[pratikbuilds]

Summary

Closes #1117. A worker parked on an approval looked like it was still working: the thread showed nothing, the pill said "working", and the drawer's Needs you row had no buttons. Approvals now render in the thread where the decision happens, as DESIGN.md describes.

 <Workbench> (apps/web/src/pages/workbench-page.tsx)
   <BenchPill>
-    "<worker> is working …"
+    "<worker> needs you"            while a live ask is pending
   <chat-thread-messages>
     <WorkbenchMessageRow>…
     <PendingOpeningMessage>
+    <ThreadApprovals>               (bench/thread-approvals.tsx)
+      <ApproveBlockView>            existing card: Allow once / Always allow / Ask every time / Deny
     <WorkingLabel>
   <BenchDrawer> → Information → Needs you
-    row + static "Waiting" pill
+    <ApprovalRow>                   the call it gates + Approve / Deny (bench/approval-row.tsx)

The thread only shows asks raised by a live agent address. An ask whose run is gone (e.g. after a hub restart) can't be answered, so it stays in the drawer only, where a 409 now reads "already resolved or its worker stopped" instead of "The server answered 409."

A parked run sends no mail and ticks no inbox event, so nothing refetched the approvals list. Polling now follows the pending turn:

 useBenchWorkerStatus
   approvals query
+    refetchInterval: pending turn ? 5s : off     (shared by every approvals surface)

 answer an approval (thread card or drawer)
+  drop the answered row from the cache
+  markTurnPending(bench)                        the run resumes and may park again
   invalidate approvals

Dropping the row first matters: with the stale row still cached, the resumed turn settled immediately and the next ask in the same reply never surfaced.

Evidence

e2e/browser/approval.test.ts, against scripted inference (aimock) on an isolated database:

thread: worker calls run_shell twice in one turn
  1st ask   → full command shown in the thread, pill says "needs you"
  Allow once → 2nd ask surfaces in the thread without reload
  Allow once → worker's reply lands, no card left

drawer: worker calls run_shell once
  Information → Needs you shows the command → Approve → reply lands
  → "Nothing is waiting on you."

• Before: the thread test fails waiting for the command in the thread. With the cache drop removed, it fails waiting for the 2nd ask.
  After: 2 pass, 0 fail.

Typecheck, lint, fmt and the apps/web unit tests (186) pass. bun run check fails at the test step in five packages this PR doesn't touch (approvals, deferred-tools, url-path, workflows, worker): they have no test files, and bun test exits non-zero when it finds none.

Merge Danger

Door: two-way

UI and e2e only. No schema, API, or data changes.

Blast Radius: workbench-thread

Approvals polling runs every 5s per bench only while a turn is pending, as the timeline poll already does. createChatApprovalActions had no callers before this PR.

---

Notes for review:

• The new test has its own signup → bench helper and waits for an enabled composer before typing. The open PR with the shared startWorkbench helper in e2e/lib/first-run.ts covers the same race; this test can switch to it once that lands.
• e2e/lib/hub.ts boots the browser-suite hub on the raw DATABASE_URL, i.e. the dev database from .env. Running these suites beside bun run dev puts two hubs on one database, and the dev hub picks up the suite's sidecar allocations. That's why the worker sometimes never launched in these suites. Worth its own issue.