September 28, 2026 · fresh Linux verification
Current main still drops usage
Same history: 2,000 versus 12,000 tokens
The fixture contains 10,000 Codex tokens and 2,000 Claude tokens. A 517 MiB Codex tool-output record sits between valid usage events. Claude has a 65 MiB record that contains usage. Current main loses the entire Codex file; applying the PR retains all expected usage. Neither server crashes in this reproduction. The remaining main bug is missing usage and high memory, rather than the original unhandled readline crash.
Baseline: main at test start, 72330e22c0. A final check found one newer commit, b528a70110, changing only Windows terminal startup; the usage reader is unchanged. After: the PR’s published commits 79634e89 and 099777b8 applied cleanly to that same main, producing local test commit c4e5a5f744. No source changes were pushed. A direct read on the unchanged published PR head also preserved the large-file usage.
Before and after screenshots
These are actual client screenshots. The bordered test caption states the known fixture total and the exact server response; it does not replace or change product data.
Recorded full rescans
Full-resolution video
Full-resolution video| Same 582 MiB history | Current main | Main + PR |
|---|---|---|
| Expected / observed total | 12,000 / 2,000 | 12,000 / 12,000 |
| Click to displayed test result | 1.514 s | 1.779 s |
| Server scan | 1,337 ms | 1,732 ms |
This case demonstrates accuracy, not a speedup. Main ends sooner with a failed Codex read; the PR completes the read. Each recording changes file modification times to force a full rescan, uses the same 90-day view and runs sequentially. Filesystem caches were not cleared. Previews omit initial navigation and the sidebar; playback is real time. The timer is test instrumentation, measured from the refresh click to displaying the received result. No browser page errors occurred.
Regression checks
379 tests passed: 231 server usage, JSON selector and transcript-import tests across 12 files; 148 shared/client/web usage tests across 15 files. Server typecheck, scoped lint and whitespace checks passed.
| Real browser action | Exact tokens | Result |
|---|---|---|
| 7 days | 12,000 | Pass |
| 30 days | 12,000 | Pass |
| 90 days | 12,000 | Pass |
| duplicate file | 12,000 | Pass |
| unfinished append | 12,000 | Pass |
| completed append | 13,000 | Pass |
| cached refresh | 13,000 | Pass |
| browser reload | 13,000 | Pass |
| Actual server restart | 13,000 | Pass |
All fixture timestamps are current; agreement between date windows checks query switching, not filtering across historical boundaries. The partial append contains half a JSON record; completing it adds exactly 1,000 tokens. Cache refresh after the server restart took 3 ms.
Focused tests cover UTF-8/CRLF boundaries, incomplete tails, malformed large records followed by valid records, file rewrites, Codex model switches and fork-copy suppression, Claude pricing metadata, Grok usage allocation, duplicate keys and import allocation/depth limits.
Fresh performance and parity checks
39 real Codex/Claude histories, 391 MiB and 15,215 usage records: complete reader results matched exactly. Five alternating runs per file gave summed median times of 871.2 → 740.0 ms (15.1% less reader time). This is not a page-load measurement. Private histories were copied read-only for testing; temporary snapshots were deleted and their contents were not published.
| Separate 517 MiB Codex reader test | Main | Main + PR |
|---|---|---|
| Outcome after the appended usage record | Read fails; no records | All 11,000 Codex tokens |
| Median elapsed read | 1074 ms | 1430 ms |
| Maximum process RSS | 1626 MiB | 203 MiB |
Three fresh Node processes per version, alternating order. Timings are medians; memory is the maximum observed process RSS. OS caches were not cleared. These are separate from the browser recordings. A successful read and a failed read are not equivalent performance outcomes.
Scope and merge risk
The fix is still needed, and its local integration with current main passed this test set. The published PR branch remains unchanged and behind main; these local results do not constitute fresh CI on an updated remote branch.
The main regression risk is incorrect usage projection or changed transcript-import behavior through the shared JSON selector. Focused tests exercise both. Selected metadata and nesting remain uncapped, so the server is not protected against every maliciously large input. General incomplete-scan reporting remains separate in #11321. This does not resolve every provider accounting issue.
Execution was on Linux web. Native Windows/macOS, Electron shell, mobile and remote/tunnel flows were not tested in this pass. CodeRabbit’s existing review rates merge risk Minimal and security risk Low; Macroscope still requests human review. No PR was merged or closed.