Codex “Waiting for network”: what it means and what to check
A careful, evidence-based way to distinguish a Codex network wait from browser success, authorization, or a provider incident.
Resources
AI Network Guides & Research
Practical troubleshooting and protocol research for ChatGPT, Codex, Claude Code and AI connectivity.
Troubleshooting
Start with the symptom.
Codex keeps reconnecting: what the symptom can—and cannot—tell you
Separate a Codex reconnect symptom from DNS, TLS, WebSocket, route, and provider evidence without guessing root cause.
ChatGPT works, but Codex fails
Why a working ChatGPT page does not prove the native Codex path, protocol, proxy context, or session is healthy.
Claude Code timeout: diagnose the missed deadline without guessing
Separate DNS, connection, TLS, proxy, first-byte, streaming, and application evidence after a Claude Code timeout.
Claude starts responding, then the stream stops
Why initial Claude reachability and stream continuity are different observations, and how to verify safely.
Cursor website works, but the client does not
A bounded way to investigate Cursor client connectivity without treating a working website as proof of the application path.
How to triage GitHub Copilot network failures
Separate Copilot client, proxy, certificate, provider-status, and entitlement evidence without guessing a root cause.
Gemini connectivity: local path or service incident?
Use bounded evidence to distinguish Gemini transport symptoms from API, account, quota, and provider-service conditions.
Why AI tools fail on public or hotel Wi-Fi
Recognize the captive-portal bootstrap window without automating login or blaming the AI provider.
The browser works, but the AI agent fails
Why browser success does not prove a native AI agent’s DNS, TLS, proxy, WebSocket, or address-family path.
Research & Concepts
Understand the evidence boundary.
HTTP 401 or 403 from an AI service: authentication or connectivity?
Why a measured HTTP 401 or 403 is an authorization response—not automatic proof of provider unavailability.
Provider outage or local network failure? A disciplined way to tell
Compare local protocol evidence, official provider status, and one controlled alternate path without overstating root cause.
Why WebSocket can fail when HTTPS works
Understand the distinct WebSocket handshake and persistent channel without claiming that Open Probe V0.1 measures it.
TLS inspection and AI coding agents
How certificate replacement and inspected traffic can affect AI agents, and how to investigate without disabling TLS validation.
Fake-IP DNS: synthetic address or real failure?
Why a synthetic DNS address is not automatically a leak or outage, and which evidence is still needed.
IPv4 works while IPv6 is degraded
Why dual-stack networks can produce different outcomes and why Open Probe V0.1 does not yet compare the two paths.
Captive portal or AI provider failure?
Use link, gateway, portal, general-access, and provider evidence in the right order.
HTTP/2 versus HTTP/1.1 in AI streaming failures
Why negotiated HTTP version is useful context but not a root-cause verdict.
Changing egress IP: failure or normal multi-egress?
Why an exit-address change is an observation, not automatically an AI route failure or quality score.
What a browser AI network scanner cannot measure
Understand browser evidence limits and when a local, read-only Open Probe run adds defensible context.