Skip to content
AI usage dashboard

Recovering missing current-day Codex usage

2026-10-07. Windows 11 Pro x64 10.0.26300, statistics timezone Asia/Shanghai, application 0.2.1, Node 24.21.0, Cargo 1.98.1; existing dependency locks. Scripts, logs, consistent backups and independent calculations are under root build/codex-today/; native sources remain read-only.

Cause and correction

Before correction, the statistics database’s latest Codex details were on 2026-10-06. Today’s native rollouts already contained individual token_usage_record entries. The existing parser could read all six usage fields from 0.160.1 / 0.162.0-alpha.2; both retain latest_fallback rather than being registered as completely verified versions. Codex was enabled. The most recent startup collection was interrupted with source_window_budget_exhausted; complete lines retained for next scan. Registered file dates ended at 10-05; today’s files had not been visited.

Scanning previously sorted ascending by date. One historical file reaching the default 32 MiB read window ended this source’s scan. It still preceded newer files in later scans, so historical backfill delayed current-day import. The correction retains bounded reads/interruption rules and enables framework file rotation for Codex using persisted last-visit times. Unvisited files come first; ties use descending date paths. Other adapters retain their original order. Confirmed events, fingerprints, generations and complete-line cursors still commit together; failures do not consume files early. No database clearing, original rollout edits or addition of cumulative repeats.

Regressions and native comparisons

  • Three new codex_file_fairness.rs cases cover today’s small file alongside large historical data, and reopening the database after a newer large file exhausts its window so an unvisited small file comes first and all files eventually complete. The desktop parallel entry also tests forwarding this rotation rule. Both initial cases failed before correction (exit 101); all three passed afterward.
  • The first targeted run covered file rotation, Codex incremental reads, parallelism, commit rollback and interruption: 19 passed, exit 0. Repeat reads add no duplicate calls; incomplete lines, replacement/truncation/rename and rollback rules remain.
  • SQLite Online Backup includes valid WAL data. First run the corrected Windows release LLMUsage.exe --scan-once --data-dir <backup-directory> on a consistent backup, enabling only Codex. The first run still had an unfinished file window; the second source scan succeeded. Process exit 0 alone was not treated as complete source success.
  • An independent script deduplicated existing native per-call records by stable response identity and selected completion times on Beijing date 10-07. Cutoff 2026-10-07 14:01:37.124 +08:00: 719 calls, input 69,722,873, cache read 65,866,240, cache write 0, output 408,863, reasoning output 115,244, total 70,131,736. Cache read belongs to input and reasoning to output; these six values are not all added together. All 719 identities, timestamps and six fields matched imported rows individually.
  • With the application stopped, the corrected client’s single-writer lock and native manual collection restored the actual statistics database. Both processes exited 0; Codex’s second scan succeeded. The same cutoff matched every row. All 4,650 pre-existing Codex detail digests/conflict flags remained; added historical backfill was counted separately. PRAGMA quick_check=ok; foreign_key_check returned no rows.
  • Final client comparison at 2026-10-07 14:12:53.983 +08:00: 781 calls, input 75,709,478, cache read 71,722,496, cache write 0, output 435,339, reasoning output 127,681, total 76,144,817; all identities/times/six fields matched. Another native manual scan retained 781 calls/76,144,817 tokens in that cutoff and all 4,650 original details. Both source scans succeeded with no updates/errors. Their 47/two added events include later calls/new backfill; ongoing-session additions are not duplicate collection.

This verifies existing local rollouts and database recovery only. Ongoing sessions’ post-cutoff usage is separate. GUI display, release installation/upgrade, other Codex versions and calls never written to disk were not verified by this comparison.

Commands and delivery checks

Targeted command:

Terminal window
cargo test --manifest-path desktop/src-tauri/Cargo.toml -p llm-usage-core --locked --test codex_file_fairness --test codex_incremental_v12 --test scan_commit_rollback --test scan_controls --test parallel_scans

npm run build:desktop exited 0 for Windows release/NSIS, with artifacts in the existing desktop/src-tauri/target/release/. Building does not establish installation or publication.

The first full check found an oh-my-pi fork test depended on existing result order: old file unchanged, new file complete. Rotation was explicitly restricted to Codex, whose verified files can be read independently. Other adapters retain discovery/conflict order, and the regression assertion was unchanged.

Final npm run verify exited 0: Rust 1,028 passed, eight platform-dependent tests ignored; frontend 22 and scripts four passed; Svelte zero errors/warnings; fmt, Clippy, documentation and frontend build passed. npm run test:headless exited 0: 11 checks on the actual executable with isolated synthetic data, three calls/75 tokens. Final documentation lint and git diff –check exited 0, without temporary artifacts in tracked files.

Initial terminal startup failed with CreateProcessAsUserW failed: 5. Retrying with the execution environment’s permissions restored it. This tool-startup error was not a product-test result. No new GUI/IPC, cross-platform CI or release-installation acceptance.