[Juliusolsson05]

Fixes #1373.

Plan: docs/plans/2026-09-27-undo-close-live-children.md (first commit).

Problem

Undo Close restores a closed pane under a fresh session id. Its LIVE children (orchestration children and linked panes that stayed open) kept orchestrationParentId / orchestrationRootId / linkedParentId pointing at the dead id. Restore remapped pointers only on the rows it restored, and on the undo stack. The renderer's orchestration visibility gate compares ids for equality, so the restored parent could not list, read or prompt its own surviving children. They also rendered as top-level rows. Replace and Reload Agents already remap every session.

Found by review c of #1369, which carries main's tombstones of CLOSED children on Undo Close. This PR is the live half; the two PRs share no code, only the file.

What merges

• Both restore commits (a single pane and a project) map every live session through the restore's old→new ids with the existing remapMetaLineage.
• Why not remapSessionsRelationships: it drops a pointer whose target is not in the record. A child of a different pane that is still closed (still on the undo stack) would then lose its link, and undoing that pane later could not relink it. remapMetaLineage keeps unmapped ids, and returns the same object when nothing changed, so untouched rows keep their identity.

Tests

• undoCloseLiveChildren.renderer.test.tsx:
  • restoring a pane points its live orchestration child and its linked child at the new id;
  • a child of another still-closed pane keeps its pointer;
  • an untouched row keeps its identity;
  • restoring a project points a live child in ANOTHER project at the members' new ids (parent and root separately).
• Red check: both tests fail on main, and each is red when either commit site is removed, or when unmapped pointers are dropped.
• Gates: on d0205863 (main merged), npx tsc -b is clean and the workspace actions suites and lib/undoClose.test.ts pass: 37 files, 260 tests.

Merge order

Merge after #1369. Its Undo Close carry rewrites main's child→parent status hints, which completes this PR's one review residual (a stale hint for up to 250 ms).

Related

• fix(orchestration): closed children follow their parent's replacement #1369 carries main's tombstones of CLOSED children and main's child→parent hints on Undo Close; this PR is the live half.
• bug(sessions): reloading a live child while its parent is closed drops the child's parent pointer for good #1379: the reverse direction (a live child reloaded while its parent is closed) is filed separately.

🤖 Generated with Claude Code

[Juliusolsson05]

Review disposition (head d0205863). a, b and c (pi) are all MERGE-READY in round 1.

Noted, handled elsewhere

• c (minor): main's child→parent status hint still names the dead parent id after an undo restore. A just-prompted child can then read as finished for up to 250 ms, until the status cache expires. Main's hint map is main-owned, so no renderer-side fix exists. fix(orchestration): closed children follow their parent's replacement #1369's carryParent, which Undo Close now calls, rewrites that map, so landing fix(orchestration): closed children follow their parent's replacement #1369 completes it. It is bounded and self-healing. Before this PR the restored parent could not see the child at all.
• c (pre-existing, the reverse direction): reloading or replacing a LIVE child while its parent is closed drops the child's pointer in remapSessionsRelationships, so a later undo has nothing to relink. Filed as bug(sessions): reloading a live child while its parent is closed drops the child's parent pointer for good #1379. This PR's keep-unmapped rule on the undo path is the right rule; the replace path's drop is the mirror image.
• c: children closed before the parent stay invisible to the restored parent until fix(orchestration): closed children follow their parent's replacement #1369 lands (main's tombstones). That is fix(orchestration): closed children follow their parent's replacement #1369's half, by design.

🤖 Generated with Claude Code