Changelog.

Cleaner Sync Settings

• Replaced the old Sync dashboard grid with a normal Settings-style layout.
• Added a clearer iCloud Drive Sync toggle.
• Added a guided passphrase flow: turn Sync on, enter a passphrase, then click Start Syncing.
• Pressing Enter in the passphrase field now starts or unlocks Sync.
• Moved detailed sync information into cleaner rows for Folders, Actions, Stats, Storage, Snapshots, Diagnostics, and Advanced.

Better Folder Sync Clarity

• Folder selection now reads as a per-Mac choice, so it is clearer which folders sync on the current Mac.
• Excluded folders are treated more like local-only content for future sync work.
• Sync now handles folder, note, notebook, tag, and attachment scope more consistently.

Less Device Clutter

• Removed the old trusted, pending, and revoked device management experience from the main Sync UI.
• Old experimental producer entries are ignored and pruned when using the new Sync format.
• Diagnostics now show read-only producer information only when needed for troubleshooting.

Safer Sync Foundation

• Added a new Sync format marker so Canto can tell old experimental sync data apart from the new privacy-first format.
• Added clearer reset handling when old experimental iCloud Sync data is detected.
• Improved sync head and snapshot validation so Canto can better distinguish current sync data from stale or old-format files.
• Improved snapshot metadata so future setup and recovery flows have a stronger foundation.

Quality and Reliability Improvements

• Sync setup now avoids confusing duplicate error banners.
• Legacy sync blockers now guide you toward the reset action instead of failing silently.
• Disabling Sync on one Mac is treated as a local action rather than a device-revocation workflow.
• Sync diagnostics are calmer and easier to understand.

Attachments and Storage

• Added attachment syncing as part of the encrypted vault.
• Improved large attachment handling so media-heavy vaults sync more gently.
• Added lightweight attachment behavior so files can stay local only when needed.
• Added storage visibility and cleanup controls for synced attachments and snapshots.

Agent Chat and AI Workflow

• Replaced the older Ask AI note/notebook popups with Agent Chat.
• Made note AI actions, notebook AI actions, slash commands, and skills feel more unified.
• Improved long local AI runs, edit recovery, and slash composer stability.
• Reduced AI UI clutter by keeping more workflows in one place.

Faster Local AI

• Added Agent performance settings.
• Added Parallel Inference Slots for running multiple AI tasks when memory allows.
• Added Snapshot Cache on Focus to prepare note context sooner.
• Added Draft Decoding with an optional helper model for faster long plain-writing tasks.
• Improved local AI responsiveness, context preparation, and common note-edit routing.
• Kept Low Memory Mode respected by scaling performance features back when needed.

Reliability and Polish

• Improved model loading and fallback behavior on memory-limited Macs.
• Improved stuck-generation detection so valid lists, formatting, and repeated structure are less likely to be stopped early.
• Improved MCP Server settings copy and external-tool behavior.
• Improved sidebar/status-bar sync indicators.

🧰 Two Tiers of Tools — Safe by Default

Connected AI tools get a small, stable set of verbs by default, and a more powerful set only if you opt in.

• Vault tools (always on) — 17 named tools for the everyday work: search, read, create, update, organize, tag, follow links, attach files
• Automation tier (opt-in) — a single dispatcher tool (canto_invoke) that lets a connected agent open notes, send prompts, capture screenshots, and drive Canto's UI for testing or automation. Off by default
• Per-tool permissions still apply — clients like Claude Desktop will still ask you to approve individual tools the first time they're used

📁 Nested Folders in the Sidebar

Folders inside folders are now a real, drag-and-drop part of the sidebar.

• Drag a folder into another folder to reparent it
• Right-click → New subfolder under any folder
• Bulk move multiple notes and folders together via a redesigned "Move to folder" modal with a full expandable tree and search
• Invalid drops show a clear cue instead of silently failing
• Friendly guard rails prevent cycles or runaway nesting depth, with a readable toast instead of an error
• Backups now restore the full nested tree instead of flattening it

🧮 Diagrams, Math, and Wikilinks Survive the Round Trip

When an outside AI tool sends Markdown into your vault, Canto now turns it into the same rich blocks you'd get from typing it yourself.

• Mermaid diagrams become real, editable Mermaid blocks with a live preview
• Inline math ($…$ and $$…$$) renders properly with KaTeX, both in notes and in the agent chat panel
• Wikilinks ([[Note Title]]) resolve to real internal links so your knowledge graph stays connected
• Inline images and attachments render as actual images or tidy attachment pills, not raw text
• Round-trips are lossless — exporting and re-importing a note no longer degrades these blocks

✏️ Inline Math Without the Modal

Inserting and editing inline math now happens right in the line of text. The old popup window is gone.

📎 Attachments by Path, Not by Payload

External AI tools used to attach files by sending the entire blob, which hit request-size limits fast. Now they just hand Canto the local path on disk.

• Up to 100 MB per file, with friendly errors when something is too large
• Symlinks resolved safely so files always come from where you expect
• Images render inline; everything else gets a clickable pill

🔋 The "AI Ready" Pill Stops Lying

After flipping Low Memory Mode off, Canto now brings the chat model and vision worker back on its own — and the sidebar pill goes through real states instead of pretending nothing changed.

• New "AI Sleeping" pill (gray) when the chat model has been unloaded to save memory
• New "Warming up…" pill (yellow) while the model is coming back
• New "Restoring AI…" pill (blue) during the post-Low-Memory-Mode restore
• No more doomed first prompt — sending a message no longer fails just because the model was secretly unloaded
• External models (Ollama, cloud) are reflected honestly instead of incorrectly going gray

🛡️ Smarter "Stuck Run" Detection

The safety net that watches for hung AI runs no longer trips on legitimate short repetitions like a row of --- dividers or heading rules. It now waits for sustained repetition before stepping in, so genuinely stuck runs are still caught and healthy ones are left alone.

🧹 One Less Big Model on Disk

Canto used to keep a separate ~1 GB classifier model around just to decide which tool to call and to name new chats. That model is gone — your active chat model handles both jobs now. Less to download, less to keep loaded.

• Chat list shimmer — when Canto is naming a new conversation in the background, that row briefly shimmers so you can tell why it says "New chat" for a moment

✨ More Small Improvements

• Folder paths in MCP tools — connected AI tools can create or update a note by giving a folder path like Projects/Canto/Notes; Canto creates any missing folders for you
• Tree or flat folder listings — connected tools can ask for a real nested folder tree (great for outline views) or the legacy flat list, depending on what fits them best
• Quick prompts updated for nested folders — the vault chat panel's built-in suggestions now know about subfolders and offer "organize this into a subfolder" presets

⚡ Improvements

• Low Memory Mode actually saves memory now. Flipping the master toggle reaches into nearly every memory-hungry corner of Canto at once, typically freeing more than a gigabyte of RAM — often closer to two on machines running a chat model and the vision server side by side:
  • Chat models use a more compact representation of their working memory on supported Qwen 3.5 models. Same answers, roughly half the cost.
  • The vision worker spawns with a tighter context window, a sensible per-model GPU offload, and a more compact cache.
  • Canto's local document database keeps a smaller per-connection cache, and grows back when you flip the switch off — no restart needed.
  • The semantic-search (MiniLM) cache halves at runtime, and entries older than 10 minutes evict on next access instead of pinning forever.
  • When you switch away from Canto, the editor revokes leaked image/video URLs, disposes preview-only editors, and asks the runtime to clean up — so the OS can reclaim memory while you're elsewhere.
  • Diagnostic events flush immediately instead of buffering for batch sends.
• Memory pressure triggers a real cleanup pass. When your Mac itself starts running short, Canto follows up its eviction wave with a forced cleanup so freed memory shows up in the dashboard immediately rather than minutes later.
• The memory total finally matches Activity Monitor. The "Canto X GB" readout in the dashboard header and status bar now sums every Canto process, including the vision worker — instead of only the main app process. The number Canto shows and the number Activity Monitor shows are now the same number.
• Clearer state badges in the Memory dashboard. Each row used to display INACTIVE for two completely different things (the master switch is off, vs. this feature is loaded but resting). Those are now distinct labels — IDLE for sleeping subsystems, LM OFF for "the master switch is off" — with matching dot colors, so you can tell at a glance whether you actually need to flip anything.
• The Low Memory Mode toggle takes effect instantly. Flipping it now fans out across services in real time — the vision worker restarts with new settings, the database resizes its cache, embeddings re-evaluate their cap — instead of waiting for the next operation to pick the change up.
• Chat model auto-fits when GPU memory is tight. If loading the chat model with full GPU offload fails because of memory, Canto retries once with half offload before giving up. A marginal launch that used to fail now succeeds.
• Memory toolbar pills are easier to scan. The Context window and Weight pinning rows wrap their controls cleanly so the new tunables don't push other elements around.

🛟 Why this release matters

If your Mac has felt borderline tight running Canto — beachballs at launch, the AI quietly never loading, or that nagging suspicion the memory readout was lying — 0.5.12 is the release for you. It doesn't ask you to upgrade your hardware; it asks Canto to be a better guest on the hardware you already have.

Improvements

• Tidier activity cards — the state chip (working / done / failed) now sits in the corner instead of pushing the title around, and the timestamp and duration share one clean line.
• Collapsible step contents — when a step touched a long list of blocks or files, you can expand it on demand instead of having it dominate the panel by default.
• Standalone thought messages get their own card, so a pure reflection from the AI doesn't visually merge into the next answer.

Behind-the-scenes

• General housekeeping and dependency cleanup so the app builds cleaner and ships smaller.

Slash commands are now on for everyone

In 0.5.9 we shipped the new / command system but accidentally only flipped it on for development builds. The "Try it now →" button on the old Ask AI banner went nowhere for most users. That's fixed: type / in chat and the menu actually opens now. Free users get three commands; Pro unlocks the full catalog and multi-step skills.

Link reading is fast again

If you opened the Memory dashboard in 0.5.9, you may have noticed a yellow caveat next to Link RAG saying "sqlite-vec unavailable — JS fallback." That meant the AI was using a slow fallback every time it read links from your notes. The fast native version now loads correctly in the packaged app. Expect snappier link reading, especially on notes with lots of links.