[BUG] 2.1.261 macOS arm64: session hangs, main thread 100% CPU in JIT loop while waiting on a background subagent (sample attached)

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code

### What's Wrong?

Mid-session (about 32 minutes in, right after dispatching a background subagent), the TUI stops repainting and accepts no input. The process stays alive but is wedged:

- Main thread `R+` at 98–100% CPU continuously; the other 10 threads idle (`ps -M`).
- RSS flat at ~750 MB (physical footprint 401 MB, peak 423 MB) — not a memory leak.
- Zero child processes. No hook, MCP server or statusline process running.
- All three HTTPS sockets to the API in `CLOSED` state; a healthy sibling session on the same machine shows 5 `ESTABLISHED`.
- System memory fine (81% free, swap 1.1 GB). Two other Claude Code sessions on the same Mac kept working normally.
- On an earlier occurrence (2026-08-27, second Mac) `SIGTERM` was ignored and `SIGKILL` was required. `claude --resume` recovers the transcript.

Same signature as #89062 (regression since ~2026-08-21, subagent-heavy sessions) and the older #51560. Per #89062, `BUN_JSC_useDFGJIT=0` still hangs and `BUN_JSC_useJIT=0` crashes at startup, so JIT tuning is not a workaround.

### `sample 5` — main thread, 3637/3637 samples in one identical stack
Load address 0x1041f0000 (binary `2.1.261`). Offsets from load address, outermost → innermost:

```
+0x112efd4 +0x104eb84 +0x1053528 +0xdfa30c +0xfe7b40 +0x79bf4c +0x79b9d8
+0x15ab938 +0x15ab4d4 +0x15abfa0 +0x11286f4 +0x13a0d38 +0x6ed9d4 +0x1a73c8
+0x2bde880 +0x28affe4 +0x35418b4
[JIT 0x116316224] [JIT 0x1163161f4]
+0x588f44 +0x31d337c +0x3074cbc +0x2dd5068 +0x2dd9b58 +0x35418b4
[JIT 0x1173ead10]
+0x3564f90 +0x3564f90 +0x3564f90 <- same frame three times
[JIT 0x115905978] [JIT 0x1169eace8] [JIT 0x115cb3fbc] [JIT 0x1163a7060]
+0x219819c +0x2198318 +0x30d4224 +0x30d4224 <- same frame twice
leaf (JIT, 100% of samples within 0x115904880–0x115904944):
2247 0x1159048dc | 1162 0x1159048b4 | 66 0x115904930 | 57 0x11590493c | 51 0x1159048a4 | 43 0x115904880
```

The leaf never leaves a ~200-byte JIT region, i.e. a tight loop in JavaScript (regex backtracking / O(n²) string or array processing), not a lock or a syscall. Full `sample` output attached under Additional Information; the process was still spinning after the 5 s window.

### Ruled out on this machine
- Fullscreen renderer: reproduced with `"tui": "default"` (and earlier with `fullscreen`).
- Hooks / MCP: no child processes at freeze time; project hooks measured 45–100 ms each.
- Memory pressure: 81% free, RSS flat.
- Terminal emulator: seen in Zed, VS Code integrated terminal and Warp.

### What Should Happen?

The parent session should keep rendering and accept input while a backgrounded subagent runs, and surface the subagent's completion. If the main thread is starved for more than a few seconds, a watchdog should log the JS stack (and SIGTERM/SIGINT should still work) instead of requiring `kill -9`.

Ask: symbolicate the offsets above against the 2.1.261 arm64 build.

### Error Messages/Logs

```shell
No error message: the TUI simply stops updating. Process state at capture time (~32 min after launch, ~25 min after the freeze):

PID PPID %CPU RSS STAT ELAPSED TTY COMM
10745 10578 100.0 750512 R+ 32:17 ttys001 claude

ps -M -p 10745 (main thread vs others):
USER PID TT %CPU STAT PRI STIME UTIME COMMAND
jm 10745 s001 98.6 R 31T 0:07.37 12:45.65 claude
10745 0.0 S 31T 0:01.33 0:08.95
10745 0.0 S 31T 0:00.20 0:00.06
... (8 more idle threads)

lsof -i -p 10745 (frozen):
TCP 192.168.0.101:51528->160.79.104.10:https (CLOSED)
TCP 192.168.0.101:50169->165.66.149.34.bc.googleusercontent.com:https (CLOSED)
TCP 192.168.0.101:50158->160.79.104.10:https (CLOSED)
lsof -i -p 8598 (healthy session, same machine, same minute): 5 x TCP ...:https (ESTABLISHED)

===== sample 10745 5 -f (main-thread excerpt; full 120 KB file available on request) =====
Analysis of sampling claude (pid 10745) every 1 millisecond
Process: 2.1.261 [10745]
Path: /Users/USER/*/2.1.261
Load Address: 0x1041f0000
Identifier: 2.1.261
Code Type: ARM64
Parent Process: zsh [10578]
Date/Time: 2026-09-06 06:58:19.163 +1000
Launch Time: 2026-09-06 06:25:43.724 +1000
OS Version: macOS 26.6.2 (25G83)
Physical footprint: 401.6M
Physical footprint (peak): 423.3M

Call graph:
3637 Thread_124058 DispatchQueue_1: com.apple.main-thread (serial)
+ 3637 start (in dyld) + 6992 [0x185dc44e4]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x112efd4 [0x10531efd4]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x104eb84 [0x10523eb84]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x1053528 [0x105243528]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0xdfa30c [0x104fea30c]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0xfe7b40 [0x1051d7b40]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x79bf4c [0x10498bf4c]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x79b9d8 [0x10498b9d8]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x15ab938 [0x10579b938]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x15ab4d4 [0x10579b4d4]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x15abfa0 [0x10579bfa0]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x11286f4 [0x1053186f4]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x13a0d38 [0x105590d38]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x6ed9d4 [0x1048dd9d4]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x1a73c8 [0x1043973c8]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x2bde880 [0x106dce880]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x28affe4 [0x106a9ffe4]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x35418b4 [0x1077318b4]
+ 3637 ??? (in ) [0x116316224]
+ 3637 ??? (in ) [0x1163161f4]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x588f44 [0x104778f44]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x31d337c [0x1073c337c]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x3074cbc [0x107264cbc]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x2dd5068 [0x106fc5068]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x2dd9b58 [0x106fc9b58]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x35418b4 [0x1077318b4]
+ 3637 ??? (in ) [0x1173ead10]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x3564f90 [0x107754f90]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x3564f90 [0x107754f90]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x3564f90 [0x107754f90]
+ 3637 ??? (in ) [0x115905978]
+ 3637 ??? (in ) [0x1169eace8]
+ 3637 ??? (in ) [0x115cb3fbc]
+ 3637 ??? (in ) [0x1163a7060]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x219819c [0x10638819c]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x2198318 [0x106388318]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x30d4224 [0x1072c4224]
+ 3637 ??? (in 2.1.261) load address 0x1041f0000 + 0x30d4224 [0x1072c4224]
+ 2247 ??? (in ) [0x1159048dc]
+ 1162 ??? (in ) [0x1159048b4]
+ 66 ??? (in ) [0x115904930]
+ 57 ??? (in ) [0x11590493c]
+ 51 ??? (in ) [0x1159048a4]
+ 43 ??? (in ) [0x115904880]
+ 5 ??? (in ) [0x115904928]
+ 3 ??? (in ) [0x1159048b0]
+ 2 ??? (in ) [0x1159048d0]
+ 1 ??? (in ) [0x115904944]
[other 10 threads (mi-scavenger, Bun Pool 0-8): all idle in __ulock_wait / __ulock_wait2 / kevent64]

Binary Images:
0x1041f0000 - 0x107fef9bf +2.1.261 (???) <4C4C4401-5555-3144-A13C-E92490FFBCBC> /Users/*/2.1.261
0x11b874000 - 0x11b98fb4f +.bun-501-b163af53257b65ef.node (0) /var/folders/*/.bun-501-b163af53257b65ef.node
```

### Steps to Reproduce

No deterministic reproducer; it happens 30 min to 2 h into sessions that use subagents (same as #89062). Exact sequence of the captured occurrence:

1. `cd` into a project whose `.claude/settings.json` defines PreToolUse, UserPromptSubmit and SessionStart command hooks (bash + jq); start `claude` in auto permission mode, classic renderer (`"tui": "default"`), from Zed's integrated terminal (also seen in VS Code terminal and Warp).
2. Prompt: "pull the latest version" → Claude runs one Bash call (`git pull --ff-only … && git log … && git status`) and answers. Status line shows "1 shell still running".
3. Run a project slash command (`/pm-loop`) whose instructions make Claude run three short Bash calls, `CronList`, then dispatch a **background** subagent via the Agent tool (`project-manager`, model Opus 5 with 1M context). The UI shows "Backgrounded agent" and Claude prints "Waiting for the tick's completion notification", spinner "Waiting for 1 background agent to finish".
4. The screen never updates again. The background agent's result is never surfaced. Ctrl+C does nothing.
5. From another terminal: `ps -M -p ` shows the main thread at ~100% CPU; `sample 5` shows one identical stack for every sample (above).
6. Only `kill -9 ` ends it; `claude --resume` recovers the transcript.

So the hang happens while the parent turn is idle waiting on a backgrounded subagent (plus one still-running background shell), i.e. in the background-agent / task-notification path, not while streaming. Possibly related to the 2.1.261 changelog entry "Fixed sustained high CPU usage when a background agent could not be resumed".

### Claude Model

None

### Is this a regression?

Yes, this worked in a previous version

### Last Working Version

Unknown. Did not happen before ~mid/late August 2026 on either of two Macs; #89062 dates the same regression to ~2026-08-21 (2.1.231+). Reverting to a much older build was tried but that build was unusable for other reasons.

### Claude Code Version

2.1.261 (Claude Code) — native install; also reproduced on 2.1.252, 2.1.258, 2.1.259

### Platform

Anthropic API

### Operating System

macOS

### Terminal/Shell

Other

### Additional Information

Environment
- Hardware: MacBook Air M1 (MacBookAir10,1), 8 cores, 16 GB. Also seen on a second Mac cloned from the same config.
- macOS 26.6.2 (25G83); earlier occurrences on macOS 26.3 (25D125).
- Install: native (`claude.ai/install.sh`), binary `~/.local/share/claude/versions/2.1.261`, `autoUpdates: false`.
- Terminal: Zed 1.18.1 integrated terminal (login -> zsh -> claude). Previously reproduced in VS Code integrated terminal and Warp, so terminal-independent.
- Renderer: `"tui": "default"` at the time of this capture; earlier freezes with `"tui": "fullscreen"`.
- Permission mode: auto. Output style: Brief. Status line: a bash+jq script (~120 ms).
- Project-level hooks (all bash+jq, measured): PreToolUse (no matcher, 79 ms; Bash matcher, 97 ms), UserPromptSubmit (1.3 s), SessionStart (2.7 s). No MCP servers configured in this project. Plugins enabled: superpowers, typescript-lsp, security-guidance (official marketplace).
- Frequency: several times per day across sessions in this and sibling projects; always during or after subagent use; two other sessions on the same machine at the same time were unaffected.

The main-thread excerpt of `sample 5 -f` (1 ms interval, 3637 samples) is in the Logs field above, with the Binary Images UUID for symbolication. The full 120 KB sample (all threads) is saved locally and I can attach it on request.

A second, older sample from 2026-08-27 (different Mac, 2.1.24x) showed the same shape: 100% CPU main thread, flat RSS, SIGTERM ignored. Happy to run any diagnostic build or `--debug` flags on the next occurrence — it recurs within hours.

Not a duplicate of #83237 (there the main thread is idle and mi-scavenger spins; here the main thread spins and every other thread is idle).

Start with the attached sample and the 2.1.261 arm64 binary, then compare the reported stack and subagent sequence with #89062 and #51560. No source file, test, or deterministic reproducer is named; progress would require symbolication or a reliable reproduction, followed by verifying that background-agent sessions keep repainting and accept input.

Vom Indexierungsmodell aus dem Issue-Text verfasst.