[Juliusolsson05]

Fixes #1281 (the terminal half; the agent half shipped in 2ab60d45). Refs #1283: this closes item 3, "attach counts not tied to a page", for plain terminals. #1311 (6cf6f054) already did it for agent PTYs. Items 1 (orchestration tombstones) and 2 (Codex window lease) remain open there.

Plan: docs/plans/2026-09-26-terminal-attach-survives-respawn.md (first commit).

Problem

A plain terminal's live output depended on a manager flag, terminalAttached. cleanupSessionState deleted it when the shell exited.

• The flag didn't survive a respawn. TerminalLeaf attaches only once per session id. A shell respawned under the same id (recover, from a delivery, a control call or a wake) buffered its bytes but never forwarded them. The xterm stayed on the dead shell while keystrokes reached the new one unseen, until a remount.
• Nothing could release the flag. There was no terminal detach at all, and the IPC handler ignored which renderer page attached, so a reloaded or crashed renderer could never give a reference back.

What merges

• SessionManager: terminalAttachCounts replaces the flag. It is a per-view count, because two lanes can show one shell. Process cleanup leaves it alone, the same rule the agent PTY count follows since Same-id wake of a mounted terminal-view agent leaves the xterm frozen on the dead process #1281's agent half. The new detachTerminal releases one reference.
  • attachTerminal returns null, and takes no reference, only for a non-terminal id.
  • A shell that is down still takes a reference, so its respawn forwards to the mounted leaf.
• IPC: the per-page ownership table that fix(sessions): a raw PTY view keeps receiving output across a same-id wake (#1281) #1311 wrote for agent PTYs becomes one factory, rawViewOwnership, used for both agent PTY and terminal views, with one shared page-document map.
  • Terminal attach and detach carry the page document.
  • A reload or destroy releases the page's terminal references.
  • A dead page's late detach is ignored, and its late attach takes nothing.
• Preload: attachTerminal now sends the document; a new detachTerminal does the same.
• TerminalLeaf detaches in its effect cleanup. An attach that resolves after unmount releases itself.

Tests

• sessionManager.terminalAttachRespawn.test.ts (new, 4 tests, drives the real manager with a fake PTY; the manager is imported once at module scope):
  • a still-attached view keeps forwarding across exit and a same-id respawn, and its detach ends it;
  • a detach while the shell is down is remembered;
  • a view that attached while the shell was down gets the respawn's output;
  • two views in one renderer, one detached, still forward. Two windows cannot both receive one shell's bytes, because the router gives a session one owner.
• ipc/session.test.ts, "terminal attach ownership": reload and destroy release a page's references; a stale page's detach is ignored and its attach takes nothing; a non-terminal attach takes no reference; one detach passes through per attach.
• TerminalLeaf.retention.renderer.test.tsx: unmount releases; a late attach releases itself.
• Red checks: all new tests are red on origin/main.
• Gates: npx tsc -b is clean. Full suite (Node 24): 7247 passed, 1 skipped, 1 failed. The failure is control.system.test.ts, which times out the same way on origin/main.

Reviews: one round (a, b, c) plus a and b's verification, all MERGE-READY; the disposition is in the PR comments.

🤖 Generated with Claude Code

[Juliusolsson05]

Review disposition (round 1 + verification, head cb76010c)

Finding	Verdict	Disposition
a, b, c: the attach-while-shell-down path (a missing session takes a reference) had no test; removing it survived	valid	Test added in cb76010c: attach after exit, then a same-id respawn forwards. It kills the mutant. Verified by a and b
a: the new manager test hit the default 5 s timeout on a cold, contended run	valid	Root cause: the manager graph was imported inside the first test (~4 s). It is now imported once at module scope, and the tests take ms. Verified by a
c: the "second window's lane" test comment overclaims, because the window router gives a session one owner	valid	Comments corrected in the test and in sessionManager.ts: the count is about views within the receiving renderer
a (suspicion): an attach whose IPC reply is lost while the page lives could leave a reference	not reproduced	A page reload or destroy still releases it. No Electron path was found that loses only the reply

Verification: a and b are MERGE-READY at cb76010c; c was MERGE-READY in round 1.