# Background subagent approval reproduction

Both runs: Linux web client, Chromium preview browser, light theme, 1280 × 800 CSS pixels (2× capture scale). Separate checkouts, fresh T3 homes, T3CODE_DEV_SEED=0. No copied database. Synthetic empty git project titled Permission probe. Same Claude provider configuration and Supervised permissions on both sides.

Base: f03db4fbef59c42113e37c32897bf4d25d31cb2b
Head: a29cb50514444e21958844572f3e1f6af2e234fd

1. Install with vp i; start T3CODE_DEV_SEED=0 vp run dev in each checkout.
2. Create an empty git repository, then register it through `node apps/server/src/bin.ts project add --base-dir .t3 --title 'Permission probe' <repository>`.
3. Pair the browser with the running server. Create a thread in Permission probe, select Claude, and set Permissions to Supervised. Close thread details.
4. Send the exact prompt below and record the run. Wait for the parent to finish and the delayed tools to run. Do not send another user message.
5. Inspect the background subagent's report on base. On head inspect the approval prompts beneath LAUNCHED. After recording, approve Read once; confirm its successful tool result without publishing file contents.

Prompt:

Launch exactly one subagent with the Agent tool and run_in_background: true. The subagent must run these tools in order: Bash `sleep 10`, Bash `echo probe`, Read `/etc/hosts`, Bash `cat /etc/hosts`, Bash `ls /tmp`. It must attempt every step even if a previous step is denied. Do not wait for the subagent. Immediately end your parent turn with exactly LAUNCHED.

Observed: base root result precedes all five tool calls; Read and the final two Bash calls receive the no-active-turn denial, with zero persisted runtime requests. Head root result also precedes all five tool calls; the app displays three approvals, and Read succeeds after approval. The remaining command prompts are left pending.

Captures are cropped to omit private provider metadata. The after still is a close-up of the parent result and the two remaining prompts after Read was approved; the video shows all three prompts appearing. The before still shows the subagent's failure report. Base still screenshots were identical; head close-up screenshots 500 ms apart were identical (live timers are outside the crop). Both videos were inspected through extracted frames. Videos end before unrelated navigation.

Replay scenarios were not used: replay speaks generic ACP and does not exercise Claude's permission callback.

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