Common problems
What to do when sign-in, keys, repositories, investigations, chatbots, notifications or the log watcher don't work as expected.
Find your problem below. Most fixes need an owner or admin, because they change settings.
Signing in#
"This Google account has not been invited"#
- Ask an owner or admin to invite your email under Your account › Team. You get an invitation email from Cerberius Agents. Check your spam folder, or ask them to click Resend invitation in your ⋯ menu there.
- Check the email they added is exactly your Google account's address.
- If you have several Google accounts, make sure you picked the right one on Google's screen.
"Your Google email address is not verified"#
Verify the email address in your Google account settings, then sign in again.
I'm signed in but see "No platforms yet"#
- As an approver or viewer: you haven't been given any of your organisation's platforms. Ask an owner or admin to switch one on for you under Your account › Team › ⋯ › Platforms and access….
- As an owner or admin: your organisation has no platform yet. Click Add platform on the Overview.
- If you were removed from the organisation, you'll need to be invited again.
The agent's key#
The Queue says "Finish setting up"#
Owners and admins see this amber card until the platform has what the agent needs. The steps are in the order they can be done. Click each step that isn't ticked:
- Choose the AI key for the agent opens Settings › Agent.
- Connect something for the agent to look at opens the Get started page. Any one is enough: code, a cluster and its logs, a database, Microsoft 365 or Google Workspace.
- Choose a key for your repositories shows only when repositories were added without a key. It opens Settings › Code.
- Connect a ticket system is optional. It opens Settings › Ticket system.
Step by step opens the Get started guide. Not now hides the card on this browser.
"Choose an AI key"#
Settings shows this label when the agent has no Claude key.
- 1
Add a Claude key under Organisation › Keys. See Keys.
- 2
Go to Settings › Agent. In the AI key row, choose the key under Claude key.
- 3
Click Save.
If you see The key chosen before is no longer in the library, someone deleted it. Choose another one and save.
"That doesn't look like a … key"#
Claude keys must start with sk-ant-. OpenAI keys must start with sk-. Check you copied the whole key, with no spaces, and picked the right Kind of key.
Repositories#
Under Settings › Code, the Status column shows whether each repository's key can read it. To check again, open the repository's ⋯ menu and click Check access. A repository that can't be read shows Can't read, with the reason under it:
| Message | What to do |
|---|---|
| Readable | All good. The default branch is filled in for you. |
| No key | Open the repository's ⋯ menu, click Edit…, and choose a key. |
| Not found. Check the name, and that the token has access to this repo. | Check the name matches GitHub exactly, such as alion/booking-api. Check the key's Repository access includes this repository. If your GitHub organisation approves tokens, ask an owner to approve it. |
| Not found. If the repo is private, add a GitHub token first. | Pick a GitHub key for this repository. |
| GitHub rejected the token. It may be expired or mistyped. | Make a new token and use Rename or replace… on the key under Organisation › Keys. |
| GitHub refused access: the token lacks permission for this repo. | Give the token Contents: Read-only (or more) for this repository. |
| GitHub's rate limit was hit. | Wait a while and try again. |
| Choose a GitLab key for this repository first. | Pick a GitLab key for the repository. |
| Could not reach GitHub / GitLab | Usually temporary. For your own GitLab server, check the GitLab server address on the key. |
For a GitLab key, a failed check often means a missing scope. Use read_repository and read_api to read, or api for merge requests.
An investigation failed#
Failed tickets show in the Queue's Waiting for you table with an Investigation failed label, and the Overview panel counts them under Failed. Open the ticket. At the top, Now says The agent couldn't finish, and explains why in plain words. Click Technical details to see the full error. The same reason shows on the agent's line in What happened, marked couldn't finish.
Nothing is sent to the requester when an investigation fails. When you've fixed the cause, click Investigate again.
| What it says | What to do |
|---|---|
| The agent stopped because this ticket reached its cost limit. | Click Remove the limit for this ticket to carry on, or raise Most per ticket (USD) under Settings › Limits and investigate again. |
| The agent ran out of time before it finished. | Try again. If it keeps happening, raise Most minutes per run. |
| The agent stopped because it reached one of its limits | Check Settings › Limits. |
| The agent could not download your code. | Check each repository's key under Settings › Code, and use Check access in its ⋯ menu. |
| The AI service refused the request. | Check the agent's Claude key under Settings › Agent, and that your Claude account has credit. |
| The agent could not be started. | Usually temporary. Try again in a few minutes. |
| The agent finished, but its answer was not in the expected shape. | Investigating again usually fixes this. |
| The agent did not change any files, so there was nothing to put in a pull request. | The approved fix produced no changes. Tell the agent what it should change, in Write › To the agent. |
| The agent stopped because of an error on our side. | Try again. If it keeps happening, send us the Technical details. See Getting help. |
A ticket says "Held back"#
The log watcher's daily limit was reached, so this ticket wasn't investigated yet. On the ticket, Now says Start the investigation, with "Held back: the log watcher's daily limit was reached." The Queue lists it as Waiting to start. Open it and click Investigate when you're ready, or raise the daily limit under Settings › Log watcher.
Replies#
Approving says "Nothing is sent"#
The platform has no ticket system connected. On these platforms the draft is titled The agent's conclusion, and approving just records that you agree and marks the ticket done. To send replies to requesters, connect a ticket system under Settings › Ticket system. See Ticket system.
"The reply to … didn't arrive"#
Your ticket system refused the reply, or didn't answer. At the top of the ticket, Now says The reply to Eugene didn't arrive, with the error under it. In What happened, the ticket system's line says it didn't take the reply, in red. Fix the problem on your ticket system's side, then click Try again.
On the reply's own page (opened from the Overview), the same problem shows as The ticket system did not accept the reply. and the reply's status is Not delivered. While the service is still retrying on its own, that page shows Retrying (attempt N) with the last error.
Chatbots#
"No AI key for chatbots yet"#
Chatbots need a Claude API key or an OpenAI key. Add one under Organisation › Keys, then choose it on the chatbot's Install tab. On the chatbot, the warning reads No Claude API key yet or No OpenAI API key yet, for the model you chose.
"Chatbots need a Claude API key (it starts with sk-ant-api)"#
You've chosen a Claude subscription token (starting sk-ant-oat). It can run the agent but not chatbots. Create a Claude API key and choose that for the chatbot instead.
The chatbot preview says the key was refused, or is rate-limiting#
- The … API key was refused: the key is wrong, revoked or expired. Replace it under Organisation › Keys.
- … is rate-limiting this API key right now (or its credit has run out): add credit to your account with that provider, or wait a little.
- … is overloaded or unavailable right now: the provider is having trouble. Try again shortly.
Visitors never see these details. They see a short apology instead. See Chatbots overview.
Notifications#
Send test fails#
| Message | What to do |
|---|---|
| Telegram said: chat not found. Send any message to your bot first, then try again. | Open your bot in Telegram and send it a message, or write in the group it's in. Check the chat id. |
| Telegram did not accept the bot token. | Check the token with @BotFather and paste it again with Rename or replace… in the channel's ⋯ menu. |
| Find my chat says No chats yet | The bot hasn't seen a message yet. Send it one, then click Find my chat again. |
| This bot sends its messages to a webhook, so its chats can't be listed. | Type the chat id yourself, or make a new bot just for Cerberius Agents. |
| Slack did not accept the webhook address. | The webhook was probably removed in Slack. Make a new one and paste it with Rename or replace…. |
| That isn't a Slack webhook address. / A Slack webhook address starts with https://hooks.slack.com/services/ | Copy the Webhook URL from your Slack app's Incoming Webhooks page. |
The test works, but real notifications don't arrive#
A channel only gets messages from platforms that switch it on. Go to each platform's Settings › Notifications, switch on what the channel should hear about, and click Save. A channel labelled Hears nothing has everything switched off on that platform. See Platform notifications.
Log watcher#
"Your logs are not connected"#
The log watcher reads your logs from Loki. You'll see this warning at the top of the Log watcher page when:
- Loki isn't connected yet. Connect it under Settings › Live system: click Connect a cluster or logs (or Connect another) and choose Grafana Loki. See Loki logs.
- Loki is connected, but the platform has no Kubernetes namespaces listed and no custom search. Add your namespaces under Settings › Live system › Kubernetes, in Namespaces it may look at, or write your own search on the Log watcher page.
It finds nothing#
- Look at the top of the Log watcher page. If it says Not checked yet, the first check runs within 5 minutes of switching it on. Otherwise Last checked shows when it last looked.
- If the label at the top says Can't read the logs, read the The last check failed message under it. It often points to the Grafana Loki connection under Settings › Live system.
- Check that your namespaces are the ones your apps log under, and that your logs mark errors (for example with the word "error" or a stack trace).
- If a ticket for an error is already open, the same error only adds to its count. It doesn't open a new ticket.
See Log watcher.
Still stuck?#
See Getting help.