Antibot Strategy & Stealth Surfaces

Crawlberg detects WAF/bot-mitigation signals, classifies them with hot-reloadable rules, and can escalate through the configured dispatch chain. Customize policy with AntibotStrategy, DispatchProfile, retry policy, domain state, and optional caller-supplied bypass providers.

Architecture

Three layers compose the antibot system:

1. WAF Classifier — Matches HTTP responses against crates/crawlberg/rules/waf_fingerprints.toml and returns a WafSignal with vendor, fingerprint_id, and weight.
2. Decision Hook — AntibotStrategy trait pair: pre_request (warm external state) and post_response (inspect response, decide next action).
3. Dispatch Policy — DispatchProfile combines escalation strategy, retry policy, classifier, domain state, budget, and optional bypass provider.

When a WAF signal is detected, the engine consults the strategy and retry policy to decide whether to retry, escalate to the browser tier, route through a caller-supplied bypass provider, or stop. Decision::RotateProxy is present but not implemented; it logs and falls through to Accept.

The AntibotStrategy trait

Implement this trait to intercept every request/response pair:

#[async_trait]
pub trait AntibotStrategy: Send + Sync {
    /// Called once per attempt, before the tower-stack fetch fires.
    /// Return Err to abort this attempt; the retry policy decides what happens.
    async fn pre_request(&self, url: &str) -> Result<(), AntibotError>;

    /// Called once per successful HTTP response, after WAF classification.
    /// Return a Decision that overrides the retry policy for this attempt.
    async fn post_response(
        &self,
        response: &HttpResponse,
        waf_signal: Option<&WafSignal>,
    ) -> Decision;
}

The Decision enum controls what happens next:

pub enum Decision {
    /// Continue normally — hand the response to the retry policy.
    Accept,
    /// Retry the same tier after `backoff` (Duration).
    Retry { backoff: Duration },
    /// Rotate to a different proxy and retry (not yet implemented).
    /// Logs a warn!() and falls through to Accept for now.
    RotateProxy,
    /// Force escalation to the Browser tier, bypassing the retry policy.
    EscalateBrowser,
}

Errors during strategy execution are returned as:

pub enum AntibotError {
    /// The pre_request hook failed.
    PreRequest(String),
    /// The post_response hook failed.
    PostResponse(String),
}

BrowserMode variants

The BrowserConfig.mode field controls when and how the browser is used. BrowserMode::Stealth is the new variant:

Variant	Escalation	Stealth surfaces	Use case
Auto (default)	HTTP first, escalate to browser on WAF/403	None	Balanced; handles most sites
Always	Skip HTTP entirely	None	JS-heavy SPAs
Never	No browser fallback	None	Performance-critical; no WAF expected
Stealth	Browser tier for every request	JS patches, native TLS spoofing, UA/viewport defaults	Stealth-hardened mode for challenging sites

BrowserMode::Stealth behaves like Always for request routing (every page goes through the browser tier) but additionally enables:

• chromiumoxide JS patches (crate::stealth::apply_stealth_patches) that spoof navigator.webdriver, navigator.chromeFlags, and navigator.plugins.
• Native-backend TLS fingerprint spoofing (JA3 randomization).
• Stealth-aware default user-agent when no explicit user-agent is set.
• Forced viewport (1920×1080) to avoid detection via unusual screen sizes.

Previously, a BrowserConfig.stealth: bool field existed but had a bug: JS patches always ran regardless of the flag. This field has been removed pre-v1. Use BrowserMode::Stealth instead.

Worked example: per-vendor backoff

Wrap DefaultAntibotStrategy to inject custom backoff rules per WAF vendor:

use std::time::Duration;
use async_trait::async_trait;
use crawlberg::{
    AntibotStrategy, Decision, DefaultAntibotStrategy,
    AntibotError, http::HttpResponse, WafSignal,
};

#[derive(Debug)]
struct VendorBackoffStrategy {
    inner: DefaultAntibotStrategy,
}

#[async_trait]
impl AntibotStrategy for VendorBackoffStrategy {
    async fn pre_request(&self, url: &str) -> Result<(), AntibotError> {
        // Delegate to default (no-op)
        self.inner.pre_request(url).await
    }

    async fn post_response(
        &self,
        resp: &HttpResponse,
        signal: Option<&WafSignal>,
    ) -> Decision {
        match signal.map(|s| s.vendor.as_str()) {
            Some("cloudflare") => {
                // Cloudflare: aggressive backoff
                Decision::Retry {
                    backoff: Duration::from_secs(10),
                }
            }
            Some("datadome") => {
                // DataDome: skip to browser
                Decision::EscalateBrowser
            }
            _ => {
                // Everything else: use default logic
                self.inner.post_response(resp, signal).await
            }
        }
    }
}

// Wire it up
let strategy = std::sync::Arc::new(VendorBackoffStrategy {
    inner: DefaultAntibotStrategy::new(),
});

let profile = DispatchProfile::builder()
    .antibot_strategy(strategy)
    .build();

Defaults

Without an attached strategy, the engine uses DefaultAntibotStrategy (defined at crates/crawlberg/src/types/antibot.rs:132-164):

• pre_request is a no-op.
• post_response returns Decision::EscalateBrowser when a WAF signal is present, Decision::Accept otherwise.

This matches the pre-Cluster-5 escalation logic, so existing code continues to work unchanged.

WAF detection corpus v0.3

Crawlberg classifies WAF fingerprints via TomlClassifier at crates/crawlberg/rules/waf_fingerprints.toml. The rules currently cover Cloudflare, DataDome, PerimeterX/HUMAN Security, Imperva, AWS WAF, Akamai, F5, and generic block patterns.

Fingerprints match response headers, body substrings, or status code/header combinations. The classifier supports hot reload via TomlClassifier::watch() so rule updates can land without restarting the process.

Dispatch and EWMA state v0.3

The default retry and dispatch layer can combine WAF signals, transient errors, and EwmaDomainState observations. EWMA state lets callers track per-domain outcomes and feed a LearningRetryPolicy without requiring a database.

Tuning dispatch and escalation (Rust)

DispatchProfile exposes fields for fine-grained control over the retry and escalation loop. This section is Rust-only — language bindings use the built-in defaults.

Build a profile using the fluent builder, which accepts optional trait-object implementations:

use std::sync::Arc;
use crawlberg::{
    CrawlConfig, DispatchProfile, EscalationStrategy,
    FixedBudget, EwmaDomainState, LearningRetryPolicy,
    TomlClassifier,
};

let profile = DispatchProfile::builder()
    // Start with a custom escalation strategy: HTTP → Bypass → Browser
    .strategy(EscalationStrategy::BypassThenBrowser)

    // Use a learning retry policy backed by per-domain EWMA state
    .domain_state(Arc::new(EwmaDomainState::new()))
    .retry_policy(Arc::new(LearningRetryPolicy::new(
        Arc::new(EwmaDomainState::new())
    )))

    // Cap escalations at $2.50 per job
    .escalation_budget(Arc::new(FixedBudget::new(250)))

    // Limit total attempts across all tiers
    .max_total_attempts(15)

    // Use a custom WAF classifier (or omit to use the built-in corpus)
    .waf_classifier(Arc::new(TomlClassifier::builtin()))

    .build();

// Attach to the crawl config
let config = CrawlConfig::builder()
    .dispatch(profile)
    .build();

Key tuning parameters:

• strategy: Select the escalation chain (None, BrowserOnly, BypassFirst, BypassOnly, BypassThenBrowser). See the strategy table above.
• retry_policy: Implement RetryPolicy for custom per-attempt decisions. The default SimpleRetryPolicy uses static error-to-directive mappings; LearningRetryPolicy consults a DomainStatePort for priors. Both live in crate::defaults::dispatch.
• domain_state: Track per-domain block rates via DomainStatePort. The in-process EwmaDomainState is provided; xberg-enterprise supplies a Postgres-backed impl for multi-process deployments.
• escalation_budget: Enforce per-job spend caps via EscalationBudget. FixedBudget tracks atomic counters; UnlimitedBudget is the default.
• max_total_attempts: Hard limit on fetch attempts across all tiers. Guards against buggy retry policies. Default: 10.

Custom WAF rules and hot-reload

The canonical WAF fingerprint corpus lives in crates/crawlberg/rules/waf_fingerprints.toml. Load the builtin rules at compile time, or supply custom TOML files for testing or rule updates without restarting.

Loading rules

Use TomlClassifier for TOML-backed fingerprinting (Rust-only):

use crawlberg::waf::TomlClassifier;
use crawlberg::waf_rules_from_path;
use std::sync::Arc;

// Built-in canonical corpus
let classifier = Arc::new(TomlClassifier::builtin());

// Custom rules from a file
let custom_rules = waf_rules_from_path("custom_waf_rules.toml")?;
let classifier = Arc::new(TomlClassifier::from_rules(custom_rules));

Hot-reload rules on file change

Watch a TOML file for changes and atomically swap rules without restarting:

use std::sync::Arc;
use crawlberg::waf::TomlClassifier;
use crawlberg::waf_rules_from_path;

let classifier = Arc::new(TomlClassifier::builtin());

// Start watching for file changes, debounced 500 ms
let _watcher = classifier.watch("/etc/crawlberg/waf_rules.toml")?;

// Swap rules atomically; concurrent classifiers see new rules on next call
let new_rules = waf_rules_from_path("/etc/crawlberg/waf_rules.toml")?;
classifier.swap(new_rules);

The WatchHandle returned by watch manages the background file watcher. Dropping it stops watching. Keep it alive for as long as you want to receive updates.

This pattern is Kubernetes-friendly: mount a ConfigMap as a file, watch it, and reload rules when the cluster operator updates the ConfigMap. The atomic-swap semantics mean in-flight classifications continue with the old rules; new requests use the new rules.

Bypass providers

Crawlberg exposes the BypassProvider trait and BypassResponse type for caller-owned integrations. Providers are responsible for authentication, request shaping, response decoding, cost metadata, and error mapping.

Crawlberg does not ship Bright Data, Zyte, ScrapingBee, or other vendor adapters in the core crate.