[Bug]: Remote stop button gives no feedback and leaves thread stuck in Thinking

[ImBIOS]

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate.

Area

apps/web • apps/mobile • packages/client-runtime • apps/server

Description

Using a remote T3 Code server, clicking the Stop button gives no UI feedback and the thread stays stuck in Thinking / Working for Xm Ys indefinitely.

Screenshot attached: prompt list me all ~/dev/projects/* → Working for 6m 26s + Thinking, bottom composer shows red stop ■ still active, environment myrehat-dev.asia-southeast1-a.c.myrehat-... · Local checkout · main, model Muse Spark 1.2 Contrib..., Xhigh · Build, Full access.

Expected: Stop shows Stopping... immediately (disabled) and within seconds drops Thinking/Working, even on relay/tunnel. Actual: no visual change, subsequent Stop clicks also silent, thread never recovers until server restart or manual projection_thread_sessions stopped + projection_turns interrupted repair.

Steps to reproduce (remote)

1. Start remote server (npx t3 --share / --tunnel / npx t3 behind Tailscale on myrehat-dev...).
2. Connect from app.t3.codes or mobile via pairingUrl: token (remote mode, not local).
3. Open thread with Local checkout, send a prompt that streams (list me all ~/dev/projects/* and it…, or any long agent run).
4. While Thinking/session.status === "running", click the red Stop button in the composer.

Actual

• Button does not morph to Stopping..., not disabled, no toast, no spinner change (100–400ms relay RTT makes it obvious; local 10–20ms masks it).
• Timeline keeps Working for ... ticking forever.
• DB: projection_thread_sessions.status = 'running' with active_turn_id pinned to a turn already interrupted/completed (completed_at set), or NULL with pending turn forever (#4713 detection query hits):

  SELECT s.thread_id, tu.state, tu.completed_at
    FROM projection_thread_sessions s
    JOIN projection_turns tu ON tu.turn_id = s.active_turn_id
   WHERE s.status = 'running'
     AND tu.state IN ('interrupted','completed','error')
     AND tu.completed_at IS NOT NULL;

  Widened query also catches active_turn_id IS NULL + pending turn (report at 2026-08-24T23:30:31Z in Thread session stuck in running after turn interrupt — stop button becomes a no-op #4713).
• Every further Stop appends thread.turn-interrupt-requested in orchestration_events (accepted in orchestration_command_receipts) but produces no follow-up thread.session-set; the stop handler has nothing to act on.

Expected

• Optimistic Stopping... within one frame on every surface (web composer, mobile ThreadRouteScreen), disabled until terminal session-set stopped/interrupted or error.
• Even when buildThreadTurnInterruptInput omits turnId (stale activeTurnId), the client and projection still flip latestTurn.state → interrupted optimistically.
• If provider interrupt succeeds yet CLI stays alive, server escalates after ~5s to stopSession → session-set stopped, so derivePhase → disconnected/ready and MessagesTimeline clears Working.
• Relay blips surface as toast instead of silent no-op (reportFailure:false currently swallows EnvironmentRpcUnavailableError while phase === "available"|"offline").

Evidence & analysis

Investigated in worktree t3code-a1ef8329 (6a9d9f9):

• UI: apps/web/src/components/ChatView.tsx:1246 interruptThreadTurn = useAtomCommand(threadEnvironment.interruptTurn, {reportFailure:false}), 5487-5510 onInterrupt has no isStoppingTurn state. Contrast handleStopBackgroundWork 4677-4740 which correctly tracks isStoppingBackgroundWork, disables Button 4543 and shows "Stopping..." 4546.
• Input guard: ChatView.logic.ts:141-150 buildThreadTurnInterruptInput includes turnId only if session.status==="running" && activeTurnId!==null. Remote snapshots lag (THREAD_RESUME_MAX_GAP 1000, SHELL_COALESCE_WINDOW 50ms ws.ts:841), so Stop often persists thread.turn-interrupt-requested without turnId. Both threadReducer.ts:271-292 and ProjectionPipeline.ts:1441-1476 return unchanged/no-op for turnId===undefined.
• Projection: session-logic.ts:1894-1905 derivePhase = running iff session.status==="running"|"starting"; MessagesTimeline.tsx:210,1313 WorkingTimelineRow only clears when isWorking===false. threadReducer latestTurn.state==="interrupted" alone does not clear it.
• Server reactor: ProviderCommandReactor.ts:1228-1321 only tears down on failure (recoverInterruptFailure → stopSession). On success it waits for provider session-set. When ClaudeAdapter.ts:query.interrupt() never settles / OpenCodeAdapter.ts:2932 10s abort times out / CodexAdapter ignores, session never leaves running. Serial interruptTurn (threadCommands.ts:177-182 concurrency:{mode:"serial"} per (env,thread)) then queues further clicks.
• Family: pingdotgg/t3code#4713 (40+ accepted interrupts no effect, SIGTERM workaround), #4589 (threads stop updating), #7820, #2644, #7349 (bounded Claude query.interrupt 10s + getContextUsage 1s — but only for failure path), #7589, #7987.

Impact

Major: remote is primary use (Tailscale/tunnel/mobile). User cannot cancel, cannot archive (useThreadActions.ts:214 refuses archive while running), mobile outbox gates on running (use-thread-outbox-drain.ts:317) so queued messages never deliver. Only way out is SIGTERM --resume <cursor> or manual DB session-set stopped + restart.

Environment

• T3 Code v0.0.37-nightly.20260829.1218 6a9d9f9 (local) / remote myrehat-dev Linux aarch64, provider opencode Muse Spark 1.2 Contrib...
• Client app.t3.codes / desktop wrap, remote relay.
• Also reproduced for claudeAgent, Codex, Cursor per Thread session stuck in running after turn interrupt — stop button becomes a no-op #4713.

Suggested fix (non-negotiables)

1. Web + mobile optimistic isStoppingTurn (mirror isStoppingBackgroundWork) → Stopping... disabled.
2. Relax turnId===undefined guard to fallback to activeTurnId ?? latest unsettled turn.
3. Server 5s fallback stopSession when interrupt succeeds but session still running (success-still-running), with bounded getContextUsage.
4. interruptTurn timeout / singleFlight so promise doesn't hang; surface transport errors for Stop.
5. Tests + projector coverage.

Logs

• state.sqlite orchestration_events thread.turn-interrupt-requested accepted without follow-up thread.session-set.
• projection_thread_sessions running + projection_turns interrupted/pending as above.

Managed via forkhub as fix-remote-stop-no-feedback-stuck-thinking-7h3k9p2m. See ImBIOS/.forkhub INTENT.md for full non-negotiables and implementation notes.