Cut Mid Object

An agent calls search_orders, which matches 900 rows. It then reports order totals that do not exist anywhere in the database, and confidently cites order numbers that are one digit off.

MAX_TOOL_RESULT_CHARS = 20_000

def run_tool(name, args, tool_use_id):
    result = TOOLS[name](**args)              # list[dict]
    payload = json.dumps(result)
    if len(payload) > MAX_TOOL_RESULT_CHARS:
        payload = payload[:MAX_TOOL_RESULT_CHARS]
    return {"type": "tool_result", "tool_use_id": tool_use_id, "content": payload}

What is producing the fabricated values, and what is the fix?

json.dumps does not guarantee key order across runs, so repeated calls disagree and the model averages them.

The slice lands inside a record, so the model receives a JSON document that stops mid-object. Truncate on a record boundary and append a marker saying how many of the 900 rows are shown.

20,000 characters is too small for this workload. Raise the cap to 400,000 characters so the full 900-row result fits in a single tool result.

The cap counts characters rather than tokens, so the real budget varies with the content of the rows. Switch it to a token count.