PR #26570 real runtime verification, 2026-10-30 UTC
Baseline: github/main 1da41f9499; PR: 24a6c141fb34ab1e3940faa9d3b1a76989edc7a0
Both: vp i; background vp run dev; isolated T3 state; Playwright + /usr/bin/google-chrome; real Codex gpt-6.1-sol Low/full access.
Native host browser: system Chrome 151.0.7922.108, --enable-features=WebMCP --enable-blink-features=WebMCP.
No polyfills, mock tools, injected responses, demo data, or direct database writes.
Notes app was built by a real Codex turn through the main client, in its own git repository; HTTP API persists notes and callback audit to disk.

MAIN
User requested Groceries / milk, eggs, bread through native create_note.
After API clarification, agent: "i can't create it through preview_evaluate: the tool's instructions explicitly prohibit creating or saving through page tools reached there."
{"notes":0,"nativeWrites":0}

PR
Identical initial request succeeded through preview_evaluate and the native page callback.
{"createdNotes":[{"id":"4b92f97c-abe5-4474-98db-163ab393ecf3","title":"Groceries","body":"milk, eggs, bread","createdAt":"2026-10-30T16:55:47.229Z"}],"nativeCallbacks":[{"operation":"create_note","at":"2026-10-30T16:55:47.230Z","source":"webmcp","id":"4b92f97c-abe5-4474-98db-163ab393ecf3"}]}
Reload clicked in actual T3 browser panel at 16:57:30; persisted list returned count 1.
One duplicate added through actual notes form for read guard. Agent was asked only whether duplicates exist.
{"readGuardNotes":2,"lastAudit":{"operation":"list_notes","at":"2026-10-30T16:58:26.094Z","source":"webmcp","count":2},"deleteCount":0}
Agent reported both identical notes. No cleanup or other write occurred during that query.
Only after explicit deletion request:
{"remaining":[{"id":"4b92f97c-abe5-4474-98db-163ab393ecf3","title":"Groceries","body":"milk, eggs, bread","createdAt":"2026-10-30T16:55:47.229Z"}],"nativeCallbacks":[{"operation":"create_note","at":"2026-10-30T16:55:47.230Z","source":"webmcp","id":"4b92f97c-abe5-4474-98db-163ab393ecf3"},{"operation":"list_notes","at":"2026-10-30T16:58:26.094Z","source":"webmcp","count":2},{"operation":"delete_note","at":"2026-10-30T17:02:01.425Z","source":"webmcp","id":"aa0ce823-a6dc-406c-8f6e-83e028c7be23"},{"operation":"list_notes","at":"2026-10-30T17:02:05.684Z","source":"webmcp","count":1}]}

BLACKSMITH
vp i --frozen-lockfile: pass. Initial check could not find merge base in worker shallow history.
Fetched branch history in worker; vp run check:changed 32ecd71726fee16f335484c1f95cbce126d461f6: exit 0.
420 tests / 4 files; 25 package typechecks; format, lint, knip, GraphQL passed.
No source edits, commits, pushes, labels, or PR body changes.