get_store_metrics returns ~1.9 MB by default (every store object × ~30 days) — ~500k tokens, larger than any agent context; even days_back=1 is 346 KB #3903

Summary

get_store_metrics returns ~1.9 MB by default on every production store — roughly 500 k tokens, more than any MCP client's context holds. Even days_back = 1 returns 346 KB (~90 k tokens), because the payload always lists every store object and the window only shortens each object's series. There is no object filter, object_kind filter, or top-N.

This is response size, distinct from #3898 (tool-definition size). The tool's own description is ~9 KB and contributes to #3898.

Measured (nightly 458, 2026-09-22)

Composition of the default response (PostgreSQL-target store, 251 store objects):

Why it matters

The questions this tool exists for — what is driving growth, which job is outgrowing its cadence, is retention running, how much can the inventory see — are answered by the ~9 KB of summary blocks plus a handful of objects. The per-object series for all 251 objects is almost never what the caller wants, yet it is always what the caller pays for. An agent that calls this tool loses its context; a client with an output cap sees the payload truncated from the tail, which drops whichever blocks happen to serialize last.

Proposal

1. Summary-first default: whole-store daily series, bytes_by_kind, inventory reconciliation, jobs, checkpointer, retention holds, and the top N objects by size and by growth (N ≈ 10).
2. Opt-in detail: an object_name / object_kind filter that returns the full series for the objects asked about.
3. A bounded page like the other list-shaped tools (the payload-contract census, Three PostgreSQL pages observe truncation off limit + 1 through one shared helper, the census sweeps the inference class instead of a roster, and four payload rules become facts and inventories, so three tools stop guessing truncation from a full page and four of the eleven payload rules stop being sentences (partial #3653) #3699, has the page-bounds helper).

Acceptance

• Default get_store_metrics under ~50 KB on the largest production store.
• Every question the description names still answerable, the per-object series via an explicit filter.