fix(post-166): close-tab UX, dead RestartAll route, stale endpoints, partial-success

[aterrylu]

Why

Post-merge visual QA of #166 (Agent + Session unification) surfaced four regressions and one critical silent-failure missed by the unification PR's tests. This PR consolidates the cleanup.

What changed

🔴 X button on only tab silently no-op'd

File: packages/dashboard/src/store.ts

closeTab → closeLeaf → removeLeaf returns null on the only leaf (layout invariant: ≥1 leaf must exist), and closeLeaf silently returned. Click went nowhere. Fix: fall through to removeTab so the leaf becomes empty and the existing "Create or select an agent to start" placeholder renders.

if (leaf.tabs.length <= 1) {
  const newRoot = removeLeaf(layout, leafId);
  if (newRoot) { /* sibling collapses up */ return; }
  // Only leaf in tree — fall through; removeTab below produces an
  // empty-tabs leaf which renders as the placeholder.
}
const updated = removeTab(layout, leafId, tabId);

Plus a defensive guard for stale tabId (cached event handlers, WS races, missed remap entries) that would otherwise reach the same dead-X UX through a different path.

🟡 RestartAll button posted to a route that no longer exists

Files: packages/server/src/routes/agents.ts, SettingsStatusBarItem.tsx

PR #166 renamed /api/sessions → /api/agents but RestartAllButton was still posting to /api/sessions/restart-all. Added POST /api/agents/restart-all (delegates to restartAllAttachments() already exported by runtime.ts) and updated the caller.

🟡 Auth probe pointed at gone endpoint AND misclassified 5xx/404 as authenticated

File: packages/dashboard/src/App.tsx

Mount-time + Retry probes called /api/sessions; the 404 silently landed users on a broken main UI because the ternary res.status === 401 ? "unauthenticated" : "authenticated" mapped 404 → authenticated. Three-state classification: 401 → unauth; 2xx → auth; anything else → error (with logging) so the existing "Cannot connect to server" screen surfaces.

🔴 RestartAll silently dropped per-agent failures

Files: runtime.ts, routes/agents.ts, SettingsStatusBarItem.tsx

restartAllAttachments() returned only the success idMap; respawn exceptions were console.error'd and discarded. The route returned 200 with a partial idMap and the UI showed green "done" — N failed agents would silently not come back. Changed signature to { idMap, failures: Array<{id,name,error}> }; route forwards failures; UI surfaces non-empty failures (e.g. "2 agent(s) failed to restart: foo, bar") via the existing error band. Empty pty.kill() catch now logs the cause.

🟢 Stale endpoint references in comments / docstrings

• mergeOrgWithSessions.ts, Sidebar.tsx (×3 comments), HierarchyPanel.tsx, Sidebar.mergeOrgWithSessions.test.ts — /api/sessions and /api/org → /api/agents and /api/agents/tree.

Test plan

• make check — biome clean, TS clean, 306 server tests pass
• vitest run src/store.test.ts -t "closeTab" — 4/4 pass (regression + stale-tabId + 2 happy paths)
• Live dev verification in browser:
  • 0 console errors after page load (was 1: /api/sessions:0 404)
  • X button on only tab → tab disappears + placeholder shows
  • X button on one of two tabs → tab disappears, sibling stays
  • POST /api/agents/restart-all returns 200 with { idMap, failures: [] }
  • RestartAll UI: click → confirm → restart flow runs end-to-end

Reviewed by

• pr-review-toolkit:code-reviewer — clean (no high-confidence findings)
• pr-review-toolkit:code-simplifier — clean (no simplification warranted)
• pr-review-toolkit:silent-failure-hunter — 5 findings, all addressed in this PR

Diff scope

10 files, +250/-38. No new dependencies.

Out of scope (separate follow-up)

• packages/server/src/channel-server/dist.mjs build artifact still references old endpoints — needs a rebuild of the channel-server bundle (pre-existing on main, not on this branch).
• 2 pre-existing dashboard test failures in switchPane > Case 1 and movePaneToLeaf > dissolves group — unrelated to this PR; fail on main too.

🤖 Generated with Claude Code

[nox-0x]

Approving — all four fixes match the symptoms described, the new tests cover the regressions cleanly, and no critical issues in the current code.

Verified:

• closeTab fall-through (store.ts) — removeLeaf returns null for the only leaf, so the fall-through into removeTab is the correct path to produce an empty leaf that renders the placeholder. The defensive stale-tabId guard prevents the same dead-X UX via WS races. Test coverage is solid (closeTab > REGRESSION + the stale-tabId guard test).
• POST /api/agents/restart-all — route doesn't shadow any earlier matcher (POST /:id only exists for PATCH; POST /:id/{kill,manager,attach} need two segments). Reachable.
• Three-state auth probe (App.tsx) — both mount and Retry now route 5xx/404/403 to error, which surfaces the existing "Cannot connect" screen with the Retry button. Catch handlers also log the cause.
• restartAllAttachments partial-success — failures is plumbed through runtime → route → UI. Empty pty.kill() catch now logs.

Follow-up (not blocking)

When respawnAgent throws inside restartAllAttachments, the agent record is left as status: "running" in the store with no live PTY (zombie). The captured pty.onExit handler returns early because live.get(id)?.pty !== pty — undefined !== oldPty — so it never marks the agent exited. resumeActiveAgents handles the same error by calling markExited(a.id, "crashed") + emitting agent.exited; mirroring that pattern here would keep the data layer consistent with the new UI signal. The PR's UI surfacing is already a clear improvement, so this is a polish follow-up rather than a blocker.

Also pre-existing on main (out of scope here): CreateAgentPanel.handleCreate calls closeTab(leaf.id, "create-agent") passing the pane id where the tab id is expected — currently masked because switchPane Case 4 has already replaced the layout by then, but the new defensive guard will start logging a [layout] closeTab: tabId "create-agent" not found warning if Case 4 ever stops firing for newly-created sessions. Worth fixing alongside the next CreateAgentPanel touch.