PR #28598 verification, 2026-11-01, disposable Linux Boat VM
GitCafe: https://git.cafe/pingdotgg/t3code/pulls/28598
Main: 8760d063e903406c8e8f695b0852c0a15fc1e00f
PR: e5b0a18854339417f09a12d6986b44ece3419cf8

Setup
$ git worktree add ../main github/main
$ cd ../main && vp i
$ T3CODE_DEV_SEED=0 setsid nohup timeout 7000 vp run dev > /tmp/main-dev.log 2>&1 &
main dev-runner: serverPort=16733 webPort=8693 baseDir=/home/user/main/.t3

Used repository Playwright 1.60.0 with /usr/bin/google-chrome.
Paired via the app, added the real /home/user/main git repo as a project,
selected Codex gpt-6.1-sol, submitted real turns through the composer.
Large text output was truncated below the 8 MiB threshold by Codex.
A real view_image(detail=original) call on a local PNG produced an
oversized response_item transcript line of 16,789,137 bytes.
The transcript was produced by Codex itself; no synthetic JSONL or mocks.
Copied that completed transcript into an isolated Codex home and selected
that home through Settings > Providers > CODEX_HOME path, so both revisions
scan identical bytes without counting this verification agent's own work.
Transcript bytes: 17,229,738
SHA256: c57a3c1fc23866883348f57de57429f2c03d039235fd2037b8a4252d2487d99b

Opened Usage > Tokens in the actual web app. Used Node inspector precise
coverage on the running server, without modifying application source.
Main cold scan: createTranscriptJsonReader=1, write=66, parseProjected=1.
This reproduces JSON tokenization of an oversized line with no usage marker.
Main usage: 9 records, 1 session, uncached input=18,220,
cached input=196,224, output=440; processed tokens=214,884.
Malformed records=0. Resume offset=17,229,738.

Stopped main by its captured process-group PID before starting PR.
$ git fetch https://github.com/hieu-lee/t3code.git hieu-lee/usage-skip-unrelated-oversized-lines
$ git switch -c hieu-lee/usage-skip-unrelated-oversized-lines FETCH_HEAD
$ git rev-parse HEAD
e5b0a18854339417f09a12d6986b44ece3419cf8
$ vp i
$ T3CODE_HOME=/home/user/t3code/.t3 T3CODE_DEV_SEED=0 setsid nohup timeout 6500 vp run dev > /tmp/dev.log 2>&1 &
PR dev-runner: serverPort=13773 webPort=5733 baseDir=/home/user/t3code/.t3

PR plain JSONL cold scan through Usage > Refresh usage:
createTranscriptJsonReader=0, write=0, parseProjected=0, checkMarkers=65.
Used an uncached filename for a fresh scan; bytes were unchanged.
All 9 records, parser state, resume offset and malformed count exactly
match main (record comparison normalizes cache dictionary model/session IDs).
The user-facing total remains 214,884 processed tokens, displayed as 215K.

Compressed-history neighbor:
Compressed the same real transcript with Node zstd, removed its plain copy
from the scanned directory, and refreshed Usage in the running web app.
createTranscriptJsonReader=1, parseProjected=1: compressed streaming retained.
All 9 records and parser state match the plain scan; malformed records=0.

Provider/append neighbor on PR:
Added /home/user/t3code as a project through the app.
Submitted two real Codex gpt-6.1-sol turns in the same new thread.
The first executed timeout 10 node -p 2+2 and completed NEIGHBOR_TURN_ONE_OK.
The second executed timeout 10 node -p 3+3 and completed NEIGHBOR_TURN_TWO_OK.
Usage > Refresh usage after first turn: 11 records, 2 sessions.
Usage > Refresh usage after second turn: 13 records, 2 sessions.
Captured actual server.getUsageSummary WebSocket responses:
uncached input=28,986; cached input=263,424; output=562.
Processed tokens=292,972 (293K displayed); cost=$0.0899344 ($0.09 displayed).
Repeated refresh: identical buckets and totals, no duplication.
Malformed records=0. Incremental thread file resume offset=179,165.
Cost/Tokens switching and repeated refresh are visible in the 12-second clip.

Focused checks through blacksmith-check (remote worker, dependencies installed):
$ vp test run apps/server/src/usage/usageTranscriptReader.test.ts apps/server/src/usage/usageTranscriptStreaming.test.ts apps/server/src/usage/UsageService.test.ts
3 files passed, 81 tests passed. Worker stopped afterwards.
Existing head CI: 29 successful, 0 failing, 0 pending.
No source changes, commits or pushes were needed.

Limits
Verified Linux web client and real Codex via the local proxy.
Did not exercise native desktop/mobile, other providers, remote/tunnel modes,
or a real oversized usage-bearing line whose marker crosses read boundaries
(the focused suite covers the split-marker and oversized-usage replay cases).
No production-scale wall-clock speedup claim is made from this corpus.