# Osaurus Beta 0.1.6

**URL:** <https://community.osaurus.ai/t/osaurus-beta-0-1-6/53>\
**Category:** Releases\
**Tags:** windows\
**Created:** [October 15, 2026, 12:25am UTC](https://community.osaurus.ai/t/osaurus-beta-0-1-6/53 "2026-10-15T00:25:05Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![sonofnun](https://avatars.discourse-cdn.com/v4/letter/s/4bbf92/32.png) [@sonofnun](https://community.osaurus.ai/u/sonofnun)\
**Post date:** [October 15, 2026, 12:25am UTC](https://community.osaurus.ai/t/osaurus-beta-0-1-6/53/1 "2026-10-15T00:25:05Z")

</div>

## Osaurus for Windows 0.1.6

### New

**Agent browser**

- A new `browser_use` tool hands a goal to a helper agent that drives a real browser built into the app. The helper reads pages through their accessibility tree and uses real mouse and keyboard input.
- A browser pane docks beside the chat. You can watch an agent cursor move to each element, see a ripple on clicks, and read a short line describing what it’s doing.
- Before a consequential action (submitting, typing on a site you didn’t name, and so on), the page dims around the element and you’re asked to approve it.
- In the chat, a browser run is one card: a live picture of the page, the steps in plain words, the sites it visited, and Watch / Stop buttons.
- **Saved logins** are kept in an encrypted vault, and you choose which agents can use each one. The model never sees a password, and passwords are removed from everything it reads.
- **Hand-off:** you can take over when the agent is stuck, with a countdown to the agent’s deadline. You can also open the browser ahead of time to sign an agent in to sites.
- Sites you’ve approved actions on are remembered, and you can see and revoke them under Behavior.
- Every agent’s browser profile comes with a pinned, hash-checked uBlock Origin Lite and Strict tracking prevention.
- Pages inside frames from other sites are handled, and each action is judged by the site of the element it touches.

**File history: every change an agent makes can be undone**

- Every tool call that can change files is recorded, and the saved copies of the previous versions are encrypted. That includes shell commands, plugins and MCP servers, which are compared before and after.
- **Changes pane:**
  - A timeline grouped by message, with rewind points. A rewind can also take the chat back to that point.
  - A Files tab, with diffs shown in place: syntax colours, changed words highlighted, and collapsed context you can open.
  - Image comparison (Swipe / Onion skin), changed cells in workbooks, and paragraph counts in documents.
  - Keyboard driven, with Undo / Redo.

- **Conflict handling:** if you edited a file after the agent did, you choose per file between Keep it as it is, Keep both, or Revert anyway. Every revert can itself be undone.
- **A file’s History** lists every version, who made it and which message asked for it, and any version can be restored.
- **Review mode** widens the pane over the chat. You can step through “This turn” or “This chat”, revert one file at a time, and send line comments back to the agent.
- **In the chat:** each reply gets one receipt with Review and **Undo this turn**. Tool headers show change counts, and the tab shows a file badge.
- The agent is told when you revert something it changed.

**Privacy Filter: reviews, an on-device model, and marks in the chat**

- The whole request is now scanned, not just the latest message: earlier turns, the system prompt and tool arguments are included. A send is blocked if anything that should have been replaced survives.
- Each sensitive value keeps the same placeholder number for the whole conversation, and the mapping is saved with the conversation.
- **Review before sending:** when new sensitive values are about to leave the PC, a card asks first. You can untick a value to send it as typed, or cancel and get your draft back in the composer.
- **Marks in the chat:**
  - A shield on your message shows what left as a placeholder.
  - Protected values are marked in your message, and values restored in the reply are marked where they appear.
  - A “protected automatically” receipt opens a flyout naming the provider and what it received.

- An optional **on-device detection model** (Rampart, running on ONNX Runtime) is downloaded in the background and checked against a pinned hash.
- **Control Center → Privacy filter:** an overview, rules with a dry-run tester, per-provider switches, model install, and a rule editor.
- Text that Osaurus or a tool catalog wrote is no longer scrubbed. This fixes false matches such as public URLs and “n8n” being treated as addresses.
- The provider is now told what a placeholder is, so it stops refusing or asking for the “real” value.

**Documents and file tools (now match macOS)**

- `file_write` creates **.xlsx, .docx, .pdf and .pptx** files, chosen by file extension.
- `file_edit` edits Office files and PDFs in place, and only commits a result that re-reads correctly.
- New `file_copy` tool.
- PDF reading:
  - page markers and `pages` ranges;
  - a `mode: "structure"` option;
  - line breaks and gaps between words are kept, so form answers no longer run together with their questions.

- `file_search` can page through results with `offset`.
- Text edits accept several edits in one call (`edits[]`).

**Automation runs are visible**

- Every watcher, schedule and self-booked run now opens as its own read-only tab under its agent, so you can watch it as it runs.
- Each automation run now starts as a fresh session: it no longer sees earlier runs that its tab doesn’t show.

**Memory page, rebuilt**

- One column: an agent picker, search, your About-me lines, the agent’s memories grouped by kind, and Forgotten at the end.
- Memories are edited in place: hover for Edit and Delete, Enter opens a memory, F2 edits it.
- Memories can now be **deleted permanently**. Delete asks whether the agent may learn this again.
- “Add a memory” composer.
- “Read now” runs in the background, shows progress, and says why waiting messages couldn’t be read.

**Agent editor**

- New “Working folder for remote runs”: an authenticated remote run of the agent works in a folder you pick. Only `file_write` and `file_edit` are opened, and they always ask first. I couldn’t tell whether this ships in the withheld build, because its main caller is the phone, so check before announcing it.

### Improvements

- **MCP tools cost far fewer tokens.** In Auto selection, MCP and plugin tool definitions are only loaded when the agent needs them. One 129-tool server had been adding 92.8k tokens to every step.
- **ChatGPT/Codex models** come from the live catalog: the model list, each model’s reasoning levels and display names, the lighter request format (Responses Lite) where the catalog says so, and images removed for models that don’t accept them.
- **Claude Sonnet 5.5 and Haiku 5.5:**
  - Every Sonnet 5.5 request with thinking turned off used to fail, and now works.
  - Haiku 5.5 is now treated as the newer “adaptive” generation, so thinking on/off and effort behave correctly.

- **GitHub skill import** streams its progress one skill at a time:
  - Large repos no longer report “failed” while the import is still running.
  - Re-running resumes, skipping skills already installed.
  - The dialog shows the outcome for each skill.

- Memory picks its writer model in your Providers-page order.
- `memory_remember` explains how to fix a refused save, so models stop looping on it.
- **Workspace sync:** a member’s connection stays up, it reconnects with backoff, and accounts that aren’t in a workspace don’t hold a connection open.
- Managed **llama.cpp** updated to b11583.

### Fixes

- Tool schemas containing regex features OpenAI’s Responses API can’t compile no longer fail every turn of an agent that carries them.
- Fixed a crash in NVIDIA’s driver during GPU detection on discrete NVIDIA cards (37 crashes seen). That code path no longer runs on those machines.
- About-me saves no longer lose lines that contain an email address or phone number. Removing a line and re-adding it works.
- Settings pages no longer sit 12–26 px lower when a hidden notice bar is on the page.
- Wrapping toolbars no longer drop their last button onto the content below at 125% scaling.
- Pane toggles in the title bar are no longer the loudest control in the window.
- A tool group that recovered after a failed call shows a caution mark, not a red failure mark.
- Delegation:
  - An answer written at the step cap is now delivered.
  - A nudged child’s draft and final reply are no longer shown glued together.

- Screen readers could not reach the chat after an agent had browsed.
- An unanswered browser hand-off now reaches the agent as a timeout instead of “user declined”.