- PR #1714: Replace tf-playwright-stealth with playwright-stealth
- PR #1721: Respect <base> tag in html2text for relative links
- PR #1719: Include GoogleSearchCrawler script.js in package data
- PR #1717: Allow local embeddings by removing OpenAI fallback
- Fix: Extract <base href> from raw HTML before head gets stripped
- Close duplicates: #1703, #1698, #1697, #1710, #1720
- Update CONTRIBUTORS.md and PR-TODOLIST.md
Pass the normalized absolute URL instead of the raw href to
can_process_url() in BFS, BFF, and DFS deep crawl strategies.
This ensures URL validation and filter chain evaluation operate
on consistent, fully-qualified URLs.
Fixes#1743
- Add TheRedRad to CONTRIBUTORS.md for PR #1694
- Document force_viewport_screenshot in API parameters reference
- Add viewport screenshot note in browser-crawler-config guide
- Add viewport-only screenshot example in screenshot docs
- Replace raw eval() in _compute_field() with AST-validated
_safe_eval_expression() that blocks __import__, dunder attribute
access, and import statements while preserving safe transforms
- Add ALLOWED_DESERIALIZE_TYPES allowlist to from_serializable_dict()
preventing arbitrary class instantiation from API input
- Update security contact email and add v0.8.1 security fixes to
SECURITY.md with researcher acknowledgment
- Add 17 security tests covering both fixes
Strip markdown code fences (```json ... ```) from LLM responses before
json.loads() in agenerate_schema(). Anthropic models wrap JSON output
in markdown fences when litellm silently drops the unsupported
response_format parameter, causing json.loads("") parse failures.
- Add _strip_markdown_fences() helper to extraction_strategy.py
- Apply fence stripping + empty response check in agenerate_schema()
- Separate JSONDecodeError for clearer error messages
- Add 34 tests: unit, real API integration (Anthropic/OpenAI/Groq
against quotes.toscrape.com), and regression parametrized
- Use class-level tracking keyed by normalized CDP URL
- All BrowserManager instances connecting to same browser share tracking
- For CDP connections, always create new pages (cross-connection page
sharing isn't reliable in Playwright)
- For managed browsers, page reuse works within same process
- Normalize CDP URLs to handle different formats (http, ws, query params)
When using create_isolated_context=False with concurrent crawls, multiple
tasks would reuse the same page (pages[0]) causing navigation race
conditions and "Page.content: Unable to retrieve content because the
page is navigating" errors.
Changes:
- Add _pages_in_use set to track pages currently being used by crawls
- Rewrite get_page() to only reuse pages that are not in use
- Create new pages when all existing pages are busy
- Add release_page() method to release pages after crawl completes
- Update cleanup paths to release pages before closing
This maintains context sharing (cookies, localStorage) while ensuring
each concurrent crawl gets its own isolated page for navigation.
Includes integration tests verifying:
- Single and sequential crawls still work
- Concurrent crawls don't cause race conditions
- High concurrency (10 simultaneous crawls) works
- Page tracking state remains consistent
- Add storage_state.json to all KEEP_PATTERNS levels
- This file contains unencrypted cookies in Playwright format
- Critical for cross-machine profile portability (local -> cloud)
- Add --password-store=basic and --use-mock-keychain flags when creating
profiles to prevent OS keychain encryption of cookies
- Without this, cookies are encrypted with machine-specific keys and
profiles can't be transferred between machines (local -> cloud)
Also adds direct CLI commands for profile management:
- crwl profiles create <name>
- crwl profiles list
- crwl profiles delete <name>
The interactive menu (crwl profiles) still works as before.
- Add Section 11 "Cancellation Support for Deep Crawls" to deep-crawling.md
- Document should_cancel callback, cancel() method, and cancelled property
- Include complete example for cloud platform job cancellation
- Add docs/examples/deep_crawl_cancellation.py with 6 comprehensive examples
- Update summary section to mention cancellation feature
- Add should_cancel callback parameter to BFS, DFS, and BestFirst strategies
- Add cancel() method for immediate cancellation (thread-safe)
- Add cancelled property to check cancellation status
- Add _check_cancellation() internal method supporting both sync/async callbacks
- Reset cancel event on strategy reuse for multiple crawls
- Include cancelled flag in state notifications via on_state_change
- Handle callback exceptions gracefully (fail-open, log warning)
- Add comprehensive test suite with 26 tests covering all edge cases
This enables external callers (e.g., cloud platforms) to stop a running
deep crawl mid-execution and retrieve partial results.
Replace hardcoded version string with import from crawl4ai.__version__
to ensure health endpoint reports correct version.
Fixes#1686
Reviewers: @chansearrington
Replace hardcoded version string with import from crawl4ai.__version__
to ensure health endpoint reports correct version.
Fixes#1686
Reviewers: @chansearrington
When using raw: URLs without a base_url, redirected_url was incorrectly
set to the entire raw HTML string (potentially 300KB+) instead of None.
Changes:
- async_crawler_strategy.py: Don't fall back to url for raw:/file:// URLs
in fast path, browser path, and HTTP strategy
- async_crawler_strategy.py: Skip page.url assignment for local content
(would return "about:blank")
- async_webcrawler.py: Don't fall back to url for raw: URLs in crawl
result and cached result paths
- Add comprehensive test suite for redirected_url handling
* Fix: Use correct URL variable for raw HTML extraction (#1116)
- Prevents full HTML content from being passed as URL to extraction strategies
- Added unit tests to verify raw HTML and regular URL processing
Fix: Wrong URL variable used for extraction of raw html
* Fix#1181: Preserve whitespace in code blocks during HTML scraping
The remove_empty_elements_fast() method was removing whitespace-only
span elements inside <pre> and <code> tags, causing import statements
like "import torch" to become "importtorch". Now skips elements inside
code blocks where whitespace is significant.
* Refactor Pydantic model configuration to use ConfigDict for arbitrary types
* Fix EmbeddingStrategy: Uncomment response handling for the variations and clean up mock data. ref #1621
* Fix: permission issues with .cache/url_seeder and other runtime cache dirs. ref #1638
* fix: ensure BrowserConfig.to_dict serializes proxy_config
* feat: make LLM backoff configurable end-to-end
- extend LLMConfig with backoff delay/attempt/factor fields and thread them
through LLMExtractionStrategy, LLMContentFilter, table extraction, and
Docker API handlers
- expose the backoff parameter knobs on perform_completion_with_backoff/aperform_completion_with_backoff
and document them in the md_v2 guides
* reproduced AttributeError from #1642
* pass timeout parameter to docker client request
* added missing deep crawling objects to init
* generalized query in ContentRelevanceFilter to be a str or list
* import modules from enhanceable deserialization
* parameterized tests
* Fix: capture current page URL to reflect JavaScript navigation and add test for delayed redirects. ref #1268
* refactor: replace PyPDF2 with pypdf across the codebase. ref #1412
* Add browser_context_id and target_id parameters to BrowserConfig
Enable Crawl4AI to connect to pre-created CDP browser contexts, which is
essential for cloud browser services that pre-create isolated contexts.
Changes:
- Add browser_context_id and target_id parameters to BrowserConfig
- Update from_kwargs() and to_dict() methods
- Modify BrowserManager.start() to use existing context when provided
- Add _get_page_by_target_id() helper method
- Update get_page() to handle pre-existing targets
- Add test for browser_context_id functionality
This enables cloud services to:
1. Create isolated CDP contexts before Crawl4AI connects
2. Pass context/target IDs to BrowserConfig
3. Have Crawl4AI reuse existing contexts instead of creating new ones
* Add cdp_cleanup_on_close flag to prevent memory leaks in cloud/server scenarios
* Fix: add cdp_cleanup_on_close to from_kwargs
* Fix: find context by target_id for concurrent CDP connections
* Fix: use target_id to find correct page in get_page
* Fix: use CDP to find context by browserContextId for concurrent sessions
* Revert context matching attempts - Playwright cannot see CDP-created contexts
* Add create_isolated_context flag for concurrent CDP crawls
When True, forces creation of a new browser context instead of reusing
the default context. Essential for concurrent crawls on the same browser
to prevent navigation conflicts.
* Add context caching to create_isolated_context branch
Uses contexts_by_config cache (same as non-CDP mode) to reuse contexts
for multiple URLs with same config. Still creates new page per crawl
for navigation isolation. Benefits batch/deep crawls.
* Add init_scripts support to BrowserConfig for pre-page-load JS injection
This adds the ability to inject JavaScript that runs before any page loads,
useful for stealth evasions (canvas/audio fingerprinting, userAgentData).
- Add init_scripts parameter to BrowserConfig (list of JS strings)
- Apply init_scripts in setup_context() via context.add_init_script()
- Update from_kwargs() and to_dict() for serialization
* Fix CDP connection handling: support WS URLs and proper cleanup
Changes to browser_manager.py:
1. _verify_cdp_ready(): Support multiple URL formats
- WebSocket URLs (ws://, wss://): Skip HTTP verification, Playwright handles directly
- HTTP URLs with query params: Properly parse with urlparse to preserve query string
- Fixes issue where naive f"{cdp_url}/json/version" broke WS URLs and query params
2. close(): Proper cleanup when cdp_cleanup_on_close=True
- Close all sessions (pages)
- Close all contexts
- Call browser.close() to disconnect (doesn't terminate browser, just releases connection)
- Wait 1 second for CDP connection to fully release
- Stop Playwright instance to prevent memory leaks
This enables:
- Connecting to specific browsers via WS URL
- Reusing the same browser with multiple sequential connections
- No user wait needed between connections (internal 1s delay handles it)
Added tests/browser/test_cdp_cleanup_reuse.py with comprehensive tests.
* Update gitignore
* Some debugging for caching
* Add _generate_screenshot_from_html for raw: and file:// URLs
Implements the missing method that was being called but never defined.
Now raw: and file:// URLs can generate screenshots by:
1. Loading HTML into a browser page via page.set_content()
2. Taking screenshot using existing take_screenshot() method
3. Cleaning up the page afterward
This enables cached HTML to be rendered with screenshots in crawl4ai-cloud.
* Add PDF and MHTML support for raw: and file:// URLs
- Replace _generate_screenshot_from_html with _generate_media_from_html
- New method handles screenshot, PDF, and MHTML in one browser session
- Update raw: and file:// URL handlers to use new method
- Enables cached HTML to generate all media types
* Add crash recovery for deep crawl strategies
Add optional resume_state and on_state_change parameters to all deep
crawl strategies (BFS, DFS, Best-First) for cloud deployment crash
recovery.
Features:
- resume_state: Pass saved state to resume from checkpoint
- on_state_change: Async callback fired after each URL for real-time
state persistence to external storage (Redis, DB, etc.)
- export_state(): Get last captured state manually
- Zero overhead when features are disabled (None defaults)
State includes visited URLs, pending queue/stack, depths, and
pages_crawled count. All state is JSON-serializable.
* Fix: HTTP strategy raw: URL parsing truncates at # character
The AsyncHTTPCrawlerStrategy.crawl() method used urlparse() to extract
content from raw: URLs. This caused HTML with CSS color codes like #eee
to be truncated because # is treated as a URL fragment delimiter.
Before: raw:body{background:#eee} -> parsed.path = 'body{background:'
After: raw:body{background:#eee} -> raw_content = 'body{background:#eee'
Fix: Strip the raw: or raw:// prefix directly instead of using urlparse,
matching how the browser strategy handles it.
* Add base_url parameter to CrawlerRunConfig for raw HTML processing
When processing raw: HTML (e.g., from cache), the URL parameter is meaningless
for markdown link resolution. This adds a base_url parameter that can be set
explicitly to provide proper URL resolution context.
Changes:
- Add base_url parameter to CrawlerRunConfig.__init__
- Add base_url to CrawlerRunConfig.from_kwargs
- Update aprocess_html to use base_url for markdown generation
Usage:
config = CrawlerRunConfig(base_url='https://example.com')
result = await crawler.arun(url='raw:{html}', config=config)
* Add prefetch mode for two-phase deep crawling
- Add `prefetch` parameter to CrawlerRunConfig
- Add `quick_extract_links()` function for fast link extraction
- Add short-circuit in aprocess_html() for prefetch mode
- Add 42 tests (unit, integration, regression)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Updates on proxy rotation and proxy configuration
* Add proxy support to HTTP crawler strategy
* Add browser pipeline support for raw:/file:// URLs
- Add process_in_browser parameter to CrawlerRunConfig
- Route raw:/file:// URLs through _crawl_web() when browser operations needed
- Use page.set_content() instead of goto() for local content
- Fix cookie handling for non-HTTP URLs in browser_manager
- Auto-detect browser requirements: js_code, wait_for, screenshot, etc.
- Maintain fast path for raw:/file:// without browser params
Fixes#310
* Add smart TTL cache for sitemap URL seeder
- Add cache_ttl_hours and validate_sitemap_lastmod params to SeedingConfig
- New JSON cache format with metadata (version, created_at, lastmod, url_count)
- Cache validation by TTL expiry and sitemap lastmod comparison
- Auto-migration from old .jsonl to new .json format
- Fixes bug where incomplete cache was used indefinitely
* Update URL seeder docs with smart TTL cache parameters
- Add cache_ttl_hours and validate_sitemap_lastmod to parameter table
- Document smart TTL cache validation with examples
- Add cache-related troubleshooting entries
- Update key features summary
* Add MEMORY.md to gitignore
* Docs: Add multi-sample schema generation section
Add documentation explaining how to pass multiple HTML samples
to generate_schema() for stable selectors that work across pages
with varying DOM structures.
Includes:
- Problem explanation (fragile nth-child selectors)
- Solution with code example
- Key points for multi-sample queries
- Comparison table of fragile vs stable selectors
* Fix critical RCE and LFI vulnerabilities in Docker API deployment
Security fixes for vulnerabilities reported by ProjectDiscovery:
1. Remote Code Execution via Hooks (CVE pending)
- Remove __import__ from allowed_builtins in hook_manager.py
- Prevents arbitrary module imports (os, subprocess, etc.)
- Hooks now disabled by default via CRAWL4AI_HOOKS_ENABLED env var
2. Local File Inclusion via file:// URLs (CVE pending)
- Add URL scheme validation to /execute_js, /screenshot, /pdf, /html
- Block file://, javascript:, data: and other dangerous schemes
- Only allow http://, https://, and raw: (where appropriate)
3. Security hardening
- Add CRAWL4AI_HOOKS_ENABLED=false as default (opt-in for hooks)
- Add security warning comments in config.yml
- Add validate_url_scheme() helper for consistent validation
Testing:
- Add unit tests (test_security_fixes.py) - 16 tests
- Add integration tests (run_security_tests.py) for live server
Affected endpoints:
- POST /crawl (hooks disabled by default)
- POST /crawl/stream (hooks disabled by default)
- POST /execute_js (URL validation added)
- POST /screenshot (URL validation added)
- POST /pdf (URL validation added)
- POST /html (URL validation added)
Breaking changes:
- Hooks require CRAWL4AI_HOOKS_ENABLED=true to function
- file:// URLs no longer work on API endpoints (use library directly)
* Enhance authentication flow by implementing JWT token retrieval and adding authorization headers to API requests
* Add release notes for v0.7.9, detailing breaking changes, security fixes, new features, bug fixes, and documentation updates
* Add release notes for v0.8.0, detailing breaking changes, security fixes, new features, bug fixes, and documentation updates
Documentation for v0.8.0 release:
- SECURITY.md: Security policy and vulnerability reporting guidelines
- RELEASE_NOTES_v0.8.0.md: Comprehensive release notes
- migration/v0.8.0-upgrade-guide.md: Step-by-step migration guide
- security/GHSA-DRAFT-RCE-LFI.md: GitHub security advisory drafts
- CHANGELOG.md: Updated with v0.8.0 changes
Breaking changes documented:
- Docker API hooks disabled by default (CRAWL4AI_HOOKS_ENABLED)
- file:// URLs blocked on Docker API endpoints
Security fixes credited to Neo by ProjectDiscovery
* Add examples for deep crawl crash recovery and prefetch mode in documentation
* Release v0.8.0: The v0.8.0 Update
- Updated version to 0.8.0
- Added comprehensive demo and release notes
- Updated all documentation
* Update security researcher acknowledgment with a hyperlink for Neo by ProjectDiscovery
* Add async agenerate_schema method for schema generation
- Extract prompt building to shared _build_schema_prompt() method
- Add agenerate_schema() async version using aperform_completion_with_backoff
- Refactor generate_schema() to use shared prompt builder
- Fixes Gemini/Vertex AI compatibility in async contexts (FastAPI)
* Fix: Enable litellm.drop_params for O-series/GPT-5 model compatibility
O-series (o1, o3) and GPT-5 models only support temperature=1.
Setting litellm.drop_params=True auto-drops unsupported parameters
instead of throwing UnsupportedParamsError.
Fixes temperature=0.01 error for these models in LLM extraction.
---------
Co-authored-by: rbushria <rbushri@gmail.com>
Co-authored-by: AHMET YILMAZ <tawfik@kidocode.com>
Co-authored-by: Soham Kukreti <kukretisoham@gmail.com>
Co-authored-by: Chris Murphy <chris.murphy@klaviyo.com>
Co-authored-by: unclecode <unclecode@kidocode.com>
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Fixes#1680
The HTML2Text class was ignoring the <base> tag, causing relative links
to be resolved against the page URL instead of the base URL specified
in the <base href="..."> attribute.
Added <base> tag handling in both HTML2Text and CustomHTML2Text to update
self.baseurl when the tag is encountered, ensuring proper link resolution
according to HTML standards.
Co-Authored-By: Claude <noreply@anthropic.com>
Fixes#1704
When AsyncPlaywrightCrawlerStrategy is instantiated without a logger,
self.logger remains None, causing crashes when logger methods are called
(e.g., during screenshot/PDF/MHTML export).
Changed to initialize a default AsyncLogger(verbose=False) when no logger
is provided, preventing NoneType errors throughout the strategy.
Co-Authored-By: Claude <noreply@anthropic.com>
Fixes#1711
The GoogleSearchCrawler's script.js file was not included in the package
distribution, causing FileNotFoundError when using the crawler.
Added crawl4ai/crawlers/google_search/*.js to setuptools package-data
in pyproject.toml to ensure the JavaScript file is bundled with the package.
Co-Authored-By: Claude <noreply@anthropic.com>
Fixes#1658
The `_get_embedding_llm_config_dict` method was returning a fallback
OpenAI config when `embedding_llm_config` was None, which prevented
the use of local sentence-transformers embeddings.
Changed the method to:
- Return None when no embedding config is provided
- Update return type from Dict to Optional[Dict]
- Update docstring to clarify the behavior
This allows `get_text_embeddings` utility to correctly switch to
the local sentence-transformers implementation when no LLM config
is provided.
Co-Authored-By: Claude <noreply@anthropic.com>
Fixes#1553
The code imports from `playwright_stealth` module (e.g., StealthConfig,
stealth_async, stealth_sync with capital S) which is only available in
the `playwright-stealth` package, not `tf-playwright-stealth`.
Changed dependency from:
- tf-playwright-stealth>=1.1.0 (wrong package, exports lowercase names)
to:
- playwright-stealth>=2.0.0 (correct package, exports StealthConfig)
This fixes the mismatch between declared dependency and actual imports
in the codebase.
Co-Authored-By: Claude <noreply@anthropic.com>
O-series (o1, o3) and GPT-5 models only support temperature=1.
Setting litellm.drop_params=True auto-drops unsupported parameters
instead of throwing UnsupportedParamsError.
Fixes temperature=0.01 error for these models in LLM extraction.
- Extract prompt building to shared _build_schema_prompt() method
- Add agenerate_schema() async version using aperform_completion_with_backoff
- Refactor generate_schema() to use shared prompt builder
- Fixes Gemini/Vertex AI compatibility in async contexts (FastAPI)
