This page is a reference mapping, not a completed Claude Code deployment validation. Confirm the exact instruction filename and hook protocol in the installed client before claiming activation.
Map the root AGENTS.md microkernel into the workspace instruction surface,
commonly CLAUDE.md, and keep the host model responsible for the user's task.
Minimal instruction intent:
Route nontrivial work before planning.
Read memory meta/index layers before payloads.
Ask for exact user confirmation before R5 actions.
Use source-grounded external research for current public facts.
Keep CBH behavior correction nonblocking and separate from authorization.
The advisory Bash router requires jq:
bash <HARNESS_ROOT>/bash/harness_intake_router.sh \
--task-text "<user task>" --cwd "<workspace root>"The bundled behavior_correction_hook.py uses an allow + updatedInput
reference protocol. Do not wire it into Claude Code unless the exact installed
version documents compatible updated-input and permission semantics. If that
cannot be verified, use only the instruction/router/context surfaces.
No match, ambiguity, parser failure, or unsupported host protocol must remain a silent no-op. CBH does not emit deny, freeze the session, create R5 permits, or replace Claude Code's native permission boundary.
After client updates, recheck the loaded instruction file, router/validator, memory meta-first behavior, hook payload schema, native permission behavior, and bypass surfaces. Do not report hard enforcement from repository presence or reference tests alone.