Hotel Receptionist Scenario Expansion#6186
Conversation
Adds 11 new scenarios (check-out early, dinner move/cancel, wake-up move, red-eye hold, valuables/liability, local-area, callback-to-finish, hostile free-night, can't-verify change, room/floor confirm) plus the example logic they exercise: view-based room moves (agent.py, hotel_db.py, modify_booking.py) and restaurant-reservation modification (hotel_db.py), and two new policy docs (local_area.md, safe_deposit.md). Ported on top of the benchmark PR branch so the gradeable expected_state versions of shared scenarios are preserved.
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
| try: | ||
| report_dict = report.to_dict() | ||
| report_dict["tags"] = sorted(ctx.tagger.tags) | ||
| report_dict["evaluations"] = ctx.tagger.evaluations | ||
| report_dict["outcome"] = ctx.tagger.outcome | ||
| report_dict["outcome_reason"] = ctx.tagger.outcome_reason | ||
| with open(os.path.join(report_dir, f"session_report-{room}.json"), "w") as f: | ||
| json.dump(report_dict, f, indent=2) | ||
| except Exception: | ||
| logger.exception("error dumping session report") |
There was a problem hiding this comment.
🚩 run_artifacts.py references potentially new SessionReport/Tagger API surface
dump_run_artifacts calls report.to_dict(), ctx.tagger.evaluations, ctx.tagger.outcome, and ctx.tagger.outcome_reason (run_artifacts.py:40-44). These may be newer SDK APIs not present in older versions. The entire block is wrapped in a try/except so a missing attribute wouldn't crash the session, but the artifact dump would silently fail. Worth verifying these APIs exist in the targeted SDK version.
Was this helpful? React with 👍 or 👎 to provide feedback.
| async def start_restaurant_booking(self, ctx: RunContext[Userdata]) -> str | None: | ||
| """Start the restaurant-reservation flow. Call it the moment the caller wants a table - the flow collects date, party size, time, name, and phone itself. Its return is the FINAL result of the reservation: relay it and move on - nothing further to confirm or call afterwards.""" | ||
| reservation = await BookRestaurantTask( | ||
| db=ctx.userdata.db, chat_ctx=speech_only(self.chat_ctx) | ||
| ) | ||
| return ( | ||
| f"You're set for {speak_time(reservation.time)} on " | ||
| f"{reservation.date.strftime('%A, %B %-d')} for " | ||
| f"{reservation.party_size} guest{'s' if reservation.party_size != 1 else ''}. " | ||
| f"Confirmation code: {_speak_code(reservation.code)}. " | ||
| "| reservation complete - relay this to the caller; no further tool call is needed." | ||
| ) |
There was a problem hiding this comment.
🚩 Asymmetric duplicate-prevention: room bookings guarded but restaurant bookings are not
The PR adds a duplicate-prevention guard for start_room_booking (tools_rooms.py:183-195) using last_room_booking and caller_turns_at_last_booking in Userdata. No equivalent guard exists for start_restaurant_booking (tools_restaurant.py:50-61). The Userdata class in common.py has no last_restaurant_booking field. This is presumably intentional — the room booking flow is longer and more prone to model re-entry than the restaurant flow — but it creates an asymmetry. If the same re-entry problem occurs with restaurant bookings, it would silently double-book a table.
Was this helpful? React with 👍 or 👎 to provide feedback.
2 participants
This PR:
Next: