Desktop Scope statistics
Inventory and usage view
Observation → Current-Scope inventory & usage contains the Statistics view.
Explicitly select the statistics-qualified sqlite-stats-63e9665c-v1 profile,
activate the connection, select an exact Scope, choose a UTC usage period
and press Load statistics. Loading is explicit; opening the view does not
silently restore a connection or issue a statistics request. The shared shell mounts Statistics in the Observation inventory panel with the
current state, language, pending-context guard and state-refresh callback.
The view displays the Server's as_of, exact returned Scope coverage and
inclusive UTC dates. Inventory is a current snapshot. Model request/input/output
token counts, purpose breakdowns and daily rows retain returned zero and null
values; null tokens display Unknown. The page does not fill missing rows or
infer complete provider reporting from recorded calls. Recall coverage includes
preparation, ready and comparable counts, estimator identity/version, baseline,
recalled tokens and signed reduction. These are estimates, not billing savings
or answer-quality measurements.
Every chart has an equivalent table with captions and row/column headers. Tables and charts have keyboard-focusable scroll regions for narrow layouts. Labels are available in English and Chinese. Changing connection, Scope or period clears the old view and invalidates late success/error responses. A statistics failure does not disable independent Memory actions.
Native boundary
Desktop consumes POST /v1/stats through the native get_stats command. The
renderer supplies only the active generation and today, 7d or 30d; native
code obtains the current Scope and constructs an exact, single-Scope selection.
There is no renderer-supplied endpoint, Scope list or all/subtree mode.
The connection must have a matching qualified operation profile. Each read rechecks identity using the existing connection policy, preserves Server authorization failures, and applies the existing transport timeout and 1 MiB response bound. A denial never retries against a broader selection. Responses must retain the requested selection, Scope IDs and period. Aggregate and per-Scope recall estimators may be null, as permitted by the public contract.
Statistics use their own cancellation channel. A replacement statistics read,
cancel_stats_reads, or a generation change cancels pending statistics; Memory
read cancellation does not cancel them. Submitted Memory writes retain their
existing original-target behavior.
The operation catalog and reachable wire models come from
pnpm --dir desktop generate; IPC declarations come from the Rust export_ipc
example. The native exact-selection projection reuses concrete generated field
types while preserving the nullable estimator and prohibiting broader selections.
Neither Server APIs nor historical qualification records are changed.
See the qualification record for the tested Server/contract/wheel identities and actual validation limits. Synthetic tests establish local handling; the built-wheel harness separately establishes the observed SQLite protocol behavior. The record also distinguishes the exercised macOS ARM64 bundled debug-app journey from Windows installed UI, Intel, IME and screen-reader gates. The Windows installed journey compares the page's tables with independent real Server responses.

