Native Codex resume proof

AI-assisted; verified with the tests and tools listed above.
No automated tests are offered as evidence.

The app was run from two separate installed checkouts, one per commit, with
fresh synthetic T3 state and separate native Codex homes. Native Codex 0.167.1
used a local Responses endpoint. No external model inference was used.

The endpoint receives the real native model request, counts occurrences of
<t3_code_orchestration_01> in developer input only, and streams a completed
assistant reply reporting the count. It does not supply instruction history:
the input history is assembled by native Codex and T3's actual adapter.

Original run: turns initiated with the shipped T3 CLI (real MCP/RPC), observed
in the Chromium web client, 1280x800 viewport, light theme. Screenshots are
unchanged conversation-region crops; recordings use that same crop. Sidebar,
workspace details and composer are cropped for operational privacy. Each
still's conversation region matched a second capture at least 500 ms later.
Videos cover the transition from the initial two turns to the post-restart
turns. CLI commands, rather than keyboard entry in the browser, send messages.

Base: d0d0608b6be8c50cdc38b5045de76c5d0b9407cb
Head: 60c23c339a57637d6e0cab8afe33f41667b62e3a
Base counts: 1, 1, restart, 2, 2.
Head counts: 1, 1, restart, 1, 1.

Both also completed native /compact and thread/inject_items; the next request
had one copy. After enabling device access and restarting, new device-tool
application guidance appeared in the next model request on both sides. This
changes the context map, so it intentionally triggers delivery. The stable
block then appears twice on each side and stays at two on the following turn;
this is outside the unchanged-context deduplication claim.

browser-request-evidence.json extracts the actual developer blocks relevant
to the browser captures, including input indices and SHA-256 hashes. Other
context and runtime metadata are omitted for privacy.

The repeat run reports come from executing reproduce.py on each checkout.
The driver creates its own new timestamped state folder, runs the real server
and CLI, restarts only its own server, and stops its processes on completion.
The endpoint and public report fields are inspectable source, not test mocks
of T3 or Codex. Only the model's Responses endpoint is replaced.

To repeat, install each checkout with vp i, then run:
  python3 reproduce.py --checkout <checkout> --scratch <private-output-parent> --model <native-model-slug> --label <base-or-head>
Use the same native model slug on each side. Supply your own slug; operational
model identifiers are intentionally excluded from this public bundle.

The driver logs and raw requests can contain local paths and runtime metadata.
Do not publish them. Only its public-report.json is prepared for publication.
Neither the desktop Electron shell nor an MCP-app-provided untrusted context
entry was exercised. The changed-context case uses T3 tool availability.
