[scode]

scode-shepherd's runner has a strong reviewer model review every PR a worker model reports done. When the run comes
from a scode-build-goal goal file, the goal the worker reads includes the user's choice of reviewer ("one
fresh-context gpt-6.1-sol high reviewer per PR"). On a real run two worker sessions took that as their own job: they
had a sol reviewer look at their diff before reporting done, and wrote "independent gpt-6.1-sol high review reported
no findings" into their reports. The runner then ran the same review again. That paid for a strong-model review twice,
and the worker's report claimed a review that the worker had set up and summarized itself, which reads like the
user's review gate passing when it had not run yet.

The worker is now told that the runner's review is the one the goal asks for, and not to have another model or agent
review the PR itself. Review steps the repository asks of every author, such as having a fresh reader check commit
messages, still apply. The plan the supervisor writes for the run keeps the goal text verbatim, so the fix is in the
runner's own instructions to the worker rather than in an edited copy of the goal.