CHANGELOG.md

Changelog

All notable changes to kuri are documented here.

[0.4.11] — 2026-07-25

Fixes — kuri no longer takes over your machine

Driving the Simulator used to seize the user's foreground window and move their cursor. Every input command called sim_window.activate before doing anything, because CGEventPost(kCGHIDEventTap, …) injects into the global event stream — events land wherever focus happens to be, so Simulator.app had to be raised first for a tap to hit the right thing. That made kuri unusable on a machine somebody is actually working on.

• Input is delivered to Simulator.app by pid — CGEventPostToPid instead of the global HID tap. Events go straight into Simulator's own queue, so no window is raised and the cursor is never warped. Covers tap, doubletap, longpress, swipe, gesture, touch, key, key-sequence, batch
• type no longer routes through AppleScript. It shelled out to System Events keystroke, which is delivered to whichever app is frontmost — so ios type was doubly hostile: it had to steal focus to be correct, and if it ever ran without doing so it would type your text into whatever you had open. Now Unicode CGEvents addressed to the pid, which needs no virtual-keycode table either. osascript is gone from this path
• uitree, find and wait-for-ui no longer activate at all. They only read the accessibility tree, which works fine on a background app. wait-for-ui was the worst offender — it polls every 250ms, so it re-stole the foreground on every poll for the length of the wait
• button and background no longer activate. They already used AXPress, which never needed focus
• open-sim launches in the background (open -g). Opening a simulator is a setup step, not a request to be interrupted
• --activate restores the old behaviour per command, for the case where a gesture genuinely needs Simulator.app to be key. Off by default

Fixes — real devices

• ios terminate --device could never have succeeded. It built devicectl device process terminate --device <udid> <bundle-id>, but devicectl's terminate takes --pid and accepts no bundle id at all — every invocation died on devicectl's own argument parser. launch --device now reads the launched pid from --json-output and prints pid=N; terminate --device --pid N does the direct thing; terminate --device <bundle-id> resolves the bundle id to a running pid by matching device info processes against the app's on-device bundle URL. A launch that reports success without an identifier is now an error rather than a silent zero, which would later terminate an unrelated process
• ios list-apps --device silently hid every system app. devicectl device info apps defaults to developer apps only and says nothing about it, so a command documented as "list installed apps" returned a handful of entries on a phone with hundreds — and exited 0. Now passes --include-default-apps --include-app-clips. The same defaulting broke bundle-id lookups, so terminate-by-bundle-id could not resolve a system app either
• ios list-apps on the simulator no longer demands --udid. It resolves the booted simulator like launch, screenshot and uitree already did; requiring it made list-apps the odd command out for no reason a caller could infer

Fixes — diagnosis

• "Simulator.app is not running" when Simulator.app was running. The accessibility tree hangs off a window, and a device booted with simctl boot does not open one — so a running-but-windowless Simulator produced an error that sent you to restart an app that was already up. Now a distinct SimulatorHasNoWindow error carrying the actual remedy, and doctor reports window presence rather than just the process

Tests

• zig build e2e-ios-device — a new end-to-end suite against physically attached hardware: inspection, the install → list-apps → launch → terminate → uninstall round trip both by pid and by bundle id, and assertions that the XCUITest-only commands still refuse cleanly while a real device is attached. Skips with a reason when nothing is plugged in, when no bundle id is configured, or when the screen is locked — phones re-lock on their own timeout, and SpringBoard refuses every launch while they are, which is the environment rather than a defect. Verified: 24 passed, 0 failed against an iPhone 16 Pro Max
• A real-device command contract group in e2e-ios, needing no hardware. 21 hermetic cases pinning the silent-success class fixed in 0.4.10: every --device command must fail loudly against a fake udid, missing arguments must exit 2 rather than 1, and the XCUITest-only commands must exit 3 with an explanation. Two of them assert the absence of devicectl's argument-parser complaint, which is what distinguishes "the device is missing" from "we called devicectl wrong" — the exact bug fixed above
• e2e-ios now degrades instead of failing on preconditions a machine cannot supply. The Accessibility grant, a Simulator window and the Xcode toolchain are each probed and skipped with a reason, which is what lets the suite be a CI gate rather than a red build on a runner that can never hold a TCC grant
• More simulator coverage — list-apps, status-bar override/clear, ui appearance set-and-read-back, set-location/reset-location, log --last, terminate. Verified: 54 passed, 0 failed against a booted simulator
• devicectl's JSON shapes are now unit-tested against fixtures — a missing pid must not decode as 0, and a process match must not be made on a coincidental path prefix (/var/Demo.appendix is not inside /var/Demo.app)

CI

• The e2e suite finally runs in CI. A new mobile-macos job builds kuri-mobile, runs its unit tests, boots a simulator and runs both suites. Until now nothing caught a regression on the device path, which is how the silent --device no-op survived from 0.4.6 to 0.4.10
• It picks and boots the simulator through kuri-mobile itself rather than xcrun simctl — partly to exercise list-devices and boot for real, and partly because bare xcrun resolves through xcode-select, which is the exact indirection whose failure mode this project exists to avoid
• The job never opens Simulator.app, so it runs headless and the accessibility cases skip
• kuri-mobile's unit tests now also run on the Linux job; they had their own build.zig and the root test step never reached them

[0.4.10] — 2026-07-25

Fixes

• Every --device command silently did nothing and reported success. devicectl.zig shelled out to bare xcrun devicectl, which resolves through xcode-select — and when that points at CommandLineTools, devicectl does not exist there. Combined with an unchecked exit status, ios launch --device exited 0 having done nothing at all. This is the same bug class fixed for simctl in 0.4.6; devicectl was missed. It is now invoked by absolute path through the Xcode resolver with a checked exit status, so a failure surfaces devicectl's own diagnostic
• listDevicesJson passed --json-output -, but devicectl only writes JSON to a file on disk — never stdout — so this would have created a file literally named -. Replaced with the human-readable listing

Features — physical iOS devices

• install / uninstall / list-apps now work on real devices via devicectl instead of erroring or being simulator-only
• New device-scoped commands — device-info, device-processes, lock-state, displays, reboot. lock-state in particular addresses a common silent failure: automation that appears to do nothing because the screen is locked
• iOS surface is 41 commands, 12 of which work against physical hardware

Scope, stated honestly

• devicectl exposes no screenshot and no UI hierarchy, so tap/swipe/type/uitree/screenshot remain simulator-only no matter what is plugged in — those need XCUITest. The registry's scope field now reflects this exactly: 29 simulator-only, 7 simulator+device, 5 device-only

[0.4.9] — 2026-07-25

Tests

• zig build e2e-ios — an end-to-end suite that drives a real booted Simulator through the built binary, so it exercises the artifact that would actually ship rather than a test-only code path. 20 cases covering the registry, doctor, install, launch, uitree, find, wait-for-ui and screenshot
• Deliberately excluded from zig build test, which must stay runnable without a device. With no simulator booted the suite prints SKIP and exits 0, so it can sit in a pipeline without becoming a flaky gate
• Defaults to Settings (com.apple.Preferences), present on every simulator, so it is reproducible on any machine. Point it at your own app with KURI_E2E_BUNDLE_ID, KURI_E2E_LABEL and KURI_E2E_APP
• Timeout regression guard — asserts wait-for-ui returns within 3x its requested deadline, locking in the 0.4.7 fix where a 3s timeout waited 13.5s
• Assertions are behavioural, not just exit codes: find must emit a tap= centroid, uitree must contain the expected label and device-pixel bounds, and screenshot must produce a file with valid PNG magic rather than an empty one
• Shell-escaping tests — the Android quoting added in 0.4.8 had no coverage. Now tested against command-injection payloads (x; rm -rf /, $(whoami), backticks), embedded single quotes, and an exhaustive sweep asserting no byte in 1..126 can leave an unbalanced quote

Verified

• ios install had never been exercised until now; it installs a real .app bundle in ~2.9s and the app launches and reports its accessibility tree correctly

[0.4.8] — 2026-07-25

Features — kuri-mobile Android reaches parity

• android tools — Android now has the same meta tool iOS gained in 0.4.7. Both render from one shared module (common/toolinfo.zig), so the two platforms describe themselves identically and an agent can consume ios tools --json and android tools --json with a single parser. Combined surface is 63 commands (36 iOS, 27 Android)
• wait-for-ui / find — the same polling and query primitives as iOS, backed by the uiautomator hierarchy. find exits non-zero on no match so it works directly as an assertion; wait-for-ui uses a wall-clock deadline rather than counting sleeps
• batch — several actions over one adb session (tap:120,400 type:hi press:enter wait:500 label:Done), so serial resolution happens once instead of per command
• gesture/drag and touch — multi-point paths via input motionevent (Android 11+). input swipe only ever interpolates between two points, so shaped drags were previously impossible
• Lifecycle & state — uninstall, clear (wipe app data), openurl/navigate, keyevent (raw keycodes), current-activity, logcat (bounded -d read), screen-info, getprop, dumpsys
• More key names — recents, wakeup, sleep, search, camera, escape, backspace, notification

Internal

• Tool-registry rendering moved to common/toolinfo.zig. Shared invariant tests (verifyTable, verifyJson) run against both platform tables, so a new entry is checked for a summary, a known category, globally unique name/aliases, and JSON round-trip wherever it is added
• Android shell commands that carry user text (package names, URLs, log filters, dumpsys sections) are now single-quote escaped, so a value containing shell metacharacters cannot break out of the constructed command
• iOS tool entries carry an explicit scope rather than relying on a per-platform default

Not implemented

• android install and android record-video need the adb SYNC protocol (file push/pull), which this client does not implement. Both are called out in the help text rather than left to fail confusingly

[0.4.7] — 2026-07-25

Toolchain

• Zig 0.17.0-dev — the project, both build files and CI now target 0.17.0-dev.813+2153f8143. b.args became Run.addPassthruArgs() and Allocator.dupeZ became dupeSentinel. The nightly is pinned exactly, since dev tarballs are not permanent. Measured against 0.16.0 on this codebase: cold builds are ~36% slower (40.3s vs 29.6s), warm rebuilds and binary size are unchanged

Features — kuri-mobile

• ios tools — the command surface as data. Both the help text and --json render from one comptime table, so the dispatcher, the docs and the machine-readable listing cannot drift apart. Each entry carries name, aliases, positional shape, flags and a scope of simulator/device/simulator+device, so an agent can distinguish "not supported here" from "not configured yet" without probing
• doctor — checks the preconditions that otherwise get misdiagnosed at the point of use: developer-dir resolution, simctl presence, the macOS Accessibility grant, whether Simulator.app is running, booted device count, and adb reachability. Each failure prints its own remedy. Exits non-zero only on blocking problems
• wait-for-ui — block until an element appears, or disappears with --absent. Polls the accessibility tree rather than the clock, so it returns as soon as the UI is ready and fails loudly when it never is. Replaces sleep-and-hope, which is the main source of flakiness when driving a simulator
• find — print every element matching a label, with tap-ready centroids. Exits non-zero on no match, so it works directly as a test assertion
• batch — several actions in one process (tap:120,400 type:hi key:return wait:500 label:Done). The win is setup cost: device resolution and window focus happen once for the whole sequence rather than once per command
• gesture/drag — drag along a multi-point path, for motions whose shape matters and that a two-point swipe would flatten
• touch down|move|up — raw touch primitives for gestures the named commands don't cover
• key-sequence — press several named keys in order; the whole sequence is validated before anything is pressed, so a typo can't leave the UI half-driven
• App & device state — install, uninstall, erase, open-sim, set-location, reset-location, record-video (bounded, SIGINT-finalised so the file stays playable), and keyboard on|off to control whether the software keyboard appears

Fixes

• wait-for-ui overshot its timeout by ~4x — the deadline summed sleep intervals while ignoring the ~0.5s cost of each accessibility poll, so --timeout 3000 actually waited 13.5s. Now measured against a monotonic clock: 3s requested, 4.2s observed (one final poll)
• kuri-mobile --version reported 0.0.1 — it now tracks the release it ships in

[0.4.6] — 2026-07-25

Features — kuri-mobile iOS, driverless

• ios uitree — dumps the running app's real accessibility tree (labels, identifiers, values, device-pixel bounds) with no XCUITest bundle and no on-device agent. Works by enabling ApplicationAccessibilityEnabled inside the simulated runtime, which populates Simulator.app's host-side AX hierarchy
• Tap by label — --label targets an element by accessibility label, identifier or value (exact match, then substring) instead of raw coordinates
• Input — doubletap, longpress, swipe/scroll/pan, type, and key with modifier support (key return --cmd)
• Hardware buttons — ios button home|lock|volup|voldown|action driven through the host accessibility bridge
• Lifecycle & environment — ios background, ios privacy, ios ui, ios status-bar, ios log, ios list-devices

Fixes

• Release pipeline never published — build-macos-x86 pinned macos-13, a runner label GitHub has retired. The job never got a runner, sat queued until the 24-hour job limit cancelled it, and skipped both publish jobs. Every tag from v0.3.3 to v0.4.5 died this way. macOS now cross-compiles both arches on one macos-latest runner, and every job has an explicit timeout-minutes
• Release notes generation crashed — the notes/checksum script split on a literal \n (backslash-n) rather than a newline, raising IndexError on the first successful run. Never observed because the job had always been skipped
• kuri-mobile was never shipped — the root build had no reference to it, so releases omitted the binary that kuri execs as a sibling. It is now a path dependency of the root build and lands in every tarball
• kuri-mobile did not build off native macOS — an inferred error set collapsed to MacOsOnly on non-macOS targets, and cross-compiling to x86_64-macos could not find ApplicationServices. Both arches and both Linux arches now build
• Use-after-free in the accessibility walker — CFArrayGetValueAtIndex follows CoreFoundation's Get Rule, so releasing the parent array freed elements still in use. This crashed uitree and made ios button intermittently fail. Elements are now retained at the match point
• Silent tool-resolution failures — ios list-devices exited 0 printing nothing when xcode-select pointed at CommandLineTools. Developer tools are now resolved to absolute paths with checked exit status
• Version strings — kuri, kuri-browse and kuri-fetch reported 0.4.1/0.4.0 regardless of the release they shipped in

Known limitations

• macOS assets are unsigned and un-notarized until Apple credentials are configured as repository secrets
• tap/swipe/type/uitree remain Simulator-only; real devices need XCUITest
• ios privacy cannot reset camera authorization — simctl privacy has no camera service

[0.4.5] — 2026-05-27

Fixes

• CDP session auto-recovery — When Chrome performs a cross-process renderer swap (e.g., navigating to Instagram), it detaches the old CDP target and assigns a new webSocketDebuggerUrl. Kuri now detects dead clients, re-fetches /json/list, and reconnects automatically instead of silently failing (#172)
• Dead client detection — CdpClient now tracks whether its WebSocket connection is permanently broken, distinguishing transient disconnects from stale target URLs

[0.4.4] — 2026-05-24

Features — beyond parity (142 HTTP endpoints)

• Action caching — /cache/set, /cache/get, /cache/clear, /cache/list — cache ref mappings so repeated workflows skip LLM inference entirely
• Set-of-Marks screenshot — /screenshot/som — screenshot with numbered bounding-box overlays on interactive elements for vision models
• Hybrid snapshot — /snapshot?include_screenshot=true — a11y tree + screenshot in one call for multimodal agents
• Smart diff — /snapshot/changes — returns only added/removed lines since last snapshot (10-100x fewer tokens)
• Recording export — /recording/export — convert recorded actions into replayable /batch JSON

[0.4.3] — 2026-05-24

Features — full agent-browser parity (135 HTTP endpoints)

• Batch execution — POST /batch runs multiple commands in one HTTP call, returns array of results
• Enhanced wait — /wait now supports text, url, state=networkidle, visible=true params; new /wait/function and /wait/download
• Element state queries — /element/state?ref=e0&check=visible for quick boolean checks without full snapshot
• Semantic locators — /find-element?text=Submit / ?role=button / ?label=Email / ?placeholder=Search / ?testid=login-btn
• Page state — /page/state lightweight observation (48 tokens vs 2,124 for full snap)
• Dialog handling — /dialog/auto, /dialog/accept, /dialog/dismiss
• Mouse control — /mouse/move, /mouse/down, /mouse/up, /mouse/wheel
• Touch events — /tap, /swipe
• Clipboard — /clipboard/read, /clipboard/write
• Form operations — /clear, /selectall, /setvalue, /multiselect
• Element queries — /boundingbox, /getattribute, /inputvalue, /evalhandle
• Emulation — /timezone, /locale, /permissions
• React inspection — /react/tree, /react/inspect, /react/renders, /react/suspense
• Recording — /recording/start, /recording/stop
• Performance — /vitals (LCP, CLS, FID, TTFB, FCP, domInteractive)
• Navigation — /pushstate, /bringtofront, /frame, /mainframe
• Injection — /addstyle, /expose, /initscript/remove, /setcontent
• Network — /request/detail, /response/body, /download
• Diff — /diff/url (compare two URLs side by side)

Token efficiency

• 7-12% fewer tokens than agent-browser on real pages (@e0 vs [ref=e0])
• /page/state is 44x lighter than full snapshot (48 vs 2,124 tokens on Google Flights)

[0.4.2] — 2026-05-24

Fixes

• React/Vue click compatibility — click, check, and uncheck actions now use CDP Input.dispatchMouseEvent (mousePressed/mouseReleased) instead of DOM .click(), producing trusted isTrusted:true events that React 18/19 event delegation recognizes (#164)
• React/Vue type/fill compatibility — type and fill actions now use per-character CDP Input.dispatchKeyEvent instead of setting target.value directly, so React controlled inputs fire onChange correctly (#164)
• Click coordinate bug — fixed JS operator precedence in bounding-rect center calculation that caused r.y + r.height/2 to be string-concatenated instead of added
• Server type/fill default — realistic mode (CDP key events) is now the default for the HTTP server /action endpoint; opt out with realistic=false

Testing

• React e2e fixture — added test/react-form.html for validating form fill + click against React 18 controlled components

[0.4.1] — 2026-05-23

Release and install fixes

• Installer parsing fix — Adds --color=never to installer grep calls so ANSI escapes do not corrupt macOS sed parsing.
• README install polish — Moves the copyable install command and direct download links to the top of the README for easier release consumption.
• Release metadata sync — Runtime version string and npm package metadata aligned for the v0.4.1 patch release.

[0.4.0] — 2026-05-22

kuri-browser — Standalone Browser Engine

• Native rendering engine — Full DOM tree, CSS cascade with layout/paint, real <table> layout, font shorthands (border/padding/margin/list-style), text metrics with per-character glyph widths calibrated against headless Chrome
• QuickJS runtime — JavaScript evaluation, fetch/XHR bridge, cookie-aware navigation state, form extraction, session-backed form submission, HAR capture for browser flows
• CDP server — Minimal CDP WebSocket router, CDP discovery server, compressed screenshot fallback, parsed DOM selectors
• Agent actions — Click, type, snap, scroll, navigate, eval, back/forward via CDP-compatible commands
• Parity tracking — Pixel parity benchmark harness vs Chrome, example.com parity tracking (98.45% wrapper / 86.37% direct), per-char glyph width calibration

kuri-mobile — iOS + Android Device Control

• Zig-native device automation — Driverless: no on-device app, no Bun, no Gradle, no Xcode-time builds
• Android — ADB host protocol, XML UI tree parser, device listing, tap/swipe/type via input commands
• iOS — Simulator control via xcrun simctl, real device listing via usbmuxd, CGEvent-based tap/swipe/pan/type into Simulator.app

Server

• Bearer-token API auth — All endpoints protected by configurable bearer token, new kuri token CLI command for token management
• Signal-safe Chrome lifecycle — Chrome process shutdown uses signal-safe paths, preventing orphan processes

CI

• Startup smoke tests — CI now validates the bearer-token authentication wall
• Mobile skills discovery — Help/version regression guard for kuri-mobile CLI

[0.3.3] — 2026-04-25

Fixes

• Auth profile reliability — macOS keychain-backed auth profiles now resolve security correctly, and profile metadata round-trips escaped JSON safely
• Session persistence safety — bridge export/import now uses real JSON serialization/parsing instead of fragile string scanning
• Redirect and localhost hardening — URL validation now normalizes localhost aliases and re-validates redirect hops in both HTTP fetch paths
• CDP stability — stale buffered events no longer satisfy later waitForEvent() calls, and unsupported external CDP endpoint shapes are rejected up front
• Packaging correctness — HAR status/duration output is fixed, Chrome binary discovery checks PATH, and the npm installer rejects unsupported platforms instead of treating Windows as Linux

Release

• Notarized macOS artifacts in GitHub Releases — tagged releases now mirror the signed/notarized macOS tarballs alongside the self-managed release channel

[0.3.2] — 2026-04-24

Release channel

• Self-managed stable channel — installers and manifests now resolve binaries from the release-channel branch instead of GitHub Releases
• Channel-only release flow — tag publishing updates the raw GitHub channel manifest and asset paths without creating a GitHub Release entry
• macOS notarization kept in path — stable macOS tarballs remain signed and notarized, with raw GitHub download URLs exposed directly in the README and channel manifest

[0.3.1] — 2026-04-23

Maintenance

• Zig 0.16 migration stabilization — build, test, and startup paths updated for Zig 0.16 across local and GitHub Actions environments
• CI portability fixes — Linux libc linking, Chrome startup, and validator compatibility regressions resolved
• Benchmark refresh — README benchmark section updated with a fresh kuri rerun from bench/token_benchmark.sh
• Version sync — runtime strings, package metadata, and docs aligned to 0.3.1

[0.3.0] — 2026-03-20

Human Copilot Mode

• open [url] — one command to launch visible Chrome with CDP and auto-attach. The human sees the browser, the agent rides alongside. No headless, no bot detection issues.
• HEADLESS=false — kuri server mode now supports visible Chrome. Default remains headless for backward compat.
• stealth — anti-bot patches (UA override, navigator.webdriver=false, fake plugins). Persists across commands via session.

Agent-Friendly Output

• All commands now return clean, flat JSON instead of raw CDP responses:
  • go → {"ok":true,"url":"..."}
  • click → {"ok":true,"action":"clicked"}
  • eval → raw value (no triple-nested JSON)
  • text → real newlines (not escaped \n)
  • back/forward/reload/scroll → {"ok":true}
• Agents no longer need jq '.result.result.value' to parse output.

Popup & Redirect Following

• grab <ref> — click + follow popup redirects in the same tab. Hooks both window.open and dynamically created <form target="_blank"> (Google Flights pattern).
• wait-for-tab — poll for new tabs opened by the page.
• Tested end-to-end: Google Flights → Scoot booking page landed successfully.

Compact Snapshot (20x token reduction)

• Default snap output is now compact text-tree: role "name" @ref
• Noise roles filtered by default (none/generic/presentation/ignored)
• --interactive mode for agent loops (~1,927 tokens on Google Flights)
• --json flag restores old JSON format for backward compat

Token Benchmark

• Full workflow benchmark: go→snap→click→snap→eval
• kuri: 4,110 tokens vs agent-browser: 4,880 tokens — 16% savings per cycle
• Reproducible: ./bench/token_benchmark.sh [url]

Security Testing

• cookies — list with Secure/HttpOnly/SameSite flags
• headers — security response header audit (CSP, HSTS, X-Frame-Options)
• audit — full security scan (HTTPS + headers + JS-visible cookies)
• storage — dump localStorage/sessionStorage
• jwt — scan all storage + cookies for JWTs, base64-decode payloads
• fetch — authenticated fetch from browser context (uses session cookies + extra headers)
• probe — IDOR enumeration: probe https://api.example.com/users/{id} 1 100
• set-header / clear-headers / show-headers — persist auth headers across commands

Install

• curl -fsSL https://raw.githubusercontent.com/justrach/kuri/main/install.sh | sh
• bun install -g kuri-agent / npm install -g kuri-agent
• GitHub release workflow with optional Apple notarization (add APPLE_* secrets)

CI

• Fixed QuickJS Debug-mode crash on Linux (-Doptimize=ReleaseSafe in CI)

[0.2.0] — 2026-03-17

kuri-agent CLI

• Scriptable Chrome automation via CDP — stateless, one command per invocation
• Session persistence at ~/.kuri/session.json (cdp_url, refs, extra_headers)
• Commands: tabs, use, go, snap, click, type, fill, select, hover, focus, scroll, viewport, eval, text, shot, back, forward, reload
• Accessibility tree snapshots with ref-based element targeting (@e0, @e1, ...)

Compact Snapshot Format

• Text-tree format: role "name" @ref — replaces verbose JSON
• Noise filtering: skip none/generic/presentation roles
• --interactive / --semantic / --all / --json / --text flags

[0.1.0] — 2026-03-14

Initial Release

• kuri — CDP HTTP API server (Chrome automation, a11y snapshots, HAR recording)
• kuri-fetch — standalone fetcher with QuickJS JS engine, no Chrome needed
• kuri-browse — interactive terminal browser (navigate, follow links, search)
• 230+ tests, 4-target cross-compilation (macOS/Linux × arm64/x86_64)
• Zero Node.js dependencies, 464 KB server binary