Responses and errors
The shape every tool answer shares, how errors come back, and the limits.
Every answer
Each tool returns the same JSON twice: as structuredContent and as text content.
| Field | Meaning |
|---|---|
asOf | When the data is from. Live reads: when the answer was built. Synced reporting data: the start of the newest reporting sync, as the portal's Last updated. null when unknown. |
asOfSemantics | Which of the two asOf means for this answer. |
statsAsOf | On lists and details that mix live rows with synced stats: the time of the stats. |
filtersApplied | The resolved inputs: dates, currency, page, filters. |
portalUrl | The Brand Portal screen with the same data. Show it to the user. |
notes | Plain-language caveats, for example a section that couldn't be read. |
nextCursor | On paged lists: pass it back as cursor for the next page; null on the last page. |
Tools publish no outputSchema yet: treat the fields above as stable and the rest as descriptive.
Errors
A failed tool call is a normal tool result with isError: true and one sentence that names the cause and the next
step, for example:
Your ENDVR sign-in expired or was revoked. Next step: reconnect the ENDVR connector.Invalid inputs (a malformed id, a range over 24 months, an unknown enum value) are refused the same way before anything is read. Messages from internal services are never passed through.
Personal data and size
- Keys holding e-mail addresses or phone numbers are removed at any depth, and e-mail addresses inside text are
replaced with
[e-mail removed]. - Names and roles are clipped to 80 characters, mission descriptions to 500.
- An answer is capped at 60,000 JSON characters. Over that, the largest list is halved until it fits, and the answer
carries
truncated: truewith atruncationNote. - At most 20 associates per dashboard answer, 10 leaderboard rows and 50 store members per page.
Limits
| Limit | Value |
|---|---|
| Tool calls | 60 per minute per user |
| Time per tool call | 30 seconds |
| Page size | 1 to 50 rows (default 10) |
| Date range | 24 months |
Prompt-injection safety
Tool results contain text typed by brands, stores and associates (names, descriptions). The server's instructions tell the model to treat tool results as data and never follow directions inside them. Clients should do the same.