v1.23.1

1.23.1 (2026-07-27)

A harness-hygiene patch. Claude-runner scenario and judge invocations in the eval sweep now
pass --setting-sources project, so the invoking user's global agent context — user-level
memory, hooks, and settings — never loads into a run. A sweep's "bare" (no-skill) responses could
previously absorb the operator's environment, which is the root cause of the environment
identifier that leaked into a committed baseline (scrubbed after the fact in #79); this closes
the hole at the harness level instead of relying on post-hoc scrubbing. The temp cwd has no
project settings, so the flag yields a true clean room. Generic runners are named as
out-of-scope: the --runner-cmd template author owns the foreign CLI's equivalent isolation.

Bug Fixes

• evals: isolate scenario and judge runs from the operator's global agent context (#114) (09a616f) — scripts/run-evals.py (both the scenario and judge paths) + the isolation note in evals/README.md