Switching sessions no longer leaves the conversation blank, citations become chips while an answer streams, and the Pick models step of onboarding holds up with a roster of hundreds of models. Highlights: - Every session opens on its own latest content. If the session you left was scrolled down, the one you opened could show nothing until you scrolled a long way up, because it took over the previous session's scroll position. Each session now opens at its end (a short one whole, from its top) and stays there while its answers finish laying out, until you scroll. A jump to the top with the scroll bar stays where you put it. A switch also comes to rest sooner: in a session-switch benchmark the median fell from 0.8–1 s to 0.1–0.3 s. - Streaming answers keep all their text. Coming back to a session that is still streaming, or scrolling the answer out of view and back, no longer loses parts of it. A stream in one session no longer scrolls or blanks another, and the conversation follows a turn to its end even when its answer arrives all at once, as it does when a single model uses tools. - Citations appear as chips while the answer streams, instead of raw tags such as [custom-sakana-fugu-max] that all changed text, width and size when the answer finished. The finished answer looks the same as the last streamed frame, and a chip clicked mid-stream opens that peer's trace. - Split on and Solo chips, the inspector's split cards and the compare sheet's spoken units keep the names of the models an answer cites as words: "minimax-m3 reports that…" where they used to read "reports that…". Tooltips and VoiceOver give the full name, and also name the models a claim is attributed to. - The Pick models step of onboarding works with a large roster such as OpenRouter's. The title, a search field and a "models per prompt" count stay in place while the list scrolls; each provider's header shows "N of M selected"; models are grouped by vendor, Anthropic, OpenAI and Google first; and batch, free and dated-snapshot listings fold under the model they belong to. The pre-selected model is the first ordinary model in each provider's list, never a variant or OpenRouter's Auto Router. - The inspector shows the newest turn when you are at the end of the conversation, even when that turn is shorter than the window, and its Latest control no longer stays visible doing nothing. Security: opening a conversation that isn't in a project used to keep the previous project's scope. Sending from it gave the models that project's file tools and project context, so files from the other project could be read and sent to your providers, and it moved the conversation into that project. A conversation's models now get only its own project's tools and context, in every window. Files you attach to a message yourself are still sent. Changed behaviour worth knowing: a conversation now changes project only when you move it, with the titlebar project menu or the project picker (⌘⇧O), which moves an open conversation at once and offers "Move to No Project". Sending a message no longer moves it. With "Automatically switch projects when opening a session" on, opening a conversation that isn't in a project switches the sidebar to No Project, so ⌘N then starts a conversation with no project. Conversations are named once, from their first completed answer, instead of being re-titled after every message, and a conversation you renamed is never renamed for you. Also fixed: polycode://settings/… links now open Settings, and a polycode:// link no longer leaves an extra window behind; two citations side by side in a table cell no longer split the cell; and the filter fields in Settings → Providers and the mode editor now use the same search field, where a filter of only spaces shows every model instead of none.

The inspector follows what you are reading, the sidebar can delete a whole group, and a long list of surfaces that failed quietly now say what went wrong. Highlights: - The Provider Trace sidebar follows the conversation: scroll back through a session and the inspector shows the trace for the turn you are reading — its peers, latencies, costs and agreement bar — instead of staying pinned to the newest exchange. A turn takes the sidebar when its top edge reaches the top of the conversation, so the two always describe the same answer. The header names the turn ("Turn 12 · 11:48", its prompt on hover) and offers Latest to jump back. Scrolling away to read something older and then receiving a stream no longer yanks the sidebar off what you are looking at. - Trace selection is no longer sticky. View trace, citation chips, Sources pills and Split chips now bring their exchange to the top and let the sidebar follow position, rather than pinning the inspector until your next prompt. Activating one from the exchange you are already on moves nothing. - Delete a whole group from the sidebar: right-click a project header for Delete N Sessions, which empties the project but keeps its folder access and exclude patterns, or Delete Project and N Sessions, which also removes the project. "No Project" and "Archived" offer the sessions-only item. The folder and its files on disk are never deleted. The confirmation counts what will actually go — for a project that is more than the header shows, because the header counts only unarchived sessions, so a group displaying six can state nine and name the split. - Answers wrap at a readable width: on a wide window an assistant answer is capped at 720pt instead of running the full column. At 1470pt with the inspector closed a line used to reach about 178 characters; the user bubble above it had been capped at the same width for months. At the default 1200pt window with the inspector open, nothing moves. - Manage models puts your picks first: your primary, then the rest of your enabled models, then everything else, with a rule marking where your picks end. With an aggregator like OpenRouter, four enabled models used to be scattered among 447. The order holds still while the drawer is open — ticking a box never moves a row out from under the cursor. - Failures say what failed and what to do next. Onboarding, Settings and the project picker now share one treatment, with the provider's or the system's own message kept beneath as selectable detail. Eight sheets previously rendered their own error line in five different styles. An offline Mac no longer shows a raw Error Domain=NSURLErrorDomain … dump, Settings → Usage no longer reports $0.00 when it simply could not load, and pressing Send with no provider configured now says so instead of doing nothing at all. Changed behaviour worth knowing: Settings → Projects → Remove now deletes the project's sessions. It used to revoke the folder grant and leave them behind under "No Project", permanently — granting the same folder again creates a new project, so those sessions could never be reassociated with it. The confirmation states how many sessions will go before you commit, and the folder on disk is still never touched. If you were using Remove to stop scoping file tools at a folder while keeping its history, use Delete N Sessions on the sidebar header for the opposite trade, or simply leave the project in place. Also fixed: Reduce Motion now suppresses the animations it should; Increase Contrast gives primary buttons their solid accent fill with a legible label in dark appearance; narrowing the Settings window no longer truncates tab titles; the tab strip and segmented pickers show a focus ring when tabbed to; the composer's mode menu names a degraded mode in words; and the inspector's provider trace no longer clips at its narrowest width.

Modes can now choose their models, the inspector shows you where the panel actually disagreed, and a sidebar that rearranged itself while you read it no longer does. Highlights: - Modes pick their own models: each provider on a mode's panel now has a Models row — chips you add and remove one at a time, with a searchable picker grouped by what's already in the mode, what you've enabled in Settings, and everything else the provider serves. A mode's picks replace your defaults rather than narrowing them, so a mode can send a model you never enabled globally; choosing the mode is the choice. Pin only the primary's own model and that provider sends the synthesis call and nothing else. #quick uses this same path now — its small-model picks are visible and editable instead of hardcoded. - The editor tells you what will run: a "This mode runs" card names the synthesis model, the peers that will actually dispatch, and the round setting — resolved the way a real turn resolves it, so muted providers, unavailable models, and the primary's own model are all accounted for. - Roles and Debate: any panel member can be a Critic, arguing the strongest case against the likely consensus while still answering in full. Debate adds one round in which every peer sees the others' first drafts and revises before synthesis — about 2× the fan-out cost, stated in the editor. Critics and round 2 are labeled in the inspector. - Where the panel split: the agreement bar now measures agreement, not attendance — each part of the answer classified from its citations as uncontested, split, or solo. A card per split point shows the claim, the providers on each side, and each side's closest passage. - Diff vs answer, and Use as answer: compare any provider's full response beside the synthesized answer, aligned sentence by sentence, and adopt it as the displayed answer with one click. The synthesized answer is kept and can be restored; exports carry an attribution line. - The sidebar holds still: clicking a session in another project no longer jumps that project's section to the top under your cursor. Section order is now purely a matter of recent activity; the active project keeps its accent dot wherever it sits. - Failed turns say why: a turn that fails now shows the provider's own error text in the bubble instead of an empty one, keeps it across relaunch, and reads it aloud with VoiceOver. Also fixed: the Modes editor's Primary dropdowns went blank for a provider you had disconnected or a model the registry does not list, and could silently drop #quick's primary model on edit-open when Anthropic was connected — both now show what the mode holds and keep it. Mode rows name providers by their connection name rather than an internal handle. A mode left on "Provider default" no longer sends the literal string default as a model ID. Note: the local store migrates from schema version 1 to 3 on first launch of this build. Sessions are preserved in place; the pre-migration backup remains the rollback path. There is no downgrade to 1.3.0.

Help that opens inside the app, a Providers page rebuilt around what actually happens when you hit Send, and MCP tools that work the way their settings say they do. Highlights: - Help without leaving the app: Help → Polycode Documentation (or ⌘?) now opens a native Help window — sidebar, search, and every guide and reference article rendered in place instead of bouncing you to a browser. Articles are downloaded as a whole set and cached, so Help reads offline and opening an article never issues a request that says which one you opened. The backup / restore / migrate article also ships inside the app itself, readable on a fresh install that has never been online — which is exactly when you might need it. "Learn more" links in Settings now land on the right section of the right article, in-app. - Providers, consensus-first: the page now leads with what your next prompt will do — the primary model that writes the final answer, a strip of every model answering in parallel, and a warning when the primary would run alone. Each connection row gains a Mute switch (keep the provider connected, pause its models in parallel), an estimated cost per exchange with its assumptions one click away, and a Manage models drawer replacing the old drill-in page. - MCP tools in every session: connected MCP servers now offer their tools even when no project folder is active — previously a no-project session silently sent none, so models truthfully claimed they had no tools despite a healthy connection. Server rows can be renamed and re-scoped in place without dropping the connection. - Read-only means what it says now: the MCP read-only switch governs which models are offered a server's tools — it never skips approvals, a per-tool Deny now genuinely blocks tools on read-only servers, and models answering in parallel no longer receive tools from servers that could write. Each server row shows who receives its tools at a glance. - One tool budget per answer: a multi-model answer now checks in at the configured number of tool calls in total instead of per model, and Stop reliably ends every model's tool use. Repeated identical tool calls across models are counted — and stored — once, and the inspector labels a repeat served from cache as "cached" instead of a meaningless 0ms. - Regenerate streams into an empty bubble: while a regenerated answer streamed, the old answer used to sit above it — and regenerating again stacked a third copy under both. It was a mid-stream display artifact only (the stored answer was always the new one), and it's gone. - Titlebar polish: the session pill centers on the window and holds still when the sidebar moves, a stray hairline next to it is gone, and the Settings window no longer stacks two title bars. Corrected in this release: the privacy policy said the app makes no calls to Polycode's own infrastructure, and that polycode.me links open only in your browser. The Help window changes both — the app now fetches its documentation from polycode.me — so the policy now says exactly what that request carries: no conversation content, no identifier, the whole documentation set rather than the article you opened, cached locally, removed when you delete your data. As always: local-first and private. Conversations are stored on your Mac, credentials stay in the Keychain, and every prompt goes directly to your configured providers — never through a Polycode server.

A performance fix for long conversations, a legibility pass over every muted string in the app, and honest wording in the setup wizard. Highlights: - No more beach ball on long chats: opening a session containing a very large answer could freeze the window for seconds. The markdown parse behind it was quadratic in message length — a 21 KB answer cost about 20 seconds. The same message now mounts in 0.64 s, a 31× improvement, with identical output. Everyday session switching was already fast and still is; the stall was specific to unusually large messages. - Muted text is legible: the faintest text step — sidebar timestamps, message counts, the composer placeholder, the token estimate, provider descriptions — failed the WCAG AA contrast bar on every surface in both light and dark. It now clears it everywhere, along with warning text and the accent used for inline citation links. Every small grey string in the app changes tone as a result. - Fewer interruptions on tool-heavy answers: Polycode used to ask "Continue?" once per model every 25 tool calls — three models could mean six prompts for one question. It now asks at most once per message, and the budget is configurable in Settings → Tools. - Closing a window stops its answer: closing a window mid-stream used to leave the request running against your providers with no way to reach or stop it. Closing or deleting now ends the work, and the exchange is marked cancelled instead of sitting blank forever. Replacing an answer marks it replaced, and an interrupted Regenerate no longer leaves an empty bubble. - Regenerate can read your files again: regenerating inside a project shipped with no tools at all, so models were handed a map of your repository and no way to open anything in it. Tools now work on regenerate exactly as they do on send. - Model pickers offer only the models you've enabled — with OpenRouter connected, that menu used to be 336 rows long for 3 enabled models. Corrected in this release: the setup wizard said your data "stays on your Mac". What stays on your Mac is the stored record — the content of every prompt goes to each provider you configure, every time you submit. The wizard now says "Stored on your Mac, sent to your providers", which is what the rest of the app has always said. As always: local-first and private. Conversations are stored on your Mac, credentials stay in the Keychain, and every request goes directly to your configured providers — never through a Polycode server.

Polycode's biggest update yet: a new streaming markdown engine, desktop widgets, per-message model targeting, and a long list of reliability fixes. Highlights: - Live styled markdown: headings, bold, code blocks, and tables now render fully styled as the answer streams in — no more plain text until the end. A rebuilt typography system (clearer heading hierarchy, roomier rhythm, proper code-block cards in light and dark) and a character-precise cursor come with it. Long transcripts scroll noticeably smoother. As always: local-first and private. Conversations stay on your Mac, credentials stay in the Keychain, and requests go directly to your providers.