Two changes that together kill the "prompt is too long: 1991577 tokens"
crash on large Greenplum catalogs.
1. Overview table cap (build_overview_postgres):
The agent re-injects the full overview into the system prompt every
turn. On a multi-thousand-table catalog (typical for GP data lakes)
this was megabytes per call and quickly overran even Kimi-K2's 256K
window. Now we list at most 400 tables and replace the rest with
per-schema counts plus a hint to call list_tables for specifics.
2. Friendly Fireworks-400 rewrite:
Detect "prompt is too long" / "maximum context length" in the
Fireworks 400 body and turn it into an actionable message asking
the user to /compact, instead of leaking raw API JSON to the chat.