Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeout

What happened

Gemini 4 Argon runs in my workshop through Gemini CLI on Vertex AI. A headless test answered in about 11 seconds. Then the interactive CLI took 3 minutes 54 seconds to answer a greeting, and after an update a session sat for 10 minutes 44 seconds before I cancelled it.

The first theory was wrong

The obvious suspect was the size of the start-up prompt: 137 skill descriptions and a 32,875-token first greeting. Disabling the skills did not fix it. The session that wrote the diagnosis said so plainly: the evidence pointed at request handling or serving latency, not a simple "too many skills" explanation.

What it actually was

Later, during MythREAL work, the same kind of stall came back as repeated retries, and this time the CLI's debug console named the error: UND_ERR_HEADERS_TIMEOUT. The CLI gives a response 60 seconds to start sending headers. A hard Argon request can take longer than that before the first byte, so the CLI gave up and retried, over and over. It looked like a hang.

A local patch raised the header and body allowance to 900,000 ms (15 minutes), only for direct Vertex gemini-4-argon generation requests. Ten focused transport checks passed, and the resumed session completed a request that ran for more than three minutes.

If you hit this

• Look for UND_ERR_HEADERS_TIMEOUT in the console before you blame quota, an outage or your prompt size.
• Google's Argon migration guide recommends a client timeout of at least 15 minutes for difficult prompts (vendor guidance, not my test).
• The upstream report is google-gemini/gemini-cli issue #29747. My fix is a local compatibility patch, not an upstream release.

Limits of this report

• One observed session and a local patch, not an upstream fix.
• This measures the harness, not the quality of Gemini 4 Argon's answers.
• Provider-side latency was not measured on its own.