Found reviewing #365 against #366.
#365 recommends #356 as the next item and describes its first step as pending work:
- Detect. Add an
INCONSISTENT_NODE_TYPE check keyed on node_id across records … Baseline it — 63 keys is a backlog, not a PR.
#366 ships exactly that — the check, the node_id keying, and the 294-row baseline.
Merge both and the reconcile is false on arrival: it recommends as "next" a step already taken, and section 11 describes a detector that exists. That is precisely the drift the next-tasks skill exists to prevent ("never read NEXT_TASKS.md and relay it verbatim"), reintroduced by the reconcile itself.
The recommendation should move to step 2 — decide, then normalise — which is the part that is actually still open, and the harder part: CHEMICAL is clearly wrong for a gradient, STATE vs BIOLOGICAL_PROCESS is a real question, and where a family genuinely means two things the fix is two node_ids rather than one type.
Also worth recording in section 7 while the file is open: automatic claude-review was turned off (#371) while its token has no quota.
Found reviewing #365 against #366.
#365 recommends #356 as the next item and describes its first step as pending work:
#366 ships exactly that — the check, the
node_idkeying, and the 294-row baseline.Merge both and the reconcile is false on arrival: it recommends as "next" a step already taken, and section 11 describes a detector that exists. That is precisely the drift the
next-tasksskill exists to prevent ("never read NEXT_TASKS.md and relay it verbatim"), reintroduced by the reconcile itself.The recommendation should move to step 2 — decide, then normalise — which is the part that is actually still open, and the harder part:
CHEMICALis clearly wrong for a gradient,STATEvsBIOLOGICAL_PROCESSis a real question, and where a family genuinely means two things the fix is twonode_ids rather than one type.Also worth recording in section 7 while the file is open: automatic
claude-reviewwas turned off (#371) while its token has no quota.