Skip to content

Praxi-Bot 0.1.0 alpha 3

Pre-release
Pre-release

Choose a tag to compare

@emreocell emreocell released this 20 Nov 06:49
2fac862

Third public pre-release of Praxi-Bot. Alpha: expect bugs. Built from 2fac862 (main).

  • A monthly spending limit with a pace forecast, and edit/regenerate for the last chat message.
  • A calmer first run: local model servers are detected, a fresh install should no longer show "service not responding", and the app can start its local service when you send a message while it is offline.
  • More security hardening around sensitive fields, attachments and the new "your own computer" tools.

"Verified" below means covered by automated tests and CI. Most items have not been tried by hand on a real machine; see "Known limitations".

Highlights

Desktop app

  • Edit your last message and resend it, or regenerate the last answer. Earlier versions are kept under an "Earlier version" fold and are not sent to the model. Side effects of the earlier run are not undone, so regenerating after a run that used write tools needs a second click.
  • Praxi ghost is the default main bot; a "Make main" action; new Praxi app icons.
  • The left rail lists only bots; chats live in the bot panel.
  • Unread dot per bot and a "New messages" divider in the open chat.
  • Keyboard shortcuts dialog (open with ? or the command palette); hover action row on messages with visible copy feedback.
  • A "What's new" card shows these release notes once after an update.
  • If you send a message while the local service is offline, the app starts it and sends once. For a paired server it tells you the server is offline and keeps your message.
  • Microphone errors are now specific, and read-aloud stops when you start dictating.
  • Visual and accessibility fixes: Send stays visible in a narrow chat column, inline errors are announced to screen readers, table and dialog alignment.

Security

  • Sensitive-field checks (password, OTP, secret-like names) run again right before a browser fill or key press, not only at approval time.
  • Attachment text is labelled untrusted and a chat with a document attached starts "untrusted-exposed", so egress and command approvals tighten as they do after a web read. Office files are refused above entry-count and size limits, extraction has a 20 s deadline, and a visible note shows when a file was only partly read.
  • pc_* tools (a bot on a paired always-on server acting on your computer: files, shell, desktop, browser) are opt-in per bot, only usable from human turns typed in the app or on the host's admin desktop, and each action needs an approval card. Per-family owner switches on the PC are all off by default.
  • Bots can create, update and delete other bots from chat (bot_list/create/update/delete) but never with more permission, tools or providers than the calling bot has, and never with a full preset.
  • Pause/recover: the host admin can clear a pause without losing input rights; Stop also aborts runs parked on a pause.
  • Prompt-injection eval pack with an "unsafe actions" gate runs nightly in CI (development tooling, not a user feature).

Host & server

  • Optional monthly spending limit (USD, off by default): calls are refused before they are sent if month-to-date cost plus a conservative estimate would exceed it. Details under "Spending limit" below.
  • Startup health is reported honestly in phases and shown in the host chip; the API is served during leadership startup; channels are stopped first on shutdown.
  • Linux host setup is more explicit: non-interactive package operations, logged stop/start during repair and upgrade, host mode detected from the effective systemd command.
  • Optional model tiering with durable escalation records (off unless configured).
  • Workflows: save a successful run as bot-scoped, model-free automation and replay it by explicit human action. Secrets are kept out of chat and run history; replay still asks per step; paired phones cannot read or run workflows.

Live screen

  • Hosts default to 1920x1080 and the watch stream is sharper. Existing servers need a host upgrade (or a changed Xvfb screen size) to pick up the new default.
  • bot proxy and a live-link measurement script for diagnosing slow relayed connections.

Chat

  • Chat management follow-ups: no ghost chat after delete, rollback per field on a failed rename or pin.
  • Queued runs count as live in the desktop, so the status line matches what is running.

Install and CLI

  • First-run detects local model servers (Ollama, LM Studio, llama.cpp on 127.0.0.1 only) and offers one-click cards with the number of installed models; clearer messages when Ollama is down or has no models (bot providers detect-local prints the same as JSON).
  • Fresh-install fix (#362): the desktop now treats the service's "setup required" health reply as alive, so first run should open the setup screen instead of "service is running but not responding".
  • Clearer CLI messages for first run, setup-required and help/errors.

Docs

  • Redesigned README; guide pages for day-to-day host operation and recent desktop, host and eval features.
  • node scripts/check-prereqs.mjs checks contributor prerequisites (Rust, Node, pnpm, C toolchain, Linux packages).

Thank you

Outside contributions are welcome; see the contributing guide.

Spending limit

  • Counts only models with a known price. Calls to unpriced models are neither counted nor blocked, and the Usage page shows how many calls were not counted.
  • The estimate is conservative, so a run can be stopped slightly before the true spend reaches the limit. The month is the host's local calendar month.
  • An unreadable limit file fails closed (paid calls blocked until you set the limit again). The 80% warning is stored, so a restart does not repeat it. Only the owner can change the limit, and the controls are hidden when a remote computer is selected.

Known limitations

  • This is an alpha. Expect bugs and rough edges.
  • The Windows installer is unsigned, so Windows SmartScreen may warn.
  • The fresh-install "service not responding" fix (#362) is covered by automated tests, but has not yet been confirmed on a fresh machine with the full app.
  • Voice is limited to push-to-talk dictation and optional read-aloud; hands-free voice and the mobile companion app are not included.
  • Linux server packages are attached by CI after the tag, so they may appear a little later than the Windows installer.
  • The monthly spending limit applies to priced models only.
  • The pc_* "your own computer" tools were verified with fakes and tests, not against a real Chrome or a real paired server. If a browser worker type refuses the interception, it stays paused (fails closed).
  • Most desktop UI changes were checked by tests and token checks, not by eye in a running app on every theme. The same goes for the host changes: a live service upgrade and a live systemd host are unverified.
  • Live-view sharpness and the 1080p default were not judged on a real host and link; relayed connections can still be slow.
  • Under the Full autonomy mode, after untrusted content a delete or write that is neither network egress nor command execution may run without asking. This is a known gap in the injection eval pack and is flagged for a decision.
  • Name-based sensitive-field detection looks at DOM attributes, not the computed accessible name.
  • Composio apps, server connections and always-on flows have not been exercised against all real services.

Upgrading

  • Back up your data folder first: %APPDATA%\praxi\praxi-bot\data.
  • Before a schema upgrade the app copies bot.db into a backups folder inside the data folder (the newest three are kept). This release adds one migration (superseded messages for edit/regenerate).
  • bot backup and bot restore are available in the CLI.

Install

Windows: download Praxi-0.1.0-alpha.3-2fac862c-setup.exe from the assets below and run it. The installer is unsigned, so Windows SmartScreen may warn; choose "More info", then "Run anyway" only if you downloaded it from this release page and the checksum below matches. On first run the app asks you to connect a provider (an API key, or a local model server such as Ollama).

Linux server: after CI attaches the packages, run install-linux.sh from the release assets (read it before running it).

Verify

SHA-256 of the Windows installer (PowerShell: Get-FileHash <file> -Algorithm SHA256):
eb6f119e8a7add5b868ec0d96db516b525d6b7b7faa943ad1149f91cb1164a64 Praxi-0.1.0-alpha.3-2fac862c-setup.exe

Full changelog

v0.1.0-alpha.2...v0.1.0-alpha.3