[Bug] Extended Thinking causes 7–8 minute stalls with no output and silent token consumption

Bug Description

Support Request — Claude Code Extended Thinking Stall / Token Loss

To: Anthropic Support
Issue tracker: https://github.com/anthropics/claude-code/issues
Date: 2026-04-21

---

Subject

Claude Code (claude-sonnet-4-6): Extended Thinking causes 7–8 minute silent stalls with no output, consuming 55 000+ tokens without delivering a response.

---

Environment

| Parameter | Value |
|-----------|-------|
| Claude Code version | 2.1.116 |
| Model | claude-sonnet-4-6 |
| OS | Linux Debian 13, x86-64 |
| Terminal | VS Code integrated terminal |
| Fast mode | Unavailable (not enabled on account) |

---

Description

After submitting a prompt, Claude Code enters Extended Thinking mode (streamMode=thinking) and produces no visible output for 7–8 minutes. The user sees only a spinner. No partial responses, no progress indicator, no timeout message. Tokens are consumed silently in the background.

The issue was reproduced 4 times in a row with the same prompt on 2026-04-21.

---

Reproduction

Reproduced with any sufficiently complex prompt (development task for a Django project). Observed in 4 consecutive sessions between 11:56 and 12:29 (UTC+10) on 2026-04-21.

---

Evidence from debug logs

Key log entries from session_20260421_221059.log:

2026-04-21T12:11:32Z [DEBUG] [API REQUEST] /v1/messages
                             x-client-request-id=4447af45-28fc-4968-818d-4fd36865de6c
                             source=repl_main_thread

2026-04-21T12:11:52Z [DEBUG] Stream started - received first chunk

      ← 451 seconds of silence (no stream events) →

2026-04-21T12:19:23Z [WARN]  Streaming stall detected: 451.4s gap between events (stall #1)
2026-04-21T12:19:23Z [WARN]  Streaming completed with 1 stall(s), total stall time: 451.4s

2026-04-21T12:19:31Z [DEBUG] autocompact: tokens=55571 threshold=167000
                             ↑ 55 571 tokens consumed — no visible response delivered

--- Second request immediately after ---

2026-04-21T12:20:03Z [DEBUG] Stream started - received first chunk

      ← 455 seconds of silence →

2026-04-21T12:27:38Z [WARN]  Streaming stall detected: 455.3s gap between events (stall #1)
2026-04-21T12:27:38Z [WARN]  Streaming completed with 1 stall(s), total stall time: 455.3s

Sessions 1–3 (user cancelled manually after 50 sec – 2 min):

[onCancel] focusedInputDialog=undefined streamMode=thinking

All three sessions: identical cancellation pattern, streamMode=thinking confirmed.

---

Impact

• 4 sessions, all failed to deliver a response within an acceptable time
• 55 000+ context tokens consumed in session 4 with no useful output
• No error message shown to user — only a spinner
• Cannot work on the project using claude-sonnet-4-6 in Claude Code CLI

---

Expected behaviour

One of the following:

1. A response within a reasonable time (e.g. ≤ 60 seconds for standard tasks)
2. A visible progress indicator showing that Extended Thinking is active and how long it has been running
3. A configurable timeout after which the thinking phase is aborted and the user is notified
4. A way to disable Extended Thinking for Claude Code CLI sessions

---

What I have tried

• Repeated the same prompt 4 times — same result every time
• Waited 7–8 minutes (sessions completed but no response appeared in terminal)
• The issue is not related to context size (tokens=55 571, threshold=167 000 — autocompact did not trigger)
• Fast mode is unavailable on my account — cannot test whether it bypasses Extended Thinking

---

Attached files

Four debug log files are available:

• session_20260421_215632.log — session 1 (cancelled at 2 min)
• session_20260421_215911.log — session 2 (cancelled at 50 sec)
• session_20260421_220029.log — session 3 (cancelled at 2 min, +30 sec slow first byte)
• session_20260421_221059.log — session 4 (waited to completion, stalls 451 s + 455 s)

Logs are located at:
/home/<user>/DevProjects/EvrikaSchool/claude_logs/

---

Question for support

1. Is this behaviour expected for claude-sonnet-4-6 in Claude Code 2.1.116?
2. Is there a setting to limit or disable Extended Thinking budget in Claude Code CLI?
3. Will tokens consumed during a stalled/incomplete thinking session be refunded or not counted?
4. Is there a known regression in 2.1.116 related to Extended Thinking stalls?

Environment Info

• Platform: linux
• Terminal: vscode
• Version: 2.1.116
• Feedback ID: 50a7c6ca-268a-4dfa-94d4-435a0958e52e

Errors

[]