Luda 0.3.4

Luda now verifies the affected application output before claiming a preference change is complete. Selecting a dark-theme row without changing light content and controls is reported as incomplete; selection-only file tasks still stop at highlighting.

The entry skill is the exact accepted 1,999-word artifact. All seven reference guides, runtime code, MCP schemas, native input and cursor behavior are unchanged. This release adds reusable recorded and interactive evaluation runners, genuine screenshot fixtures, separate evaluator criteria, and preserved evidence/failure history.

Validation newly run for this release with explicitly selected gpt-5.6-sol, Codex CLI 0.155.1, and no reasoning override:

• Two consecutive unchanged 23/23 batches: 34 recorded decisions and 12 interactive scenarios.
• Two fresh actual-Luda XFCE confirmations: independent xfconf, AT-SPI and screenshots verify Greybird → Greybird-dark, dark application content/controls, visible Style page, and accurate GUI-only agent reports.
• 1,111 local repository tests: 1,109 passed, two skipped. All six CI workflows passed for the source commit.
• Core wheel, source distribution, plugin archive and fresh wheel installation contain all eight exact skill files.

The reconstructed interactive batches retained four and five recovered fixture errors. Live confirmation two recovered one initial desktop_doctor BUSY response; live one had no MCP errors. Pre-agent MCP startup failures, the initial Thunar capture failure, corrected preflights, semantic adjudication and cleanup are preserved separately. No unassessable run counts as a pass. Test-owned desktops, temporary credentials links, inputs and downloaded themes were removed; existing credentials and user resources were preserved.

The supplied historical results used CLI 0.145.0 and are documented separately. This adaptive finite suite is not a claim of universal GUI reliability, untested-model support, or isolated sentence efficacy. See evaluation and limits and archived evidence.

Install the core wheel or source distribution for the runtime. The core plugin ZIP contains configuration and the complete skill, not the Python runtime. Editor Bridge 0.2.0 remains a separate optional distribution/plugin; core installation does not install it. These are GitHub release assets, not a PyPI publication.

Source commit: e3863fb24dd28bda5910a8c2382a13afd4ee1064

Core source archive SHA-256:
fc4530d3e374fa3430fa3f6756f2fff60d02ce7f38551abcf7d677115c879c7a

Installed/distributed entry skill SHA-256:
e90eae580e187882b7d6308eb35c77ff19a5a483857a5a0202a52ca7d6f88260

release.json pins the source commit and package versions; SHA256SUMS covers all attached assets.

Luda 0.3.3

Make selection verification explicit in the result the agent receives.

Successful desktop_choose results now add verification_scope="selection" and conditional next_step guidance, including when the requested item was already selected. Existing result fields, effect="verified", and selection behavior are preserved. Failures and uncertain results do not gain verified-selection guidance.

The feedback distinguishes selection-only requests from requests to apply, open or execute something. It asks the agent to check the requested application effect first, since some controls apply changes on selection. Only when needed should the agent use an advertised activation action or Apply/Open control, then verify the outcome. Luda does not automatically activate every selection.

Validation:

• 1,068 unit tests ran successfully with two environment-dependent skips; three wheel-build tests passed separately.
• Real XFCE regression covers both changed-selection and already-selected feedback, independently checks that selecting does not apply the theme, then verifies activation does apply it. This check is included in CI.
• Three fresh fixed-version agents all applied Greybird-dark according to independent XFConf readback. One reported verified completion; two honestly remained uncertain about the visual appearance.
• A fresh file trial began unselected and ended selected without opening the file.
• The reported false-completion failure was not reproduced in two comparable baseline agents. An earlier baseline with a missing settings daemon and an inconclusive file-selection oracle run are both retained. These observations do not establish an improvement rate or guarantee agent success.

The complete report records all eight attempts, prompts, outcomes, limitations and reproducible commands.

Upgrade using the installation guide, rerun setup for your selected agents to update the skill, and restart their MCP connection. The optional Editor Bridge remains at 0.2.0.

Luda 0.3.2

Clarifies selection versus application in the main skill, semantic-controls guide and desktop_choose tool description.

A verified selection may only highlight a row. When the requested outcome requires applying or opening the choice, use an advertised activation or Apply action when needed, then verify the resulting application state. Selection behavior itself is unchanged.

Validation: an isolated real XFCE Appearance test confirmed that choosing a theme leaves the actual theme setting unchanged and invoking its advertised activation action changes the setting. A fresh-agent decision exercise correctly distinguished applying a theme from selecting a file without opening it. Skill validation and focused contract tests passed; upstream installer and pinned Codex registration CI passed.

Upgrade using the installation guide, then rerun setup for your selected agents to install the updated skill. The optional Editor Bridge remains at 0.2.0.

Luda 0.3.1

Fixes two desktop reliability bugs reported against 0.3.0 and improves the skill's recovery guidance.

• Preserve the selected session's desktop identity through launch and reconnect so Gio can correctly apply OnlyShowIn/NotShowIn. No desktop identity is guessed when the session omits it.
• Support GTK composite table cells whose named renderer is a descendant of the canonical cell. Inspection records canonical identity, position and ancestry; selection rejects replaced cells, changed ancestry and reordered rows.
• Keep essential coordinate guidance in the main skill. Verify the exact requested outcome, report unavailable options without substitution, and change approach after repeated interaction failures.

Validation: the composite-cell failure was reproduced before the fix with both a regression test and real GTK. The fixed GTK test confirms selection through an independent application callback. Real Gio tests cover desktop-entry visibility. The full suite ran 1,061 tests successfully with two environment-dependent skips; the skipped wheel-build test passed separately in the build environment. A fresh-agent decision exercise followed the revised skill, but this is not an end-to-end guarantee of agent behavior. Real GTK/Gio regressions are now included in CI.

The desktop environment must still provide XDG_CURRENT_DESKTOP. Missing theme packages remain an environment prerequisite; Luda does not install themes. No Silo changes are included.

See the installation and upgrade guide. Re-run setup for your selected agents after upgrading so they receive the updated skill. The optional Editor Bridge remains at 0.2.0.

Luda 0.3.0

Luda 0.3.0 provides a complete Linux installer for the runtime, MCP registration and skill installation.

• Configure selected agents, all seven supported clients, or prepare a runtime for later setup. Agents do not need to be installed yet.
• Delegate skill and MCP installation to pinned Vercel skills and add-mcp packages, with a private Node runtime provisioned automatically.
• Preserve unrelated settings and skills, reject malformed MCP configuration without modifying it, and migrate unchanged legacy Luda skill copies.
• Support user/project scope, supported profile overrides, plugin export and optional desktop readiness checks.

Start with the README and installation guide. Runtime-only installation does not require a running desktop or agent credentials.

Validation: 1,042 unit tests covered across runtime/build environments, real bundled installer checks for all seven targets, combined bootstrap installation, and seven passing pinned Codex registration/discovery checks. See validation details and limits.

The optional Editor Bridge remains at 0.2.0 and is packaged separately. No desktop tool behavior changes are included in this release.

Luda 0.2.0

Luda 0.2.0 installs the Linux desktop runtime, registers its computer-use tools, and installs the accompanying skill together. The same installer works on a user's machine and during VM image construction.

Install for your agent

On the Linux desktop machine, from your desktop/agent account:

git clone --branch v0.2.0 --depth 1 https://github.com/0xpolarzero/luda.git
cd luda
sudo bash scripts/install.sh --user "$(id -un)"

Select one or several agents, confirm the destinations, then restart/reconnect those agents. For a noninteractive image build, create the intended account first and run as root:

bash scripts/install.sh --user YOUR_ACCOUNT --agent codex --yes

No running desktop or agent credentials are needed at build time. The runtime, complete skill and MCP configuration are prepared for that account. Start the desktop before using the tools. With Codex's built-in SSH remote-project connection, configure the account used by the backend inside the VM.

• Installation, upgrades and verification
• Image build and first-boot recipe
• Supported clients and custom integrations

Changes

• Combined tools-and-skill setup for seven documented Linux client profiles; multiple client selection, detection, project/user scope, explicit account ownership, repeatable installs and managed skill updates.
• Portable Agent Plugins export for custom clients. Runtime provisioning and agent configuration require no Silo-specific integration.
• Visible agent cursor where supported; automatic background/independent input where compatible and foreground/shared input when needed. Foreground fallback can interrupt human mouse/keyboard activity or bring windows forward.
• Browser native-input handoff restores real widget focus while preserving the active tab. The separately installed Editor Bridge correctly checks shared keyboard state for clipboard operations.
• Existing configuration and edited skills are preserved. Installation success is distinguished from a running desktop and actual client discovery.

Downloads

The core source archive includes scripts/install.sh and the documentation; extract it and use the same commands as a Git checkout. The core wheel is for environments whose Linux dependencies are already provisioned. Plugin ZIPs contain configuration and skills, not the runtime. Editor Bridge remains a separate optional package; its 0.2.0 wheel requires core 0.2.0. Check SHA256SUMS; release.json identifies the exact source commit.

Packages are published on GitHub. No PyPI publication is claimed.

Support and validation

This release targets native Linux X11, with Ubuntu 24.04 and XFCE as the tested starting point. Python 3.12+ is required. Wayland/Xwayland are unsupported. The installer does not create a desktop, provision SSH, install agents or authenticate them.

Local validation includes 1,033 unit tests (one skipped), real Codex MCP configuration and skill discovery with a pinned CLI, a fresh selected-account installation with system prerequisites already provisioned, an unchanged repeat run, and an upgrade from the downloadable source archive. The installed launcher also passed desktop readiness and MCP calls in an isolated XFCE session, with all owned processes cleaned up. Live desktop and browser checks verify independent application effects, focus/input cleanup and file results. Configuration adapters are not a claim of end-to-end acceptance for every supported client/version; toolkit/window-manager behavior is not universally qualified.

All six CI workflows passed for source commit 9cfd3b98bb7e12d7718b3e129865a1f81f565d39:

• Distribution packages
• Codex plugin registration
• Optional desktop media
• Native application workflows
• Desktop contracts
• Optional owned browser

Luda 0.1.0 preview

Luda gives coding agents tools and a skill to see applications, use controls, enter text, and check results on a Linux desktop.

This preview targets native X11. Ubuntu 24.04 with XFCE/XFWM4 is the tested starting point; Wayland and Xwayland are unsupported.

Start here

• Install Luda
• Connect Codex, Claude Code, Cursor, Gemini CLI, or OpenCode
• Package it into a Linux environment
• Read the complete agent skill

The runtime runs on the Linux machine that owns the desktop. Agent registration and skill discovery are explicit, per account or project. Luda does not provision desktops or manage VMs.

Choose your download

• luda-0.1.0-py3-none-any.whl: core Python runtime and complete core skill.
• luda-0.1.0.tar.gz: core source distribution.
• luda-plugin-0.1.0.zip: core plugin configuration and skill; install the runtime separately.
• luda_editor_bridge-0.1.0-py3-none-any.whl, its source archive, and luda-editor-bridge-plugin-0.1.0.zip: optional, separately installed Editor Bridge. Core does not include or enable it. Install matching core and add-on wheels together in its own environment.
• SHA256SUMS and release.json: checksums, source commit, and package verification metadata.

The Editor Bridge verifies supported rich-text edits in applications whose developer registered the ProseMirror adapter. It has its own tools, skill, and temporary browser. It does not work on arbitrary websites. Most desktop tasks do not need it.

Validation

Core: 888 local tests, with one build-environment-dependent skip. Separate add-on: 64 unit tests and 157 live GUI assertions across four real MCP suites. Final release wheels were independently installed; the add-on unit suite also passed against its installed wheel. Complete skills, separate tool discovery, checksums, and real Codex plugin registration were checked. CI covers native application workflows, desktop contracts, optional browser/editor behavior, optional media, plugin registration, and distribution packaging.

Other agent recipes follow their official documentation; they are not claims that every client/version was exercised end to end. Input verification is distinct from saving, and supported accessibility behavior depends on the application.

Packages are provided here on GitHub; no PyPI publication is claimed.