ResponseParser._validate_response accepts a number result via isinstance(result["value"], (int, float, np.int64)). But NaN and inf are floats, so they pass that check, get wrapped in a NumberResponse, and are returned as the answer — with no error.

This is the silent-wrong case: when generated code aggregates over an empty result — e.g. df["sales"].mean() after a filter that matched zero rows — pandas returns nan. The user (and any downstream charting code) then receives nan as a confident numeric answer with no indication anything went wrong.

The fix adds a finite check after the existing numeric isinstance check:

if isinstance(result["value"], float) and not np.isfinite(result["value"]):
    raise InvalidOutputValueMismatch(
        "Invalid output: Numeric result is NaN or infinite (likely an aggregation over empty data)."
    )

np is already imported; np.isfinite covers both NaN and inf, and np.float64 subclasses float so it's covered too. Booleans/ints are unaffected.

Includes a regression test (no LLM call) asserting NaN and inf both raise InvalidOutputValueMismatch while a normal number still parses.

Found while building a fault-injection tester for agent tools.