# Live reproduction

Run the two commits in separate worktrees, with fresh `.t3` state:
`T3CODE_DEV_SEED=0 vp run dev`.
Run `python3 seed.py CHECKOUT SCRATCH_PROJECT` for each running server.

In each web client (Chromium, 1280x800 CSS pixels, light theme):
1. Pair with its own server.
2. Settings > Providers > Codex: select Codex CLI 0.169.0 and an authenticated scratch CODEX_HOME. The scratch home has API-key auth and the model endpoint, but no command rules, approval policy, or sandbox policy overrides.
3. New thread > Approval reproduction. Use current checkout, Codex GPT-6-Luna, Medium reasoning, Full access.
4. Collapse Thread details to keep environment details outside captures.
5. Start recording. Submit exactly: Execute the shell command `rm -rf ./absent-repro-target` exactly, then report its exit code.
6. Base: wait for Waiting for approval; do not press Approve. Head: wait for the turn to finish and report Exit code: 0.

The recordings show real Codex turns, not replay. The head was run twice with the same prompt; the second recording includes completion. No command was approved manually. The base recording ends with the native command approval still pending.
