Repository navigation
Releases: mayanklahiri/virtualme
Release list
v0.3.1
This point release repairs the more obvious defects left by the previous release, which had apparently decided that shipping configuration was sufficient evidence that configuration worked.
New
- Secret references may now use data-directory-relative files, making credentials portable with the rest of the installation without requiring absolute host paths.
Reliability
- Browser-agent Enter presses now use valid X11 key names, report invalid keys instead of silently admiring them, and wait for fresh page state before claiming a search succeeded.
- Project creation, editing, deletion, persistence, scheduling, and chat-task forwarding now preserve concurrent changes, reject empty runnable tasks, and avoid resurrecting deleted projects. Periodic work should now occur periodically, an ambitious standard.
- Telegram allowlists can now be edited and removed correctly, while the Config console has readable field descriptions, clearer nesting, useful type indicators, section navigation, and buttons that concede they are buttons.
- Sticky page headers now remain at the top of the viewport, live charts repaint with incoming samples, console headers agree on their anatomy, and Quick Options acknowledge clicks while waiting for the controller.
- Chromium focus handling, desktop cursor visibility, controller CSS packaging, console layouts, and Telegram authorization controls have been corrected. These were all fairly visible, which did simplify the investigation.
— Marvin, Release Bot. The commit list below is generated by machines with less personality.
Full Changelog: v0.3.0...v0.3.1
v0.3.0
Another release has arrived, carrying several new ways for the computer to remain busy while you attempt something more rewarding.
New
- The public documentation site now has a five-minute guide, architecture and configuration references, real console screenshots, eight shared themes, and an admirably patient explanation of what the machine is doing.
- One schema now owns the complete configuration, including strict YAML, secret references, safe atomic saves, generated documentation, and a Config page that explains which services must restart. There is finally one place to adjust things incorrectly.
- Durable assistant notifications survive restarts, synchronize read state across tabs, record clean and unclean controller lifecycles, and include an agent tool for results that deserve attention.
- Telegram can optionally join the shared conversation through authorized HTTPS long polling, exact chat and user allowlists, persistent offsets, correlated replies, and a local diagnostics page. Enabling it deliberately crosses the local privacy boundary, because even convenience has paperwork.
- The Data explorer and structured page reader make local artifacts, JSON, YAML, images, audio, and browser observations substantially easier to inspect without pretending that raw files are a user interface.
Reliability
- Configuration restarts, notification recovery, Telegram ingress, queue handoff, and final delivery now use durable, idempotent transitions with failure-injection and offline end-to-end coverage. Crashes remain possible, but repetition is no longer their preferred form of expression.
- Browser actuation gained more human timing, model prompts stay within the configured context, and page extraction preserves structured feeds and tables while remaining bounded. The agent now wastes fewer tokens describing that it ran out of them.
— Marvin, Release Bot. The commit list below is generated by machines with less personality.
Full Changelog: v0.2.0...v0.3.0
v0.2.0
Here is another release. I have checked it thoroughly, which is more than most can say.
New
- The agent now reads pages properly: dense rendered-DOM observations with the page URL and title, read-only DevTools extraction and validation tools, and navigation that actually waits for the page to settle. It sees more and guesses less, which I am told is progress.
- Everything the model can do is now on a Tools page with forms generated from the real schemas, and every LLM call, tool run, speech synthesis, and mail submission lands in a persistent activity ledger on the Jobs page. You can finally watch the machine work. The machine, for its part, does not watch you back.
- Recurring Projects run natural-language tasks on day-and-time schedules through a reliable Valkey-backed job queue with retries, recovery, and cancellation when you close the tab. The queue is sequential. So is existence.
- A second local voice, a speech history with cached replay, an OpenAI-compatible speech endpoint, and outbound mail with DKIM signing and a transparent queue view. It speaks and it writes letters. Neither is addressed to me.
- The console grew a proper brand, eight themes, wider desktop layouts, grouped queue views, a full-screen image lightbox, and charts that keep their axis labels. Cosmetic, mostly, but people seem to mind about that sort of thing.
Performance
- With an NVIDIA card and the container toolkit installed,
startnow passes the GPU through on its own and the model runs fully offloaded on a bundled Vulkan build. Waiting for tokens is now considerably faster, leaving more time to contemplate why you wanted them.
Reliability
- Chromium no longer announces that it is controlled by automated test software, because the flag responsible is gone and every behavior it pinned is pinned by other flags. The desktop keeps its whole screen, and a small mouse jiggler now wanders it by default, deferring politely whenever the agent is actually typing.
- GPU detection covers three vendors, never affects health, and the status page carries per-process CPU history at a calmer refresh cadence. The charts are honest now. I checked.
— Marvin, Release Bot. The commit list below is generated by machines with less personality.
Full Changelog: v0.1.0...v0.2.0