[Bug]: Grok Composer hangs on Stop, resume replay floods thread, and long session/load stalls

[mwolson]

Before submitting

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

Area

apps/server

Steps to reproduce

Reproduced on T3 Code desktop with Grok Build / Grok Composer 2.5 Fast on a multi-turn thread (main @ ffae5410e).

A. Stop / working state stuck after Grok finishes

1. Open a Grok-backed thread with several prior turns.
2. Send a tool-heavy prompt.
3. When Grok has effectively finished (tools done, prose streaming) but T3 Code still shows working/Stop, press Stop.
4. session/prompt can remain stranded for many minutes while the composer stays active; follow-up turns may not dispatch.

B. Resume replay floods the live turn

1. Use a long-lived Grok thread, restart T3 Code, and reopen the thread.
2. Grok session/load replays historical session/update chunks marked _meta.isReplay: true.
3. T3 Code projects those replay chunks into the new turn and the composer shows old tool calls / prose again.

C. Long session/load after restart

1. Resume a long Grok native session after restart.
2. Replay finishes but the session/load RPC may never return (observed ~90s+).
3. Resume can stall or fall back to session/new, orphaning pending messages.

D. Timeline / missing prose on rebound segments

1. On threads where Grok reuses assistant segment ids across turns, responses can sort above the user message that started the turn, making tools/prose look missing below the prompt.

Expected behavior

• Turns settle when Grok returns to idle, even if the standard ACP session/prompt RPC stays stranded.
• Stop cancels the in-flight Grok prompt and clears composer state without requiring an app restart.
• Resume ignores replay-marked session/update traffic and completes load promptly (normal RPC or replay-idle synthesis).
• Assistant timeline keeps stable chronological order across rebound/replay segments.

Actual behavior

• Grok can emit _x.ai/session/prompt_complete while session/prompt never settles, leaving the composer stuck.
• Stop may return success without unblocking a wedged session/prompt RPC.
• Replay chunks after restart attach to the live turn and flood the composer.
• session/load can hang after replay on long sessions.
• Reused segment ids can leave prose/tools visually above the prompting user message; streaming markdown can jump near completion.

Impact

Major degradation or frequent failure

Version or commit

main @ ffae5410e (current upstream; reproduced on 0.0.27 desktop)

Environment

• Linux x86_64 desktop AppImage
• Grok Build CLI via ACP (grok-composer-2.5-fast)
• Long-lived native Grok session across multiple turns and restarts

Logs or stack traces

# Stalled prompt (~27 min); Stop did not unblock the RPC before app restart
RpcClient.session/prompt duration ~1,630,000ms, exit Interrupted (process restart)
interruptTurn Success, but session/prompt kept running until restart

# Same thread after fix in #3156: prompt_complete race settles the turn
RpcClient.session/prompt duration ~39s, exit Interrupted
sendTurn Success

Workaround

Restart T3 Code and send a follow-up message. Avoid restarting long Grok threads when possible.