LimitWarnerProcessor

The LimitWarnerProcessor is a standalone history processor that injects warning instructions into the next model request as configured limits approach.

It is useful when you want the agent to finish efficiently before:

• request count reaches its maximum
• current message history gets too close to the context window
• cumulative run token usage reaches a budget cap

Basic Usage

Python

from pydantic_ai import Agent
from pydantic_ai_summarization import create_limit_warner_processor

processor = create_limit_warner_processor(
    max_iterations=40,
    max_context_tokens=100000,
    max_total_tokens=200000,
)

agent = Agent(
    "openai:gpt-4.1",
    history_processors=[processor],
)

How It Works

1. Removes any warning parts it generated on earlier turns
2. Checks configured limits against the current RunContext
3. Measures current context size from the cleaned message history
4. Appends a new trailing ModelRequest whose UserPromptPart carries the warning text (a separate user turn, not extra system text on the last message)

The warning is ephemeral for context-window pressure: once compaction reduces the history size, the old context warning is removed and not re-injected.

Iteration and total-token warnings are monotonic: if those metrics are still above threshold, the processor injects an updated warning again on the next turn.

Configuration

Python

from pydantic_ai_summarization import LimitWarnerProcessor

processor = LimitWarnerProcessor(
    max_iterations=30,
    max_context_tokens=90000,
    max_total_tokens=180000,
    warn_on=["iterations", "context_window", "total_tokens"],
    warning_threshold=0.75,
    critical_remaining_iterations=2,
)

Ordering with Other Processors

Processor order matters:

• Put LimitWarnerProcessor after trimming or summarization processors if you want warnings to reflect the post-compaction history.
• Put it before them only if you explicitly want a pre-compaction warning pass.