The appointments are in the list. The sorting runs without an error message. Yet 2 December appears before 18 January, because the program compares character strings instead of date values. A coding agent can explain this bug at length and still fix it in the wrong place. The interesting question is therefore: does it hold the entire path together, from the observation to a change that can be verified?
Kimi Code has listed K3.1 Pro as available since 23 October 2026. The model introduction aims the offering at programming and longer agent tasks, among other things. As of the sources on 25 October, the full weight files had only been announced for a later date. Service access and a local model package remain two different states.
A reproducible bug beats a big wish
A local sample file is enough for a first assignment. Create three appointments, each with a date and an ID. Deliberately use the day.month.year notation so that an alphabetical comparison cannot pass for a true ordering by date. A fourth line without a date forms an additional state that your assignment must handle explicitly.
termine = [
{"id": "A", "datum": "02.12.2026"},
{"id": "B", "datum": "18.01.2026"},
{"id": "C", "datum": "18.01.2026"},
{"id": "D", "datum": None}
]The desired order can be described in advance: B and C before A, and the undated appointment in an agreed position. For identical days, the input order is preserved. This rule matters more than an example output that happens to fit. It says how the program should behave with further inputs.
Also give the agent the sorting section that is actually faulty. A guess about the cause is no substitute for reading the existing code. The assignment can require it first to name the specific point of comparison and then to change only the part needed for it.
Ask for three results separately
Have the diagnosis, the change and the evidence delivered separately. The diagnosis states which form of data is currently being compared. The change shows the limited intervention. The evidence refers to the same inputs and the rules set in advance. This breaks a long agent answer down into three deliverables that can each be checked.
A suitable instruction is: “Reproduce the wrong date order with this data. Change only the sorting logic. Keep identical days in their original order and put missing date values at the end. Then show the original case and a further case across a month boundary.”
Add what must remain untouched: filters, display and stored records. Otherwise an agent may deliver a seemingly elegant solution that rebuilds the interface or the data format along the way. A larger intervention is not automatically a more careful one.
A small reference path without a model call
You can follow the underlying rule independently of the agent. In Python, an existing date string can be converted into a date value with datetime.strptime. The undated case needs its own sort key. This is one possible reference set-up, not a claim about the code of any particular project:
from datetime import datetime
def schluessel(eintrag):
wert = eintrag["datum"]
if wert is None:
return (1, datetime.max)
return (0, datetime.strptime(wert, "%d.%m.%Y"))
ergebnis = sorted(termine, key=schluessel)
assert [x["id"] for x in ergebnis] == ["B", "C", "A", "D"]This teaching example covers valid strings and None. Invalid calendar days are a further requirement and are not silently repaired here. If you want to support them, you should also specify the desired error message or handling.
What a small success actually shows
A passed sorting case would be a success for this task. It proves neither error-free agent work nor general suitability for a large repository. That is exactly why the small task is useful: you can observe whether a requirement gets lost along the way and whether the agent judges its change against the original problem.
Keep the request, the actual patch and the test call that was run separate from one another. A written claim that “all tests passed” only becomes evidence when the call and its result are actually available. Here, the practical way into longer AI work consists of a manageable bug, clear boundaries and a result you can understand yourself.
Sources and status
Sources as of 25 October 2026. Kimi Code — What’s New, K3.1 Pro · Introducing Kimi K3.1 Pro.
