First, check what is actually affected.
Open the Codex status view or Claude status view. Match the product and time with the source. If no incident is listed, continue with local checks: a public feed cannot verify the health of your individual session.
Errors, timeouts or no first response
Record the exact error and when it happened. Check whether other independent tasks fail on the same product. A status marked “monitoring” means the provider is watching a mitigation. Wait for an explicit recovery update before treating that incident as resolved.
For APIs, distinguish a rate-limit response from a service error. Anthropic documents its limits and response headers in the official rate-limit reference. The API’s rules are separate from consumer subscriptions. Avoid rapid automatic retries that can create additional requests.
Chat responds, but a tool will not run
Note the failing tool and whether it is a local command, file operation, MCP server or hosted feature. Check its permission message, connection and environment. A tool permission failure is not evidence that the underlying model lost reasoning ability.
Claude Code’s /doctor checks installation and configuration. Its official troubleshooting guide separates installation, connection and performance problems. Use the documented fix that matches your error, and preserve unfinished work before changing your setup.
Claude Code: disconnected MCP or authentication errors
If an MCP server disconnects and automatic reconnection gives up, open /mcp to inspect that connection. Claude Code v2.1.273 added a notification pointing there. The same release improved 401/403 guidance for Bedrock, Vertex, Foundry and Claude app gateways: follow the message for the affected credential or gateway administrator instead of assuming every error needs /login. These are client diagnostics, not evidence of a model-quality change. Read the version-specific release notes ↗
It becomes slower during a long session
Separate time spent waiting for the model from time spent reading files, running commands or retrying tools. Save a concise task description and compare a fresh session with the required context. Keep the selected model and reasoning setting the same. A faster fresh session suggests a session-specific difference; it does not by itself identify the cause.
If behavior changed after a client upgrade, compare your installed version with Codex and Claude Code releases. If the answers are worse but response speed is normal, use the quality comparison guide.
Context fills unexpectedly quickly? Claude Code v2.1.273 fixed the context meter counting advisor-tool turns at roughly twice their actual size, which could trigger compaction early. Check whether your installed version includes that fix and whether your session uses the affected tool. This does not establish that every long-session slowdown has the same cause. See the documented fix ↗
Usage limit reached or allowance falling quickly
Check the allowance and reset time displayed in your account. Save that time in the personal reset timer. Public Codex resets and banked credits follow different rules from your ordinary usage window; see the Codex usage-limit guide.
Repeated failed attempts can be relevant to consumption, but token usage alone does not reveal which model answered. Keep usage records separate from task-quality results.
What to include in a reproducible report
- Product and entry point: browser chat, CLI, desktop app, API or another client.
- Client version, selected model and reasoning setting, when visible.
- Pacific timestamp, exact error, affected tool and a minimal task.
- Expected outcome, actual outcome and whether a fresh session reproduces it.
- A matching official incident or release, if one exists.
Remove secrets and private project content before posting a report. Keep your own full logs locally. NextReset currently links official records; it does not accept private diagnostic uploads.
Back to AI health ↗