Prestige IDE 0.2.0 · Preview, scheduled runs, notebooks

Five new features, all running on your own models.

Preview

• A Preview tab shows the app you're building, live from its dev server. Start dev server guesses the command (npm run dev and friends, Django, a static page), runs it in a terminal and picks up the address it prints.
• The agent can look too: its new preview_page tool screenshots the page (headless Edge or Chrome) and reads its console errors, so a vision model like Qwen3.6 35B can check its own UI changes. Local addresses only.

Scheduled runs

• Agent jobs on a timetable while the app is open, daily at a time or every few hours, like Prestige's Missions: nightly tests, a dependency check, a morning summary.
• Nobody is there to approve anything, so a run declines what would ask and reports what it would have done. Your settings-file allow rules still apply, e.g. run_command(npm test*).
• Each result is saved as a session, with a notification. Open them with the clock button or /schedule.

Notebooks

• .ipynb files read as numbered cells with their outputs, and notebook_edit replaces, inserts or deletes a single cell, keeping the file valid. Edits show a diff first and work with Changes and rewind.

Output styles

• /style picks how replies read: Default, Terse, Explanatory, or Learning (it leaves small TODO(you) pieces for you to write). Add your own in .prestige/output-styles/<name>.md.

Works with Prestige

• When the models aren't running, Start (in the status line) brings up the Workstation with its own start-all.ps1.
• A matching Prestige update opens Prestige IDE from its top bar and keeps the stack running while the IDE is open.

Includes the 0.1.1 security fixes. Installed copies update themselves.

Prestige IDE 0.1.1 · Security fixes

Security and reliability fixes, found by new tests and a model bake-off.

Security

• Reading files outside the project now asks first, and fetching a web page asks once per site. A repo with hidden instructions could otherwise have the model read files elsewhere on your PC and send them out.
• Commands with a hidden second line, or with > / < redirection, now always ask, even after you've approved their start ("npm test").
• An approved command prefix only matches whole words ("npm test" no longer covers "npm testx").
• Deny rules in settings files also catch commands chained inside others ("npm test; git push").

Fixes

• Switching between llama.cpp models no longer hangs when the previous model has gone idle.
• Tool calls some models write as plain text (a tool name, then <path>…</path>) are now understood.
• Two quick saves of the same settings file can no longer fail.

Changes

• Thinking is off by default for new sessions: in the bake-off it scored as well and ran about 3× faster. The toggle is still on the message box.

Prestige IDE 0.1.0

The first release of Prestige IDE: an agentic coding IDE that runs on your own local models through the Custom AI Workstation.

• An agent that reads, searches, edits and runs PowerShell, with permission modes (ask, accept edits, plan, bypass) and approvals with diffs
• Changes pane with revert, checkpoints and rewind, Monaco editor, real terminals
• MCP tools from the Workstation's tool server, read-only subagents, pictures for vision models, thinking on/off
• Git: branch chip, switching, worktrees, pull/push, /commit, /pr
• Settings files with allow/deny rules and hooks
• Updates itself from these releases

Install: download Prestige.IDE_0.1.0_x64-setup.exe below and run it (installs for your user, no admin prompt). Needs the Workstation running for models.