OpenAware v0.4.0 — independent agents and video sources

OpenAware v0.4.0 adds independent AI agents and video sources.

• Add up to four agents, each with its own local LM Studio, Ollama or llama.cpp connection, chat, rules and caption memory.
• Assign up to four feeds per agent from a shared library of sixteen. Capture is shared; model requests contain only the agent's assigned sources. Start/Pause affect one agent; Stop all releases every feed.
• Add videos from your PC, direct HTTP(S) video links, or web-player pages such as YouTube, Facebook, TikTok, Instagram and X. Open player gives access to playback controls and site sign-in. Website mode watches the rendered page; platform authentication, protected playback and codec compatibility are not guaranteed. Browser-window capture is the fallback for unsupported players.
• Empty seats retain the glowing Needs agent outline. The compact workspace keeps draggable, resizable panes and side-by-side monitor defaults.
• Operator actions remain restricted to the active verified operator and its assigned desktop, with review and approval for every step. Videos are never native action targets.

Validation results and installer identity are recorded in the implementation handoff. Tests use generated videos, localhost pages and mock model servers. Actual platform accounts, personal feeds and sustained model throughput require separate testing. Live previews update continuously; AI uses bounded image samples.

Windows x64 developer prerelease. No model or NVIDIA VSS server is bundled. The installer is unsigned. Verify the downloaded installer against SHA256SUMS.txt.

OpenAware v0.3.3 — source controls and agent desk

This developer prototype makes source setup visible and adds the agent desk below the feeds.

• + Add source now sits in the Live workspace header and remains in the first monitor/source header after selection. The empty workspace prompt opens the same picker.
• Empty desk chairs have a cyan glowing agent outline labeled Needs agent. A selected or probing model cannot fill a seat. Split desk shows the existing Observer and Operator roles with an explicit Shared model label.
• Desk and Video memory share the lower pane. Switching views retains memory search and filters. Header controls stay reachable at minimum pane sizes.
• Default monitor previews retain their proportions while the board grows vertically to fit the desk. A stable scrollbar gutter prevents resize feedback near the overflow threshold. Custom docking and compact layouts remain available.

Validation: strict TypeScript, 201 unit/integration tests, 14 packaged Windows desktop workflows, independent source review and synthetic UI inspection. Windows/Ubuntu application and planning CI passed for the release source. Installer extraction verifies its application and bundled CLI match the tested unpacked build. SHA256SUMS.txt identifies the installer.

Seats visualize existing roles sharing one verified model binding; this does not add independent provider connections, separate per-seat monitoring jobs or action authority. Operator retains its separate page and per-step approval. Readiness reflects verification, not a provider heartbeat. No model is bundled. The installer is unsigned; installation itself, personal monitor/camera capture, mixed-DPI behavior, native input effects, Bionic compatibility and sustained model throughput remain outside this slice's acceptance evidence.

OpenAware v0.3.2 — adversarial audit fixes

OpenAware v0.3.2 repairs 23 reproduced defects found by adversarial testing. Fixes preserve chat drafts and selected sources across page changes, handle IME Enter correctly, keep newer drafts when replies finish, prevent AI frame starvation during continuous preview, release stalled Stop/capture operations, and reject malformed provider, image and CLI inputs.

Verification: 193 unit/integration tests, 14 Windows packaged desktop workflows, strict TypeScript, four-page native inspection, independent source review and Windows/Ubuntu application/planning CI for the tagged source. The audit report lists reproductions, regression evidence and remaining coverage.

The Windows x64 installer is unsigned. Its application and external CLI payloads match the locally tested unpacked build. Installation and shortcut creation have not been exercised. No model is bundled.

Installer SHA-256: 37b196047431b00a0b7b9dc9e3d21aef29944c31c4b692ab1fe181752078dd63

This remains a developer prototype. Real monitor/camera hardware performance, mixed-DPI input, sustained model accuracy and Bionic compatibility remain unverified by this audit. Local workflows require no NVIDIA VSS server; NVIDIA/Metis support and more than four simultaneous sources remain planned. A finite audit cannot establish that every possible bug is absent.

OpenAware 0.3.1 — direct llama.cpp

OpenAware now connects directly to a local llama.cpp server alongside LM Studio and Ollama. Choose Connections → llama.cpp, use the local server origin (default http://127.0.0.1:8080, without /v1), discover the exact model ID, and run the synthetic vision test before explicitly starting monitoring.

The same bounded temporal observations, caption search, historical text summaries and opt-in CLI workflows use the selected binding. Provider controls are compact and readable. This is original OpenAware code; no NVIDIA VSS server, implementation, container or weights are bundled.

Validation: strict TypeScript, 152 unit/integration tests, 11 final packaged desktop workflows, independent source review, installer payload extraction, and manual Connections inspection. Synthetic fixtures verify protocol and UI behavior. Actual llama.cpp inference and sustained performance remain unverified because no local llama.cpp server is running on the test computer. Start your own compatible vision GGUF model with its matching multimodal projector. No model or runtime is bundled or downloaded automatically.

Source commit: e73ff0ca24b6466990f85c92db35e0d72b76fcef. Application checks and Planning checks passed on Windows and Ubuntu for that exact commit.

See the workflow guide and implementation evidence.

This remains an unsigned Windows x64 developer prerelease. Installer installation and shortcuts have not been exercised. Real cameras/monitors, mixed DPI, native action effects, Bionic compatibility and useful real-time AI throughput remain acceptance gates. Operator actions still require interactive approval for each step.

Installer: OpenAware.Setup.0.3.1.exe (114,925,663 bytes).
SHA-256: 2b1d6f7df6f1c671b96764e487e87cd534e1b93883742c59346f2201c9b32399.

OpenAware 0.3.0 — local video workflows

OpenAware now runs original VSS-inspired video workflows through its existing local Ollama and LM Studio adapters. No NVIDIA VSS server is required.

• Temporal image samples with timestamped captions and bounded session memory.
• Semantic alert rules with scoped sources, freshness checks and cooldowns.
• Video memory pane for caption search and asynchronous text-only historical summaries.
• Opt-in authenticated agent CLI (--enable-cli) and an installable Codex workflow skill.
• Fitted side-by-side monitors, additional feed rows and movable/resizable panes; computer actions remain on Operator.

See the workflow guide for setup and commands. The Windows installer bundles the CLI at resources/cli.cjs; CLI use requires Node.js 24 or later. The desktop application uses its bundled Electron runtime. No model is bundled.

Validation: strict TypeScript, 147 unit/integration tests, all 10 final packaged desktop workflows, installer payload/hash inspection and four-page manual inspection passed. Mock/synthetic fixtures prove workflows, not real model accuracy. The current desktop session exposed one Entire screen entry; both physical monitors are not claimed verified here.

Actual local model testing recognized a synthetic image in about 18–20 seconds. Three-image inference and a historical summary reached their 60-second limits. Fast continuous monitoring needs further model/hardware calibration. This is sampled image analysis, not full NVIDIA VSS compatibility or native every-frame video understanding. No NVIDIA implementation, containers or weights are bundled. Installer execution, real cameras, mixed DPI, native action effects, Bionic and sustained performance remain unverified.

The source commit 50b0f8848a0bf685e09fc6d8a34bc913a31d53ad passed Application checks and Planning checks on Windows and Ubuntu.

Unsigned Windows x64 developer prerelease. Verify the installer against SHA256SUMS.txt.

OpenAware v0.2.3 developer prototype

OpenAware now has one flat workspace. Every monitor, camera, assistant, actions and activity pane moves and resizes directly. The extra status/Layout rows and nested feed frames are removed; common controls share the top bar and pane headers use compact arrangement menus. Chat keeps a conventional history and composer.

Two selected monitors start side by side automatically. Adding sources, returning to Overview and Reset dashboard layout restore the same default. Three or four desktop feeds use two-column rows; cameras appear below desktop feeds. Custom arrangements scroll when needed to preserve readable pane sizes. Layouts and source configuration remain session memory; a new launch starts with no feeds.

All 104 unit/integration tests and all 10 packaged Windows workflows pass. Verification and boundaries are recorded in the v0.2.3 implementation ledger. Generated fixtures check geometry, Reset/navigation, settings and source scope, movement/resizing, and live capture continuity. Manual UI checks covered all four pages and two actual monitor previews, including default/Reset/navigation. Public screenshots contain generated content only. Real model inference and native action effects remain unverified.

Windows x64 developer prerelease. Unsigned installer; no model bundled. SHA256SUMS.txt names the uploaded installer asset. The installer itself has not been exercised. Capture authority, model providers, the four-source capacity and native action approval are unchanged.

The application checks and planning checks passed on Windows and Ubuntu at release commit c891914991d6d99912bf453c8658b576b0b31eca.

OpenAware v0.2.2 developer prototype

OpenAware v0.2.2 adds movable and resizable workspace panes, readable lower live feeds, and a compact chat interface.

The assistant now uses a conventional message history and composer. Repeated captions and oversized empty-state prompts are removed; less frequent analysis, motion, privacy and source-scope settings use compact controls.

Drag a pane by its grip and drop it beside, above, or below another pane. Drag horizontal/vertical split dividers to resize; keyboard Move menus and divider controls offer the same layout operations. Main panels and individual desktop/camera feeds can be arranged independently. Stack and Side by side controls provide quick feed arrangements; Reset layout restores the starting layout. Global Stop and top navigation stay fixed.

Moving or resizing keeps existing live capture connected and preserves source identity, settings, and assistant drafts. Layout preferences remain in session memory. This prototype still allows four simultaneous sources total across monitors, windows, and cameras.

The blue OpenAware tray eye and isolated desktop capture correction from v0.2.1 are retained. Double-click the tray icon to reopen the dashboard; right-click for Show, Stop, and Quit. Preview needs no AI model or Start watching.

All 91 unit/integration tests and all 10 packaged Windows app workflows pass. Verification and limits are recorded in the implementation ledger. UI tests use synthetic feeds and model fixtures; native capture tests use only generated windows. Personal monitor/camera hardware, mixed DPI, real model behavior, native input, Bionic, installation, sleep/lock, and sustained performance retain their acceptance gates.

Windows x64 developer prerelease, unsigned distribution, no model bundled. SHA256SUMS.txt names the uploaded installer asset. The roadmap and existing planning issues remain open until their full gates are verified.

The application checks and planning checks passed on Windows and Ubuntu at release commit 301c6f611a606d5a5f5dc932792d7a25dadd6d61.

OpenAware v0.2.1 — Live desktop fix and clear tray icon

OpenAware v0.2.1 corrects selected desktop capture and replaces the unclear tray square with a recognizable blue eye/O icon.

• Separate fixed capture workers let selected monitors/windows run together without granting desktop media access to the dashboard. Previews preserve native dimensions and decoded-frame timestamps; Stop cancels pending starts and destroys capture owners.
• The tray icon has transparent raster sizes for Windows scaling, a tooltip identifying OpenAware, and matching app/installer branding. Double-click it to open; right-click for Show, Stop, and Quit.
• Privacy-mask and frame freshness checks now prevent a resized or newer preview from slipping into an older in-flight sample.

To view two screens, use Add source → Monitor → select your monitor → Connect, then repeat for the second monitor. Preview requires no AI model or Start watching. Background keeps the configured session running; its tray controls restore or stop it.

Verification evidence and limits are recorded in .omx/logs/implementation-v0.2.1.md. Native regression uses only two generated fixture windows, verifies exact source routing and background continuity, denies direct modern/legacy desktop capture, and checks Stop cancellation and cleanup. These fixture results do not certify personal monitor/camera hardware, mixed DPI, real model behavior, native input, Bionic, or installation.

Windows x64 developer prerelease; no model bundled. Installer distribution is unsigned. SHA256SUMS.txt names the uploaded installer asset. Existing roadmap issues remain open until their full acceptance evidence is complete.

OpenAware v0.2.0 desktop and tray prototype

OpenAware v0.2 follows the original interface concept with top navigation, a focused desktop preview, a separate camera strip, and the assistant, computer actions and activity alongside it.

Background monitoring keeps the current session running while the dashboard is hidden. Use the system tray to Show OpenAware, Stop all capture and actions, or Quit. Show preserves the background preference; Window mode clears it. The --background and --headless launch flags start an idle tray session with no selected feeds in the same interactive Windows desktop session. Sources and model configuration remain memory only.

This is an unsigned Windows x64 developer prerelease. Download the installer and SHA256SUMS.txt below. No model is bundled. Existing action steps still require native approval.

Validation includes strict TypeScript, 65 unit/integration tests, eight real Electron workflows, production build, and packaged synthetic/tray workflows. The actual UI screenshots use synthetic sources. Mock model responses and installed tray menu callbacks are used for automation; no personal screen or camera, actual native input, or trade was exercised. The installer itself has not been installed during verification.

Real model inference, real camera/monitor negotiation, physical tray interaction, mixed DPI, sleep/lock behavior, Bionic, cloud adapters and unattended automation remain open acceptance gates. See the prototype status, background guide, and v0.2 evidence ledger.

OpenAware v0.1.0 developer prototype

OpenAware's first runnable Windows x64 developer prototype couples two AI roles: an Observer for selected live sources and an Operator for reviewed computer actions. Both use the explicitly selected local vision model.

• Up to four monitor, window, camera, OBS virtual-camera, or synthetic source slots; continuous previews with sampled AI analysis.
• Local LM Studio and Ollama discovery, selection, synthetic vision check, source-scoped chat, and independent motion events.
• Privacy masks before analysis, bounded in-memory content, capture freshness, and Stop all.
• Operator proposals of up to eight click, typing, or keypress steps. Each native input requires a separate review and fresh checks on an unmasked selected monitor. Stop cannot undo an event already sent.

Validation: strict TypeScript check; 54 unit/integration tests; three Electron UI workflows; one packaged executable workflow using synthetic feeds and a mock model; production build and Windows NSIS packaging; 29-document planning validation. No personal feed, real model inference, or actual computer input effect was used in these checks.

The installer is unsigned. Its packaged executable was tested; installation itself was not exercised. No model is bundled. Start a local vision model server, then use Connections to discover, select, and run the synthetic probe. Preview and motion-only watching can run independently of model inference.

This is an experimental prototype. Real camera/OBS/monitor capture, device loss, decoded-frame delivery, sustained multi-source performance, focus/DPI behavior, and actual native input effects need live acceptance tests. Bionic embedding/MCP pairing, cloud providers, persisted encrypted settings/history, voice, broker integration, unattended workflows, and signed updates remain future work. The full 26-slice roadmap stays open until its gates are evidenced.

The MIT-licensed source, third-party notices, setup instructions, and verification ledger are in the repository. SHA256SUMS.txt accompanies the installer.

Application CI and planning CI passed on Windows and Ubuntu for source commit 6df933e1a5b4da5b9d245cf5af10ec029e31b3c3. Windows CI also passed the three synthetic Electron UI workflows.