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

Surface: real web client, Chromium, light theme, 1280x800 CSS viewport,
2x capture scale. Each commit ran its own Node server and Vite client,
with T3CODE_DEV_SEED=0 and its own fresh database. The same synthetic git
repository and same provider configuration were used. Native CLI: 1.0.77.
No tests or mocked protocol fixtures were used as evidence.

The replay scenario runner was ruled out by its documented exclusion of
provider-specific behavior. The bug depends on native background task markers.

Driver: reproduce.py. Run from each prepared checkout:
EVIDENCE_DIR=<private-scratch> REPRO_PROVIDER=<provider> REPRO_MODEL=<model> \
  python3 reproduce.py <base-or-head> <project-id> <success-or-failure>
Provider and model are intentionally supplied outside this public script.
The requested provider/model were used, with low reasoning, shared isolation,
full access, and asynchronous delegation. The provider needed an isolated
configuration with models_base_url and xai_api_base_url pointing to the
existing authenticated API endpoint; this resolved the installed CLI's
startup authentication mismatch without changing either app checkout.

Child instructions, success:
Use only native terminal tools. Run `sleep 5; echo EARLY_BG_DONE` in the
background. Run `sleep 15; echo FOREGROUND_DONE` in the foreground and wait.
Never poll, read, or kill the background task. Then reply exactly
EARLY_SUCCESS_DONE and end.

Child instructions, failure: identical, with the background command replaced
by `sleep 5; echo EXPECTED_BG_FAILURE; exit 7`.

Parent instructions: delegate once, end the turn, await automatic notification,
then call task_status and report notification summary and task_status.summary.
After both runs settled, send this same final check on both sides:
Read task_status for the ORIGINAL delegated task (not the provider-continuation
 task). Report its summary verbatim now that all child runs have settled.
Do not start any new child.

The inspection recordings capture that live final check, then opening the
child and expanding Worked for / Ran 2 commands. Both use the same sequence.
They are shown at 2x speed and cropped to the conversation to omit private
runtime identifiers and machine details. The first failing execution itself
finished before its recording began. The final stills were checked against
another screenshot at least 500ms apart: zero changed pixels in the crop.

Observed success variant:
Base: root answer EARLY_SUCCESS_DONE exists. Child has runs 1,2,3; run 2 has
no assistant answer. Original task summary is the empty-result placeholder,
including on the final recheck after all runs settle. Native run 3 publishes
a separate correct continuation task; it does not repair the original task.
Head: root answer retained in original notification and task_status.summary.
Child has runs 1,2 and each has an assistant answer. Native run 2 publishes
its own continuation task. Final original-task recheck is EARLY_SUCCESS_DONE.

The successful-background variant was repeated once to obtain a recording
from execution. That repeat did not reproduce the base's corrupted summary
(the timing-dependent bug occurred in 1 of 2 matched success runs). Both head
runs preserved the original answer. Do not infer deterministic reproduction.

The failed-background variant was run once per commit: both returned
EARLY_SUCCESS_DONE. Both child histories contain the failed background command
and successful foreground command. This variant did not prove a base/head
difference; its paired recordings and transcript are supporting coverage only.

runtime-transcripts.json contains selected real stored command, assistant,
run and task_status output fields. Private model metadata and machine paths
were omitted. This is supporting runtime evidence, not test output.