The spinner keeps turning, the token counter does not move, and no error message appears. Five minutes pass, then fifteen, sometimes more than twenty. You press Esc, send the same thing again, and the answer arrives in seconds.
That is the symptom in issue #26224 in the CLI's official repository, opened in February 2026 and still active in November, with reports of 5 to 20 minute stalls on consecutive prompts. It is not the model thinking, not your account and not your prompt: it is a connection that died without warning, and no watchdog noticed.
What is happening when the CLI hangs with no error?
The connection died silently. No FIN, no RST, no HTTP error response. The process keeps waiting for bytes that will never arrive, and the only clock that can interrupt that wait is API_TIMEOUT_MS, which defaults to 600 seconds in the client.
That number is not forum folklore. It sits in the code of the client that runs in your terminal, in the gap between the request and giving up: timeout: parseInt(process.env.API_TIMEOUT_MS || String(600 * 1000), 10). If the connection dies before the headers arrive, nothing fires before that.
| Where it stalled | What should warn you | What actually happens |
|---|---|---|
| Before the headers | a first-byte watchdog | none is armed, only API_TIMEOUT_MS is left |
| Mid-stream | the stream idle watchdog | the stall still waits for the same backstop |
After an api_error | an immediate retry | the retry happens, but only after the time runs out |
I raised CLAUDE_SLOW_FIRST_BYTE_MS and nothing changed. Why?
Because that variable aborts nothing. According to the consolidated report in issue #88178 in the same repository, CLAUDE_SLOW_FIRST_BYTE_MS defaults to 30 seconds and only writes to the log and emits telemetry. It does not cut the request, does not trigger a retry and does not tell the user.
It is the difference between an alarm that rings and an alarm that only writes to a file. If you spot the variable in the log and raise it, you keep staring at the same spinner.
| Variable | What it controls | Default |
|---|---|---|
API_TIMEOUT_MS | how long the client waits for a response | 600000 (10 minutes) |
CLAUDE_SLOW_FIRST_BYTE_MS | threshold at which a slow stall is logged | 30000 (30 seconds), without aborting |
Why does it hang for 15 minutes instead of 10?
Because in the desktop app the injected value is larger. The author of #88178 measured it with ps eww along the process chain: the app launches the CLI with API_TIMEOUT_MS=900000 in the environment, and process environment variables take precedence over the env block in the settings file.
Nine hundred seconds is exactly 15 minutes, and the stalls measured in that report match the number: 900s, 939s and 753s, all ending in api_error Request timed out, all followed by a retry that answered in seconds. Stack two stalls in the same session and you have the 20 minutes in the title of #26224.
Practical consequence: if you use the desktop app and set API_TIMEOUT_MS in settings, your value may be ignored. Check with ps eww on the CLI process before blaming your configuration.
How do I get out of the stall in seconds?
The manual exit is the same in both reports and it is the fastest: interrupt and resend. What you lose is the turn that was being generated, not the session.
- Press
Escto interrupt the stalled turn. - Resend the same instruction, or send any short message that forces a new request. In #26224, a simple follow-up often unblocks the pending prompt.
- If it stalls again at the same point, the problem is on the network path or the provider, not in your prompt. Switch route instead of insisting.
Step 3 matters most. A stall that always comes back at the same stretch, with the clock landing near a round number, is the signature of a dead connection, and resending will not fix the cause.
How do I reduce the chance of stalling again?
- Lower
API_TIMEOUT_MSto something between 90 and 120 seconds in the CLI environment. You trade a long wait for a fast error that already comes with a retry. - Split long turns. The longer the connection has to stay open, the more exposure to a silent stall.
- Run
/compactbefore large tasks, so the next turn carries less context. - Open long sessions with
--debug-fileand look for the stall line. Without the log, the symptom stays invisible. - Keep a second route configured. When the first one stalls and you only find out 15 minutes later, having somewhere to go changes your day.
Verboo Code exists precisely to be that second route: the same terminal, command and MCP workflow, with unlimited tokens per subscription. If your provider's connection stalls, you switch engines with /model without changing tools or losing the session.
![[CC] stuck with no response for 15 minutes: how to break out](/_next/image?url=https%3A%2F%2Fstorage.googleapis.com%2Fverboo-landing-page%2Fblog-uploads%2F1783232445352-verboo-code-cover.png&w=3840&q=75)
