# Runtime reproduction

AI-assisted; verified with the tests and tools listed above.

Both sides ran the real development server through `vp run dev:server --no-seed`, on loopback, in separate worktrees with fresh file-backed SQLite databases. `T3CODE_DEV_SEED=0` was set. Synthetic projects and threads were created through HTTP/RPC, never copied from live state. No unit or integration test result is used as evidence.

The replay scenario runner was unsuitable: generic replay cannot emit Claude native background-task completion. The native Claude adapter produced the continuation on both sides. The one-shot reader-worker gate is the supplied deterministic scheduling technique, adapted to these revisions. It holds the actual user's lineage SELECT until the genuine continuation reaches intake. It does not replace the continuation or change intake locking.

The same driver starts a native background Bash request to a driver-owned localhost fixture, waits for the initial turn to complete, holds the continuation after preflight, dispatches the user follow-up, and releases the lineage SELECT on agent intake. It then cancels ordinal 3 on both sides. Before, ordinal 3 belongs to the user and is cancelled. After, ordinal 3 belongs to the continuation, remains never-started, and the user's run 2 continues through a foreground fixture request to completion.

`before.cast` and `after.cast` contain timestamped selected driver output merged with intake/worker milestones. The MP4 files render those recorded lines at their original 1x timing; the PNGs show their final frames. These are server-output recordings, not browser screenshots. Full sanitized runtime transcripts, server logs, persisted run observations and intake traces accompany them. Paths and private runtime model configuration are redacted; all other captured data is synthetic. The recordings use the same 1440x1120 viewport, monospace font and dark palette.

## Reproduce

In each commit's own worktree, run `vp i`, apply `runtime-gate.patch` (its context offsets differ on base), and copy `reproduction-driver.ts` to `apps/server/scripts/proof24335.ts`. Set `PROOF_MODEL` to an available native Claude model alias privately and `PROOF_OUT` to a new evidence directory, then run `node apps/server/scripts/proof24335.ts`. The published driver externalizes the private model selection; otherwise its logic is the same as executed. The driver starts/stops its own server and logs its owned PID. Every run creates a fresh T3 home. No VITE origin overrides are used.

Base: merge-base [3a8c250cb0](https://github.com/pingdotgg/t3code/commit/3a8c250cb0c8056f2f563091fd412fe5db809fa1). Head: [8864fe682b](https://github.com/pingdotgg/t3code/commit/8864fe682b506b1c86df4cd3a95dc098ba6b469f).

Limitations: this demonstrates deterministic overlap, not spontaneous race frequency. Browser, desktop, mobile, relay/tunnel and other providers were not exercised. All owned servers were stopped and their ports checked closed before cleanup. No source changes were committed or pushed.