Changelog.

🗺️ Big Documents, Actually Understood

• Canto can now map a long PDF's chapters and sections (from its bookmarks or printed table of contents) and knows exactly which pages each one spans.
• Asking about a specific chapter or section now jumps straight to the right pages instead of searching the whole document piece by piece.
• Very large files (hundreds of pages, tens of megabytes) are read in careful, sized chunks so a single oversized request can no longer overwhelm the conversation.
• A live progress card shows reading progress for long, multi-part reads instead of a silent wait.
• If Canto starts repeating the same question about a document, it now corrects itself with a helpful nudge instead of giving up entirely.
• Documents that are still being indexed in the background can still be read by chapter and page range right away.

🧾 See What a Run Will Use, Before It Runs

• The research plan now shows a clear ✓ / ✗ checklist of exactly which sources (attached file, Library, Notes, Web, Past chats) a run will use, including your lock.
• Sources from the same file are now grouped together ("3 files · 46 passages") instead of listing every passage separately.
• Sources that turn out to come from a file you never attached are quietly flagged as "outside scope."

🧯 When a Run Gets Too Big, You Get Options

• If a research run gathers strong sources but the final answer becomes too large to finish writing, a new recovery card lets you:
  • Get a shorter answer using just the best sources ("compact synthesis")
  • Save everything it gathered as a reference list, even without a written answer
  • Retry with a smaller, lighter source set
  • Add what it found as a new section on an existing note
• The error message you see now explains what actually happened, instead of a generic failure.
• Quick follow-up questions after a research note now run with a lighter, tighter source budget by default, so they finish more reliably.

🎨 A Friendlier "Save as Note" Dialog

• Choosing a note format (Note, Brief, Matrix, Notebook) is now four simple icon tiles with a smooth animated highlight, instead of a cramped row of text.
• Advanced tuning options (fidelity, tone, etc.) are now tucked behind a single expandable section, so the dialog feels simpler at a glance.
• The dialog has a softer, frosted-glass look that matches Canto's light and dark themes.

🛠️ Smaller Fixes & Refinements

• Reading an attached file now correctly shows "Reading your attachment…" instead of the confusing "Reading your library…" message.
• Short, simple documents that used to trip up the repeated-question loop now read through cleanly.

Four Ways to Read a Research Note

Research notes now have a floating view switcher:

• Human — the clean, readable write-up.
• Audit — a claim-by-claim look at what's asserted and how well each point is supported.
• Trace — the step-by-step record of what was searched, read, and gathered.
• Diff — a "what changed" summary of what was preserved, added, or reframed from the original answer.
• Each note also carries a lineage panel showing where it came from, notes derived from it, and open questions still worth chasing.

Claims You Can Trust at a Glance

All on by default, and adjustable in Settings → Research Fidelity:

• Low-confidence claims are gently qualified instead of being stated as hard fact.
• Sources are graded by how directly they support each point.
• Key ideas are preserved — important concepts are pinned and any drift during rewriting is flagged.
• Sensitive topics are framed responsibly — clinical and psychological subjects are presented as conceptual exploration, not diagnosis.

Follow-Ups That Actually Go Somewhere

• Typed follow-up tasks with an Act picker send each next step down the right workflow with a click.
• Guided sessions walk you through a few quick questions, then use your answers to do the work — and you can pause and resume later.
• A "what now?" chooser offers clear action chips after a research run, with Save as note front and center.
• Notes remember their origin — a note created from a task links back to its source, and the source learns which notes grew out of it.

Answer in Your Language

• A new response-language control in every composer (welcome, Canto Chat, and per-document chat).
• Leave it on Auto to match your question, or pin a language so every reply, note title, and follow-up sticks to it.
• Reliable with non-Latin scripts like Chinese, Japanese, and Korean.

Smooth on Modest Macs

• No more out-of-memory crashes — sync, web search, and Memory Links now watch available memory and back off gracefully.
• Smarter under pressure — Canto releases the heaviest things first (like hidden web-search pages) and checks in more often when memory runs low.
• Problem notes step aside — if a single note repeatedly fails to index, it's quarantined so the rest of your library stays searchable.
• Read entire books — full-length PDFs, EPUBs, and Word documents now index cover to cover, not just the opening chapters.

Clearer Errors, Never a Blank Turn

• Empty replies explain themselves — when a local or external model returns no visible text, Canto tells you what likely happened and what to try.
• Errors stay contained — a hiccup rendering one chat can no longer take down the whole conversation.
• Note-creation deep links — when the assistant makes a note, the chat timeline shows a direct Open note ↗ link.

Fix Diagrams With One Click

• A broken Mermaid diagram now has a Fix with AI button.
• Canto first tries fast, deterministic clean-ups for common mistakes, and only calls on the model if the diagram is still broken.
• The corrected diagram is written straight back into your note.

A Calmer Composer

• Compact icon controls — persona, response-language, and scope pills collapse into tidy icons so the toolbar stays on one line.
• Still obvious when it matters — a control set to something other than its default gently tints and spells out the choice in its tooltip.
• A resizable docked chat — the chat panel beside a note can be dragged to the width you like, and remembers it.

Reliability Round-Up

• Web search stays inside Canto — the hidden browser used to read pages no longer occasionally pops links open in Safari.
• Citations survive skill-made notes — notes created by a skill keep their sources as clickable links.
• Related Notes never gets stuck — the spinner no longer hangs during vault indexing and refreshes on its own once it finishes.
• Task-driven note saves just work — saving from an explicit action no longer leaves a chat stuck mid-stream.
• Saved research matches the live run — process traces and source counts line up with what you saw, and restored chat tabs keep their real titles instead of reverting to "New chat".
• Steadier vault import — imports handle stray folder references gracefully, and Vault Mode transcripts read in the right order.
• Better folder moves — the Move-to-folder dialog can create a new subfolder under the folder you picked and drops your item right into it.
• Cleaner action menus — the post-research and task pickers were refreshed with consistent icons and tighter layout.

🔬 Deep Research That Recovers

• Graceful timeouts — deep-research runs now have sensible, well-separated time budgets and quietly retry brief hiccups instead of spinning forever.
• A real fallback — if a deep-research run times out without gathering anything, Canto automatically falls back to a normal web search so your question still gets answered.
• Honest reporting — when something does take too long, Canto can tell you where it stalled rather than showing a generic failure.

📡 Connect Canto From Your Other Devices

• Opt-in local-network mode — turn it on and AI apps on your phone, tablet, or another computer on the same network can reach Canto (off by default).
• Always protected — local-network access only turns on with a required access key in place, so nothing can connect without it.
• Clear addresses — when enabled, Canto lists the exact network addresses other devices should use.
• A bigger setup menu — a new searchable client picker in Settings → MCP Server covers 20+ popular AI clients (Claude Desktop, ChatGPT, Cursor, VS Code, Cline, Continue, Windsurf, Zed, Goose, and more), generating a ready-made connection snippet for your app, OS, and connection type.
• Smoother reconnects — clients holding a stale session from an older Canto version no longer hit "session not found" errors.

🌐 Self-Hosted Search On Your Network

• SearXNG on your LAN — point Canto at a SearXNG search server running on your home network or your own machine.
• Careful by design — private/LAN addresses are allowed for this one setting while risky internal targets stay blocked.
• Clearer setup help — the settings now note that private/LAN addresses are supported, remind you to enable JSON output on your instance, and give sharper errors from the "Test" button when something's misconfigured.

🖇️ Steadier Attachments & Images

• "File unavailable," calmly — in the Asset Library, an attachment whose original can't be found is flagged with a quiet pill instead of a broken thumbnail or crash, with more efficient availability checks.
• Images tell you what they need — dropping an image into chat now checks up front whether your current model can see it, and clearly prompts you to download a vision model when one is required.
• Kinder to iCloud files — files that haven't downloaded yet are recognized as present-but-offline rather than mistaken for missing.

🔌 More Reliable External AI Endpoints

• No more mid-thought timeouts — connections to local endpoints (Ollama, LM Studio, oMLX, and similar) now watch for activity instead of cutting off after a fixed time, giving a working model the time it needs.
• Long tool calls survive — when an endpoint streams a large or slightly malformed tool request, Canto repairs and completes it instead of dropping the whole turn.

🛠️ Smaller Fixes & Refinements

• Honest source counts — the chat timeline now shows how many sources were actually new after removing duplicates, and fetch/read steps no longer mislabel results as "0 sources".
• Relative timestamps in chat — the Canto Chat sidebar shows friendly relative times ("2h ago") based on when you last opened or used each conversation.
• Fusion respects your choice — turning on result fusion with a single search engine now uses that engine instead of quietly falling back to another.
• Consistent web-tool summaries — Canto Chat and per-document chat now describe web searches and page reads in the same clear language.

🔓 Everything unlocked

• All product features are now free — Workspace, Memory Links, Research Mode, the Canvas, web search, personas, and more.
• A license now buys unlimited on-device AI usage rather than unlocking features.
• Memory Links works out of the box for everyone, and deactivating no longer wipes it.
• One clear, consistent free-query meter across chat, research, and inbox actions.

🌐 Web search, your way

• Choose providers for three jobs: Find (locate pages), Read (extract text), Research (deep reports).
• New built-in browser renders real pages to read JS-only or blocked sites — no key, no permission, no setup.
• Bring your own provider: Tavily, Keiro (cheaper, AI-native), Serper, self-hosted SearXNG, an experimental DuckDuckGo finder, and an always-on key-free reader.
• Result fusion fans a query out to several engines and merges the best results — counts as one web step.
• Automatic reader fallback when a page won't open, plus provider tags on every source.
• Deep research now gathers up to 50 sources (bounded by your model's context).
• Note & Notebook chat get web search too, with clickable source badges.

🧠 Chat that remembers

• Canto can search and cite your past conversations, fully on-device.
• Pick how recall works: Keyword, Semantic, or Hybrid (recommended).
• Emerald [C#] citations open the cited chat and flash the exact turn.
• Proactive chips suggest related past chats as you type.
• The current conversation is always excluded; you can also exclude any chat.
• Canto Chat is now a tab you can keep, reopen, split, and restore after relaunch — with fast full-text history search.

🤖 Personas, commands & skills — one Agents library

• Personas, slash commands, and skills now live in a single Settings → Agents tab (browse, create, fork, import/export).
• Personas are one-click chat presets — built-ins include Companion, Critique, and Researcher — with their own prompt, tools, and model settings.
• Slash commands in chat — type / for commands & skills; new built-ins like /summarize-thread, /critique-last, /eli5, and /action-items.
• Run skills step by step from the composer, with approval for any step that writes.
• Personas are available in per-document chat too.

✍️ One composer, everywhere

• Welcome, Canto Chat, and document chat share the same composer.
• @ mentions render as clean inline pills (notes, attachments, past chats).
• Smarter placeholders, sticky scroll (with a "Jump to latest" pill), drop overlays that don't get stuck, and right-click spellcheck suggestions.

🔬 Sharper, deeper research

• Writes a tailored query for each source (vault, linked folders, web, past chats).
• Closes its own gaps — chases down thin results before writing the answer.
• Your past chats and linked folders can both be first-class research sources.

🖼️ Images that stay yours

• Online images are saved into your vault on insert, so they always render, sync, and export — no broken hotlinks. AI follows the same rule behind an approval step.

🧭 A real guided first run

• New vaults open with an accurate, hands-on tour plus a disposable Sample Project (notes, tags, a daily note, templates, a notebook, and a canvas) so everything feels alive immediately. Existing vaults are never touched.

💾 Portable, lossless files

• .canto, .cnb, and .cwsp files now carry everything faithfully — notes, tags, attachments, history, chats, agents, links, and settings.
• Attachments are deduplicated, secrets and downloaded models are left out, and backups use the same format.
• Very large vaults export reliably now, thanks to a new container format — older files stay compatible.

🔌 Connecting outside tools

• More reliable connections to external AI clients like Claude Desktop, with updated setup snippets and a legacy fallback.

🛠️ Fixes & refinements

• Workspace PDF previews render correctly instead of going blank.
• Linked folders keep re-indexing even when system file-change events go silent.
• Past-chat citations open in-app instead of as web links (including older notes).
• Theme-consistent right-click menus, cleaner Library vs. Workspace search, and a tidier welcome-screen window drag.

📚 Citations That Stay Dense Through the Save

• Automatic citation recovery — when saving a research chat, dropped source markers are automatically re-matched to the sentences they belong to, even when sentences were merged or reworded.
• Optional AI citation placement — a new switch in Settings → Notes runs an extra AI pass that attaches gathered-but-uncited sources to the claims they support. Every suggestion is verified against the source before it lands, and it's skipped on very small models.
• Fuller related notes — the "related notes" list at the bottom of a note is less likely to come up sparse.

🧮 Math That Renders the Way You Wrote It

• Equations stay intact — formulas now display correctly in chat instead of occasionally arriving as raw symbols.
• Multi-line formulas carry over — aligned equations and piecewise definitions move from chat into a saved note without breaking apart.
• Tricky labels just work — labels with underscores or special characters inside a formula now render reliably instead of throwing an error.

📐 Diagrams That Match Your Theme

• Themed to your vault — flow diagrams now follow your active Canto theme (light or dark) instead of clashing with the note around them.
• Self-repairing — common syntax slips are quietly fixed on a quick retry, so a small mistake no longer turns a diagram into a wall of raw code.
• Cleaner guidance — the AI is steered away from the patterns that caused diagram trouble, so they come out right more often to begin with.

💾 Steadier Research Saves

• Thinking models stay on task — reasoning-heavy local models (including Ollama and Ollama Cloud) no longer spend their whole budget thinking and hand back an empty draft.
• Drafts finish themselves — a save that gets cut off at the model's length limit now continues automatically instead of saving a truncated note.
• Watch the finish happen — the final pass (citations, roles, evidence ledger) shows a live, step-by-step indicator with a real progress bar, instead of a frozen "Almost ready…" pulse.
• Follow-ups dig deeper — research keeps its sources open after the first pass so a follow-up question can fill the gaps, with a sensible cap on extra web lookups per turn.

🛠️ Smaller Fixes & Refinements

• Shift+Enter line breaks stick — line breaks added in the chat composer are preserved exactly as typed instead of silently collapsing.
• The window is draggable again — the chat header bar and the update banner are reliable grab-and-move zones.
• Tool steps finish in order — when a model runs several tools at once, each result matches the right step, so earlier steps no longer get stuck on "running".
• Live reasoning scrolls neatly — a model's streaming thought card scrolls within its own panel instead of dragging the whole conversation along.
• Disabled sources stay quiet — a source you turn off shows as gently skipped rather than flagged as a failure.

📐 Diagrams That Hold Together

• Tables and text diagrams finish drawing. The safety check that watches for stuck output used to occasionally cut off the neat, repetitive-looking parts of an answer (table rows, the lines and boxes of a text diagram). Those now get much more room, while genuinely stuck output is still caught right away.
• Diagrams export faithfully. Flow diagrams now export to HTML and PDF looking exactly as they appear in the app, instead of sometimes losing their labels.
• Captions that carry. Diagram captions show up consistently under the diagram, both in the app and in exported files.

🎨 A Calmer, Cleaner Canvas

• Smoother to move things. Dragging, resizing, and panning now track your cursor immediately, without the slight lag that made precise placement feel imprecise.
• A toolbar that only shows what applies. Color swatches and an incomplete "connect" button that did nothing on certain items have been removed, so the selection toolbar only offers actions that actually work on what you've selected.
• Connections work as before. You still connect ideas by dragging from one item to another.

🛠️ Smaller Fixes & Refinements

• Export menu stays on top. The Export dropdown on a research note no longer gets clipped behind the bar beneath it, so every option is reachable.
• Cleaner Evidence Ledger. Each claim now shows its text without stray citation markers mixed in, and the progress bar counts up smoothly without appearing to overshoot.
• Truer to the chat. A handful of small rendering touch-ups mean a saved research note lines up more closely with what you watched stream into the chat.

📚 Research notes that look published

When you save a research answer, you now get a polished document instead of a wall of text.

• A real title cover with the date and a live count of how many sources informed the work.
• Native charts — genuine bar, line, pie, and scatter charts with axis labels, value labels, gridlines, and legends, able to pull their data straight from a table or the evidence ledger.
• Callout boxes — highlight key points in tidy info, insight, warning, or success callouts.
• Captioned diagrams that render reliably, with stray or mis-typed diagrams quietly repaired.
• One clean reference list — duplicate bibliographies are stripped, and the evidence ledger tracks every claim.

Export that matches what you see

Export any research note to HTML, PDF, or Markdown and the file looks exactly like the one on your screen — diagrams, charts, math, tags, and table of contents included. Exports are theme-aware, so a light vault produces a light document and a dark vault a dark one.

⚡ Turbo — faster local AI

An optional new engine runs supported local models up to about twice as fast, with no change in answer quality. Standard stays the default.

• One tap to switch — supported models show a purple "⚡ Turbo" pill.
• Live speed readout — watch real-time tokens-per-second, prefill progress, and thinking state in the sidebar.
• Clear trade-offs — Turbo is a separate download and is text-only, so image input switches off while it's active; the app spells this out before you commit.
• Manage downloads — a new dialog lets you delete the Standard build, the Turbo build, or both.
• Smoother long answers — streaming keeps up with big drafts instead of turning choppy.

💭 Watch the model think

A model's private reasoning now streams into a quiet, collapsible thought card instead of leaking into the answer. You can watch it think, then it tucks itself away once the real answer begins — and that reasoning never ends up in your copies, exports, or follow-up turns.

🖋️ Richer, calmer chat

• Diagrams that draw as they stream, holding the last good frame instead of flickering.
• Collapsible toggles and callout boxes in replies, with smooth open/close animations.
• Crisper math, and pasting a table full of dollar signs no longer gets mistaken for an equation.
• Expressive when it helps, concise when it should be — replies match the depth of your question.
• Gentle motion — new messages slide in and the view follows along, all respecting your reduced-motion setting.

🔎 Smarter, fairer research

• Hears both sides — on genuinely debatable questions, Canto seeks out opposing evidence and files it in the Evidence Ledger's "Against" column. Neutral topics are left alone, so there's never manufactured controversy. A simple Auto / On / Off control puts you in charge.
• "Expand on [W101]" — refer to a source by its citation marker and Canto re-reads exactly that source instead of guessing with a new search.
• Clearer save feedback — a real progress bar counts claims as they're extracted, and the "Created…" confirmation is a link that opens your new note.
• Save more than once — turn a single rich chat into several documents; the save bar stays ready with a "Save another as" option.

🕸️ Smarter connections everywhere

• Consistent suggestions — the related-notes list, the knowledge graph, and chat grounding now all rank notes the same way, so they stop contradicting each other.
• Short AI summaries of each note power cleaner, more relevant related-note matches.
• Natural graph navigation — scrolling now pans the knowledge graph, and Ctrl/Cmd+scroll (or a trackpad pinch) zooms toward your cursor.

🔌 A bigger door for connected tools

For people who link Canto to external AI clients, connected tools can now read note details and chat transcripts and drive more of the app (chat, sidebar, Vault Manager, Settings, and graph) — all still on your terms.

🛠️ Fixes & refinements

• Local Ollama models now report the context size they actually loaded, so what you see always matches reality.
• On notes with a cover and a long table of contents, the tag row no longer overlaps the contents.
• Clicking into a note's body below a collapsed diagram or the tag row now places your cursor correctly.
• A safety cap prevents weaker models from hanging on the faster Turbo engine.
• Collapsible sections export to Markdown respecting whether they were open or closed.

🎛️ The Right Settings in the Right Place

• Context-window controls moved to Settings → Agent, right next to auto-compaction, so everything about your conversations lives together.
• Choose Adaptive (Canto sizes the context window to your machine) or Manual (pick the exact size yourself), with plain-language descriptions for every option.
• A clear Reload model button applies your changes immediately.
• A brand-new Settings → Advanced panel gathers the deeper, optional knobs — Low Memory Mode, GPU/CPU balance, and idle-model timers — into one clearly labelled place, with a "Reset to defaults" link one click away.
• Behind the scenes, Canto now tidies up local models it loads while checking them, so they don't quietly pile up in memory on smaller machines.

📚 Polished Research, Everywhere

• The Document Agent now produces research notes to the same high standard as Canto Chat.
• Numbered citations in agent-written research turn into tidy, clickable citation pills.
• Sources gather into a single, clean References section instead of scattered or duplicated lists.
• The full bibliographic detail a model writes for a source is preserved, so references read properly rather than getting trimmed to a bare link.
• Diagrams in your notes can now carry a short caption underneath, so it's clear what each one shows.

🏷️ Chat Titles That Actually Fit

• Auto-generated chat titles now reflect what the assistant actually answered, not just your opening question.
• The model's behind-the-scenes "thinking" never leaks into a title anymore.
• Opening Agent Chat no longer pops the slash-command menu open on its own — it waits until you type "/".

🎨 A Calmer, Cleaner Window

• The sidebar now reads as its own gentle "island" while the editor, tabs, and status bar settle onto a soft, neutral page surface.
• The sidebar smoothly crossfades as you switch between Canto Chat, the Welcome screen, and the editor, rather than snapping.
• The cleaner surfaces look great across all of Canto's light and dark themes.

🛠️ Smaller Fixes & Refinements

• Update notification pills are easier to click precisely, and the Restart button can't be triggered twice while an install is already underway.
• While a new version is installing, the button shows a spinner and an "Installing" label, then recovers gracefully if anything goes wrong.
• Canto recognizes your existing References, Sources, and Footnotes sections more reliably when adding to a note.
• Code blocks and their toolbars now have crisp borders in light themes, matching the polish of dark mode.