PR: https://git.cafe/pingdotgg/t3code/pulls/26613
Head: c142b1b4b8015667fa3b50d75cf59ac579c63355
Main: 616c426f5f199865e34974777740dff678ffb7f5
Linux; repository Playwright 1.60.0; /usr/bin/google-chrome.

Commands: git worktree add ../main github/main; vp i (both checkouts).
Only one dev server at a time, started with setsid nohup vp run dev.
Main: server 16733/web 8693, home /home/user/main/.t3.
PR: server 13773/web 5733, explicit home /home/user/t3code/.t3.
Observed real SQLite StatementSync.all calls with a NODE_OPTIONS preload.
The observer called the original method unchanged and logged only SQL and row counts.
Refreshes below were clicks on the web usage page and used server.getUsageSummary.

MAIN 16:01:30 initial: 2 message payload rows fetched.
MAIN 16:02:00 refresh: 2 unchanged message payload rows fetched again.
PR   16:04:48 initial: 2 message payload rows fetched.
PR   16:04:59 refresh: 0 message payload rows fetched; fingerprint listing only.
PASS: original main and PR OpenCode buckets deep-equal, including costs and sessions.
Original: input 27373, output 7, total 27380, 1 session, cost $0.054816.

Created project pointing at real git repository via Add project > Local folder.
Created Codex gpt-6.1-sol thread in app; response: codex-usage-beta.
Added OpenCode provider through Settings > Providers.
Selected localproxy/gpt-6.1-sol in model picker and sent real turn in app.
Response: opencode-usage-beta. No mocks, fixture writes or demo provider used.
One earlier real OpenCode CLI turn supplied the shared before/after database.
PR   16:07:57 refresh after app turn: 2 new payload rows fetched, old rows reused.
New: input 64718, output 16, total 64734, 2 sessions, cost $0.129596.
PR   16:08:30/41 unchanged refreshes: 0 payload rows fetched.
PR   16:08:43 7-day window: 0 payload rows fetched, same total/cost/sessions.
PR   16:08:44 90-day window: 0 payload rows fetched, same total/cost/sessions.
OpenCode source status ok, skippedFiles=0, malformedRecords=0.
Cost/Tokens switching and return to 30 days worked. Codex usage remained in RPC output.

blacksmith-check: vp test run packages/provider-opencode/src/server/usage.test.ts
7 passed.
blacksmith-check: vp test run apps/server/src/usage/UsageService.test.ts
42 passed.
Existing head CI: 29 successful, 0 failing, 0 pending.
No source changes, commits or pushes.

Limits: database has 4 real message rows, not a multi-GB history.
No large-history speedup claim; observed whole-summary warm durations were 13ms main,
15ms PR before append. These tiny runs prove payload reuse, not timing improvement.
No native mobile, Electron shell, macOS/Windows, cold disk, or runtime database-replacement exercise.
Codex Limits showed JSON-RPC -32600 on both main and PR; historical usage and real turns worked.
Current diff only has in-memory caching; persistent-cache claims in description are stale.