code-review plugin ends its turn while waiting on subagents in one-shot (-p) runs, posting nothing

Where: plugins/code-review, run through claude-code-action v1 with
prompt "/code-review:code-review //pull/ --comment".

What happens: the review fans out to subagents (CLAUDE.md compliance,
bug scan, summary). In some runs the parent finishes with a final message
such as "Agent A reports no CLAUDE.md violations. Waiting on the remaining
three." and the session ends there: no review comment is posted, the
action reports success, and only a downstream assertion notices that
nothing was written. In other runs the parent reaches the posting stage
and tries "gh api --method POST .../pulls//reviews", which is denied
unless the workflow allowlists it, so the findings stay in the log.

Evidence: onurcelep/labgrid-tui runs 35097982872 (findings produced,
write denied) and 35110682311 (ended while waiting on subagents,
num_turns=20). Execution logs are attached as run artifacts on both.
The same plugin version passed on the same repo fifteen minutes earlier
(run 35095623737), so this is run-to-run behaviour, not a version change.

Expected: in a non-interactive run the parent waits for its subagents by
tool result and always posts, or posts an explicit "review incomplete"
comment when it cannot; and the posting path uses a tool the action
already authorises (the sticky comment) instead of a raw gh api call.

Workaround in use: --allowedTools for gh pr view/diff/comment/review and
the read helpers (git diff/log/show, grep, wc), which fixes the write
path but not the early stop.

Start in plugins/code-review and reproduce the behavior through claude-code-action v1 using the referenced runs 35097982872 and 35110682311, reviewing their attached execution logs. Trace the subagent wait and posting paths; done means non-interactive runs wait for all results, post a review or explicit incomplete comment, and use an already-authorized sticky-comment tool.

Written by the indexing model from the issue text.