[Bug]: CrawlerRunConfig.dump()/load() drops deep-crawl filter and scorer arguments, so a DomainFilter allows every domain #2424

crawl4ai version

0.9.5 (also on main and develop)

Expected Behavior

CrawlerRunConfig.load(cfg.dump()) gives back a config that crawls the same way, as #1436 made true for URLPatternFilter.

Current Behavior

to_serializable_dict() rebuilds an object from attributes named after its constructor arguments. Several deep-crawl filters and scorers keep their arguments only in private slots (_allowed_domains, _keywords, ...), so the arguments are dropped:

Crawl4aiDockerClient sends crawler_config.dump() to the server, so a deep crawl run through the Docker client with a DomainFilter silently leaves the allowed domain.

Is this reproducible?

Yes

Inputs Causing the Bug

A CrawlerRunConfig whose deep_crawl_strategy uses DomainFilter or any of the scorers above.

Steps to Reproduce

Run the snippet below.

Code snippets

import asyncio
import json

from crawl4ai import CrawlerRunConfig
from crawl4ai.deep_crawling import (
    BestFirstCrawlingStrategy, BFSDeepCrawlStrategy, DomainFilter, FilterChain, KeywordRelevanceScorer,
)

cfg = CrawlerRunConfig(deep_crawl_strategy=BFSDeepCrawlStrategy(
    max_depth=2, filter_chain=FilterChain([DomainFilter(allowed_domains=["docs.example.com"])])))
back = CrawlerRunConfig.load(json.loads(json.dumps(cfg.dump())))
url = "https://evil.example.net/page"
print(asyncio.run(cfg.deep_crawl_strategy.filter_chain.apply(url)))   # False
print(asyncio.run(back.deep_crawl_strategy.filter_chain.apply(url)))  # True, expected False

cfg2 = CrawlerRunConfig(deep_crawl_strategy=BestFirstCrawlingStrategy(
    max_depth=2, url_scorer=KeywordRelevanceScorer(keywords=["api", "guide"], weight=0.7)))
print(cfg2.dump()["params"]["deep_crawl_strategy"]["params"]["url_scorer"])
# {'type': 'KeywordRelevanceScorer', 'params': {'weight': 0.699999988079071}}  <- keywords missing
CrawlerRunConfig.load(json.loads(json.dumps(cfg2.dump())))
# TypeError: KeywordRelevanceScorer.__init__() missing 1 required positional argument: 'keywords'

OS

Windows 11 (the code path is platform-independent)

Python version

3.13

Browser

Not involved

Error logs & Screenshots (if applicable)

The TypeError above for KeywordRelevanceScorer, DomainAuthorityScorer and CompositeScorer; no error for the others.

I have a fix that follows the #1436 approach (read-only properties exposing the original arguments, so __slots__ and memory use don't change), with a round-trip test for every filter and scorer. I'll open a PR against develop referencing this issue.

Investigated with AI assistance; I ran the reproduction and checked the output myself.