Upstream 2026.03.12

[handstandsam]

Summary

• Multi-Agent V3 features: Exception handling, reflection/self-correction, task decomposition, cross-app memory, enhanced recording, and progress reporting (inspired by Mobile-Agent-v3 research)
• New modules: Added trailblaze-accessibility, trailblaze-accessibility-app, trailblaze-capture, trailblaze-compose, trailblaze-playwright, and example apps (android-sample-app, compose-desktop, playwright-native)
• Android on-device support: New documentation and tooling for on-device execution
• Trail & Blaze architecture: "Blaze once, trail forever" workflow — AI explores dynamically, then replays deterministically with zero LLM cost
• Documentation updates: Updated CLI docs (configurable ports, mcp command, run directory support, agent selection), new benchmarks, expanded MCP docs, and new tool reference docs (pasteClipboard, requestViewHierarchyDetails, setActiveToolSets, setClipboard)
• Python SDK: Added Python SDK with venv under sdks/python/
• Expanded .gitignore: Comprehensive ignore rules for Python venv artifacts
• New trails and test fixtures: Extensive new trail files for Android host, on-device, and emulator testing scenarios

Test plan

• Verify existing trails continue to pass on Android emulator
• Test new Multi-Agent V3 features (exception handling, reflection, task decomposition)
• Validate new modules build successfully (trailblaze-accessibility, trailblaze-compose, trailblaze-playwright, etc.)
• Test trailblaze run with directory argument
• Verify MCP server starts correctly with trailblaze mcp
• Run Playwright native benchmarks