Diagnose AI Access Issues by Layer
Start with the last step that worked
Turn “this tool doesn’t work” into a description you can reproduce. For example: the homepage loads and the identity provider completes verification, but the product still shows you as signed out. That pinpoints the boundary: the initial page load and authentication succeeded. Next, check the redirect URL, browser session, and workspace selection instead of switching routes right away. Note the last successful action, the first failed action, and the exact message shown. If there is a specific error, save it rather than recording only a vague error name.
A page loading only shows that its resources or cache are available. Sending a message may connect to a different service; file uploads, conversation history, and continuous text delivery may each use separate connections. A fast homepage does not prove the generation path is stable. Conversely, if the response is still updating but the conversation list is not, record a UI sync issue rather than assuming the generation service is down. If only an older conversation fails, test with a new conversation that contains no personal information and keep the old one for comparison.
Separate regional eligibility, identity, and transport
Regional eligibility determines whether a product is available to users in a given location and to the relevant account or organization. Providers may consider the apparent source IP, account details, billing location, and organization region; check the provider’s own policies for specifics. The network egress address is only one part of what a service sees when a request arrives. It does not change account ownership, contractual permissions, or payment eligibility. If you see an explicit regional restriction, consult the official guidance for that product entry point. Do not assume that because a basic webpage loads, models, APIs, or enterprise features from the same brand are also available.
The identity layer answers “whose request is this?” The permissions layer answers “what can this identity do?” Being able to sign in does not mean you joined the right organization; seeing an organization’s interface does not mean an administrator assigned you the necessary license. The transport layer determines whether a request can reach its destination and the response can return intact. DNS failures, certificate errors, and proxy timeouts usually happen before the application checks permissions. If you receive a recognizable product error, save the response and determine whether it came from the product or an intermediary gateway before choosing which layer to investigate.
Understand egress IPs and risk checks
A single network egress point may carry traffic from many users. A service sees the request’s source and behavior; it does not treat every session as the same environment just because the client shows the same region name. Repeatedly moving between connection paths during one sign-in may also trigger another verification request. If prompted to verify, use your usual device and an approved network environment to complete the official process. If the account is explicitly restricted, switching egress points repeatedly will not restore access and can make the original issue harder to trace.
Change only variables that test a hypothesis
When setting up a comparison, decide which conditions will remain unchanged. If you suspect browser blocking, keep the account, network, and steps the same, but try a clean browser profile. If you suspect an unstable connection, keep the client, workspace, and test text the same, and repeat the same action in an approved connection environment. If you change the account, device, and route all at once, even a successful result will not show which change helped. During troubleshooting, disable extra automatic retries so requests from the previous attempt do not keep running in the background and skew the next result.
Also note the scope of the issue: is it limited to this device, affecting everyone in the same organization, or reflected in an official status notice? Scope is more useful for diagnosis than a single speed test. If only this device is affected, start with its clock, proxy settings, and browser logs. If the whole organization is affected, check admin policies and project status first. If the provider confirms an outage, note which features are affected, wait for recovery, and retest in the original environment. Do not mistake a temporary success for the result of a configuration change.
Platform Eligibility and Product Entry Points
Check the entry point, then confirm availability
When searching official documentation, use the full product and entry point name rather than the brand alone. A browser assistant, developer API, cloud-hosted model, and enterprise editor plugin may each have different supported regions, account types, and billing rules. Check whether a policy refers to the user’s location, billing details, or resource deployment region, and distinguish open account registration from access to a specific feature. Even if a region is listed, confirm whether the entry applies to individual accounts or organization tenants, and whether the specific model or tool you need is available there.
| Tool | Entry points that are easy to confuse | Check first |
|---|---|---|
| ChatGPT | Chat interface and developer API | Workspace, API project, and billing |
| Claude | Chat interface and developer console | Account type, organization, and model permissions |
| Gemini | Personal app, developer portal, and cloud platform | Identity type, project, and resource location |
| Copilot | Microsoft Copilot and GitHub Copilot | Subscription, organization license, and editor identity |
| Midjourney | Web app and officially supported Discord entry point | Linked identity, subscription, and task status |
| Cursor | Editor features and account management page | Current team, sign-in status, and model availability |
Check chat assistants separately from developer products
When using ChatGPT or Claude, if the chat interface works but your code reports that the project is unauthorized, check the developer console for the organization associated with the credential, available models, and billing status. A web subscription does not tell you the API quota, and a browser session is not an authentication credential for scripts. If only one model fails, compare it with a model you know is authorized in the same project and check the product’s error code. This helps distinguish a project-wide issue from a mismatched model name or missing access to that model.
Gemini troubleshooting especially requires separating the personal app, developer product, and cloud platform. A successful chat in a personal browser does not prove that the corresponding API is enabled for a cloud project. In the cloud project, confirm that you are working in the intended project, then check the API requirements for resource location, identity roles, and billing. If an organization-managed identity is authorized only for certain resources, ask an administrator to confirm and grant access; do not use a personal project for organizational data. Use input without sensitive information for comparisons, and do not copy troubleshooting samples into an unapproved environment.
Check identity linking for plugins and creative tools
For Copilot, first specify whether you mean Microsoft’s product or GitHub’s developer tool, then identify whether the issue occurs in the web app, desktop app, or editor. The identity signed in to your browser may differ from the one used by the editor extension. Even when an extension shows as signed in, check the account and the source of its license. If suggestions work in a personal repository but not an organization repository, check organization policies, authorization scope, and the current repository’s settings first. Reinstalling the extension repeatedly can hide the real permissions difference.
If Midjourney accepts a task but no image appears, check the official entry point to see whether the task was accepted, whether it has finished, and whether preview assets load. A generated image that cannot be downloaded and a task that never reached the service need different fixes. If Cursor’s account page works but only some models are unavailable in the editor, check the current team and model permissions, then determine whether the issue is local or in the remote editor environment. An online indicator in the editor proves only that some requests succeeded.
Account Setup and Session Recovery
Confirm account ownership before setup
Before starting work, decide who will manage the account: use your own identity for personal testing and the organization identity specified by an administrator for team projects. Before entering details, verify the official entry point, eligibility, and billing entity. Do not use someone else’s identity or submit information through an unofficial account-creation service. If you need to share a repository or hand off work later, ask an administrator to assign roles through the organization rather than sharing a personal account with colleagues. Settling account ownership and export permissions early makes it easier to find the right person when issues arise.
LiuVPN does not require an email address; you can create an account with a username and password. This describes the LiuVPN user panel only and is separate from authentication with AI providers. Use separate passwords for your panel account and third-party accounts; panel credentials cannot be used to sign in to third-party services. If team members need access to a LiuVPN subscription, follow the setup guide in the user panel. Do not share subscription links in AI chats, shared repositories, or public support tickets. Keep identity credentials and network subscriptions secure and use each only for its intended purpose.
Troubleshoot sign-in redirects and session storage
With single sign-on, a product sends your browser to an identity provider for verification, then redirects it back to the product. If verification succeeds but the product still shows you as signed out, confirm that you returned to the expected domain and completed the steps before and after verification in the same browser profile. The default browser, an in-app window, and a managed enterprise browser may use different session storage. Opening a link in another window is not necessarily a problem, but if the two profiles do not share a session, the return redirect may not match the original sign-in attempt.
Use an isolated browser profile for comparison: open only the official entry point you need, do not copy old site data, and avoid unnecessary extensions. If a clean profile works, check privacy extensions and site permissions in the original profile. Make changes only for the relevant site; ask an administrator to handle organization-managed settings, and do not disable security protections globally. If you need to clear site data, first make sure local drafts are saved, then clear data only for the relevant site. Afterward, restart from the product entry point rather than reusing a redirect URL from the previous sign-in attempt.
Distinguish successful sign-in, expired sessions, and authorization failures
Seeing an avatar only means the interface recognizes an identity. Check the organization and workspace in the account menu, along with the role required for the action. Similar accounts accessed through different identity providers may have different histories and licenses. If the page intermittently returns to sign-in, check whether the session expired, whether the system clock is correct, and whether the identity provider has posted a notice. If administrator approval is required, note the request status and organization name and contact an administrator; repeatedly signing out and back in usually will not change the approval outcome.
For handoffs, record only the failed step and a redacted error. Do not include passwords, recovery information, session values, or full redirect URLs containing identity parameters. You can note the public domains before and after the redirect, the browser type, and the time. If you used an unfamiliar device, sign out when finished. If you find unauthorized sign-in activity, follow the provider’s account recovery process to revoke suspicious sessions and review related credentials before investigating network issues. Record security incidents separately from ordinary sign-in failures so credential issues that need immediate attention are not mistaken for connection instability.
Page Loads and Streaming Responses
Identify where the response stops
A conversation does not wait for the server to finish generating before showing a result. A streaming API typically sends an initial response, followed by text and a final status in the service’s event format. A browser may use server-sent events or another streaming mechanism, while an editor may pass events to an extension process. During troubleshooting, check separately whether the request was sent, response headers arrived, body chunks arrived, and an end marker was received. A stall at each stage points to a different problem; the spinner alone is not enough to diagnose it.
If response headers are slow to arrive, the connection may not be established yet, or the request may have arrived and be queued. Receiving body chunks means at least part of the response was established, but generation can still be interrupted. If replies stop after a period with no new content, check intermediary gateway or client idle timeouts. If they stop soon after the response begins, look for an application error, a user cancellation signal, or a worker process exiting. If the service ends the response with a stated reason, start with that reason rather than mislabeling an output limit as a transport failure.
Inspect a comparison request in the browser’s network panel
Open the browser developer tools on the official web app, select the Network panel, and send a short message without personal information. Find the request by start time, check the status and timeline of failed entries, and compare them with the message in the product UI. If the network panel shows that a local extension blocked the request, test in a controlled clean profile. If the browser blocks the connection during certificate verification, check the system clock, trusted certificates, and organization gateway settings; do not bypass certificate errors to continue. Some page data is handled by background workers, so the error may appear in the UI after the request actually failed.
Next, test a somewhat longer, ordinary task using the same model and conversation type. If short tasks finish but longer ones are often cut off, note when the last text arrived, the final status, and the scope of the failure, then check for a fixed deadline or idle timeout. If the answer appears all at once at the end, consider intermediary buffering, client display behavior, or whether the service actually streams incrementally. Do not judge route latency only by how text appears on the page; use the same request’s records to distinguish when data arrived from when the UI displayed it.
Test uploads and background transitions separately
If plain text works but attachments fail, start with a small, non-sensitive sample that complies with platform rules. Note whether the failure occurs before upload, during post-upload processing, or during generation. Before upload, check whether the browser can read the file, whether the upload destination is reachable, and whether the organization allows uploads. After upload, check the file format, server-side processing result, and authorization for the current tool. If removing the attachment makes the request succeed, that only links the issue to the attachment path; it does not prove the file is too large or the network is congested. Compare different attachments one at a time in the same conversation to isolate the cause.
When a device goes to the background or enters sleep, the browser may pause page tasks and the system may switch connections. First keep the test in the foreground and stable, then reproduce the transition separately. After a failure, check whether the original conversation already has a result and whether the task is still running. Do not immediately resubmit an action that could incur a charge or trigger another operation. For a support ticket, you can provide redacted status details, timestamps, and request IDs. Unreviewed network logs may contain conversations, attachment URLs, and authentication information; do not send them as-is.
API Requests, Timeouts, and Quotas
Test the web app and API separately
A browser manages the web session. An API call uses a project credential issued by the provider and must follow the API documentation for the model, input format, and response mode. A working web app only proves that the web entry point is available. Before debugging an API call, check the official endpoint, API type, and project environment in the configuration. Make sure credentials are injected through environment variables or a secrets manager, and that test and production projects are not mixed. Never put credentials in frontend pages, test fixtures, or commits. Do not print all request headers just to check whether credentials loaded.
Start with a minimal non-streaming request supported by the documentation, using an authorized model and a short input without private data. Record the status, server request ID, and error code after removing sensitive fields. Once it returns fully, try streaming with the same input. If non-streaming works but streaming fails, check whether the client parses event boundaries correctly, treats each read as a complete event, or whether a gateway is buffering. A streaming response may deliver one event in multiple chunks; the application must parse the protocol rather than decode arbitrary network chunks directly.
Set separate connection, read, and total timeouts
| Parameter type | What it controls | What to check when it triggers |
|---|---|---|
| Connection timeout | Time allowed to establish a connection | DNS, gateway, and certificate handshake |
| Read idle timeout | How long to wait for the next data | Buffering, generation gaps, and proxy idle settings |
| Overall deadline | Maximum time allowed for one task | Queueing, generation time, and caller budget |
Timeout parameter names and when their clocks start vary between clients. When choosing an SDK, check whether the read timeout applies to each successive read or the whole response; connection reuse also affects timing. Set an overall deadline based on the task, with time left for follow-up processing. If an outer task cancels but the inner client keeps waiting or retrying, the call continues to use resources. Propagate cancellation to the inner client and distinguish user cancellation, timeout, and remote closure in logs instead of reporting all three as the same network error.
Read the error before deciding to retry
For an authentication error, first check whether the credential belongs to the current project, has been revoked, and is the one loaded by the process. For a permissions error, check the model and organization role; for a validation error, correct the request body to match the documentation. Do not retry these responses without changing anything. When a rate limit is reached, read the error body and any retry guidance from the server to distinguish a temporary concurrency spike from an exhausted quota. Reduce concurrency and queue requests in the first case. In the second, pause the work, check the budget and quota, and wait for conditions to change or follow the provider’s official request process.
For temporary failures that are clearly retryable, use bounded backoff with jitter so multiple workers do not send requests again at the same time. If the SDK already retries, do not let the outer scheduler resubmit indefinitely. A timeout does not mean the server did not execute the request. For actions that may create tasks or incur charges, check the task status first. Use idempotency keys only when the provider explicitly supports them; do not assume a custom request header will prevent duplicates. If you already received part of a streaming response, mark it as incomplete and track the business action separately from the text delivery status.
CLI, Editor, and CI Configuration
Verify from the process that makes the request
A CLI, desktop editor, and pipeline may run on different hosts. A remote development extension can render its UI locally while the extension process sends requests from a remote machine or container. First identify the process making the request, its network environment, and the configuration source it uses, then run basic diagnostics in that same environment. Connectivity in a developer’s browser does not prove connectivity from the execution agent. If a script wraps the command, confirm it starts in the expected working directory and reads the intended configuration, rather than inheriting settings from an earlier process.
If curl is available, use a public example site to test basic DNS, connection, and TLS paths. The command below sends no account information and does not access an AI API. The exact error can help distinguish DNS failure, connection gateway failure, and a response from the remote host. Success only means that this process can reach the example destination; it does not prove that an AI service is available or that streaming will remain uninterrupted. Then run the provider’s minimal API test in the same environment and interpret the result using the official documentation.
curl --silent --show-error --fail --output /dev/null \
--write-out '%{http_code}\n' \
https://example.com/
Check proxy scope and certificate trust
Check whether your tool uses the system proxy or requires variables such as HTTPS_PROXY to be set for the process. Not every SDK reads variables with the same names, and NO_PROXY matching rules vary between runtimes. Before making changes, check the tool’s configuration precedence to see whether command-line options, user settings, or environment variables take priority. Restart the target process after changing settings, then verify that the application is using the new configuration. An extension or task that is already running usually will not pick up new variables from another terminal.
Check the editor’s main app, extension, and remote environment separately. If extension logs show a connection failure, first determine whether it concerns the identity page, a model call, or background indexing, then check the relevant module’s configuration scope. If a corporate gateway requires a trusted certificate, use one distributed and verified by an administrator, and configure it in the runtime that actually sends the request—not just the local browser. Do not disable verification to make the error disappear. If a proxy rule would send internal addresses outside the organization’s network, stop testing and ask an administrator to review its scope.
Manage CI access in the pipeline configuration
Give CI a dedicated project identity with only the permissions it needs. Store credentials in a protected secrets manager and inject them only into approved jobs. Jobs triggered by external contributions must not expose production credentials to unreviewed code. If the service requires an approved connection source, have an administrator configure the execution agent and egress policy for the organization rather than handing a personal session to the pipeline. Log only whether configuration is present, the target environment type, and redacted service errors. Do not print all environment variables or include raw request bodies and responses in default build artifacts.
Retest with an isolated job: check basic connectivity from the same execution agent, make a minimal API call, then restore concurrency. When several jobs start together, control their combined concurrency through the scheduler rather than limiting each job in isolation. For development branches, prefer predictable test responses to validate error handling and parsing. If a real API call is necessary, restrict it to an approved environment. This makes it easier to distinguish an application regression from a shared quota issue or an execution agent connectivity failure, even when the upstream service is down.
Account Restrictions and Rate Limits
First identify what the restriction applies to
“Account banned” can describe several different situations: an account was disabled, an organization removed access, billing status changed, an action requires verification, or an API request was rejected for lack of quota. Read the subject in the product’s message: is it the account, organization, project, credential, or current request? Check official notices and console status rather than drawing conclusions from one page error. If you can access the account page but not team resources, ask an administrator to check your membership. If every member of a project is denied, check the project configuration and billing status rather than asking each member to reset their password.
Temporary verification usually requires completing the official identity flow. If it still fails, save the request ID or appeal link from the notice and contact the official channel using your existing identity. If the account is explicitly disabled, stop automated jobs and wait for guidance; do not continue the same activity with a new or borrowed account. If there is no status notice and the error occurs only for one action, use the same identity to try a routine action you know is authorized and compare the permissions. Repeatedly retrying a rejected action will not produce useful evidence and may make the provider’s investigation harder.
Reduce avoidable risk alerts
Risk checks may consider identity continuity, request patterns, and usage behavior. Shared accounts, using a personal session for background scripts, repeatedly triggering identity checks, or suddenly starting a large number of tasks may make legitimate activity look unusual. Teams should assign access to individual members, use approved APIs and project identities for automation, and keep interactive work separate from batch jobs. Revoke access promptly when someone leaves or a project ends. Do not transfer browser sessions to other devices or processes; use the product’s authentication flow when working from a new client.
If request volume suddenly increases, check whether jobs are being triggered repeatedly, multiple workers share one credential, or both the SDK and its caller are retrying. When a quota is reached, do not move the same workload to another unapproved project. Instead, pause low-priority requests, combine duplicate jobs in one queue, check whether non-sensitive results can be reused, and request a quota change through the provider’s official process. Before resuming, fix the scheduling issue that caused the spike; otherwise, the backlog may fill the queue again as soon as the limit resets.
Prepare appeal and exposure-response records
When contacting a provider, include the product entry point, time and time zone, a redacted excerpt of the exact error, the affected workspace, and any request ID that can be shared. Note which actions still work to help distinguish an account restriction from a feature-specific issue. For an organization account, have an administrator confirm ticket permissions. If the issue involves an account or billing with a third-party AI provider, LiuVPN cannot change its permissions or handle an appeal on your behalf. Record network issues separately so two different support requests do not become one problem that cannot be verified.
If you find unexplained API calls, unexpected charges, or an unfamiliar credential in a project, treat the credential as potentially exposed. Stop using it, revoke and rotate it through the official console, then check task logs, build artifacts, and shared documents for the source of exposure. Deleting a log entry does not revoke a credential that has already been exposed. After remediation, send a minimal request from an approved environment, check monitoring for activity from the old identity, and confirm that normal jobs use the new credential before restoring production traffic.
Choosing and Retesting International Routes
Test for your task instead of judging by region name
LiuVPN covers 120+ countries / 220+ routes. Visit the routes page to check regions and route types, then confirm the available options in your client. Direct, relayed, and IEPL dedicated routes describe how a path is organized; they do not determine whether a third-party AI account is eligible for use. The performance of a direct route still depends on the access network and destination. A relay adds another forwarding step, and a dedicated-route label does not mean every segment between your location and a third-party service uses the same dedicated path. Confirm regional eligibility and organizational network policies before comparing the full task.
Use the same eligible workspace for comparison. If your usual work is text chat, record sign-in, starting a new conversation, the first response segment, the final status, and whether the conversation opens afterward. If you mainly work in an editor, check sign-in, one model call, and whether cancellation works as expected. If files are involved, separately test upload and access with a sample that contains no sensitive information. Change only the connection path and keep everything else constant. After switching routes, do not reuse a request that has not finished; check its status before starting a new test.
Distinguish route behavior, account benefits, and capacity
When reproducing an intermittent issue, note the route name shown in the client, the platform, when the issue occurred, and what you observed. Do not treat one success or failure as representative of long-term performance. If multiple destinations disconnect on the same route, investigate local connectivity and client status. If only one feature in one product fails and shows a clear permissions or billing message, check the product console. If you switch from mobile data to Wi-Fi, then retest in the original environment to avoid confusing a change in access network with a third-party service issue.
There is no device limit on simultaneous connections. When comparing devices, note platform differences and do not assume that sessions and permissions sync automatically across clients. Estimate capacity from your own usage: short text requests, file uploads, image downloads, and background editor tasks use different amounts of data. Do not infer network traffic from model bills or input character counts. For a comparison, track data usage for similar tasks and exclude unrelated traffic such as background updates. Then check the plans page for monthly subscriptions and data package terms.
Document what counts as recovered
A retest record should make the environment, failed action, original signal, change made, and result clear. Include the device, entry point, and organization in the environment. Signals may include a public error code, the stage where it failed, and the time. Describe changes well enough to reproduce them, but never record secrets. To confirm recovery, repeat the original failed action and verify that the result is genuinely usable: an avatar alone does not prove the web app works, the first words do not prove streaming completed, and a submitted API task does not prove it finished. If the issue can no longer be reproduced consistently, mark it for monitoring rather than calling it fixed after a temporary success.
If the client has not passed its basic connectivity checks, follow the first-use checklist to verify subscription import and connectivity, then return to this handbook. If you have traced the issue to the LiuVPN client or a route, submit a redacted report through the user panel support ticket. If the issue concerns a third party’s eligibility, account, or model permissions, contact that provider or your organization’s administrator. Keep the conditions already tested and start the next round with an unverified step; there is no need to change the whole setup again.