Upstream sync through 2026.03.18

[handstandsam]

Summary

Major additions:

New platform support

• Playwright module (trailblaze-playwright) with native browser and Electron app automation, including screen state, ARIA snapshots, tools, and recording support
• Compose Desktop modules (trailblaze-compose, trailblaze-compose-target) for Jetpack Compose desktop UI testing via semantic tree mapping and RPC
• On-device accessibility service (trailblaze-accessibility) for running Trailblaze directly on Android without a host connection
• Video/logcat capture module (trailblaze-capture) for recording test sessions

Agent architecture

• Blaze agent with goal planning, memory, reflection, and progress tracking nodes
• Two-tier MCP agent with subagent orchestration, local LLM sampling, and dynamic tool set management
• TrailblazeNode abstraction with selectors, hit testing, and compact view hierarchy formatting
• Direct MCP agent for tool execution without HTTP round-trips

MCP server expansion

• Multi-device session management and device claim registry
• Progress session tracking with subscribe/poll support
• Async run endpoint, config/step/trail file tool sets
• MCP resource providers and content builder utilities

CLI and desktop app

• mcp install command and MCP proxy for IDE integration
• CLI report generator and describe subcommand
• Desktop app: MCP tab, recording tab, environment picker, session timeline/progress views, interactive device composable
• Desktop build workflow, ProGuard config, git-based versioning

Android sample app

• Full Compose sample app with taps, swipes, forms, lists, catalog, and settings screens
• Trail definitions for android-host, android-emulator, and connected-device configurations

New tools

• pasteClipboard, setClipboard, setActiveToolSets, requestViewHierarchyDetails
• clearAppData, launchApp, assertNotVisibleWithText, assertVisibleBySelector
• Android World benchmark tool set (contacts, files, SMS, SQLite, ADB shell)
• Removed pressBack in favor of pressKey

Trails and benchmarks

• Playwright native and Electron trail definitions (counter, forms, navigation, scroll, search)
• Compose Desktop trails (add-todo, manage-todos, widget-interactions)
• Android World benchmark trails (80+ scenarios across contacts, calendar, SMS, markor, recipes, expenses, system settings, media apps)
• Playwright native benchmark results

Documentation

• Expanded MCP and CLI documentation
• On-device Android MCP guide
• Host JVM unit test docs
• Architecture and getting started updates

Other

• install.sh script and scripts/trailblaze wrapper
• Dependency updates and build configuration changes
• Extensive test coverage across new modules