Agent conversations you can come back to
You ask Claude to restructure a draft, get halfway through the discussion, close the sidebar to check something, and the conversation is gone. Or it is still there, but only in this window, only until the next restart, and only with the agent you started it with.
Ritemark v1.10.0 turns Agent Chat from a row of temporary sidebar tabs into a conversation system you can trust. Conversations belong to the project, survive a restart, and can continue with another agent without you reconstructing the discussion from memory. Today's v1.10.1 is a maintenance release on top of it, mostly about making Codex behave the way you already expected. Both are covered here.
Conversations that belong to the project
Every non-empty conversation is now saved locally under its project. Closing the sidebar, or closing Ritemark, no longer turns a conversation into a mystery.
The conversation rail keeps your working and recent chats close. Up to five Pinned conversations stay put, and All conversations opens the complete list for the project. Opening a chat does not reshuffle your shortcuts: recents move only when there is real conversation activity, and Pin, Unpin and Delete remain deliberate actions.
The project's conversations, saved locally. Pinned ones stay where you put them.
Titles become useful on their own: a new conversation starts with a shortened version of your first prompt, gets a concise title after the first answer, and can always be renamed. Deleting one starts with a confirmation that fits the narrow Conversations pane, and Undo comes through standard VS Code notifications, so it stays valid even after the panel closes.
Conversation history stays local to this Ritemark installation and profile. Cloud sync, shared conversations and account portability are not part of this release.
Continue with another agent, truthfully
Reopening a conversation does not silently start an agent. On your next Send, Ritemark first tries the selected agent's compatible saved session; Claude, Codex and OpenCode can all resume their native sessions. If that session is missing, rejected, incompatible or unsafe after an interrupted turn, Ritemark starts a fresh one with a bounded text transcript, and says so.
That transcript carries your requests and the completed answers. It does not replay tool state, approvals, partial output, plans, hidden instructions, provider history or attachment contents. A request that never got an answer stays visible to the next agent as labelled context; it is not quietly sent again as a second command.
Switching agents is now just a choice. Pick another agent and keep writing: the selection applies immediately, your draft stays intact, and nothing runs until you press Send. There is no confirmation dialog. When Ritemark falls back to the transcript, one quiet line between the turns says which agent is continuing and that previous messages were included as context. Another provider's private session identifier is never transferred, and late output from the previous agent cannot overwrite the handoff.
One conversation, two agents. The boundary is a single line, not a dialog.
"Previous messages were included" means the transcript, not the previous agent's complete private memory. Native continuation is intentionally version- and configuration-specific.
Choose how much thinking a message gets
Supported Claude and Codex models now have an Effort control right in the composer. Leave it on Auto to use the model's default, or drag the Faster → More thorough scale when a task deserves a deliberately chosen effort level. Only choices the selected model actually supports are shown.
Effort is kept per conversation and per agent, and every sent message keeps its own snapshot.
The preference stays with that conversation and agent, and every accepted or queued message keeps its own snapshot, so changing the control later cannot alter work already waiting to run. OpenCode shows the same control only when its live session advertises compatible thought levels; selecting a conversation never starts a runtime just to find out.
Refreshed engines and a sign-in that recovers itself
v1.10.0 refreshes all three bundled engines: Codex 0.153.0, Claude Code 2.1.239 and OpenCode 1.18.21, with their SDKs pinned to match (Claude Agent SDK 0.3.239, ACP SDK 1.4.0), so an update can no longer quietly pair an old client with a new engine. The refreshed Codex accepts the current GPT-5.6 model catalog instead of failing with a stale-client error, and its file-tools host ships beside the app-server on Apple Silicon, Intel macOS and Windows. Agent Chat also opens with your real selected model in every window instead of a permanent Model placeholder.
If Claude's saved sign-in expires, the conversation no longer shows a raw OAuth error. It explains that Claude needs you to sign in again and offers Sign in to Claude in the same card, which waits for the browser, confirms success with OK, and stays actionable with Try again if you cancel. If the conversation uses an Anthropic API key instead, the card points to Update API key in AI Settings. And signing out of one provider no longer hides Agent Chat or the model selector when another is connected: you get an explicit Use Codex, Use Claude or Use OpenCode action, never a silent switch.
Windows installation you no longer have to argue with
Every executable in the Windows package is now found by inspecting file content rather than trusting a file extension, Ritemark-owned payloads are signed as Productory Services OÜ, and Inno signs its own setup loader and uninstaller during compilation. The build refuses to upload an installer if credentials, any signature, the standard-user installation, the installed payload or the uninstall check fails.
This installer was verified on a clean Windows 11 machine with Smart App Control turned on, without disabling SmartScreen, Smart App Control or Defender. Its SHA-256 is published with the release, so you can confirm that the file you downloaded is the file we tested. A Microsoft Store listing for the same installer is in preparation and becomes the recommended Windows channel once certified; the direct download stays available as the identical signed file.
Editing that stays out of your way
When Codex, Claude, OpenCode, a formatter or another tool changes an open Markdown or CSV file, the revision now appears in the visible editor without reopening the file. Your own unsaved typing stays quiet while it legitimately leads the version on disk. If both your version and the disk changed from the same starting point, Ritemark keeps both and offers Compare changes, Keep my version and Use disk version. The old always-on file-changed control and its automatic ten-second reload are gone; no background timer resolves a conflict for you.
Task lists keep their checkbox structure and checked state through edit, save, close and reopen, including Markdown with blank lines between items, nested checklists and adjacent ordinary bullets. A blank document accepts # on the first line. The desktop Help menu now holds only Support, View License and Advanced → Toggle Developer Tools. Transcribe Insights can be generated in any language or dialect, and Create insights document writes a separately named Markdown snapshot without touching the existing transcript.
v1.10.1: Codex behaves the way you expected
Today's maintenance release reaches everyone through the normal in-app update, on macOS (Apple Silicon and Intel) and Windows.
The Codex setup pane no longer loops. With no Claude account signed in, choosing Use Codex brought up a setup pane that vanished and came back every couple of seconds, so clicking the button became a race. Ritemark was checking Codex's sign-in status every two seconds while the pane was open, and each check briefly reset it. The pane now stays put. This affected macOS and Windows equally.
Codex can work with your files again on Windows. Asking Codex to read or change a file failed on every turn with orchestrator_helper_launch_failed: ... program not found, while chat itself kept working, in the default and recommended workspace-write sandbox setting. Codex runs its Windows sandbox through two small helper programs that we simply were not shipping. Both are now bundled and verified at build time, and a missing helper fails the build. If you switched the sandbox to full access as a workaround, you can set it back to workspace-write. macOS was never affected; there the sandbox is built into the operating system.
Typing, checklists and undo. Your text stays as you typed it. A moment after each edit the app used to hand the document back to the editor, which sometimes treated that as an outside change: a brief "clean-up" while typing, an Enter inside a list that did nothing, a new checklist item turning into [ ] text. Documents with front matter had this since 1.10.0. Cmd+Z (Ctrl+Z on Windows) undoes exactly your last action again and redo no longer duplicates list items; Undo from the Edit menu still uses VS Code's document history. Checklist items sit on one line with their checkbox, and an empty checklist item survives saving and reopening.
Two smaller editor fixes. Saving a document that ends in a list no longer drops the trailing newline (#254), and Home lists the documents you actually opened recently (#194).
Download
| Platform | File |
|---|---|
| macOS Apple Silicon | Ritemark-arm64.dmg |
| macOS Intel | Ritemark-x64.dmg |
| Windows x64 | Ritemark-Setup.exe |
If you already run Ritemark, v1.10.1 arrives through the normal in-app update. New users: download and start a conversation you can come back to.
FAQ
What is new in Ritemark v1.10?
Agent conversations are saved locally with the project and survive closing the sidebar or restarting Ritemark. You can pin up to five, rename any of them, and continue a conversation with another agent. The composer gets a thinking-effort control, the bundled engines are refreshed (Codex 0.153.0, Claude Code 2.1.239, OpenCode 1.18.21), and the Windows installer is signed and verified with Smart App Control on.
Where are my agent conversations stored?
Locally, under the project, in this Ritemark installation and profile. Cloud sync, shared conversations and account portability are not part of this release.
What happens when I switch from Claude to Codex in the middle of a conversation?
The switch applies immediately and your draft stays. On the next Send, Ritemark first tries the agent's compatible saved session. If none is usable, it starts a fresh session with a bounded text transcript of your requests and completed answers, and one line in the conversation says so. Tool state, approvals, partial output, plans and attachment contents are not carried over.
Does the Windows installer still trigger SmartScreen or Smart App Control?
The v1.10.0 installer was verified on a clean Windows 11 machine with Smart App Control on, without disabling SmartScreen, Smart App Control or Defender. Ritemark-owned executables are signed as Productory Services OÜ, and the SHA-256 is published with the release.
Codex says it cannot read or change files on Windows. What should I do?
Update to v1.10.1. The two helper programs Codex needs for its Windows sandbox are now bundled. If you switched the sandbox to full access as a workaround, you can set it back to workspace-write.
Why did Cmd+Z sometimes do nothing or undo the wrong thing?
Two undo histories were answering the same keystroke. Since v1.10.1, undo and redo stay with the editor while a Ritemark document is active. Undo from the Edit menu still goes through VS Code's document undo.
Where can I find previous release notes?
All Ritemark updates are on the Ritemark blog. v1.9.0 added Transcribe, v1.8.6 made plan mode hold, and v1.8.5 introduced parallel agent chats.