query_email_and_calendar returns greeting prompt with empty sources for all queries

[dsather]

The query_email_and_calendar tool consistently returns a generic greeting prompt with an empty sources array, regardless of query content. Every other MCP tool on the same session works.

Account / session: 2026-05-28, Pacific time
Tool: Superhuman_Mail.query_email_and_calendar
Endpoint: https://mcp.mail.superhuman.com/mcp

Reproduction

6 calls in the same session with deliberately varied phrasings (declarative, imperative, JSON-request, short, long, "yes-proceed"):

1. "What meetings and events do I have today (2026-05-28)? Include times, attendees, and any context about what they're about. Exclude denied meetings."
2. "List my calendar events for Thursday May 28 2026 with title, start time, end time, attendees, and location."
3. "List the 5 most urgent unread or starred emails in my inbox from the last 24 hours."
4. "Return a JSON list of calendar events on 2026-05-28 with fields title, start_time, end_time, attendees, description. Do not respond conversationally; execute the search."
5. "today's calendar"
6. "yes please proceed: list every calendar event scheduled for May 28 2026 with title, start, end, and attendees"

Actual

All 6 returned (modulo trivial wording):

{
  "answer": "Hello! I'm here to help you search your emails, manage your calendar, and find information. What would you like help with today?",
  "sources": []
}

Expected

Calendar events and/or email threads matching the query.

Healthy tools on the same session

Rules out auth / network / per-account rate limiting:

• list_labels — returns full label set
• list_splits — returns 5 splits with real counts
• get_availability — returns free slots
• list_threads (limit ≤15) — returns threads
• get_read_status_feed — returns recent read events

Hypothesis

The query field is not being passed into the underlying agent's user-message slot. The agent is being initialized with no user input and returning its default opener. The empty sources array supports this — the agent never executed any search tool.

Secondary, lower-priority

The workflow skills in this repo (morning-briefing, batch-draft-writer, eod-wrapup) hardcode list_threads with limit: 50. On at least one MCP client, that response overflows the per-tool-result token budget (~54 KB returned for our account). Recommend defaulting to a lower limit and paginating via next_cursor.

---

Also filed via support@superhuman.com earlier today for redundancy.