Logos Files

Logos 0.3.0 is a transcript release: a turn's work folds into one activity block with a sticky header, every message shows its footer only while you point at it, and selection copies content instead of chrome. Updating from 0.2.1 is a drop-in replacement - settings.json, logs, session journals and projects in %LOCALAPPDATA%\Logos are kept.

Transcript

• One card per turn. Steps of one response now collapse into a single activity block: read-only exploration (read_files, list_directory, search_file_content, web tools, MCP reads) groups into one run row, everything else into an action run. A thought or a narration line always ends a run, so the model's reasoning stays visible between groups. A finished clean run folds to Ran 6 tools - read_files, write_file +1; a failed or denied step stays open.
• Every row opens with a status badge: a green circle-check for ok, a red circle-cross for error, a hollow accent ring while running, amber ! for denied and a grey dash for stopped. They are painted, not typed, so they follow the theme at any DPI.
• Thoughts are exactly one line. The first line of a thought is measured and clipped with an ellipsis; the full text is one click away. Resizing the window re-clips them.
• A sticky bar follows a long turn. Scrolling a tall block no longer loses its header: the bar pins the block summary (and inside an opened step, that step's own title) and folds what it names on click.
• Step arguments are a disclosure, not content: the Arguments label toggles the raw call, the call itself stays selectable, and it is not repeated when it is already the step title.
• Long output is a preview of 25 lines / 3 500 characters with Show all - N more lines, because a transcript is one rich-text document and cannot nest a scroll area.
• Message footers, on hover only. A user bubble dates itself on its right edge with the send time and a copy button; an answer leads with the model that wrote it (qwen3.7-flash 2 Nov, 14:02) and its copy button, aligned to the left of its text. The strip appears only while the pointer is over that message; the line is always present, so nothing moves when it shows. Pending is the exception and stays visible, because it says the text has not reached the model yet.
• Copy is offered only where there is a conclusion. A finished turn (finish, communicate_to_user, a completed Chat answer) carries the copy button with the plain answer text; a stopped turn and an interim note do not.
• Selection and Copy All touch content, never chrome. Block headers, run rows, step titles, carets, badges, the Arguments / Show all links and the footers cannot be dragged into a selection or a copy - while narration, thoughts, arguments, output, written files, diffs and answers stay fully selectable, markdown links included.

Toolbar

• The model button shows the provider, not a generic bolt. It leads with the active provider's brand glyph (Anthropic, OpenAI, Gemini, OpenRouter, Hugging Face, xAI) tinted for the current theme; a custom endpoint has no bundled glyph and simply gets no icon. The same SVGs are shared with the Providers settings page.

Download

Both builds keep settings, logs, cache and project data in %LOCALAPPDATA%\Logos.

Notes

• Built for Windows 10/11 (x64). On Linux, run from source - see the README.
• Browser mode drives a real Chromium through Playwright, which is not bundled to keep the download small. On first use Logos offers the download (~320 MB) into a per-user folder and shows per-file progress; Settings > Browser shows the same status and button.
• Vision (Computer Use) needs a provider that supports it - OpenAI, Anthropic or Google with a paid key. OpenRouter and custom endpoints are not used for Vision.
• License: MIT.