[chkn]

Requested by Alex · project thread

Before: the dataset grid showed every field in schema order, plus the keys one level inside a field. A row whose identity sat behind a resource reference, like asgard's planTask task title, looked the same as every other row. Column order and visibility couldn't be changed, and rows were only ever named by their number.

After: a Columns button in the dataset header shows, hides, reorders and removes columns, and can add a column at any path in the rows. A path can step through instance references, so toolsContext › create_task › rootTaskId › title works. Headers can be dragged, and their menu gains "Hide column" and "Use as row label". The details pane reads "Row 3 · Web site for Evalution". The layout and row label are dataset metadata, shared and synced, while widths and expansion stay per viewer. An added column that ends at an instance's argument can be typed into in place.

This implements specs/datasets.md §Q. It's based on main, since #8 (named resource instances) is already merged there.

How:

• Q.1 Paths. src/shared/column-paths.ts adds ColumnPath, readColumnPath and resolveColumnPath, replacing the grid's CellPath/readPath. Paths stop after 16 reference hops. They serialize to ids that keep today's width keys (fieldId, fieldId/key), with an instance root written as @name.
• Q.2 Finding paths. describeRows adds instances (each instance's resource, row count and merged argument names). The new DatasetProvider.describePath lists the keys one level in, each with a sample. A JSONB query picks the rows with something at the root, and each row is then walked with the shared resolveColumnPath, so the server and grid always agree on references. The route is POST …/paths.
• Q.3 Layout (server). Dataset.columns and rowLabel are stored in a new nullable layout column on datasets. One migration covers both providers, since the local one wraps Turso. setColumns is served by PATCH …/columns, which checks that every root exists in the schema or some row. deleteField prunes the field's columns and the row label. A column of an unknown kind is stored as written.
• Q.3 Layout (client). reconcileColumns (dataset-columns.ts) fits the stored list to the dataset:
  • It drops gone fields and keys, and appends new ones.
  • It dedupes paths and keeps unknown kinds in place as dimmed "Unknown column" placeholders.
  • A dataset with no columns looks exactly as it does today.
• Q.3 Grid and picker. canMoveColumn backs onColumnProposeMove: a key column stays in its own group, and nothing lands inside an expanded group. movedLayout writes one setColumns on drop. The picker reuses the trace list's popover, with a live-swap drag that is also written once on drop.
• Q.4 Editing. Through an added column, an argument holding a plain string, number or boolean is editable. The edit replaces that instance in the update's resources. A field's key columns stay read-only, as P.2 has them.
• Tests.
  • Unit tests cover each step rule, two references, the hop limit, "n/a", reconciliation (dropped, appended, duplicate, no columns), moves and argument edits.
  • Provider contract tests cover instances, describePath, setColumns and pruning, with handler tests for validation.
  • Playwright CT covers an accepted header drag, a vetoed one, editing a referenced argument inline with the row label, the picker, and an unknown column round-tripping.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QMiMTviyLp3ciCvQf3mzkf

---

Generated by Claude Code

[chkn]

Overall this is well structured. Sharing resolveColumnPath between the server and the grid, reconcileColumns, and how canMoveColumn/movedLayout handle hidden columns all read cleanly. On the head commit, npm run typecheck passes, as do the unit and provider tests that cover this change (7 files, 210 tests). The only check CI runs on this PR is reuse, so those tests aren't run there.

Two behaviour bugs, both reproduced with a scratch test:

1. 🔴 Layout deadlock. Once no row declares an instance that a column or the row label is rooted at, every PATCH …/columns returns 400. That includes the two edits that would fix it (handleSetColumns).
2. 🟡 Clearing an argument in place. The argument is deleted, and the cell then goes read-only (groupEdits).

There's also one smaller UX gap: removing the row label's column leaves no way to clear the label. The rest are elegance suggestions: a narrowing guard that removes two casts, moving instance-name validation into parseColumnPath alongside a pathInstance helper, and restructuring a branch in reconcileColumns.

---

Generated by Claude Code