release: bump version to 0.1.48

Author: iFurySt

apps/OpenComputerUseLinux/main.go

@@ -19,7 +19,7 @@ import (
 	"time"
 )
 
-var version = "0.1.47"
+var version = "0.1.48"
 
 //go:embed runtime.py
 var linuxRuntimeScript string

apps/OpenComputerUseSmokeSuite/Sources/OpenComputerUseSmokeSuite/main.swift

@@ -35,7 +35,7 @@ final class MCPClient {
         _ = try request(method: "initialize", params: [
             "clientInfo": [
                 "name": "OpenComputerUseSmokeSuite",
-                "version": "0.1.47",
+                "version": "0.1.48",
             ],
             "capabilities": [:],
             "protocolVersion": "2025-03-26",

apps/OpenComputerUseWindows/main.go

@@ -16,7 +16,7 @@ import (
 	"time"
 )
 
-var version = "0.1.47"
+var version = "0.1.48"
 
 //go:embed runtime.ps1
 var windowsRuntimeScript string

docs/ARCHITECTURE.md

@@ -125,7 +125,7 @@
 - 开源版当前不复刻官方闭源实现里的 caller signing、私有 IPC、完整 overlay choreography 和 plugin 自安装逻辑。
 - 因为官方 `SkyComputerUseClient` 带有宿主侧 launch constraints，普通 stdio MCP client 在本机上可能被系统直接杀掉；如果要探测官方 bundled `computer-use`，`scripts/computer-use-cli` 的 app-server 模式现在只适合做工具清单和协议面观察。官方 `1.0.755` 的真实 tool call 还会经过 service-side sender authorization / active IPC client 追踪，外部 raw helper 即使走已签名 Codex binary，也可能返回 `Sender process is not authenticated`；需要真实使用官方工具时应走正常 Codex agent/tool 调用链，开源版则继续提供可直连的 `open-computer-use` MCP server。
 - 当前权限引导已经具备可运行 app、深链、拖拽辅助，以及一版更接近官方的 accessory panel 入场动画和返回 affordance；点击链路也已经补上独立 visual cursor、官方 asset fallback 和相对目标 window 的排序逻辑，并且在 overlay 可见期间会持续重申“排在目标 window 之上”，避免用户手动激活目标 app 后 cursor 被目标窗口重新盖住；但整体还没有完全复刻官方那套嵌入式 choreography / host 集成 / session approval 体验。
-- screenshot 当前通过 `ScreenCaptureKit` 捕获目标窗口，并直接以 MCP `image` content block 的 base64 PNG 返回，不再把普通 app 截图落盘到仓库或临时目录；单次 ScreenCaptureKit capture 会设置超时，超时后省略 image block 而不是卡住整个 `get_app_state`。
+- screenshot 当前通过 `ScreenCaptureKit` 捕获目标窗口，并以 MCP `image` content block 的 base64 PNG 返回，不再把普通 app 截图落盘到仓库或临时目录；编码前会按最大尺寸和目标字节数自适应缩小，避免复杂页面的大 PNG 触发 host 侧 MCP result 降级，同时 coordinate tools 继续按实际返回的 screenshot pixel 尺寸映射坐标；单次 ScreenCaptureKit capture 会设置超时，超时后省略 image block 而不是卡住整个 `get_app_state`。
 - 会话状态现在是进程内内存态，保存每个 app 最近一次 snapshot 和 element index 映射。
 
 ## 主要验证路径

docs/histories/2026-05/20260508-1508-bound-screenshot-result-size.md

@@ -0,0 +1,22 @@
+# Bound screenshot result size
+
+## 用户诉求
+
+用户发现 `open-computer-use` 的部分 tool call 在 Codex 历史记录里变成 `text` 下的 JSON 字符串，怀疑最近几次 commit 引入了额外包装，并提供了历史 session id 辅助定位。
+
+## 本次改动
+
+- **[Root cause]**: 确认 MCP server 原始 `tools/call` 响应是标准 `content: [text, image]`，嵌套只出现在 Codex 为了持久化超大 MCP result 而做的事件日志降级路径里。
+- **[Screenshot bound]**: macOS `ScreenCaptureKit` 窗口截图在编码为 PNG 前按最大尺寸和目标字节数自适应缩小，减少复杂页面触发 host 侧大结果降级的概率。
+- **[Regression tests]**: 增加截图压缩边界测试，覆盖大图会缩小、小图保持原尺寸。
+- **[Docs]**: 更新架构文档，说明截图仍以 MCP image block 返回，但会做尺寸/字节上限控制，coordinate tools 继续按实际返回的 screenshot pixel 尺寸映射。
+
+## 设计动机
+
+问题不是协议层重复包了一层，而是截图 PNG 过大时，Codex 会把整个 `CallToolResult` 序列化成一个 text preview 来保护 rollout storage。直接改 MCP shape 会破坏兼容性；更稳妥的修复是控制截图结果大小，让正常 `get_app_state` / action tool 返回保持扁平的 `text + image` 结构。
+
+## 影响文件
+
+- `packages/OpenComputerUseKit/Sources/OpenComputerUseKit/AccessibilitySnapshot.swift`
+- `packages/OpenComputerUseKit/Tests/OpenComputerUseKitTests/OpenComputerUseKitTests.swift`
+- `docs/ARCHITECTURE.md`

docs/histories/2026-05/20260508-1518-release-0.1.48.md

@@ -0,0 +1,39 @@
+# Release 0.1.48
+
+## 用户诉求
+
+用户要求提交本轮相关改动并推送发版，同时确认 Codex TOML 配置改回连接线上 `open-computer-use`，不要指向 dev 构建。
+
+## 本次改动
+
+- **[Version bump]**: 将 macOS / Linux / Windows runtime、插件 manifest、smoke suite、测试和 reverse-engineering helper 版本统一 bump 到 `0.1.48`。
+- **[Release notes]**: 更新用户可见 release notes，说明本版解决复杂窗口截图过大导致 Codex 历史事件降级成 JSON text preview 的问题。
+- **[Config check]**: 确认本机 Codex 配置中的 `open-computer-use` MCP server 使用线上命令 `open-computer-use mcp`，没有指向本地 dev binary。
+
+## 设计动机
+
+本轮修复已经影响公开 npm 包里的 macOS runtime 行为，且用户明确要求发版，因此按 release guide 走 patch release。版本源继续以插件 manifest 为准，并同步所有辅助工具与测试里的版本字符串，避免 tag 与 npm staging 包版本不一致。
+
+## 验证
+
+- `swift test`
+- `./scripts/run-tool-smoke-tests.sh`
+- `node ./scripts/npm/build-packages.mjs --out-dir dist/release/npm-staging-check`
+- `./scripts/build-cursor-motion-dmg.sh --configuration release --arch universal --version 0.1.48`
+- `node -p "require('./dist/release/npm-staging-check/open-computer-use/package.json').version"` -> `0.1.48`
+- `test -x dist/release/npm-staging-check/open-computer-use/dist/linux/arm64/open-computer-use`
+- `test -f dist/release/npm-staging-check/open-computer-use/dist/windows/arm64/open-computer-use.exe`
+- `node -e "if (require('./dist/release/npm-staging-check/open-computer-use/package.json').optionalDependencies) process.exit(1)"`
+- `ls dist/release/cursor-motion/CursorMotion-0.1.48.dmg`
+- `git diff --check`
+
+## 影响文件
+
+- `plugins/open-computer-use/.codex-plugin/plugin.json`
+- `packages/OpenComputerUseKit/Sources/OpenComputerUseKit/OpenComputerUseVersion.swift`
+- `apps/OpenComputerUseLinux/main.go`
+- `apps/OpenComputerUseWindows/main.go`
+- `apps/OpenComputerUseSmokeSuite/Sources/OpenComputerUseSmokeSuite/main.swift`
+- `scripts/computer-use-cli/main.go`
+- `scripts/computer-use-cli/README.md`
+- `docs/releases/feature-release-notes.md`

docs/releases/feature-release-notes.md

@@ -4,6 +4,7 @@
 
 | 日期 | 功能域 | 用户价值 | 变更摘要 |
 | --- | --- | --- | --- |
+| 2026-05-08 | MCP 结果大小控制 | Chrome / Electron 等复杂窗口返回截图时更不容易触发 Codex 的大结果降级，历史记录和上层工具消费里会继续保持正常的 `text + image` 结构，而不是把整个 MCP result 塞进 text 字符串。 | 发布 `0.1.48`，macOS `ScreenCaptureKit` 截图在编码为 MCP image block 前会按最大尺寸和目标字节数自适应缩小，并补测试锁住大图压缩、小图保真路径。 |
 | 2026-05-08 | Smoke 测试静默运行 | 开发者运行 `./scripts/run-tool-smoke-tests.sh` 时不再被内部 fixture 的橙色测试窗口打断，回归验证更安静，也不会让用户误以为桌面上弹出了额外应用。 | 发布 `0.1.47`，smoke suite 默认以 headless 模式启动内部 fixture，并保留 fixture 在 `list_apps` smoke 路径里的可发现性。 |
 | 2026-05-07 | 窗口不可用错误文案 | Lark / Chrome 等 no-window 场景下，开源版 `get_app_state` 的返回文本与官方 `computer-use` 当前输出一致，减少上层 prompt、测试和用户排查时的差异噪音。 | 发布 `0.1.46`，将 missing AX window 和 missing visible CGWindow 两条路径统一为 `Apple event error -10005: cgWindowNotFound`，并补单元测试锁住该官方风格文本。 |
 | 2026-05-07 | 可见窗口匹配 | `get_app_state` 在 app 没有可见窗口、窗口已最小化，或 AX 窗口无法匹配到 on-screen CGWindow 时会更接近官方 `computer-use` 的 `cgWindowNotFound` 语义，避免返回没有截图坐标基础的半残状态。 | 发布 `0.1.45`，基于官方 `1.0.770` 逆向字符串中的 `cgWindowNotFound`、`noWindowsAvailable`、`matchingWindowNotFound` 和 `AXWindowMiniaturized` 线索，要求真实 app snapshot 同时满足未最小化 `AXWindow` 与可见 `CGWindow`。 |

packages/OpenComputerUseKit/Sources/OpenComputerUseKit/AccessibilitySnapshot.swift

@@ -30,6 +30,9 @@ enum SnapshotMode {
 let accessibilityTreeMaxNodeCount = 1200
 let accessibilityTreeMaxDepth = 64
 let screenshotCaptureTimeout: TimeInterval = 5
+let screenshotResultMaxPNGBytes = 900_000
+let screenshotResultMaxDimension: CGFloat = 1280
+let screenshotResultMinScale: CGFloat = 0.25
 private let windowVisibilityRecoveryDelay: TimeInterval = 0.7
 private let axWebAreaRole = "AXWebArea"
 
@@ -412,11 +415,75 @@ private struct WindowCapture {
             return nil
         }
 
-        let bitmap = NSBitmapImageRep(cgImage: image)
-        return bitmap.representation(using: .png, properties: [:])
+        return boundedScreenshotPNGData(for: image)
     }
 }
 
+func boundedScreenshotPNGData(
+    for image: CGImage,
+    maxBytes: Int = screenshotResultMaxPNGBytes,
+    maxDimension: CGFloat = screenshotResultMaxDimension,
+    minScale: CGFloat = screenshotResultMinScale
+) -> Data? {
+    guard image.width > 0, image.height > 0, maxBytes > 0 else {
+        return nil
+    }
+
+    let original = pngData(for: image)
+    let largestDimension = CGFloat(max(image.width, image.height))
+    var scale = min(1, maxDimension / largestDimension)
+
+    if scale >= 1, let original, original.count <= maxBytes {
+        return original
+    }
+
+    var best = original
+    while scale >= minScale {
+        guard let resized = resizedCGImage(image, scale: scale),
+              let data = pngData(for: resized)
+        else {
+            break
+        }
+
+        best = data
+        if data.count <= maxBytes {
+            return data
+        }
+
+        scale *= 0.85
+    }
+
+    return best
+}
+
+private func pngData(for image: CGImage) -> Data? {
+    let bitmap = NSBitmapImageRep(cgImage: image)
+    return bitmap.representation(using: .png, properties: [:])
+}
+
+private func resizedCGImage(_ image: CGImage, scale: CGFloat) -> CGImage? {
+    let width = max(1, Int((CGFloat(image.width) * scale).rounded()))
+    let height = max(1, Int((CGFloat(image.height) * scale).rounded()))
+    let colorSpace = CGColorSpaceCreateDeviceRGB()
+    let bitmapInfo = CGImageAlphaInfo.premultipliedLast.rawValue
+
+    guard let context = CGContext(
+        data: nil,
+        width: width,
+        height: height,
+        bitsPerComponent: 8,
+        bytesPerRow: 0,
+        space: colorSpace,
+        bitmapInfo: bitmapInfo
+    ) else {
+        return nil
+    }
+
+    context.interpolationQuality = .medium
+    context.draw(image, in: CGRect(x: 0, y: 0, width: width, height: height))
+    return context.makeImage()
+}
+
 private final class AsyncResultBox<T>: @unchecked Sendable {
     var result: Result<T, Error>?
 }

packages/OpenComputerUseKit/Sources/OpenComputerUseKit/OpenComputerUseVersion.swift

@@ -1,6 +1,6 @@
 import Foundation
 
-public let openComputerUseVersion = "0.1.47"
+public let openComputerUseVersion = "0.1.48"
 
 public func resolvedOpenComputerUseVersion(bundle: Bundle = .main) -> String {
     if let version = bundle.object(forInfoDictionaryKey: "CFBundleShortVersionString") as? String,

packages/OpenComputerUseKit/Tests/OpenComputerUseKitTests/OpenComputerUseKitTests.swift

@@ -1,4 +1,5 @@
 import AppKit
+import ImageIO
 import XCTest
 @testable import OpenComputerUseKit
 
@@ -136,6 +137,29 @@ final class OpenComputerUseKitTests: XCTestCase {
         XCTAssertEqual(resolvedOpenComputerUseVersion(bundle: Bundle(for: Self.self)), openComputerUseVersion)
     }
 
+    func testBoundedScreenshotPNGDataShrinksLargeScreenshots() throws {
+        let image = try makeNoisyTestImage(width: 800, height: 600)
+        let data = try XCTUnwrap(boundedScreenshotPNGData(
+            for: image,
+            maxBytes: 50_000,
+            maxDimension: 320,
+            minScale: 0.05
+        ))
+        let size = try imageSize(in: data)
+
+        XCTAssertLessThanOrEqual(data.count, 50_000)
+        XCTAssertLessThanOrEqual(max(size.width, size.height), 320)
+    }
+
+    func testBoundedScreenshotPNGDataKeepsSmallScreenshotsAtOriginalSize() throws {
+        let image = try makeSolidTestImage(width: 32, height: 24)
+        let data = try XCTUnwrap(boundedScreenshotPNGData(for: image, maxBytes: 1_000_000, maxDimension: 320))
+        let size = try imageSize(in: data)
+
+        XCTAssertEqual(size.width, 32)
+        XCTAssertEqual(size.height, 24)
+    }
+
     func testToolDefinitionCount() {
         XCTAssertEqual(ToolDefinitions.all.count, 9)
     }
@@ -436,7 +460,7 @@ final class OpenComputerUseKitTests: XCTestCase {
 
     func testInitializeResponseContainsToolsCapability() throws {
         let server = StdioMCPServer(service: ComputerUseService())
-        let response = server.handle(line: #"{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","clientInfo":{"name":"test","version":"0.1.47"},"capabilities":{}}}"#)
+        let response = server.handle(line: #"{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","clientInfo":{"name":"test","version":"0.1.48"},"capabilities":{}}}"#)
         XCTAssertNotNil(response)
         XCTAssertTrue(response!.contains(#""name":"open-computer-use""#))
         XCTAssertTrue(response!.contains(#""tools":{"listChanged":false}"#))
@@ -445,7 +469,7 @@ final class OpenComputerUseKitTests: XCTestCase {
     func testInitializeResponseContainsComputerUseInstructions() throws {
         let server = StdioMCPServer(service: ComputerUseService())
         let response = try XCTUnwrap(
-            server.handle(line: #"{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","clientInfo":{"name":"test","version":"0.1.47"},"capabilities":{}}}"#)
+            server.handle(line: #"{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","clientInfo":{"name":"test","version":"0.1.48"},"capabilities":{}}}"#)
         )
         let data = try XCTUnwrap(response.data(using: .utf8))
         let json = try XCTUnwrap(JSONSerialization.jsonObject(with: data) as? [String: Any])
@@ -1420,4 +1444,59 @@ final class OpenComputerUseKitTests: XCTestCase {
         let length = max(hypot(dx, dy), 0.001)
         return CGVector(dx: dx / length, dy: dy / length)
     }
+
+    private func makeNoisyTestImage(width: Int, height: Int) throws -> CGImage {
+        var pixels = [UInt8](repeating: 0, count: width * height * 4)
+        for y in 0..<height {
+            for x in 0..<width {
+                let offset = (y * width + x) * 4
+                pixels[offset] = UInt8((x * 37 + y * 17) & 0xFF)
+                pixels[offset + 1] = UInt8((x * 11 + y * 43) & 0xFF)
+                pixels[offset + 2] = UInt8((x * 71 + y * 5) & 0xFF)
+                pixels[offset + 3] = 255
+            }
+        }
+
+        return try makeTestImage(width: width, height: height, pixels: pixels)
+    }
+
+    private func makeSolidTestImage(width: Int, height: Int) throws -> CGImage {
+        var pixels = [UInt8](repeating: 0, count: width * height * 4)
+        for index in stride(from: 0, to: pixels.count, by: 4) {
+            pixels[index] = 80
+            pixels[index + 1] = 140
+            pixels[index + 2] = 220
+            pixels[index + 3] = 255
+        }
+
+        return try makeTestImage(width: width, height: height, pixels: pixels)
+    }
+
+    private func makeTestImage(width: Int, height: Int, pixels: [UInt8]) throws -> CGImage {
+        let data = Data(pixels)
+        let provider = try XCTUnwrap(CGDataProvider(data: data as CFData))
+        let image = CGImage(
+            width: width,
+            height: height,
+            bitsPerComponent: 8,
+            bitsPerPixel: 32,
+            bytesPerRow: width * 4,
+            space: CGColorSpaceCreateDeviceRGB(),
+            bitmapInfo: CGBitmapInfo(rawValue: CGImageAlphaInfo.premultipliedLast.rawValue),
+            provider: provider,
+            decode: nil,
+            shouldInterpolate: false,
+            intent: .defaultIntent
+        )
+
+        return try XCTUnwrap(image)
+    }
+
+    private func imageSize(in data: Data) throws -> (width: Int, height: Int) {
+        let source = try XCTUnwrap(CGImageSourceCreateWithData(data as CFData, nil))
+        let properties = try XCTUnwrap(CGImageSourceCopyPropertiesAtIndex(source, 0, nil) as? [CFString: Any])
+        let width = try XCTUnwrap(properties[kCGImagePropertyPixelWidth] as? Int)
+        let height = try XCTUnwrap(properties[kCGImagePropertyPixelHeight] as? Int)
+        return (width, height)
+    }
 }