One verifiable mechanism in OpenAI Codex 0.166.0 is that the client can fall back from WebSocket to HTTPS when the former fails, while preserving that state for the current session. But “it only works near 5/5” cannot, by itself, prove the root cause.
What this is
The “5/5” shown in the interface is a retry budget, not a model startup sequence. In OpenAI's configuration, request_max_retries controls HTTP request retries and defaults to 4; stream_max_retries controls retries after a streamed response is interrupted and defaults to 5. Connection failures, interrupted response streams, and retryable service errors can all trigger recovery logic.
For Codex 0.166.0, one source-verified path is this: if the WebSocket handshake, connection maintenance, or forwarding fails, the client switches to HTTPS and displays the message “Falling back from WebSockets to HTTPS transport.” This version preserves session-level fallback state, so the initial wait may be long while later interactions recover normally.
A protocol switch, however, does not mean the proxy has switched. Both WebSocket and HTTPS + SSE (server-sent events, which continuously push events within an HTTP response) can carry text and tool calls, and both may pass through proxies. Whether a proxy route is taking effect depends on the process's actual configuration and network logs; it cannot be inferred from the number of reconnects.
Industry view
What deserves credit is that this article does not present “5/5” as a universal rule. Instead, it provides a verifiable version, an error message, and a troubleshooting sequence. Enterprises deploying coding agents—autonomous programs that can read files, modify code, and run tests—should also monitor the model service, network proxy, and client configuration separately.
The risk is equally clear: different clients, model providers, feature flags, and network environments may use different recovery flows. There is no protocol guarantee from Codex that fallback always happens after exactly five attempts. If teams only change proxy variables, or treat a browser's ability to connect as proof that the terminal's network path is healthy, the problem may persist.
Our judgment is that the main value of WebSocket for coding agents is reducing the overhead of coordinating multi-turn model and tool interactions, but that advantage depends on a stable network. When the connection path keeps failing, recovery waits can easily exceed the time saved through more efficient transport. Performance claims must be grounded in end-to-end logs rather than model-side response latency alone.
Impact on regular people
For enterprise IT: Log proxies, certificates, gateways, and endpoint configurations separately. Desktop clients, command-line tools, and IDE plugins may not inherit the same network settings, so initial connections and mid-stream interruptions require different diagnostic checks.
For individual professionals: When using a coding assistant such as Codex, if the first message takes too long to produce any output, record the client version, error messages, and stage of failure before changing network settings. This avoids unnecessary reinstalls or misdiagnosing account limits.
For the consumer market: When an AI assistant occasionally “spins,” it may not mean the model is thinking; it may be switching connections. As enterprises integrate coding tools into everyday systems, network observability will become an invisible cost affecting user experience.