The tab list from Playwright MCP uses '- N: (current) [title](url)' format,
but extractTabEntries only matched 'Tab N ...' format. This caused
_initialTabIdentities to always be empty, so tabs were never cleaned up.
Now supports both formats.
- Add isCdpApiAvailable() to verify CDP HTTP JSON API before using endpoint
- Chrome's chrome://inspect#remote-debugging writes DevToolsActivePort but
its CDP endpoint is incompatible with Playwright connectOverCDP (init
succeeds but all tool calls hang silently)
- Add CDP readiness probe in connect() to catch unresponsive endpoints
- Skip tab cleanup in close() for CDP mode (no bridge tabs to clean)
- Add PlaywrightMCPMode tracking for extension vs CDP lifecycle policy
- search.ts: Added deduplication break condition so if the API returns duplicate pages, it stops instead of looping forever.
- search.ts: Removed unsafe Promise.race around page.goto which caused 'context closed' errors.
- browser.ts: Added robust static cleanup hooks to guarantee process and MCP lock destruction on crashed or ctrl+c'd runs.
- The API endpoint only returns 15 results per page.
- Added a loop to fetch multiple pages until the requested limit is reached.
- Added random 1-3 second delays between page fetches to avoid rate limits.
- BOSS Zhipin aggressively detects CDP connections and kills pages
- Completely rewrote boss/search.ts to use browser: false (bypassing Playwright)
- Implemented native Python subprocess to extract cookies via browser-cookie3
- Falls back to boss-cli ~/.config/boss-cli/credential.json if available
- Uses direct Node.js fetch with anti-detection headers to avoid bans
- Add TCP port reachability check to handle stale DevToolsActivePort files
- Make discoverChromeEndpoint() async with port probe before returning
- Change connection priority: extension mode is default, CDP is opt-in
- OPENCLI_CDP_ENDPOINT: explicit CDP endpoint
- OPENCLI_USE_CDP=1: enable DevToolsActivePort auto-discovery
- Default: --extension mode (Playwright MCP Bridge)
- Add verbose logging for CDP discovery when -v flag is used
- Fix indentation inconsistency in connect() method
- Update README.md, README.zh-CN.md, SKILL.md with new connection docs
- Added forceExtension flag to CliCommand, CliOptions, browserSession, and PlaywrightMCP.connect()
- BOSS Zhipin detects CDP mode and drops connections; marked boss/search with forceExtension: true
- Also supports OPENCLI_FORCE_EXTENSION=1 env var for global override
- Fixed async discoverChromeEndpoint() placement outside Promise constructor
- Enhanced boss search with city names, experience, degree, salary, etc.
Chrome 144+ allows enabling remote debugging from chrome://inspect
without any command-line flags (--remote-debugging-port). Chrome writes
the active port and browser GUID to a DevToolsActivePort file in the
user data directory.
This change reads that file to auto-discover the CDP WebSocket endpoint,
eliminating the need for the Playwright MCP Bridge browser extension in
most cases.
Connection priority:
1. OPENCLI_CDP_ENDPOINT env var (manual override)
2. DevToolsActivePort auto-discovery (Chrome/Edge, cross-platform)
3. --extension mode fallback (original behavior)
Supports Windows, macOS, and Linux with Chrome, Edge, and Chromium.
Co-authored-by: kensei <backtime1993@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
- 30+ Chinese city names → code mapping (e.g. --city 杭州)
- Experience filter: 应届/1-3年/3-5年/5-10年/10年以上
- Degree filter: 大专/本科/硕士/博士
- Salary filter: 3K以下/3-5K/5-10K/10-20K/20-50K/50K以上
- Page number support for pagination
- Area now shows city·district·business district
- Skills column added to output
The manifest builder now statically parses TS source files to extract
args, description, columns, domain, and strategy metadata. Previously
TS adapters had empty args in the manifest, causing commander to reject
their options (e.g. --keyword) before the lazy module could load.
Also includes the xiaohongshu search DOM fallback fix from earlier.
Co-authored-by: jack4world <jack4world@users.noreply.github.com>
- Remove xiaohongshu/me (profile page async loading too unreliable)
- CLI-CREATOR: add 4 lessons to 常见陷阱 (stale dist YAML, TS IIFE
format, async polling, YAML vs TS threshold)
- Fix build: clean-yaml before copy to prevent stale overrides
- Fix docs: MCP token is configured via MCP config, not env var
- Update all command lists (README, README.zh-CN, SKILL.md)
- browser.ts: close all tabs opened during session (not just extension tabs)
uses _initialTabCount to restore Chrome to pre-command state
- README/README.zh-CN: add ⚠️ login requirement warning with examples
- SKILL.md: add login note and tab auto-close mention
- me: personal profile info (navigate + evaluate from Pinia store)
- notifications: browse notifications by type (mentions/likes/connections)
uses tap pipeline to call getNotification store action
- removed note, comments, user-posts (store action names unverified)
All 4 commands use YAML tap pipeline with Pinia store interception:
- me: personal profile info
- note --id: note detail
- comments --id: note comments
- user-posts --uid: user's published notes
- browser.ts: strip '### Ran Playwright code' suffix from MCP evaluate results
to correctly parse JSON data (was causing all evaluate-based commands to fail)
- browser.ts: auto-close chrome-extension:// tabs on session close
- pipeline.ts: add JSON string auto-parse fallback for evaluate step
- v2ex: add strategy:public + browser:false (public API, no browser needed)
- v2ex: simplify to use fetch step instead of evaluate
Tested: bilibili hot/search ✅, zhihu hot ✅, v2ex hot ✅,
hackernews top ✅, github search ✅
jackwener