CAPTCHA Handling

Built-in CAPTCHA detection with configurable handling strategies for browser automation agents.

Why Predicate Does Not Solve CAPTCHAs

Predicate is designed as infrastructure for token-efficient browser automation runtime, not as a full-service automation provider. We deliberately limit our scope to detection and verification for two reasons:

Predicate detects CAPTCHAs, pauses execution, and verifies clearance. You decide how to resolve them using your own workflows or external systems that comply with your organization's security policies.

How CAPTCHA Handling Works

When the Predicate SDK detects a CAPTCHA during browser automation, the following flow occurs:

1. Detection: The Chrome extension scans for CAPTCHA signals (provider iframes, container selectors, keywords) and reports snapshot.diagnostics.captcha with a confidence score.
2. Policy check: If captcha.detected == true and confidence >= minConfidence, the runtime invokes your configured policy.
3. Action dispatch: Based on the policy, the runtime either aborts, retries with a new session, or waits for clearance.
4. Verification loop: For wait_until_cleared, the runtime continuously re-snapshots until captcha.detected == false or timeout is reached.

Recommended Hybrid Flow (Token → OCR → Vision)

CAPTCHAs vary by site and session. A practical approach is a hybrid fallback chain that starts with token-based solvers (best coverage for standard providers), then tries OCR for image-only challenges, and finally uses a vision LLM as a last resort.

This design keeps the SDK responsibilities focused on detection + orchestration, while you control the actual solving logic and providers (e.g., a token service like 2Captcha, or an internal OCR pipeline).

High-level behavior:

• Token solver first: Works for reCAPTCHA/hCaptcha/Turnstile when a site key is available.
• OCR second: Works for simple image captchas with a visible <img> and nearby input.
• Vision last: Best-effort fallback for custom or unusual layouts.

Signal source: CAPTCHA detection is surfaced on the current page via snapshot.diagnostics.captcha. This signal is only reliable on the page where the CAPTCHA is actually rendered (e.g., after a modal opens or a challenge iframe appears).

Pseudocode (Python):

from predicate.captcha import CaptchaContext

def solve_captcha(ctx: CaptchaContext):
  if ctx.captcha.provider_hint in {"recaptcha", "hcaptcha", "turnstile"}:
      if try_token_solver(ctx):  # e.g., external provider
          return wait_until_cleared(ctx)

  if has_image_captcha(ctx):     # visible <img> + nearby input
      if try_ocr_solver(ctx):    # image → text
          return wait_until_cleared(ctx)

  # Best-effort fallback
  if try_vision_solver(ctx):     # vision LLM on captured image
      return wait_until_cleared(ctx)

  return abort("captcha_unresolved")

For questions or guidance, contact

Page-control Hook for External Solvers

When you run a CAPTCHA handler, the runtime provides a page-control hook so your solver can read or write small page state inside the same live browser session. This is what makes token injection or callback triggering possible without reloading the page.

Use it sparingly and keep the JS payload small and bounded.

from predicate.captcha import CaptchaContext

async def external_solver(ctx: CaptchaContext):
  # Minimal, bounded JS: read a site key and inject a token.
  sitekey = await ctx.page_control.evaluate_js("/* read sitekey from DOM */")
  token = await request_token(sitekey)  # your external system
  await ctx.page_control.evaluate_js("/* write token into response field */")
  return {"action": "wait_until_cleared"}

Concepts and Terminology

Policies

Policies determine what happens when a CAPTCHA is detected:

Actions

Actions are the runtime behaviors your handler can request:

Configuration Options

Default Strategies (Helpers, Not Solvers)

Predicate provides three built-in strategy helpers. These are not solvers; they configure the runtime's response to CAPTCHA detection and rely on external systems or humans for actual resolution.

Quick Start

Option 1: Abort Policy (Safest Default)

Immediately stop when CAPTCHA is detected. No resolution attempted.

from predicate import AgentRuntime, CaptchaOptions

runtime.set_captcha_options(
    CaptchaOptions(
        policy="abort",
        min_confidence=0.7,
    )
)

Option 2: Human-in-Loop (Recommended for Live Sessions)

When CAPTCHA is detected, the runtime pauses and waits for a human to solve it in the live browser session.

from predicate import CaptchaOptions, HumanHandoffSolver

runtime.set_captcha_options(
    CaptchaOptions(
        policy="callback",
        handler=HumanHandoffSolver(),
        min_confidence=0.7,
        timeout_ms=120_000,  # Wait up to 2 minutes

Option 3: Vision-Only Verification

Uses vision to confirm the CAPTCHA has cleared. Does not click or type. Useful after an external system has already solved the CAPTCHA.

from predicate import CaptchaOptions, VisionSolver

runtime.set_captcha_options(
    CaptchaOptions(policy="callback", handler=VisionSolver())
)

Option 4: External Resolver Orchestration

Calls your webhook/service when CAPTCHA is detected, then waits for clearance. Your external system performs the actual resolution.

from predicate import CaptchaOptions, ExternalSolver

async def notify_webhook(ctx) -> None:
    """
    Example hook: send context to your external system.
    Replace with your own client / queue / webhook call.
    Predicate does NOT implement solver logic.
    """
    print(f"[captcha] external resolver notified: url=

Complete Example

import asyncio
import os

from predicate import (
    AgentRuntime,
    AsyncPredicateBrowser,
    CaptchaOptions,
    ExternalSolver,
    HumanHandoffSolver,
    VisionSolver,
)
from predicate.tracing import JsonlTraceSink

CAPTCHA Detection

Detected Providers

Predicate detects the following CAPTCHA providers with high confidence:

CaptchaDiagnostics Schema

The snapshot.diagnostics.captcha object contains:

type CaptchaDiagnostics = {
  detected: boolean;
  provider_hint?: "recaptcha" | "hcaptcha" | "turnstile" | "arkose" | "awswaf" | "unknown";
  confidence: number; // 0..1
  evidence: {
    text_hits: string[];       // Matched text keywords
    selector_hits: string[];   // Matched CSS selectors
    iframe_src_hits: string[]; // Matched iframe sources
    url_hits: string[];        // Matched URL patterns
  };
};

Confidence Scoring

The detection uses a multi-signal scoring system:

• Provider iframe hit: +0.7 confidence
• Provider container selector hit: +0.5 confidence
• Keyword text hit: +0.3 confidence
• URL pattern hit: +0.2 confidence

A CAPTCHA is considered detected when confidence >= 0.7 (configurable via minConfidence).

External Resolution Guidance

If you integrate an external provider (e.g., 2captcha, Anti-Captcha) or your own internal system:

1. Predicate Only Detects and Verifies

The external system performs the actual resolution. Predicate monitors for clearance.

2. Webhook Payload Structure

Your handler receives a context object with:

3. Policy Compliance

Keep audit logs and ensure your resolution approach complies with your organization's policies (consent, allowed domains, rate limits).

4. Handler Return Value

Your handler should return or allow the default wait_until_cleared action. The runtime then confirms clearance before resuming.

5. Timeout Configuration

Set appropriate timeouts based on your external service's SLA. External services may take 30-180 seconds to resolve.

Trace Integration

CAPTCHA events are automatically emitted as verification events to the tracer:

{
  "type": "verification",
  "data": {
    "kind": "captcha",
    "label": "captcha_detected",
    "passed": false,
    "reason": "CAPTCHA detected: recaptcha (confidence: 0.85)",
    "details": {
      "provider_hint": "recaptcha",
      "confidence": 0.85,
      "evidence": {
        "selector_hits": [".g-recaptcha"],
        "iframe_src_hits": ["https://www.google.com/recaptcha/..."]
      }
    }
  },
  "step_id": "abc-123"
}

Best Practices

Related Documentation

• Agent Runtime - Full AgentRuntime API reference
• Tracing & Debugging - Trace events and debugging
• Debugging Agent Failures - Post-mortem analysis