# Queue ordering reproduction

Both sides run the real Linux T3 development server, its CLI calling MCP tools, and the dev replay provider on separate fresh databases. No tests are used. Videos render the captured terminal observations at their original timestamps; they are not web UI recordings. State observations come from read-only queries of each server projection.

Base: 47f2e91210fe80b0b0f0339d36d3a2f313de5134
Head: c2d945f6efb985457e8a589f56b3448f150780bb

The scenario documentation excludes timing reproductions. This manual reproduction uses the PR's supplied timing control: identical temporary instrumentation delays the run-end handler by 20 seconds and each replay prompt by 3 seconds. A CLI shim selects MCP auto mode for NEWER and queue mode otherwise. Neither the queue selection nor dispatch logic is altered.

## Reproduce

Create one checkout per revision and run `vp i` in each. Run `python3 instrument.py <checkout>` on each, then start each with `T3CODE_DEV_SEED=0 vp run dev`. Wait for each server to listen. In each checkout, register the workspace with `node apps/server/src/bin.ts project add . --title "Queue ordering evidence" --base-dir .t3`.

Run `python3 driver.py <base-checkout> before`, then `python3 driver.py <head-checkout> after`. The driver creates a fresh replay thread, starts ACTIVE, sends QUEUED while ACTIVE runs, waits until ACTIVE is completed with QUEUED still queued, and sends NEWER in auto mode. It observes all three runs through completion and reports start-time order.

Each run writes a new timestamped output directory beside the driver. Run `python3 render-recording.py <output-directory>` to render its captured terminal stream. The terminal text and original timestamped recording.cast are also supplied. Stop both dev stacks and remove the temporary checkouts when finished.

Observed base order: ACTIVE -> NEWER -> QUEUED. NEWER reports delivery started while QUEUED stays queued.

Observed head order: ACTIVE -> QUEUED -> NEWER. NEWER reports delivery queued and QUEUED starts first.

Natural-load frequency, repeated-arrival starvation, failed-worktree queue holds, restart continuations, and web/desktop/mobile presentation are not demonstrated.

Proof run by GPT-6.1 Sol in Codex.
