v0.5.0 - Hardware-ready, memory-tested, multi-format bundles

Highlights

• Hardware detection in winagent doctor: CPU (Intel/AMD) + GPUs (NVIDIA/AMD/Intel); Ollama auto-routes CUDA/ROCm/Vulkan/CPU
• Memory stability test in CI (30-round vision loop; measured ~1.2MB, gate 30MB)
• Multi-format bundles: py3-none-any wheel + portable zip (unzip -> one-click install script) for Windows/macOS/Linux
• docs/INSTALL.md, one-click install scripts, release CI (wheel/sdist auto-build on tags)
• MCP server: UTF-8 I/O + per-message fault isolation

Assets: wheel / portable zip / sdist.

v0.4.0 - MCP Plugin Mode

Highlights

• MCP plugin mode: hardened stdio server (protocol negotiation, notifications, spec errors), 9 tools: look / click / type_text / key / act / run_scenario / doctor / screenshot / discover
• screenshot returns MCP native image content (hosts render directly)
• docs/PLUGINS.md: drop-in configs for ZCode / Claude Desktop / Cursor + security notes + troubleshooting
• Protocol conformance test runs in CI on Windows & Linux (xvfb)
• README revamped: capabilities, architecture, plugin quickstart

Full changelog: CHANGELOG.md

v0.3.0 - Cross-platform backends, perf config, CI matrix

Highlights

• Cross-platform hands: win32 (full) / x11 (XTest, CI-tested under xvfb) / macos (Quartz, awaiting real-device verification)
• Windows 10 / Python 3.10 compat; CI matrix: Windows 3.10+3.13 GUI smoke, Ubuntu xvfb XTest real input, macOS import smoke
• Performance config: image_format png|jpeg (jpeg ~10x smaller payloads), capture_monitor (single-display savings), typing_interval_ms, max_image_width tuning
• Model residency via server-side OLLAMA_KEEP_ALIVE (no per-request params)
• bench process ops on psutil (cross-platform); tasks support platforms: gating

Full changelog: CHANGELOG.md

v0.2.0 - Benchmark, Planner, MCP

Highlights

• Benchmark engine: 10 tasks (YAML setup/teardown/steps/verify), A/B verification, trace + screenshot per run
• Baseline v0.2: 7/10 tasks at 100%, 7 field findings (docs/baseline_v0.2.md)
• Planner: natural-language goal -> step JSON via local model
• MCP server: 6 tools for any harness
• winagent doctor: one-command environment self-check
• CI: ruff + Windows GUI smoke (real mouse/keyboard, no Ollama needed)
• IME-aware typing, session-state cleanup, tolerance-retry closed loop

Full changelog: CHANGELOG.md