Replace `echo $MIDSCENE_MODEL_API_KEY` with `cat .env | grep ...`
since the CLI auto-loads .env via dotenv at runtime. The shell env
check always shows empty and confuses AI agents into manually
sourcing .env files. Also remove unnecessary setup verification
steps (node --version, xcodebuild, --help discovery).
Switch from --bridge flag to MIDSCENE_USE_BRIDGE=1 env var prefix for
all commands. This prevents AI agents from improvising command syntax
and ensures bridge mode is always activated correctly.
Remove setup verification steps that Claude would run before the actual
task. In bridge mode the user has already prepared the browser and
extension — just connect directly.
Individual CLI commands will never work for Spotlight, context menus,
and other transient UI because each command is a separate process and
macOS returns focus to the terminal on exit. Remove the individual
commands alternative and make act the only approach.
@midscene/web has two bins (midscene-web, midscene-playground) and
@midscene/ios has two bins (midscene-ios, midscene-ios-playground),
causing `npx @midscene/web` to fail with "could not determine
executable to run". Use `npx -p <pkg> <bin>` syntax instead.
- Rewrite browser SKILL.md to use @midscene/web instead of deprecated @midscene/cli
- Add `act` command documentation to all 5 platform skills
- Recommend `act` as primary approach for transient UI interactions
- Remove outdated Puppeteer fallback references from chrome-bridge
- Rename web-automation back to chrome-bridge (Bridge mode using @midscene/web CLI)
- Restore browser skill (Puppeteer mode using @midscene/cli)
- Add detailed "Handle Transient UI" sections to all platform skills with
concrete examples showing back-to-back command execution without screenshots
- Add "Be specific about UI elements" guidance to computer skill
- Add "When to Use" section to chrome-bridge with fallback guidance
Rewrite all SKILL.md files to use the new platform-specific CLI
binaries (midscene-android, midscene-ios, midscene-computer,
midscene-web) instead of the centralized @midscene/cli subcommand
architecture.
- Merge browser and chrome-bridge skills into unified web-automation
- Update command examples to use MCP tool commands directly
- Remove act/query/assert commands (AI agent now acts as the brain)
Update all 5 platform SKILL.md files to use the new CLI subcommand
architecture: `midscene do <command> -p <platform>` instead of
`midscene <platform> <command>`. Affects browser, chrome-bridge,
computer-automation, android-automation, and ios-automation skills.
Advise combining interactions with ephemeral UI (Spotlight, menus,
dropdowns, toasts, action sheets) into a single act command to prevent
them from disappearing between separate CLI invocations.
Consolidate web, computer, android, and ios plugins into one unified
plugin so users can install all skills with a single command:
/plugin install midscene-automation@midscene-marketplace
SKILL.md already contains all commands, examples, patterns, and
troubleshooting info needed by the AI. These extra files only
duplicate content and waste context window tokens.
setup.json files were static and never dynamically updated, duplicating
information already present in each skill's SKILL.md. Remove all 4
setup.json files and update references in SKILL.md and README.md.
Streamline the setup process to match agent-browse's approach:
- Remove manual pnpm install/build steps (handled by postinstall)
- Use MIDSCENE_MODEL_* environment variables instead of .env file
- Add .env.example with official Midscene configuration format
- Update documentation to show 3-step quick install
- Created plugin.json for Midscene automation with metadata.
- Added EXAMPLES.md detailing common automation tasks and workflows.
- Developed REFERENCE.md providing technical details and command specifications.
- Introduced SKILL.md outlining the skill's purpose, setup instructions, and best practices.
- Implemented setup.json to track installation prerequisites and setup completion status.
quanruzhuoxiu