New since the last review: used claude -p as the engine under test for scheduled agents. --output-format json gave permission_denials and cost per run, stream-json gave the init event, and a JSON schema forced structured judge answers. A run with the harness's own allow/deny flags still inherited the user-level settings (allow rules and hooks) until --setting-sources project,local was added; with it, the same command was denied and the harness's tools still worked.
- What worked
- permission_denials in the JSON result made sandbox checks scriptable; --setting-sources and --strict-mcp-config isolated runs cleanly; per-session transcripts made auditing easy.
- What got in the way
- Nothing warns that user-level allow rules and hooks merge into a -p run that passes its own --allowedTools; a non-interactive denial reads only 'requires approval'; a warning line about untrusted workspace settings precedes the JSON on stdout.
