feat: true sub-notes as a parent_id tree in the sidebar (Zed project-panel feel)

[chasehuh]

Closes #80.

Creating a note from inside another note now nests it. Linking to an existing note still does not. That split is the whole feature, and it revises the lock in #77 ("sub-notes are not a nested DB hierarchy") while leaving everything #79 shipped intact.

"link은 그냥 다른 곳으로 가는 거고, 그 안에서 생성하거나 하는 놈들은 sub note."

Action	Result
[[Name → Create "…", or / → New note with a title	True sub-note — parent_id = the note you were writing in. Link still inserted at the caret, no navigation away.
Picking an existing note in [[, or / → Link to note	Peer hyperlink only. No parentage changes on either note.
⌘N / sidebar +	Root note. Unchanged.

A body full of /n/… links implies no hierarchy at all — the tree comes from where a note was created, never from parsing text. That also means the wire format is unchanged: bodies still store only [Title](/n/{id}).

Schema

ALTER TABLE notes
  ADD COLUMN IF NOT EXISTS parent_id TEXT
  REFERENCES notes(id) ON UPDATE CASCADE ON DELETE SET NULL;

CREATE INDEX IF NOT EXISTS notes_user_parent_idx ON notes (user_id, parent_id);

Added through the existing idempotent ensureSchema() path, same style as deleted_at / author_handle. ADD COLUMN on a nullable column with no default is catalog-only in Postgres — no table rewrite on the live notes table.

ON DELETE SET NULL, deliberately not CASCADE. Permanently deleting a parent promotes its children to root rather than silently taking a subtree the user never selected.

Lifecycle semantics

Event	Edge	Sidebar
Archive a parent	Kept, untouched	Children render at root; restoring the parent re-nests them
Permanently delete a parent	SET NULL	Children survive, now at root
Archive a child	Kept	Moves to the Archived section (flat, as today)

Archiving deliberately does not cascade: one × click must not hide an arbitrary amount of work, and restoreNote has no record of which children it would need to restore.

Security

parent_id is a user-supplied foreign key into notes, so it is never trusted as a string. Validation runs through resolveCanonicalNoteId, which enforces tenant scoping, liveness, and alias normalization in one call. Anything that does not resolve to a live note the caller owns is a 400 — fail closed, so a create meant to nest never quietly produces a root note. Covered by tests for the cross-tenant, archived, malformed, and non-string cases.

The client only sends parent_id when the active note is still in the live list, so the realistic 400 is a genuine bug rather than a cross-tab archive race.

lib/note-tree.ts

All the logic, pure and dependency-free. Two invariants it is built around:

• Every live note gets exactly one row. A child whose parent is absent (archived) surfaces at root with its edge intact rather than disappearing — a vanished note reads as data loss, and this repo has history there (Data loss: cross-tab upsert with forceBody overwrites an actively-edited note body #51, Data loss: stale-tab LWW PUT still truncates note bodies after #51/#56 dirty guards (RCA 2026-07-30 uvk-cumd-omo) #57, fix: stop 409-rebased stale buffers from silently clobbering newer note bodies (0804 wipe RCA) #73).
• Cycle-safe by construction. The API cannot create a cycle (parent_id is set once at insert, on a row that cannot yet have children; there is no reparent endpoint), but the renderer walks user-controlled data, so a hand-edited row degrades to flat rows instead of unbounded recursion. Reachability is tracked separately from render state, so a note hidden under a collapsed parent stays hidden rather than being "rescued" to root.

Sidebar (Zed project-panel feel)

Indent + disclosure chevron, →/← on a focused row, ⌘→/⌘← for the whole tree, collapse state persisted to localStorage. Selecting a nested note expands its ancestors (auto_reveal_entries spirit). Collapse state is stored as collapsed ids, not expanded ones, so a freshly created sub-note is visible under its parent without opening anything.

The ⌘←/⌘→ binding is scoped to the sidebar. This handler sits on window, and ⌘←/⌘→ are line-boundary motions in CodeMirror — an unscoped binding would swallow them exactly the way plain ⌘B once did (#77).

A #tag filter renders flat on purpose: nesting matches under parents the filter excluded would draw structure the result set does not have. Called out in a comment so it does not read as an oversight.

CRDT

Untouched. Hierarchy is a notes row field, not document content — no Yjs changes, no new update types, no change to the crdt_managed_body 409 gate.

Verification

• pnpm test — 345 passing, up from 312 (+33)
• tsc --noEmit clean; pnpm build compiles
• pnpm lint: 9 errors / 6 warnings, byte-identical to origin/main (verified by linting a clean worktree at origin/main with the same node_modules). All are pre-existing react-hooks/refs and react-hooks/set-state-in-effect findings, three of them in files this PR never touches. This branch adds zero lint problems, but pnpm lint is not green on main today — see Residuals.
• Migration and query paths verified against a real Postgres 16 (throwaway container), not just mocks:
  • column added to a pre-migration table with existing rows untouched
  • FK confirmed as ON DELETE SET NULL / ON UPDATE CASCADE via pg_constraint; index confirmed on (user_id, parent_id)
  • ensureSchema() idempotent across repeated runs
  • hard-deleting a parent leaves the child alive with parent_id = NULL
  • createNote → listNotes → flattenNoteTree nests correctly through real SQL
  • archived parent → child at root with its edge intact → restore re-nests
  • hyphenless parent id resolves to canonical; another user's note and an archived note both refused

The verification harness was temporary and is not part of this diff.

Not verified locally

No browser smoke test: the only .env.local on this machine predates Clerk, so the app cannot boot locally. Everything above was verified through tests and a real database instead. The sidebar interaction changes (chevron, arrow keys, auto-reveal) are the part that warrants a look on the deploy preview.

Residuals

• pnpm lint fails on main today (9 pre-existing errors from the react-hooks rules). Out of scope here — fixing it means touching codemirror-editor.tsx, publish-panel.tsx, and the ref-assignment block in agentnote-app.tsx, none of which this PR is about. Worth its own issue.
• Zed parity left on the table, all non-goals in True sub-notes: parent_id tree in sidebar (Zed project-panel feel) #80: drag-and-drop reparenting, multi-select, sticky ancestor rows (sticky_scroll), single-child chain folding (auto_fold_dirs), indent guides.
• Long titles at depth 5+ clip in a narrow sidebar (text-overflow: ellipsis already applies). No depth cap.
• Carried over from Wiki-style note links + tags (Obsidian/Notion), muted Offline chrome, and ⌘B bold #77/feat: wiki-style sub-note links and a #tag system #79 and untouched here: renaming a note does not rewrite inbound links; sub-note links in a published note are inert for anonymous readers.

Operator: hard-refresh open agentnote.dev tabs after the deploy lands.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
memo	Ready	Preview	Aug 5, 2026 7:03am