Changelog

What changed, and when

Newest first. The labels in brackets (D58 and so on) are the project's decision records, where each change is written up with the evidence behind it.

From the trailer

  1. Jimmy repairs itself, and the camera works (D65)1.0.2

    • Jimmy repairs itself. Say "Jimmy, what's wrong with the camera?" and it checks itself and says what it found. Say "you can use the camera, fix it" and it repairs what it can on the spot, without a restart: its own setting, a stuck camera, the right camera. For Windows' privacy switch, which only you can change, it opens the right Settings page.
    • The camera works. The webcam took 16.5 s to open through Windows Media Foundation on the test laptop. It now opens in 0.5 s, with a DirectShow fallback.
    • No more flashing command windows at start. Every helper process starts hidden.
    • A gentler start. Jimmy runs at below-normal priority while it loads. The camera starts 15 s after voice, and search indexing 30 s after.
    • See what Jimmy is doing. A Status dock (the pill's Status button, or "Jimmy, show status") shows:
      • listening, and for how many seconds you can talk without saying "Jimmy";
      • the camera;
      • Jimmy's own CPU and RAM, with a one-minute graph;
      • the update, with download % and Update now;
      • capture, and the AI model.
      The pill itself shows a countdown ring, the camera, a call, and a download badge.
    • A smoother floating UI. Streamed answers no longer redraw the overlay for every word, and hover handling and the "Thinking…" animation got lighter.

  2. Jimmy updates itself (D63–D64)1.0.1

    • Every six hours the installed app checks for a newer release, downloads it to your data drive, checks its SHA-256, and offers to update now or at the next start. A switch in Settings turns it off.
    • A new version is published only after the clean-machine test passes.
    • The camera no longer counts as "in use" by an app that crashed while holding it, or by Jimmy itself.

  3. Jimmy-Setup.exe (D60–D62)1.0.0

    • A setup file for any Windows 10 (22H2) or 11 PC, 64-bit: installed for one user, on any drive, with no admin rights, Python or Node needed.
    • A first-start setup in Jimmy's own design that checks the PC, asks where to keep data and how much of the PC to use (Lite, Balanced, Full), and downloads Jimmy's own Python and speech model.
    • Settings for storage: days to keep, a size cap, clean-ups and start with Windows. Keys typed in the setup are stored encrypted.
    • Every change is built and tested on a fresh GitHub Windows machine: install, setup, self-test, update and uninstall.

  4. Gemini, measured live (D58)

    • A slow model now counts as a down model: one try per model, the first word within 6 seconds, then the next model in the chain (Gemini 3.8 Flash → 3.5 Flash-Lite → NVIDIA).
    • In a "which one?" question, the model's own pick comes first, where the cursor points and where "yes" lands.

  5. Gemini, acting when sure, answer cards (D53–D57)

    • Google Gemini on its free tier as a provider, with NVIDIA as a backup that answers when Gemini can't.
    • Jimmy acts without a "yes" only when it is sure which control you meant and the action can be undone, shown first with a moment to say "stop". Otherwise it asks "which one?".
    • Answers can come as cards: steps, a choice, a confirmation, a list, a table, a checklist or a chart. Plans are played back live as they run.

  6. Lighter and safer (D47–D52)

    • Background work runs in Windows' efficiency mode and waits on battery or a busy CPU; capture slows down when you're idle.
    • The agent sees fewer tools per step, checks every action, and is scored on whole tasks.
    • A policy layer in code decides what the agent may do. A plan's steps run back to back after one "yes".
    • Windows OCR, read-only recall over MCP, and several voice, GPU and storage upgrades, built behind settings that stay off until measured.

  7. Fixes from a live session; OpenAI (D44–D46)

    • OpenAI as a second provider, and keys in a .env file.
    • Each request gets its own decision trace; slow steps fall back to another model; "stop" ends a running task; switching, minimizing and maximizing windows.
    • While paused, the mic listens only for the name and stores nothing.

  8. Hands, eyes, and an agent (D41–D43)

    • Reminders, goals and memories by voice and in a Memory tab; questions about the screen can use a vision model.
    • A virtual cursor that presses and types through UI Automation after a "yes".
    • Jimmy became one agent loop that sees its own state, your lists and the window's controls, runs multi-step tasks on one "yes", and logs every decision. 95 % on 130 real commands, up from 71 %.
    • When the default NVIDIA model reached end of life, Jimmy moved to the next one automatically.

  9. Insights, a privacy curtain, and a lot of polish (D31–D40)

    • The Insights tab and instant answers to "how was my day?".
    • Cards Jimmy writes itself: welcome back, reminders, deadlines, an evening recap. Hands-free navigation after it shows you something.
    • The privacy curtain: the webcam notices when you leave, and capture and listening stop. "Remember my face", encrypted with Windows DPAPI.
    • English and Hindi only; timers; forgetting a span of history; asking without the name by looking at the screen.
    • A performance pass: search from 2.8 s to 0.5 s, with identical results.

  10. Cards, the overlay, recall and voice (D19–D30)

    • The trigger gate: RECALL and FOCUS cards, decided by a local model and closed against blind judges.
    • The overlay: a transparent, click-through pill and cards that never steal focus.
    • Search by keywords and meaning, and the timeline window. All five planned stages done.
    • "Jimmy, …" by voice, with evidence and a spoken answer; it asks back when a question is unclear.

  11. The first real hour (D17–D18)

    • A change gate that sees a one-message chat scroll, leaner stored text, and a visible mic state.
    • A faster chat model, about 1 s to the first word.

  12. Capture and the Jimmy core (D1–D16)

    • Screen text, speech and blurred screenshots into one searchable store, with the exclusion list in place before the first run.
    • The Jimmy core: one model client, memory, and a plugin seam with the capture layer as its first plugin.