How do you handle context window management in long-running OpenAI applications?

from openai import OpenAI
import tiktoken

client = OpenAI()
enc = tiktoken.encoding_for_model("gpt-5.5")

# Context window strategy 1: Sliding window
# Keep the most recent N turns, discarding the oldest
def sliding_window(messages: list, max_tokens: int = 100_000) -> list:
    total = sum(len(enc.encode(str(m["content"]))) for m in messages)
    while total > max_tokens and len(messages) > 2:
        # Remove oldest user+assistant pair (keep system message)
        removed = messages.pop(1)
        total -= len(enc.encode(str(removed["content"])))
    return messages

# Context window strategy 2: Summarisation
# Summarise old conversation instead of dropping it entirely
async def summarise_history(messages: list[dict]) -> str:
    summary = await client.responses.create(
        model="gpt-5.4-mini",  # cheap model for summarisation
        input=f"Summarise this conversation concisely:\n{messages}",
    )
    return summary.output_text

# Strategy 3: Responses API state management (avoids manual history)
# Use previous_response_id - OpenAI manages the context server-side
first = client.responses.create(model="gpt-5.5", input="Hello!", store=True)
second = client.responses.create(
    model="gpt-5.5",
    input="What did I just say?",
    previous_response_id=first.id,  # server retrieves context automatically
)

# Strategy 4: use file_search for large static documents
# Instead of pasting 100k-token documents into every prompt:
# Upload once to a vector store, retrieve only relevant chunks
vector_store_id = "vs_abc123"
response = client.responses.create(
    model="gpt-5.5",
    tools=[{"type": "file_search", "vector_store_ids": [vector_store_id]}],
    input="What does section 4.2 of our API spec say about authentication?",
    # Only the relevant section is retrieved, not the entire doc
)