SessionStats

Total LLM API requests made.

One completed API request counts once, however many stream chunks carried its usage: record_message_usage revises a request in place rather than recording each chunk separately.

Cumulative input tokens across all LLM requests.

Cumulative output tokens across all LLM requests.

Cumulative prompt tokens served from provider caches.

Cumulative prompt tokens written to provider caches.

Cumulative estimated USD cost across priceable LLM requests.

Requests with a cost estimate, including estimates of literal zero.

Wall-clock duration from stream start to end.

Per-model breakdown keyed by (provider, model_name).

Populated only when record_request receives a non-empty model_name. Empty dict means no named-model requests were recorded; print_usage_table omits the model table in that case and shows only the wall-time line (if applicable).

Per-type breakdown for assistant, nested, and hidden model spend.

Accumulate usage for one completed LLM request.

Updates session totals plus the per-model and per-type breakdowns.

Reverse the record_request call that produced recorded.

A request whose usage arrives across several stream chunks is recorded as soon as the first chunk lands, so the display can react, and then re-recorded with its running totals as later chunks arrive. Retracting the previous version first keeps that one API call counted once, with one per-model row, instead of once per chunk.

Takes the ledger entry rather than loose values so the retraction cannot drift from what was recorded: a mismatch would desync the session totals from the per-kind and per-model breakdowns silently.

Merge another SessionStats into this one (mutates self).

Used to accumulate per-turn stats into a session-level total.