PR 28422 real dev verification, 2026-11-01 UTC
GitCafe: https://git.cafe/pingdotgg/t3code/pulls/28422
Main: 5732bbee397f0a2b0e61d0c2e432f01dca6df3dc (github/main)
PR head: a79ca6246e1eeeb0f0225028e439f896ee521297 (only21mil:fix/pr-batch-starvation)

Setup
- Separate main worktree; vp i on main and exact PR head.
- One background dev server at a time, stopped by captured process-group PID.
- Playwright from apps/web dependencies, /usr/bin/google-chrome.
- Project /home/user/t3code created through the web app; real git checkout.
- Codex gpt-6.1-sol user turn completed on both versions.
- 110 distinct real GitHub PR links added through the running server WebSocket command path.
- Existing GitCafe PR link remained in the same thread, exercising a mixed-host sweep.
- No mock providers, fake responses, fixtures, direct database writes, or demo data.

Baseline explicit refresh, 03:03:20 UTC
110 GitHub summaries, 14 actual GraphQL requests, maximum 25 summary reads in flight.
All 110 refresh RPCs accepted. All summary reads succeeded.
PR explicit refresh, 03:08:56 UTC
110 GitHub summaries, 15 actual GraphQL requests.
Refreshes arrived over several queued sweeps, so this did not demonstrate an improvement.

Matched cold startup comparison
Used two copies of the same stopped app-created state after a normal cache invalidation.
No changes to production code or batching delays. Observational fetch wrapper forwarded requests unchanged;
it recorded only operation name, timestamp, PR numbers, and alias count. Native server traces provided
summary concurrency and successful completion. Both servers used their normal background sync reactor.

main, 03:11:12.087 to 03:11:16.393 UTC request launch range
110 summaries in 31 PullRequestSummary requests; maximum 25 summary reads in flight.
Batch sizes: 8,8,6,2,4,1,1,7,3,4,4,1,2,6,3,7,4,2,1,1,1,4,2,4,6,7,1,1,2,3,4
110 summary spans succeeded; 0 failed. One completed background sweep.

PR, 03:13:15.412 to 03:13:17.958 UTC request launch range
110 summaries in 22 PullRequestSummary requests; maximum 100 summary reads in flight.
Batch sizes: 2,2,1,1,2,3,2,4,2,6,11,7,6,17,15,6,5,6,2,4,5,1
110 summary spans succeeded; 0 failed. One completed background sweep.
The cold sweep used 29% fewer requests. Batching remains fragmented; this VM did not reproduce
the author's 1.6 PR/request severity or claimed 3.5x to 5x improvement.

Checks on Blacksmith at this head
Focused reactor, service and GitHub provider tests: 182 passed.
vp run check:changed against merge base 54c5f25249961a1aa0b61932329092eadbd9d6fc:
format, lint, knip, GraphQL generation, 18 package typechecks, and 184 tests passed.
The worker's shallow source snapshot initially lacked a usable merge base; fetched original
head history on the worker before running the successful check.
Hosted CI: 28 successful, 0 failing, 0 pending.

Limits
Only Linux web client exercised; native desktop, iOS, Android, Windows and macOS unverified.
Mixed GitHub/GitCafe summary reads exercised, but a live saturated 25-request unbatched pool was not.
The existing deterministic mixed-provider test passed; other host credentials were unavailable.
Startup showed a Codex status-probe timeout; a real Codex turn still completed successfully.
No commits, pushes, merges, rebases, labels, or changes to other PRs.