[Juliusolsson05]

Problem

One repository ended up as three agent-code project tabs (created Sept 3, 7 and 10), each with its own generation of agents and worktree agents spread across them. Dispatch, its row filters and the bulk modals then show three projects for one repo, and every "everyone in this project" action has to be repeated per tab.

Two causes. The New Tab path picker never dedupes: ⌘T creates a fresh tab for a folder that is already open, while the operator's projects.open has reused an existing tab for the exact directory since it shipped. And once duplicates exist, nothing can fold them back together without closing agents.

Implemented behavior

Reuse on open

• When the typed folder is already held by an open tab, the picker shows "Already open as E · agent-code" and Enter / the primary "go to tab" activate that tab without spawning. When the holder is the tab the user is standing in (the folder ⌘T pre-fills), the hint reads "Already open in this tab (G · agent-code)", the primary reads "stay here" and Enter only closes the picker; it never jumps to an earlier tab that also holds the folder.
• "new tab anyway" and Shift+Enter keep the deliberate second-tab case. Every footer button forces the action its label names, so a click that lands before the debounced hint has caught up cannot switch tabs under a "new session" label; Enter alone decides at submit time.
• A folder nobody holds behaves exactly as before. Resume rows are unchanged.
• The picker and projects.open now share one rule, findTabsHoldingDirectory: a tab holds a directory when one of its sessions runs there. Exact match; a worktree is a different directory on purpose. With several holders the picker prefers the current tab, then the first in tab order; projects.open keeps refusing with ambiguous_owner because an operator has no current tab to prefer.

Merge Project Tabs (new command, advanced use of the tab bar)

• Pick the tab to keep and the tabs to fold into it. Tabs that share a working directory with the target are pre-ticked; worktree tabs are a deliberate tick. The dialog's draft mounts with the dialog content, so it always starts from the live active tab (the surface registry mounts every modal before the workspace loads).
• Source grid panes and detached agents become Dispatch agents of the target; buried records and Dispatch row project filters are re-pointed; the tab-index snapshots of the surviving tabs' records are refreshed (removing tabs shifts every later letter); a tiled source hands its slot to the target when the target was not tiled, so the kept tab stays on screen; Spotlight and Reader references to removed tabs are cleared; the emptied tabs are removed; the active tab follows.
• No session is killed, reloaded or re-spawned. The dialog states what moves ("2 tabs, 9 agents move to E · agent-code") before anything happens, and the toast repeats it. Moved grid panes become Dispatch agents, so after the next launch they wake on first use instead of being live from the start; the command description says so.

Design decisions

• A pure planner, not a sequence of existing actions. closeTab kills what it owns and detachSession refuses a tab's last grid pane, so neither could express a merge. mergeProjectTabs(state, input) does the whole re-pointing in one function; the hook plans inside the store updater against the live state and gates every side effect (tiled tabs, Spotlight, Reader, toast) on the applied result, so a refused merge changes nothing and never reports success.
• Grid panes go to Dispatch, not into the target's grid. Tab.root is a tile tree the user built; attaching every source pane would make one tab a wall of splits, and buildDispatchGroups already lists detached sessions under their tab.
• No undo. Undo-close exists to re-spawn sessions that were killed from a captured SessionMeta. A merge kills nothing, so there is nothing to re-spawn, and reversing it would mean rebuilding the removed tabs' tile trees from a snapshot while the sessions inside have kept running and may since have been attached, buried or closed. Every moved agent stays reachable in the target's Dispatch list, so the manual reverse (new tab, attach) is always available and never lossy.
• Owned-session invariant. collectOwnedSessionIds must be identical before and after a merge; a session that lost its owner would be deleted by the next autosave. This is the load-bearing test. The planner uses the same own-property hasSessionMeta test as ownership, and a pane that already had a detached record (a pre-existing violation) is counted once.

Review

Orchestrated Claude review, two rounds. Round one: REQUEST CHANGES, one blocking finding (the dialog's seeding effect raced its prune on the first open per app run and kept the wrong tab) plus ten non-blocking findings; all addressed in e5f85358 except the narrow close-tab race, which is noted only. Round two: APPROVE WITH COMMENTS; the remaining comments (PR body, "stay here" clearing Spotlight, a buried-record test gap) are addressed in the final commit.

Linked issues

Fixes #913

Tests and verification

• mergeProjectTabs.test.ts (new): the three-tab fixture with detached, buried, tiled Dispatch rows bound to sources (including the legacy single binding), lanes, and pins; asserts every re-pointing, the refreshed indexes of the surviving target's own records, the ownership invariant, the untouched target tree, the summary, the refusals (target among sources, unknown tab, empty selection), a pane with no metadata skipped and a pane with a pre-existing detached record counted once. Tiled-tab retargeting incl. slot substitution for an untiled target and exit below two tabs.
• MergeProjectTabsModal.renderer.test.tsx (new): mounts before the workspace loads and then opens (the review's reproduction), keep-change re-seeds the ticks, a closed tab leaves the draft and a fallback target is never left ticked.
• queries.directory.test.ts (new): exact match through grid and detached sessions in tab order; worktrees excluded.
• PathPickerModal.renderer.test.tsx: Enter on the input goes to the holder; the current tab is preferred over an earlier holder; Shift+Enter opens a new tab; a button does what its label says before the hint has caught up; no holder creates as before.
• Command catalog baseline: 122 commands, 25 approved additions (after merging main, which added Use Global MCP Settings).
• tsc -b clean on both projects under Node 24; npm run test:contract satisfied; touched unit and renderer suites green.

Known limitations and follow-ups

• Resume flows (picker rows and the Conversations picker) still open a new tab; folding a resumed conversation into an existing tab needs the detached-agent resume path.
• Window adoption still appends a closed window's tabs without root dedupe; the merge command is the cure there.
• A merge into a tab whose close is mid-flight (between its kill list and its removal) files the moved sessions under a tab that is about to vanish; the window is a few milliseconds and detach/create into a closing tab already have the same shape.
• Path matching compares the stored cwd string; symlinked spellings of one folder are different directories here, as they always were for projects.open.
• Not exercised in the running app; verified from source and tests.

🤖 Generated with Claude Code

https://claude.ai/code/session_016vtjk6o4nukQp7Vfbt5TDc

[Juliusolsson05]

Orchestrated Claude review, two rounds.

Round one: REQUEST CHANGES. Blocking: the Merge Project Tabs dialog's seeding effect raced its "tab closed while open" prune on the first open per app run (surfaces mount before the workspace loads), so it kept the first tab instead of the active one and could leave that tab ticked as a source. Ten non-blocking findings: Enter in the picker could jump to an earlier holder or do nothing, a footer button could act against its label before the debounce settled, merge side effects were not gated on the applied plan, stale tab-index snapshots and tiled focus after a merge, ownership edge cases, missing WHY comments, an untested Enter path, symlink matching, and a narrow close-tab race.

Addressed in e5f85358 (all but the close race, noted in the PR body). Round two: APPROVE WITH COMMENTS. Remaining comments (PR body drift, "stay here" clearing Spotlight, a buried-record test gap) are addressed in the final commit; Shift+Enter as "new tab anyway" is noted as a product choice the hint text advertises.

Verified by the reviewer at HEAD on Node 24: tsc on both projects, the touched unit and renderer suites, and the test contract. Not exercised in the running app.