feat(agents): reopen a closed agent instead of only deleting it

[snehithareddy28]

What & why

Closes #447. Closing an agent from the strip retains it — store.ts says so outright ("Retained + flagged, NOT deleted"), and Command Center → Archived → show closed agents lists it. But the only control on an archived row was the ✕ that forgets it permanently, so the retained record had no way back onto the floor: the one deliberate way to close an agent was effectively one-way.

"Restore team" does not cover this, and it is easy to assume it does. That button is fed by restorableAgents, which only collects agents whose terminal died unexpectedly (crash / quit recovery). An agent you closed on purpose never enters that list.

Archived rows now have a reopen button. It brings the agent back with its own id, its own cwd and its prior CLI session resumed — which is what makes memory.md, the inbox and the registry entry reattach by id rather than coming back as a stranger. addAgent already un-archives by id ("an id is active xor archived"), so the row leaves the list as the card returns to the floor.

The recipe is shared, not copied. The subtle part of a respawn already existed inside useRestoreTeam, so rather than writing a second copy that could drift, it moves to hooks/respawnAgent.ts and both entry points call it:

• re-enter the isolated worktree when it is still on disk — probed, never assumed, because git worktree add would conflict on the existing path and branch, and a fresh worktree would strand the uncommitted work in the old one;
• fall back to the base repo when that worktree has been pruned, and tell the caller so the dead path is dropped instead of re-probed forever;
• resume the prior session, never re-isolate;
• treat an id whose terminal is already running as already-live rather than a phantom failure.

The module is structural and dependency-injected (no store, no window, no config imports), so the sequence is unit-testable without Electron — the same shape queueDelivery.ts uses for its gate.

Two smaller UX repairs on the same row: a reopen that cannot work now says why, on the row that failed, instead of looking inert; and the ✕ finally carries a title saying it deletes permanently.

Type of change

• Bug fix
• New feature
• Refactor / cleanup
• Docs
• Build / CI

Evidence

Before

After

---

Notes for review:

• The UI change, in words, since the evidence above is the recipe's tests: each archived row keeps its portrait, name and cwd, and gains a small reopen button before the ✕. While a reopen is in flight the button reads reopening… and every row's button is disabled, so a second click cannot spawn a duplicate. A failure renders as one coral line under that row (Couldn't reopen: …). Nothing else in the panel moves; the row becomes a small column so the error has somewhere to go. Strings are new keys in commandCenter and are translated in en, zh-CN and ar.
• useRestoreTeam's behaviour is unchanged — same flags, same per-agent tallies (restored / already live / failed — …), same roster-order insertion, same worktree-gone warning. It is ~60 lines shorter because the per-agent body is now one call.
• Reopening an isolated agent pairs with the teardown fix in fix(worktree): never auto-discard a named agent's unintegrated work on PTY exit #496: a worktree holding unintegrated work is no longer removed when the tab closes, so reopening walks straight back into it. When it was removed (clean and integrated), gitIsRepo returns false and the agent reopens on its base repo with worktree gone — using base repo on its card.
• already-live puts the card back rather than reporting an error: the roster and a live PTY had disagreed, and restoring the card is the honest repair.
• Seven tests in test/respawn-agent.test.cjs, with the two main-process calls faked. They fail on main because the module does not exist there.
• npm run typecheck and the full npm run test:focused suite (841/841) pass locally.

🤖 Generated with Claude Code