PR 21277 real dev verification, 2026-11-01 Baseline: github/main 6d584fa8150965cda908268e65a393502064f261 PR head: 6c751866114e6ef57f623fd8cbcba760056250ee Setup: separate main worktree, vp i on each revision, one background dev server at a time. Playwright 1.60.0 from repository dependencies, /usr/bin/google-chrome, 1440x1000. Project /home/user/preview-verification added through the web UI. Real Codex gpt-6.1-sol parent and native child turns; no mocked provider or seeded scenario. After baseline completion, copied its stopped development userdata into PR isolated .t3 for matching-data comparison. PR server: vp run dev --home-dir /home/user/t3code/.t3; server 13773, web 5733. Actual browser tooltip hover: RegExp.prototype.exec observation, preserving original return values and attributing calls by subagentDetailPreview stack. Same completed real child result: main input length 14935, maximum observed regex cursor 14924, flags gu PR input length 14935, maximum observed regex cursor 282, flags uy The main global replacement searches the complete result; the PR stops near the visible prefix. This is scan-bound evidence, not an end-to-end latency or CPU percentage measurement. Matching-data tooltip screenshots retain the same preview text and ellipsis. Fresh PR turns: Short Preview: completed in 21s; full SHORT RESULT: README read successfully, no ellipsis. Clicking Short Preview opened its actual progress and final response. Progress Preview: running status observed, then completed; final PROGRESS COMPLETE. Fresh long child: input length 8742, maximum observed regex cursor 281; completed in 51s. No repository files edited by these verification turns. Blacksmith: vp test run packages/client-runtime/src/state/subagentDisplay.test.ts: 5 passed vp run check:changed 968dd51704441138559a3fe9c348719f2b11e0c6: passed format, lint, knip, typecheck (7 packages), tests (5) passed. CI at exact PR head: 27 successful checks, 0 failing, 0 pending. Unverified: native mobile, Electron shell, other providers, app-owned delegation, failed/cancelled lifecycle through live app, blank fallback and UTF-16 boundaries through real provider. The focused suite covers formatter edge cases. Native Codex running tooltip did not expose progress detail, so live progress-text priority is not claimed. No end-to-end performance claim.