Orkas is an open-source (MIT), local-first AI team for your desktop: a Commander turns your goal into to-dos and coordinates 16 built-in specialist agents plus Claude Code, Codex, OpenCode, OpenClaw and Hermes, with projects, scheduled automations and connectors. Free app for macOS & Windows; built-in models or your own keys.
Windows Runtime Regression
This checklist covers Windows-only runtime behavior that is easy to miss in platform-neutral unit tests. Keep the automated cases deterministic and run the native smoke separately because it executes the bundled Whisper model.
Platform-difference risk map
| Boundary | Windows difference and user risk | Automated regression / discovery probe | Packaged/manual residual |
|---|---|---|---|
| Paths and filesystems | Drive letters, case-insensitive NTFS, junctions, reserved names, quoted/space-containing PATH entries, file locks, and long paths can reject valid files, escape a workspace, or leave data undeletable. | Path sandbox uses native case-insensitive containment and a real junction escape negative-control; conversation slugs reject reserved names; CLI lookup covers PATHEXT and quoted PATH; cleanup paths retry EPERM/EBUSY; source-data migration preserves the prior copy on publication failure and redacts early-boot errors. |
Exercise a non-ASCII/space-containing install and workspace, a path over 260 characters, OneDrive-controlled folders, and Defender-held temporary files. |
| Processes and shells | PowerShell and cmd have different quoting, encoding, shim, signal, and descendant semantics; POSIX mkdir -p can bind as PowerShell -Path and fail after partial interpretation; killing the parent alone leaks work and locks files. |
Preflight rejects high-confidence POSIX syntax (including mkdir -p) with PowerShell rewrite guidance; real PowerShell UTF-8 and npm-style .cmd argv round trips; OAuth and relaunch use bounded argv-only launch contracts; every production child-process owner has a focused suite in the serialized native lane; real parent/grandchild sentinels prove whole-tree termination. The local-CLI lane executes the governed Lark multiline/MCP stdio contracts with native Windows process semantics. |
Inspect Task Manager after repeated cancel/timeout/relaunch cycles and confirm no console flashes, descendants, or locked work directories. |
| Official connector CLIs | A package may advertise Node compatibility but omit its Windows executable, fail behind npx, emit a startup warning as false success, or leave its cache locked after exit. |
A Windows x64 canary uses the production authorization helper and direct node npx-cli.js path to verify the exact integrity-pinned WeCom, Feishu/Lark, DingTalk, and Xero packages: a child PATH containing only bundled Node and Windows system tools (never CI/system Node), a no-Node negative control, exact bundled-Node resolution, registry integrity, isolated download, executable discovery, real --help startup, clean stderr, and immediate cache removal. The owning cross-platform cases also run real offline npm lifecycle scripts with missing/shadowing system Node, space/non-ASCII runtime paths, install failure without a ready marker, and successful retry before authorization/MCP access. |
Real account authorization and business API calls require provider accounts and remain release canaries; verify browser/device login, one read, one confirmed write, logout, and reconnect for each provider without using personal production data. |
| Source startup without system Node | run.cmd must bootstrap the pinned runtime before any JavaScript can execute; a partial download, corrupt marker, bad hash, or failed directory swap can otherwise leave development startup permanently broken. |
Cross-platform wiring/manifest contract plus a real Windows PowerShell integration using an actual zip and Node executable: first install, verified cache, corrupt-marker repair, hash rejection, temp cleanup, and failed-swap rollback. | On clean x64 and Arm64 hosts with Node removed from PATH, launch run.cmd through Explorer and a path containing spaces/non-ASCII characters. |
| Window and OS integration | Negative monitor coordinates, DPI/topology changes, minimized close bounds, registry command parsing, AppUserModelID, notification policy, and protocol cold launch differ from macOS. | Window state preserves visible left-monitor coordinates, drops disconnected positions, and ignores minimized sentinel bounds; registry cleanup rejects executable look-alikes; notification probes are hidden/bounded and escape the app id; deep-link registration has argv/error coverage. | Verify mixed-DPI monitor removal/reconnect, sleep/resume, notification Group Policy, default-app changes, and cold/second-instance orkas:// activation. |
| Install and packaging | App-local VC DLL closure, PE architecture, native ABI, signing, installer elevation, and antivirus scanning can make a build pass source tests but fail after install. | PE/ABI/DLL/license/package gates; Windows shutdown and installer process ownership; NSIS compilation requires an installed public compiler; installer execution remains Windows-only. | Build the public Windows package and run packaged-app/native checks; test signed upgrade/downgrade/reinstall/uninstall as standard and admin users, including data retention and interrupted installation. |
| Performance and recovery | Process creation is slower, file handles are retained longer, and native CPU dispatch can regress latency or destabilize concurrent tests. | Windows unit workers are serialized; native Whisper has a 60-second budget/90-second hard limit; high-risk process-tree cases repeat three times in isolated Vitest processes so forced handle closure cannot contaminate the next owner. | Record cold/warm startup, handle/process count, working set, Whisper latency, and cleanup time with Defender enabled and after sleep/resume. |
Embedded-browser phone passkey canary (pending hardware)
The requester deferred Windows hardware execution on 2026-10-04. The cases below are not run. Run them against the integrated packaged app when a Windows machine is available; the isolated macOS caBLE probe is not Windows evidence. Keep official Electron. USB keys, Windows Hello enrollment, and new credential creation are outside this phone-authentication canary.
Use an isolated ORKAS_WORKSPACE_ROOT, a current Windows 11 installation with
Bluetooth, and an iPhone or Android phone with a Google passkey already saved.
Record Windows build/update, Electron version, phone OS, packaged app version,
and proxy mode without account identifiers or credential/session bytes. Windows
10 and older Windows 11 builds require a separate capability check; do not infer
phone support from the presence of WebAuthn. Microsoft's documented native
passkey management experience starts at Windows 11 22H2 with KB5030310; this
does not establish a minimum hybrid-transport version. Windows 11 24H2 also adds
per-application passkey privacy consent:
Windows passkeys.
| ID / user goal | Procedure and variation | Independent result and recovery |
|---|---|---|
| W-PK-01 / sign in with the nearby phone | In a fresh embedded-browser tab, open Google sign-in, enter the account locally, choose passkey authentication and the phone/tablet option, scan the QR, and confirm on the phone. Repeat after signing out. | Google accepts the assertion and displays the signed-in session in that same tab. A QR, capability flag, phone confirmation, or returned assertion alone is not a pass. Record both runs without copying private page content. |
| W-PK-02 / cancel and retry | Cancel the Windows authentication dialog before scanning; retry. Repeat with cancellation on the phone after scanning. | The cancelled attempt does not sign in or leave a blocking dialog. A fresh attempt can reach Google's accepted session; no delayed result from the cancelled attempt completes it. |
| W-PK-03 / unavailable credential | Scan with a phone that has no matching Google credential, then retry with the prepared phone. Do not enroll a new credential as part of this test. | The first attempt does not sign in and remains recoverable. The prepared phone succeeds on a new challenge; no wrong-account session appears. |
| W-PK-04 / Bluetooth and connection failure | Disable Bluetooth before starting, restore it and retry. In another attempt disconnect network access after scanning, allow the operation to fail/expire, restore access, and retry. | Failure never appears as login success. The dialog can be dismissed, the page remains usable, and a fresh challenge succeeds after restoration. Record the OS feedback and timeout rather than assuming a fixed duration. |
| W-PK-05 / tab and account lifetime | While phone authentication is pending, navigate the requesting tab, close it, or switch the Orkas account in separate attempts. Confirm on the phone only after teardown. | No stale authentication completes in a replacement page, another tab, or another Orkas account. No orphan dialog remains; a new tab/account can initiate an independent attempt. |
| W-PK-06 / proxy continuity | Repeat W-PK-01 with the configured system or environment proxy, including an authenticated supported proxy. Then make that proxy unavailable and retry after restoration. | Sign-in works through the intended route. A blocked proxy does not silently use a direct tunnel; the failed attempt is recoverable. Establish routing with a controlled proxy's connection counters, not private authentication payloads. |
| W-PK-07 / OS privacy consent | On Windows 11 24H2 or newer, decline the application's passkey consent on first use. Restore its access in Settings > Privacy & security > Passkey access and retry. | Denied access cannot authenticate and does not silently expand permissions. Restoring consent permits a fresh phone attempt; Google must accept it before recording recovery as passed. |
Inspect app stdout/stderr and Windows dialog outcomes for every attempt. Preserve failed results separately from retries; report warnings and unavailable variants explicitly. No Windows passkey support claim is valid until these real-device results are recorded.
Automated coverage
npm run test:runtime-path runs the focused environment regression entry through
the Electron test runner: missing/shadowing system Node, offline npm lifecycle
scripts, space/non-ASCII runtime paths, nested Skill commands, and bundled
runtime/reference-process contracts. This is a local focused entry; neither
public workflow currently invokes it. Native dependency repair runs its listed
native/ABI regression suites on Windows, macOS and Linux; Linux source smoke
also runs the Linux desktop/native lane. Run and retain the Windows-native,
stability and packaged-app checks separately; those workflows do not establish
their results.
The process-owner inventory covers shipped bin JavaScript helpers as well as
src modules. A new child-process owner requires a behavioral suite in each
applicable native lane; negative controls check discovery and failure propagation.
- Process lifecycle: OfficeCLI, ffmpeg/ffprobe, VideoStudio native helpers, package commands, and local-agent version probes use hidden windows, bounded output, prompt timeout/abort settlement, and whole-tree termination. Real parent/grandchild cases verify that a descendant cannot outlive its parent and write a delayed sentinel file.
- Windows paths: sandbox containment and skill imports compare case-insensitively without prefix false positives; a real junction cannot escape the workspace.
- Shell and launch integrity: local-agent
.cmdshims round-trip hostile argv; quoted PATH entries remain discoverable; OAuth URLs stay a singlerundll32argument; development relaunch removes frozen user/workspace environment state and safely quotesrun.cmd. Local connector CLIs bypass.cmdshims; Windows reruns the Lark multiline/MCP stdio journeys and starts every exact official package through the productionnode npx-cli.jshelper. - No-Node source startup:
run.cmdinvokes the PowerShell bootstrap before dependency preparation; the pinned manifest selects x64/Arm64, cache reuse requires marker/files/executable version, and download size/SHA plus Node self-check complete before an atomic swap. The Windows bootstrap integration injects a failed swap and requires the prior runnable payload to return. Run it throughnpm run test:windows-nativeon Windows and retain its result; the existing public CI workflows do not execute that integration. - Desktop state and installation: negative-coordinate monitor positions remain usable and minimized shutdown does not persist Windows sentinel bounds. Shutdown and installer cleanup target this installation's process tree; failure and retry cases preserve unrelated running applications.
- Document recognition: local OCR installation and execution are retired. Images use model input and scanned PDFs use bounded page rendering; a text-only model requests readable text or a vision-capable model.
- File locking: native dependency pruning uses retrying recursive removal, and asynchronous quality-report writes are serialized and drainable before a temporary workspace is removed.
- Packaging integrity: the Windows Whisper contract verifies every CLI, CPU dispatch DLL, VC runtime DLL, model, and pinned OpenBLAS license by size and SHA-256. The repository copy of every pinned license must satisfy that same contract before packaging starts.
Primary suites:
src/core-agent/test/oauth-flow.test.tssrc/core-agent/test/process-session.test.tstest/main/features/connectors/protocol.test.tstest/main/features/connectors/oauth-dcr.test.tstest/main/features/connectors/local-cli.test.tstest/main/index-process.test.tstest/main/util/shutdown.test.tstest/main/util/windows-shutdown-processes.test.tstest/main/util/windows-installer-processes.test.tstest/main/util/nsis-installer-processes.test.tstest/main/features/local_agents/version.test.tstest/main/features/local_agents/which.test.tstest/main/features/office/office_engine.test.tstest/main/model/core-agent/file-tools.test.tstest/main/features/packages.test.tstest/main/features/skills.test.tstest/main/features/video_gen.test.tstest/main/features/video_studio_native_qa.test.tstest/main/features/window_state.test.tstest/main/util/migrate-source-data-root.test.tstest/main/ipc/app-data-root.test.tstest/main/platform-process-coverage.test.tstest/main/quality/report.test.tstest/main/util/fetch-whisper.test.tstest/main/util/media_probe.test.tstest/main/util/path-sandbox.test.tstest/main/util/prepare-win-native-deps.test.tstest/main/util/windows-node-bootstrap-contract.test.tstest/windows/windows-node-bootstrap.integration.ps1(real Windows lane)test/windows/windows-local-cli.integration.mjs(real Windows x64 package lane)
Required commands
The compiled NSIS suite requires a working public makensis on PATH, or an
absolute ORKAS_TEST_MAKENSIS path. For an extracted electron-builder NSIS
toolchain, set ELECTRON_BUILDER_NSIS_DIR to its resource root; this also supplies
NSISDIR. Tests do not download a compiler or depend on a private product/
directory. Missing or failed compilation is a test infrastructure failure, not a
skipped pass. macOS compiles the production hook and all four fault fixtures;
only Windows executes the installers and verifies file-lock release and retry.
On macOS 14 or newer, use a compiler containing the upstream
UTF-32 conversion fix
and its function-order correction.
Unpatched 3.13 can abort while generating language tables; a successful
makensis -VERSION alone does not prove compilation works.
- Run
npm run typecheck. - Run
npm testfor the full JavaScript and Python resource suites. - Run
npm run test:platform-nativein a desktop session. The serialized native-host suites are followed by real Electron composition-inspect and background-browser-input canaries; a missing display fails the lane. - Run
npm run test:coverageand investigate every failure; do not accept a worker crash, unhandled rejection,EPERM, or leaked-handle warning as an unrelated test issue. - On Windows, run
npm run test:windows-native. The real PowerShell no-Node bootstrap integration runs first; the Windows x64 connector canary then downloads and starts every exact official CLI package without accessing an account; bundled Whisper must finally finish within its 60-second performance budget and before the 90-second hard limit. - On Windows, run
npm run test:windows-stability. It repeats the real persistent-process and descendant-termination cases three times. After the loop, no test-ownednode.exe,cmd.exe, ffmpeg, ffprobe, OfficeCLI, or delayed sentinel should remain.
The Windows-only commands report a skip on other hosts; that is not Windows execution evidence. Record the actual OS, commands, skips, diagnostics and process cleanup for a Windows release. These deterministic/native commands do not call a chat or generation provider; the bundled Whisper smoke performs local inference. Real connector authorization and phone-passkey journeys remain the manual checks described above.
The public Native dependency repair workflow exercises native repair and
focused Electron regressions on Windows, macOS and Linux. The Linux source smoke workflow additionally runs the Linux desktop/native lane. Windows
coverage, stability, native inference and packaged-app acceptance must be run
and recorded separately; these workflows do not establish those results.
The review found that bootstrap-node.ps1 was a production startup entry with
no test owner. It also exited the caller on a cache hit, preventing safe
in-process fixture execution. The script now returns normally and accepts
isolated root/architecture parameters while preserving its default run.cmd
behavior; the Windows integration is the behavior oracle, while the
cross-platform contract only guards wiring and pinned inventory.
Manual packaged-app checks
- Launch from Explorer with no developer-shell PATH and confirm local CLIs, connector OAuth callbacks, notifications, OfficeCLI, ffmpeg/ffprobe, and Whisper resolve without a visible console window.
- Cancel and time out one Office, video, media-probe, and package operation. Confirm Task Manager shows no descendants and the working files can be renamed or deleted immediately afterward.
- Read an image and an image-only PDF on a clean profile with package-network access unavailable. Image input and page rendering must work without a Python recognition install. A text-only model must state its limitation.
- Repeat start/cancel/timeout cycles with Microsoft Defender enabled and verify stable handle count, process count, latency, and temporary-directory cleanup.
- Install and launch from paths containing spaces and non-ASCII characters; use a workspace path over 260 characters and confirm file tools, OfficeCLI, FFmpeg, and produced-file reveal/open actions still work.
- Move the app between mixed-DPI monitors (including one left of the primary), minimize, sleep/resume, disconnect the secondary monitor, and relaunch. Confirm the window is visible and the last normal bounds remain recoverable.
- Exercise signed NSIS upgrade, interrupted upgrade, reinstall, and uninstall under standard-user and admin accounts. Confirm rollback/repair is clear and the pinned user-data container is neither redirected nor removed.