ToolOutputTrimmer summaries can exceed max_output_chars for string outputs

[CoffeeDrivenCoder]

Describe the bug

ToolOutputTrimmer can emit a string summary longer than the configured max_output_chars budget. This happens when the existing rich summary is shorter than the original tool output but larger than the configured budget.

Repro steps

With openai-agents 0.22.0:

from types import SimpleNamespace
from agents.extensions import ToolOutputTrimmer
from agents.run_config import ModelInputData

items = [
    {"type": "function_call", "name": "search", "call_id": "call_1", "arguments": "{}"},
    {"type": "function_call_output", "call_id": "call_1", "output": "x" * 1000},
]
data = SimpleNamespace(
    model_data=ModelInputData(input=items, instructions=None),
)
result = ToolOutputTrimmer(max_output_chars=40)(data)
print(len(result.model_data.input[1]["output"]))

The emitted summary is longer than 40 characters.

Expected behavior

Every supported string function-call output replacement should respect max_output_chars. Existing summaries should remain unchanged when they already fit, and outputs should remain untrimmed when the rich summary would not be shorter than the original.

Debug information

• Agents SDK version: 0.22.0
• Python: 3.12.14
• OS: Ubuntu 22.04
• Reproducible: yes

[seratch]

Thanks for reporting this. I confirmed that a string summary can be longer than max_output_chars while still being shorter than the original output.

The released contract currently defines max_output_chars as the threshold for considering an output for trimming. The string path intentionally keeps its rich preview whenever that preview is shorter than the original, while the stricter replacement-budget guarantee is documented only for structured outputs. The report does not demonstrate a model-call failure, context-limit problem, material token cost, or another supported scenario requiring a hard cap.

I am going to close this issue for now. If a real application requires a strict per-output bound and the existing call_model_input_filter extension point is insufficient, we can reconsider that narrower product requirement.