key-up when window focus shifted left it "stuck down" so the next press was swallowed. A 150 ms watchdog reconciles against the key's real physical state (GetAsyncKeyState), so it can't get stuck.
• Orb has no invisible grab box — the chat-head's square window let its transparent corners catch clicks/drags and show a hand cursor. Pointer events outside the circle are now ignored and the cursor is a plain arrow there.

Update

bagidea update

🤖 Generated with Claude Code

v0.9.4 — Reliable voice hotkey + gender-aware agents

Fixed

• Voice hotkey (Right Ctrl) no longer wedges — holding push-to-talk sometimes did nothing (then worked again after clicking elsewhere), most often in feed mode. A key-up could be missed when window focus shifted around the moment of a press, leaving the hotkey stuck "down" so the next press was swallowed as auto-repeat. A 150 ms watchdog now reconciles the tracked state against the key's real physical state (GetAsyncKeyState), so the hotkey can never get stuck in either direction.
• Agents now know their gender — voice & words match — an agent with a male voice could still write/speak about itself as female (e.g. saying "ค่ะ"), clashing with the voice you heard. Gender is now read straight off the assigned voice preset (♀/♂) and stated in the agent's persona, so it refers to itself consistently in every language (Thai ครับ/ผม vs ค่ะ/ฉัน, pronouns, honorifics) — in both chat and realtime calls.

Update

bagidea update

🤖 Generated with Claude Code

v0.9.3 — Voice fixes, smarter calls, macOS copy/paste

Fixed

• Voice push-to-talk no longer garbles Thai. It produced � characters (worse the longer you spoke) because the transcription response was decoded per network chunk, splitting multi-byte characters. Bodies are now decoded as UTF-8 whole (same fix applied to Claude summaries + the auto-translation path).
• macOS: copy/paste works (⌘C/⌘V/⌘X/⌘A) — the frameless window had no Edit menu, so the shortcuts never reached text fields. Adds a standard Edit menu. Fixes #8.

Improved

• Smarter voice calls — the call agent is now framed as your Director and gets a live office snapshot (projects in progress, proposals awaiting approval, scheduled jobs) on top of the team + notes, so it can actually talk about your work and help plan. Every call also leaves a chat-app-style record: "📞 Voice call with · HH:MM · 2m 13s".

Update with the 🔄 banner or bagidea update (macOS: git pull && ./build-mac.sh).

🤖 Generated with Claude Code

v0.9.2 — Launch with Windows by default

Fixed

• The office now launches automatically with Windows. Previously a fresh install never wrote the auto-start entry, so it didn't come back after a reboot. Now the installer enables it on first install (without overriding a later "off"), and existing installs get it turned on once on their next bagidea update. Toggle anytime with bagidea startup on|off.

Update with the 🔄 banner or bagidea update.

🤖 Generated with Claude Code

v0.9.1 — Office files + File & Media Toolkit + MIT license

A small but practical follow-up to v0.9.0.

Added

• Office-file support — the installer now bundles LibreOffice, so agents can read & convert xlsx / docx / pptx (→ csv / pdf / txt) headlessly. Fills the spreadsheet gap.
• "File & Media Toolkit" built-in skill — a protected skill that maps each task to the right bundled tool, so the office's existing power actually gets used: PDF (Read), Office files (LibreOffice), docs/books & slides (pandoc), YouTube/video (yt-dlp + transcribe, ffmpeg), images (ImageMagick), data (csv/jq). Assign it to your hands-on agents.

Changed

• Added an MIT LICENSE — the project is now properly open source.

The toolkit skill ships via bagidea update; LibreOffice + the agent CLI tools are installed at install time (fresh install or re-running the installer).

🤖 Generated with Claude Code

v0.9.0 — More brains, safer delegation, agent-built workflows

A big follow-up to Swappable Brains — more models, safer delegation, and a Workflow Builder the team can drive.

🧠 More brains

• 8 new providers — Groq, Cerebras, xAI (Grok), Mistral, Together, Fireworks via the built-in proxy, local Ollama / LM Studio that need NO API key, and Kimi (Moonshot) direct. 18 providers built in, plus your own custom ones.
• Live model lists — pickers fetch each provider's current models on connect/open, so new models always show up.

🛡 Safer, smarter delegation

• Verification loop (opt-in) — a skeptical reviewer double-checks delegated work before it reaches the CEO, handing it back once for fixes.
• Approve / reject project proposals in-place — right in the chat or the feed, no digging into 🗂 TASKS.
• The Director is locked as the office manager — survives any prompt edit.

🔀 Workflows the team can build

• Agents can save workflows into the Builder (new Build Workflow skill), and 🪄 Draft with Director turns a goal into an editable workflow.
• Example workflows are read-only (Save forks a copy), with a save-overwrite confirm and on-brand dialogs.

✨ Polish

• Protected built-in skills + searchable Skill/Tool pickers in the agent editor.
• Redesigned neon chat-head (crisp glowing ring), bagidea brains CLI, and 14-language coverage for all new UI.

Update with the 🔄 banner or bagidea update. See CHANGELOG.md for the full list.

🤖 Generated with Claude Code

v0.8.2 — Embed orb logo (real cold-boot fix)

The real fix for the dark chat-head orb after a reboot.

Fixed

• The orb's logo is now embedded in the app, so it always shows. v0.8.1 retried the HTTP fetch, but on a cold boot the very first failure could be missed (the image started loading before the retry was wired), leaving the orb dark even after the daemon came up. The logo no longer touches the network — it's baked into the binary as a data URI — so the orb comes up correctly every time.

Update with the 🔄 banner or bagidea update (this one rebuilds the shell, which is what applies the fix).

🤖 Generated with Claude Code