PR #33418 runtime verification, Linux, 2026-11-10
Baseline: github/main f12bb6229c007a12d24fd2e539222733778ac9ea
PR head: 48ae470c850ed04a1dda1b7dad9ef93a50c8e696

Commands run in the corresponding checkout:
  timeout 600 vp i
  setsid nohup timeout 7000 vp run dev > /tmp/main-dev.log 2>&1 &
  [dev-runner] mode=dev source=worktree /home/user/main serverPort=16733 webPort=8693 baseDir=/home/user/main/.t3
  Stopped main's captured process group before starting PR.
  setsid nohup timeout 5000 vp run dev --home-dir /home/user/t3code/.t3 > /tmp/dev.log 2>&1 &
  [dev-runner] mode=dev source=default ports serverPort=13773 webPort=5733 baseDir=/home/user/t3code/.t3

Data was created through the web UI, including a new Hyperdrive git project.
The matching comparison uses a copy of main's app-created data after main stopped.
Chrome: /usr/bin/google-chrome, driven with repo Playwright.
Real chat provider: Codex gpt-6.1-sol.
Title generator for the Hyperdrive comparison: shipped Codex gpt-6-luna setting.

User request:
  Compare Hyperdrive source mode versus patches for generated code edits. Give two short tradeoffs for the Hyperdrive workflow. Don't inspect or change files.

Observed in the real web app:
  main initial title: Hyperdrive Source Mode vs Patches
  main Regenerate title: Hyperdrive Source Mode vs Patches
  PR Regenerate title, same conversation: Generated Code Edits: Source vs Patches
  PR repeated regeneration recording: Source Mode vs Patches for Generated Edits
  PR initial title for a new thread with the same request: Source Mode vs Generated Code Patches
  Manually renamed the new thread to Hyperdrive manual review.
  Sent a real Sol follow-up asking which approach preserves edits.
  Received an answer recommending patches.
  Manual title remained Hyperdrive manual review after the follow-up and a full reload.

Focused existing checks, run remotely through blacksmith-check:
  vp i && vp test run packages/provider-core/src/server/textGenerationPrompts.test.ts
  Test Files 1 passed (1)
  Tests 2 passed (2)

Limit: this is model guidance, not a deterministic name filter.
A diagnostic using project name Offline Sync and a discussion about offline sync
could still include that phrase when gpt-6.1-sol generated the title.
Native desktop/mobile, other providers, relay/tunnel, and the full model eval were not exercised.