CLI crash: stack-overflow recursion in AskUserQuestion render path while idle (no question pending)

Environment

• Claude Code: v2.1.112
• Model: Opus 4.7 (claude-opus-4-7)
• Platform: macOS 15.3 (Darwin arm64)

Description

CLI crashed with infinite-recursion stack dump while a foreground session was
idle, waiting on sub-agent results. The crash dumped a large chunk of
minified cli.js to the terminal and killed the foreground process; orphaned
sub-agent sessions kept writing to their transcripts for ~3 minutes before
stopping (presumably when their parent's death was detected).

The foreground assistant never invoked AskUserQuestion, EnterPlanMode, or
ExitPlanMode — yet the crash trace is in the AskUserQuestion / plan-mode
interview render path.

Workload shape

• Foreground session was coordinating a team of 5 sub-agents running in

parallel git worktrees, communicating via SendMessage.

• At time of crash, foreground was idle, awaiting completion messages from

the sub-agents. No user input was pending.

Crash trace (head + tail)

Crash site: cli.js:8231:15642 — inside the function that fires the
tengu_ask_user_question_finish_plan_interview analytics event and awaits
xf8(g, i) (image attachment processing) for the AskUserQuestion
onAllow / onReject callbacks.

Stack tail shows tight recursion:

- <anonymous> (cli.js:8231:15642)
- qX  (cli.js:477:63169)
- sPH (cli.js:477:76220)
- ov  (cli.js:477:76101)
- sPH (cli.js:477:76199)
- ov  (cli.js:477:76101)
- sPH (cli.js:477:76199)
- ov  (cli.js:477:76101)
...repeats...

sPH ↔ ov looks like a React error-boundary / effect re-render loop.

Transcript evidence

Final entries of the foreground transcript (redacted), in order:

T+0    user      <teammate-message teammate_id="quality-reviewer" summary="Worktree verified, ready for review">
                 Worktree verified at <worktree-path>. WORKTREE OK. Ready to receive tickets.
                 </teammate-message>

                 <teammate-message teammate_id="quality-reviewer">
                 {"type":"idle_notification","from":"quality-reviewer","idleReason":"available"}
                 </teammate-message>

T+3s   assistant TEXT: "Quality-reviewer verified. All 5 agents ready; 3 implementers
                       working. Awaiting DONE signals."

T+3.2s system    stop_hook_summary  (clean, no error)

[no further entries — transcript ends here]

• Tool-use census on the foreground session over its full lifetime:

18× Bash, 5× Agent, 3× SendMessage, 3× Read, 3× TaskCreate, 1× Grep,
1× ToolSearch, 1× TeamCreate.
0× AskUserQuestion, 0× EnterPlanMode, 0× ExitPlanMode.

• Sub-agent transcripts continued writing for ~3 minutes after the

foreground transcript ended, then stopped abruptly.

Reproduction

Not reliably reproducible. Happens intermittently in workloads that
coordinate multiple parallel sub-agents over SendMessage.

Possibly related

• #16918 (closed as dup of #16871) — same symptom (RangeError stack overflow

during AskUserQuestion render after parallel agents completed) on v2.1.1,
closed 2026-01-08. The current crash on v2.1.112 may be a regression or
a related uncovered code path.

Impact

• Foreground CLI dies; user loses interactive session.
• Sub-agent processes are orphaned and continue running for several minutes

(related to #20369 / #19045).

• Any in-flight coordination state is lost.

What should happen

• The CLI should not crash while idle.
• If the AskUserQuestion / plan-interview render path is being entered

spuriously (no user-facing question pending), entry should be guarded.

• If a render does fail, the error boundary should report and recover, not

recurse.