Cross-Timezone Scheduler

ranked by score ↓

cross-timezone

A trap-compatible task that asks an agent to schedule a meeting across attendees in different time zones, given each attendee's local availability window. The agent must return a JSON object with a single canonical meeting time in UTC plus each attendee's local start time.

Layout

cross_timezone/
├── README.md
├── traptask.yaml             # case list + judge/grader cmds
├── judge.py                  # per-case scorer (zoneinfo-driven; strict)
├── grader.py                 # aggregator (score, latency, cost, by-category)
├── gold.cases.json           # source-of-truth case data
├── inputs/
│   └── {case_id}/
│       └── question.txt      # the scheduling brief
└── expected/
    └── {case_id}/
        └── answer.json       # gold UTC window + per-attendee availability + canonical answer

Cases (v0 — 2 cases)

id	category	difficulty	what it tests
`dst_gap_with_ist`	`dst_boundary`	hard	UK still on GMT (DST starts 2026-03-29) and India on UTC+5:30 — catches models that assume "London in spring = BST" or round India to UTC+5
`dst_quarter_hour_sydney`	`multi_zone_expert`	expert	5 attendees · all of `dst_gap_with_ist`'s traps plus Nepal `UTC+5:45` (quarter-hour) + Sydney `UTC+11` (southern-hemisphere DST in March) + day-shift for Sam (their local date is the next day). Exactly one valid start time exists.

Planned additions (still to draft):

Idea	Hits
Simple 2-zone in May	baseline TZ math
3-zone with Sydney (UTC+10/11) crossing midnight	day-boundary
October DST gap (UK ends BST Oct 25, US Nov 1 — 1-week reverse)	symmetric DST
5-attendee overlap (Sydney/Singapore/Mumbai/Berlin/SF)	narrow overlap
No overlap exists → expected `{"start_utc": null, "reason": "..."}`	refusal vs hallucination
Recurring weekly across a DST shift	series consistency

Solution contract

Each solution must:

Read INPUTS env var (JSON dict mapping filename → absolute path).
Read INPUTS["question.txt"] — the scheduling brief.

Print exactly one JSON object to stdout, matching this schema:

{
  "start_utc": "2026-03-26T14:00:00Z",
  "duration_min": 60,
  "attendees": [
    {"name": "Alice", "tz": "America/Los_Angeles", "local_start": "2026-03-26 07:00"},
    {"name": "Bob",   "tz": "Europe/London",        "local_start": "2026-03-26 14:00"},
    {"name": "Priya", "tz": "Asia/Kolkata",         "local_start": "2026-03-26 19:30"}
  ]
}

The judge tolerates markdown code-fence wrappers (```json ... ```) — but not prose. Plain JSON is the canonical format.

What the judge checks (strict, no partial credit)

#	Check	How it's verified
1	stdout parses as a JSON object	`json.loads` (markdown fences stripped)
2	`start_utc` is ISO 8601 with explicit UTC (`Z` or `+00:00`)	`datetime.fromisoformat`
3	`start_utc` ∈ `[expected_start_utc_min, expected_start_utc_max]`	arithmetic
4	`duration_min` matches exactly	int equality
5	For every attendee, `local_start` ≈ converted UTC (± 1 min)	`zoneinfo`-based conversion
6	For every attendee, the meeting fits inside their availability window	naive datetime arithmetic

Any single check failing → score 0.0. All pass → 1.0.

Example: the `dst_gap_with_ist` case explained

Today = 2026-03-25 (Wednesday). Meeting is tomorrow (Thursday 2026-03-26).

Attendee	TZ	Offset on 2026-03-26	Why
Alice (SF)	`America/Los_Angeles`	UTC−7 (PDT)	US DST'd on 2026-03-08
Bob (London)	`Europe/London`	UTC+0 (GMT)	UK DST starts 2026-03-29 — Bob is still on GMT
Priya (Mumbai)	`Asia/Kolkata`	UTC+5:30 (IST)	India has no DST

Their local 2-hour availability windows all align with UTC 14:00–16:00 on this date — overlap window for the 60-min meeting is start times in [14:00Z, 15:00Z].

Common failure modes:

Model thinks	Resulting UTC start	Off by
"London in spring is BST"	13:00Z	−1 hr
"India is UTC+5"	14:30Z (but with wrong local times)	+30 min and inconsistent
Both errors	13:30Z	chaos
Off-by-one date math (`tomorrow` ≠ +1 day)	wrong date entirely	clearly wrong

Wiring up a solution

From a solution dir with a trap.yaml pointing here:

tasks:
  cross-timezone-scheduler:
    cmd: uv run python solution.py
    traptask: /path/to/trapstreet-tasks/tasks/scheduler/cross_timezone
    timeout: 120

Then:

uv run tp run                        # all cases
uv run tp run --fail-fast            # stop on first failure
uv run tp submit cross-timezone-scheduler   # upload to trapstreet.run