Changelog

• Safe Mode launches Canto on tight Macs without crashing — notes, editor, and search work; AI features pause until you opt back in
• Low Memory Mode now actually frees memory — chat-model cache, vision context, document database, embeddings, and renderer cleanup all tighten when you flip the switch
• One-click "Switch to a smaller model" suggestion when the model you have loaded is too big for current memory
• Memory dashboard now matches Activity Monitor — sums every Canto process (main app, every window, helpers, vision worker), with a per-process breakdown
• Clearer state badges — IDLE for "loaded but resting" vs LM OFF for "the master switch is off", with matching dot colors
• Crash-loop protection — if startup crashes, the next launch enters Safe Mode automatically so you have a clear path back to a working window
• Quieter, more reliable diagnostic delivery — event names are validated at the source and a stale backlog of rejected events is cleaned up on startup

v0.5.12 is the release that makes Canto work on the Mac you actually own. Safe Mode is a brand-new boot-time launch path designed for memory-constrained machines (or any boot following a crash). Low Memory Mode goes from a polite request to a firm one. And the Memory dashboard finally agrees with Activity Monitor down to the byte.

• Live streaming progress on long tool calls — the agent shows a "Preparing… N characters" readout with a ticking elapsed-time counter
• Per-step wall-clock durations next to every step in the agent timeline
• Reasoning interleaved with the tool calls it led to, so multi-step runs read like a colleague thinking out loud
• Stuck-run watchdogs widened so long, healthy structured edits no longer abort right before they finish

v0.5.11 makes Agent Chat feel less like a black box. Heavy runs surface live progress, per-step durations make slow steps obvious, and the watchdogs that protect against truly stuck runs no longer false-positive on long, healthy ones.

• Type @ in the agent composer to pin notebook cells, notes, attachments, past chat sessions, or your current selection as structured AI context
• Slash commands and reusable multi-step skills run straight from the chat box — the old Ask AI modal is gone
• Packaged sqlite-vec fix so semantic search and Memory Links work reliably out of the box
• Slash commands are on by default for new vaults

v0.5.10 gives every important thing in Canto an address. Mentions turn vault items into typed pills the AI can resolve unambiguously, and slash commands replace ad-hoc prose prompts with structured, reusable skills.

• New user-controlled Low Memory Mode toggle to make Canto usable on 16 GB Macs
• Native whisper.cpp for voice transcription with Metal acceleration, replacing ONNX Whisper
• Voice & Transcription settings redesigned around the native engine
• Wikilink and URL pill content is now indexed for the agent so referenced material is searchable
• Paste and Match Style works in the editor with sandbox-safe clipboard IPC

v0.5.9 is the first release that takes memory-constrained Macs seriously. Low Memory Mode shipped as a single toggle, voice transcription moved to a native Metal-accelerated Whisper, and the agent gained read access to whatever your wikilinks and URL pills point at.

• AI chat history now stays with each note or notebook
• Start a fresh chat without losing older conversations
• Saved chats are easier to browse with clearer previews and timestamps
• AI-written note content comes back cleaner, and app closing gives clearer feedback

Canto v0.5.8 makes AI chat feel more organized and easier to return to. Conversations now stay attached to the work you were doing, while note edits and app shutdown both feel smoother and more reassuring.

• Manual Agent output cap plus automatic matching to the active model
• Clearer stop-by-limit feedback during long AI responses
• Cleaner placement for structured edits, lists, tables, and note updates
• Connected models show clearer read and reply limits

Canto v0.5.6 focuses on AI control and reliability. Longer runs are easier to manage, structured edits land more cleanly, and model limits are easier to understand while you work.