13: Multi-Model Fallback

from afk.agents import Agent, FailSafeConfig

agent = Agent(
    name="resilient-agent",
    model="gpt-4.1",                    # Primary model
    instructions="Be helpful and thorough.",
    fail_safe=FailSafeConfig(
        # Fallback chain: try these models in order if the primary fails
        fallback_model_chain=[
            "gpt-4.1-mini",              # First fallback: cheaper, faster
            "gpt-4.1-nano",              # Last resort: fastest, cheapest
        ],

        # When LLM calls fail, retry then degrade
        llm_failure_policy="retry_then_degrade",

        # Cost ceiling still applies across all models
        max_total_cost_usd=1.00,
    ),
)

from afk.agents import Agent, FailSafeConfig
from afk.core import Runner

# Start cheap, escalate if quality is insufficient
simple_agent = Agent(
    name="classifier",
    model="gpt-4.1-nano",              # Start with cheapest
    instructions="""
    Classify the support ticket. Output exactly one label:
    billing, technical, account, other.
    """,
    fail_safe=FailSafeConfig(
        fallback_model_chain=["gpt-4.1-mini", "gpt-4.1"],
        max_total_cost_usd=0.05,
    ),
)

# Complex tasks get the big model with fallbacks
analysis_agent = Agent(
    name="analyst",
    model="gpt-4.1",                   # Start with most capable
    instructions="""
    Provide detailed technical analysis with code examples.
    Be thorough and precise.
    """,
    fail_safe=FailSafeConfig(
        fallback_model_chain=["gpt-4.1-mini"],
        llm_failure_policy="retry_then_degrade",
        max_total_cost_usd=2.00,
    ),
)

runner = Runner()

# Simple task -> cheap model handles it
r1 = runner.run_sync(simple_agent, user_message="I can't log in")
print(f"Classification: {r1.final_text} (${r1.total_cost_usd or 0:.4f})")

# Complex task -> powerful model with safety net
r2 = runner.run_sync(analysis_agent, user_message="Analyze Python's asyncio event loop")
print(f"Analysis: {r2.final_text[:100]}... (${r2.total_cost_usd or 0:.4f})")

agent = Agent(
    name="breaker-demo",
    model="gpt-4.1",
    instructions="...",
    fail_safe=FailSafeConfig(
        fallback_model_chain=["gpt-4.1-mini", "gpt-4.1-nano"],

        # Circuit breaker settings
        breaker_failure_threshold=5,     # Open after 5 consecutive failures
        breaker_cooldown_s=30.0,         # Wait 30s before retrying the model

        # Failure handling
        llm_failure_policy="retry_then_degrade",
        max_total_cost_usd=1.00,
    ),
)

from afk.agents import Agent, FailSafeConfig

# Cheap model for simple classification
router = Agent(
    name="router",
    model="gpt-4.1-nano",
    instructions="Route to the correct specialist.",
    fail_safe=FailSafeConfig(fallback_model_chain=["gpt-4.1-mini"]),
    subagents=[
        # Powerful model for complex analysis
        Agent(
            name="analyst",
            model="gpt-4.1",
            instructions="Provide deep technical analysis.",
            fail_safe=FailSafeConfig(
                fallback_model_chain=["gpt-4.1-mini"],
                max_total_cost_usd=1.00,
            ),
        ),
        # Mid-tier model for summarization
        Agent(
            name="summarizer",
            model="gpt-4.1-mini",
            instructions="Summarize findings concisely.",
            fail_safe=FailSafeConfig(
                fallback_model_chain=["gpt-4.1-nano"],
                max_total_cost_usd=0.25,
            ),
        ),
    ],
)

result = runner.run_sync(agent, user_message="Analyze this...")

print(f"State: {result.state}")
print(f"Requested model: {result.requested_model}")
print(f"Normalized model: {result.normalized_model}")
print(f"Provider: {result.provider_adapter}")
print(f"Total tokens: {result.usage_aggregate.total_tokens}")
print(f"Total cost: ${result.total_cost_usd or 0:.4f}")