What it means
majordomus doctor does not check that the continuity directories exist. It runs the commands a worker would run and reports what they do. A store nothing reads is indistinguishable from no store at all, and files on disk are not evidence that anything uses them.
How it works
mj_doctor_continuity in lib/doctor.sh exercises each path:
- the resolver runs against the handover store and reports its match class and git label, or a clean absence — both healthy; a record skipped as malformed is a failure
- every prompt asset is validated through the same function
prompt renderuses, so a broken asset fails here rather than when someone needs it open-questions.mdandledger.jsonlare parsed with the same validators the gates use, and a line that does not parse is a failure- the context builder is invoked through
mj_cmd_contextand its output measured against the policy's budget - the directories each command writes into are reported
watch runs the equivalent set as drift, and check runs the store validators for the active task. The same code answers the same question in all three, so the three cannot disagree.
How to see it
majordomus doctor
# ok layout .ai/local/state/checkpoints — present
# ok records open-questions.md — every entry parses
# ok records ledger.jsonl — every line carries ts and event
# ok prompt 4 asset(s) — front matter valid, every token known
# ok resolver handovers — same_worktree_same_branch, advanced
# ok context builder — 106 lines, budget 300
printf -- '---\nname: broken\ndescription: d\n---\n{{NOPE}}\n' > .ai/repo/prompts/broken.md
majordomus doctor; echo $? # 10What it does not cover
It proves the commands run and the stores parse. It does not judge whether the records say anything useful — a repository full of empty-but-valid checkpoints passes.
It does not prove a worker ran anything. Reachability is about the tool, not about compliance.
Why it exists
This is the failure the whole tool was built to catch, turned on itself. The evidence this design came from is full of subsystems that existed on disk, were documented as active, and were invoked by nothing — enforcement declared and never wired, stores written and never read. Adding a continuity subsystem without proving it is reachable would have reproduced exactly the thing being fixed, one directory over.