🤖 fix: an open edit loses its typed changes on a workspace switch (#5801), plus two older edit races

[ThomasK33]

Follow-ups from the final readiness check of PR #5801 (edit text lives only in the composer's memory, #5672 / #5571).

1. A workspace switch drops an open edit's typed changes (introduced by #5801, blocks it)

What happens: you open an edit, type changes (text, attachments), switch to another workspace, and come back. The edit reopens with the original message text, and the typed changes are gone.

Cause: ChatInput is keyed by workspace (ChatPane.tsx ~2011), so a switch remounts it. ChatPane keeps editingState per workspace (~381-386), so the edit itself survives the switch. The memory-only edit buffer lives in the remounted ChatInput (useComposerDraft), so the buffer does not survive. When you come back, the composer refills the edit from the original message (ChatInput/index.tsx ~1448).

On main the edit text survived a switch but replaced the unsent draft. #5801 keeps the unsent draft but loses the edit changes. The accepted tradeoff of #5801 covers reload only.

Fix idea: give the edit buffer a stable per-workspace, memory-only owner that survives the ChatInput remount (next to ChatPane's editingState, or an equivalent). It must survive workspace switches, stay out of other windows, and still disappear on reload. Failing-first test: switch away and back with edit text and an attachment, and check that the edit keeps both and the unsent draft stays intact.

2. Starting a second edit while one is open leaks the first edit's notes (pre-existing on main, not part of the #5801 repair)

beginEditingMessage does not check for an open edit (ChatPane.tsx ~416-427). The new session snapshots preEditReviews: draftReviews (index.tsx ~1444), and those are the first edit's notes. Cancelling the second edit then puts the first edit's notes on the unsent draft. Main had the same snapshot.

3. A Stop restore can race with opening an edit (pre-existing on main, narrow, not part of the #5801 repair)

A restore event that arrives before the edit opens, with its async acceptedSendIds lookup resolving after (index.tsx ~1632), skips the edit check (~1656). setInput(mergedText) (~1665) then replaces the edit buffer, and a later cancel loses the restored text. Main lost it in the same race.

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high • Cost: $35.15

[ThomasK33]

Found by the normal review of #5801's repair head 72a15a919c (the fix for item 1 is on that branch). #5801 is parked. These must be fixed before it can merge:

1. Transcript-only switch during an edit loses the edit's typed changes (introduced by 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801). ChatPane swaps ChatInput for the transcript-only notice and clears the edit target, so the memory-only buffer is orphaned. Fix idea: release the buffer into the normal draft (after the unsent draft, as releaseEndedEdit does) before the composer is replaced. Thread: 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801 (comment)
2. A second Edit during an open edit overwrites the first edit's buffer (introduced by 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801). beginEditingMessage checks only a pending edit send, and beginEditDraft with a new id replaces the buffer. Fix idea: release the open edit first, or keep Edit disabled while an edit is open. Note: this overlaps item 2 above (the note leak on a second edit). A fix here can cover both. Thread: 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801 (comment)
3. hasOlderHistory guard (tradeoff). An edit whose row another renderer really deleted stays open until its send is refused (history-changed). The refresh then releases the text, so nothing is lost. Fix idea: tell an unloaded row from a deleted one, using the target's history sequence against the loaded window or a backend check. Thread: 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801 (comment)

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high • Cost: $49.02

[ThomasK33]

#5801 is parked again at e8acc69cd8, after its last allowed repair push.

That push added one rule: an edit that loses its target without a settle keeps its text and files in the draft, after the unsent draft. It covers transcript-only, a second Edit, a deleted row and target-not-found, and it addresses item 1 and the overlap with item 2. #5801 is not merged, so nothing here is resolved on main.

The normal review then found three new defects in the same design. The edit outlives a composer remount (a workspace switch), but some composer-local state does not survive with it:

1. Duplicate edit send: a remounted composer can send the same edit again while the first send is in flight. sendingCount is local state, and canSend ignores the restored session's sendInFlight. 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801 (comment)
2. Review notes: a send accepted after a switch away and back restores the notes only in the unmounted composer, so the edit's notes stay as the override. 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801 (comment)
3. Refresh retry: a failed history refresh loses its retry state on a switch away and back, so Send stays disabled with no retry button. 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801 (comment)

Re-plan before more code: keep all of the edit's per-edit state with one owner that survives the remount (the send in flight, the review restoration, the refresh retry), or end the edit on a switch and keep its contents in the draft with the same preservation rule. The second option is simpler: it gives up "the edit resumes after a switch" but keeps every typed change.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high • Cost: $54.92