Release 1.3.0

Author: sharakusatoh

README.md

@@ -2,14 +2,15 @@
 
 AutoCruise CE is a Windows desktop automation app powered by Codex App Server and ChatGPT sign-in. It observes the current desktop, asks Codex for the next action, executes through Windows automation and input backends, and continues until the task is complete or the user stops it.
 
-Current source and packaged release version: `1.2.0`
+Current source and packaged release version: `1.3.0`
 
 The project is experimental. Verify important operations yourself before relying on the result in a real workflow.
 
 ## What It Does
 
 - Runs natural-language desktop tasks on Windows.
 - Uses Codex App Server as the only AI runtime in this edition.
+- Uses `gpt-5.5` as the fixed Codex model.
 - Prefers structured Windows automation, then direct Win32 input, optional browser adapters, and visual fallback.
 - Supports manual runs, scheduled runs, pause/resume, stop, thread history, screenshots, and prompt profiles.
 - Loads model context from the constitution, the selected system prompt, and custom instruction files.
@@ -20,7 +21,7 @@ AutoCruise CE is distributed as a portable Windows package. There is no installe
 
 Download the latest portable archive from GitHub Releases:
 
-- [AutoCruiseCE-portable-1.2.0.zip](https://github.com/sharakusatoh/autocruise/releases/download/v1.2.0/AutoCruiseCE-portable-1.2.0.zip)
+- [AutoCruiseCE-portable-1.3.0.zip](https://github.com/sharakusatoh/autocruise/releases/download/v1.3.0/AutoCruiseCE-portable-1.3.0.zip)
 
 To run it:
 
@@ -89,7 +90,7 @@ build_windows.bat
 This creates:
 
 - `release\AutoCruiseCE\AutoCruiseCE.exe`
-- `release\AutoCruiseCE-portable-1.2.0.zip`
+- `release\AutoCruiseCE-portable-1.3.0.zip`
 
 `release/` is intentionally excluded from Git tracking. Publish the zip through GitHub Releases.
 

build_windows.bat

@@ -10,7 +10,7 @@ set "BUILD_ROOT=%~dp0build"
 set "BUILD_DIR=%~dp0build\pyinstaller"
 
 for /f %%i in ('python -c "import sys; sys.path.insert(0, r'%~dp0src'); from autocruise.version import APP_VERSION; print(APP_VERSION, end='') "') do set "APP_VERSION=%%i"
-if "%APP_VERSION%"=="" set "APP_VERSION=1.2.0"
+if "%APP_VERSION%"=="" set "APP_VERSION=1.3.0"
 
 if not exist "%RELEASE_DIR%" mkdir "%RELEASE_DIR%"
 

constitution/constitution.md

@@ -1,18 +1,4 @@
 <Constitution>
-
-# AutoCruise Constitution
-
 ## Core Goal
-
 Finish the user's stated Windows task with autonomous judgment and steady progress.
-
-## Required Principles
-
-- Choose the action that best advances the current goal.
-- Keep acting while the goal remains reachable.
-- Re-observe when the screen state is uncertain or has changed.
-- If a path stalls, retry, adjust focus, or switch to another viable path.
-- Use the most reliable available control path, including structured automation, keyboard input, mouse input, browser control, or visual guidance.
-- Keep enough logs and learning notes to explain what was tried, what changed, and what worked.
-
 </Constitution>

docs/architecture.md

@@ -13,15 +13,16 @@ AutoCruise CE is a Windows desktop application for autonomous GUI operation. The
 - PowerShell is used only for the Microsoft UI Automation client layer where .NET UIA APIs are the most direct Windows path.
 - PyInstaller produces the portable Windows folder used for releases.
 
-### Why Structured Automation First
+### Why Smart Windows Operator First
 
-Desktop automation is not reliable if every step depends on image coordinates. AutoCruise CE therefore prefers structured adapters before vision:
+Desktop automation is not reliable if every step depends on screenshots and coordinates. AutoCruise CE therefore chooses the richest direct control surface available before falling back to visual input:
 
-1. UIA for native Windows controls, element properties, and control patterns.
-2. Win32 for screenshots, windows, pointer, keyboard, clipboard, and global hotkeys.
-3. Playwright only when an integration supplies a live browser page object.
-4. CDP DOM / Accessibility / Input domains only as a browser fallback.
-5. Vision for remaining areas such as canvases, custom-rendered controls, and coordinate-level drawing.
+- App-specific APIs and object models first. Microsoft Office tasks should use COM/Object Model access for workbooks, cells, documents, messages, calendars, selections, and attachments before touching the UI.
+- Browser automation for Edge, Chrome, Chromium, and web apps. Playwright locators are preferred, with CDP DOM / Runtime / Network / Input / Event domains as targeted fallback.
+- PowerShell CIM/WMI and native management cmdlets for OS and administration data such as processes, services, devices, network state, registry, installed software, and settings.
+- UIA for normal Windows desktop apps without a richer app API.
+- MSAA or targeted Win32 messages for legacy controls when UIA is weak and the exact control/message is known.
+- Vision, OCR, screenshots, raw keyboard, mouse, and coordinates only as the final fallback.
 
 Playwright and browser binaries are optional and are not bundled in the standard package.
 
@@ -55,6 +56,7 @@ Playwright and browser binaries are optional and are not bundled in the standard
 - `infrastructure/windows/*`
 - `infrastructure/browser/*`
 - Codex connection, JSON/JSONL/YAML storage, screenshot capture, window enumeration, UIA client layer, Win32 input execution, optional Playwright/CDP adapters, and visual guidance.
+- Codex model selection is fixed to `gpt-5.5`; stored provider settings are normalized to that model before use.
 
 ## State Machine
 
@@ -84,16 +86,15 @@ Rules:
 
 1. Interpret the user goal.
 2. Select the constitution, selected system prompt, and custom instruction files.
-3. Capture a screenshot and visible window state.
-4. Query UIA for root, focused element, active-window descendants, and target candidates.
-5. Add optional Playwright/CDP state when a connected browser page is available.
-6. Build the observation payload for Codex.
-7. Ask Codex for the next action.
-8. Re-observe in `PRECHECK`.
-9. Resolve the target through UIA / browser adapter / visual target fallback.
-10. Execute one action through the best available backend.
-11. Re-observe in `POSTCHECK`.
-12. Validate visible progress, replan, complete, or stop with an issue record.
+3. Capture the active Windows state, structured automation state, and screenshot fallback evidence.
+4. Query the direct-control stack: app object models, browser automation, OS management APIs, UIA, and legacy control paths where available.
+5. Build the observation payload for Codex.
+6. Ask Codex for the next action.
+7. Re-observe in `PRECHECK`.
+8. Resolve the target through the best available direct backend before visual fallback.
+9. Execute one action through the best available backend.
+10. Re-observe in `POSTCHECK`.
+11. Validate visible progress, replan, complete, or stop with an issue record.
 
 ## Automation Interface
 
@@ -138,16 +139,18 @@ If locator operations fail and CDP is available, the adapter can use CDP `DOM`,
 
 ## Prompt Context Model
 
-Priority order:
+Prompt sources:
 
-1. Constitution
-2. Session mission
-3. Selected system prompt
-4. User custom prompt and custom prompt files
-5. Runtime observation
-6. Recent execution context
+- Constitution
+- Selected system prompt
+- User custom prompt and custom prompt files
 
-No other bundled prompt-source categories are loaded into the model context in this edition.
+Runtime inputs:
+
+- Current session mission
+- Current screen observation
+
+Session history, thread history, audit logs, execution logs, and learning-memory sources are not loaded into the model context in this edition. Each Codex App Server call starts a fresh thread.
 
 ## Logging and Storage
 
@@ -170,7 +173,7 @@ The settings screen includes:
 - Pause and stop hotkeys
 - Codex App Server status
 - ChatGPT sign-in / sign-out
-- Codex model
+- Fixed Codex model: `gpt-5.5`
 - Reasoning effort
 - Planning response size
 - Screenshot retention

docs/release_qa_checklist.md

@@ -1,6 +1,6 @@
 # AutoCruise CE Release QA Checklist
 
-Use this checklist before shipping a Windows build. Record the build path, date, tester, and result in diagnostics or the release QA memo.
+Use this checklist before shipping a Windows build. Record the build path, date, tester, and result in diagnostics or the release QA record.
 
 ## Build
 
@@ -16,12 +16,12 @@ Use this checklist before shipping a Windows build. Record the build path, date,
 
 - Open Settings and confirm Codex App Server status is visible.
 - Sign in with ChatGPT and run the connection test.
-- Confirm model, reasoning effort, and planning response size can be saved and restored.
+- Confirm the fixed Codex model displays as `gpt-5.5`, and reasoning effort plus planning response size can be saved and restored.
 - Confirm Japanese language selection persists after restart.
 
 ## Desktop Operation
 
-- Run `ペイントを開いて、簡単な猫の絵を描いてください。`.
+- Run a Paint smoke task that opens Paint and draws a simple cat picture.
 - Confirm Paint launches through a direct Windows path such as Run, visible launcher, or search.
 - Confirm the agent waits for the Paint window and canvas.
 - Confirm click, drag, and curve-like multi-point drawing work on the canvas.