Environment for all captures: Linux, separate detached Git worktrees at the reported commits; each runs its own real T3 dev server with T3CODE_DEV_SEED=0 and a new database. No live database was read or copied. Dependencies installed separately with vp i. Web browser: 1280x800 CSS pixels, dark theme, device scale 2. Evidence crops the same conversation region on both sides: 1440x1200 physical pixels at (528,104), resized to 720x600. This removes runtime identifiers and local environment details without changing conversation content. Seed: default development project, Device hub off, global Agent device access on, no project overrides. Settings written through authenticated server RPC. This sets the same relevant state without requiring a second linked environment merely to enable the UI toggle. Interaction on both sides: start a new Claude thread with the identical prompt in driver.sh; record from the empty conversation, navigate to the newly returned thread, wait for completed parent response. Each run makes one actual native Agent call. Parent initialization and native SubagentStart hook responses were captured by a transparent executable wrapper passing stdin/stdout unchanged to the real Claude executable. Uploaded protocol JSON contains selected device-related fields and the completed result; unrelated context, credentials, runtime identifiers and paths are omitted. The JSON is protocol evidence, not a test assertion or a synthetic rendering. A direct tools/call device_list probe using the real thread's MCP credential was then made against each real server. Both calls failed; the corresponding real responses are in the protocol JSON. No simulator was launched. Each screenshot crop was compared against a second screenshot at least 500 ms later and had zero pixel differences. Video review contact sheets and final screenshots were opened and visually inspected before upload. Multiple identical reproduction runs were made while preparing captures; published videos show complete fresh runs. Screenshots show completed runs of the same reproduction. Scripts: rpc.mjs and capture-claude.py are the runtime scripts used. driver.sh packages the same seed and launch commands with checkout, output and model supplied as arguments to avoid publishing private runtime metadata. Raw logs must remain private. The wrapper records each process to a new timestamped directory beside itself. AI-assisted; verified with the tests and tools listed above.