Provider-reported Claude and Gemini breakdown usage is recorded before output parsing and semantic validation. Failed or refused responses with reported usage remain in the estimate even when a capture allowance is released. Token totals now work before the month’s first capture. Usage/task counters and a provider-response receipt are written atomically; repeated receipt delivery does not double count. Receipts expire after 90 days; monthly aggregates remain. Network failures with no observable usage remain outside the estimate.
Expired inbox rows are removed before claiming work. Captures deleted during retrieval stop before analysis. Background breakdowns retain the entire supported source; large results are split into bounded SQLite values instead of silently shortening text. Result writes are atomic and chunks are cleaned up with the inbox. Old stored text cannot be reconstructed by this update.
Completed captures are saved on-device before matching, then use the existing durable connection/placement queue. Server-scoped deterministic inbox IDs and transactional insert-if-absent prevent duplicate text-only imports and preserve later writing after acknowledgment failure. Available browser locks coordinate inbox imports across tabs. Normal navigation and controls remain unchanged. Gemini’s backup counter is labeled as requests because it can now include failed responses.
No model switch, automatic library reprocessing, context reduction or new spending cap is introduced. Broader paid-operation reuse, server-enforced allowances, restore validation and evidence-budget/model benchmarks remain separate follow-up work. Both worker deployment and the v43 app update are needed for the complete change.