×Scott Oldford
Module 18

Fixesevery failure I have seen, and the fix

Find and fix the things that go wrong, most in under five minutes, and know when to hand it to us instead.

Time
20 minutes to read
Bookmark it.
Credits
Most fixes are free
Re running a job costs what the job costs.
The idea, then the doing

Everything here has happened to me, or to someone I helped fix it

Almost every support message I get is one of the problems on this page. The fix is usually under two minutes once you know which one you are looking at.

The method is the same every time

Name what you are seeing. Find the matching card below. Do the fix in the order written, and stop as soon as it works. If the card says ours, send us what the card asks for. The error text, the session link, the output. Send those, not a description. A description costs us a call. The output costs us nothing.

You fixfive minutes, this pageConnectors and loginsWrong account on a cardStalls and long threadsMemory correctionsAutomation schedules and modelsBrowser handoffs in Comet We fixsend the error text and the session linkAn agent errors every runBundle versionsInstall recordsA fault in an agent's filesAsk the agent three questions first:did you read Foundation, which connectors, last error You fixConnectors and loginsWrong account on a cardStalls and long threadsMemory correctionsAutomation schedules and modelsBrowser handoffs in CometFIVE MINUTES, THIS PAGEWe fixAn agent errors every runBundle versionsInstall recordsA fault in an agent's filesAsk the agent three questions first: did youread Foundation, which connectors, last error.SEND THE ERROR TEXT AND THE SESSION LINK
Find your symptom in the table below. The column tells you who fixes it.

One rule above all of them

When something is stuck, do not keep typing into the same session. Every message you add re reads the whole mess above it. Stop, open a new session, attach what you need, and ask again, shorter. Half the problems on this page are solved by that alone.

Step by step

What you do, step by step

  1. Bookmark this page.

    Then read the table once so the symptoms are familiar before you need them.

  2. When something breaks, find the row and do the steps in order.

    Stop when it works.

  3. If an agent is the problem, ask it three questions in its own session before you change anything: did you read the Foundation profile, which connectors did you use, and what was the last error you saw.

    The answers tell you which row below you are in.

  4. If the fix is ours, send the exact error text and the session link, with the agent’s answers to those three questions.

    Send those, not a description.

  5. After any fix, note it in the Project: in the Weekly Review session say, this broke, this fixed it.

    The weekly review reads it and the agents learn from it.

The fix table

Symptom, cause, fix, who

Connector stopped working, or an agent says it cannot reach a tool

You

Likely causeToken expired, a password changed, or the tool revoked the app.

Fix, in orderOpen Connectors, Connected. Find the card. If it shows an error, click Reconnect and log in again. Then run the tool’s test prompt from module 4 in a new session. If it still fails, remove the connector and add it fresh.

Agent reads the wrong account or the wrong data

You

Likely causeThe connected card is using a different login than you think. Usually a personal Google instead of the business one.

Fix, in orderConnectors, Connected, read the email on the card. Remove, reconnect with the right account. Tell the agent: read again, the connector changed.

API key tool says unauthorized

You

Likely causeWrong host, wrong credential type, or the key was rotated.

Fix, in orderSettings, Credential vault. Open the entry. Check host and type against the card in module 4. Create a new key in the tool, paste it, save, test. Keys are write only, so you cannot read the old one back, you replace it.

Task has been thinking for ten minutes with nothing new

You

Likely causeA long session, a huge file, or a tool call that never returned.

Fix, in orderClick Stop. Open a new session. Attach only the file you need. Ask again in one sentence. If it stalls on the same file twice, the file is the problem: split it or export it as plain text.

It did the task but the output is generic, not in my voice

You

Likely causeFoundation is thin on voice, or the agent did not read it.

Fix, in orderAsk the agent: did you read the Foundation profile and the voice file. If no, say: read them and redo. If yes, the voice part of Foundation needs real samples. Add five of your emails to Material and tell Foundation: update voice from the new samples.

It keeps asking me things that are in my files

You

Likely causeThe fact is buried, or the file is not in Sources.

Fix, in orderProject home, Sources. Is the file there. If not, add it. If yes, open the Foundation session and state the fact in one line. It saves it and the agents stop asking.

Memory learned something wrong and keeps repeating it

You

Likely causeA test run, a client’s details, or a figure from an example got saved.

Fix, in orderSettings, Memory, Manage. Find the line, delete it. Or in any session: that is wrong, X is true, update your memory. Confirm in a fresh session.

Automation did not run

You

Likely causeThe schedule is wrong, the model it was created on was retired, or it is paused in the Project.

Fix, in orderAutomations list. Open it. Check next run time and time zone. Click the model button and set it again, a retired model leaves the job sitting. Open the Project, right column, check it is not paused.

Automation ran but I never saw the output

You

Likely causeNotifications are off, or it is waiting for an approval you did not see.

Fix, in orderSettings, Notifications, turn on task completion and approvals for email and push. Open the automation’s thread. If it is waiting, approve or decline.

Automation is costing far more than it should

You

Likely causeThe thread is long and re reads itself every run, or the effort is on Ultra.

Fix, in orderOpen it. Copy the instruction. Delete the automation. Create it again fresh with the model on Standard. Cost drops the next run. Do this every few weeks.

Approval gate did not appear and it sent something

You

Likely causeThat tool is set to Allow in the connector, or the instruction did not include ask me before you send.

Fix, in orderConnectors, open the tool, set every write action to Always ask. Edit the automation to include the words ask me before you send. Check what went out and clean up by hand.

Comet stops on a login or a two factor code

You

Likely causeBy design. It hands credentials back to you.

Fix, in orderLog in or enter the code in the Comet tab, then tell the assistant: continue. If a passkey or hardware key blocks it, use a password login for that tool or do that step yourself.

Desktop app shows Offline in Devices

You

Likely causeThe app is closed, signed out, or a Mac permission was denied.

Fix, in orderOpen the app, check you are signed in. Mac: System Settings, Privacy and Security, confirm Accessibility, Screen Recording and Full Disk Access are on for Perplexity. Restart the app.

Installer says a pack failed its check

You first, then us

Likely causeA corrupted download, an old zip, or a pack that needs a newer bundle.

Fix, in orderDownload the zip again from the link in your welcome email. Confirm the version in the file name is the latest one we sent you. Attach it in a new session and run the install again. If the same pack fails twice, send us the installer output.

Agent is installed but does not answer its name

You

Likely causeThe skill is on the account not the Project, or a different Project.

Fix, in orderProject settings, Skills. Is it listed. If not, Add skill, upload the pack. If yes, start a new session inside this Project and say the name again.

The agent errors every run, in a fresh session, with Foundation read

Us

Likely causeA bug in the agent’s files or a version mismatch with Foundation.

Fix, in orderSend us the error text and the session link. We ship a fixed pack. You upload it through Project settings, Skills, Add skill.

Two agents loaded on one request and argued

You

Likely causeYou described a job that matched two skills.

Fix, in orderStart a new session and say the agent name first, then the job. Names stop the overlap.

Session usage is climbing fast in a normal session

You

Likely causeThe session is long, or Extended context is on with a heavy model.

Fix, in orderNew session for the next job. Keep Extended context on for Foundation only if credits are tight. Move the effort to Standard for anything you are only reading.

Credits ran out mid job

You

Likely causeAuto refill is off, or you hit the monthly spending cap.

Fix, in orderSettings, Usage and credits. Turn on auto refill with a small amount, and check the spending cap, it defaults to $200. Paused jobs pick up on their own once credits are back.

Perplexity itself is down or acting strange everywhere

Nobody, wait

Likely causePlatform, not you.

Fix, in orderCheck the Perplexity status page. Wait. Do not rebuild anything. Everything in your Project is saved.

Questions

Before you ask

Where do I start when something breaks?

The fix table above. Find the line that sounds like your problem, do the fix in order, then test in a fresh session. Nine times out of ten you are done in five minutes.

How do I know if it is me, you, or Perplexity?

Each row says who. You means the fix is a setting or a reconnect on your side. Us means a bug in an agent’s files, and you send us the output. Nobody means the platform is having a bad hour, and you wait.

How do I tell a settings problem from a broken agent?

Ask the agent in its session: did you read the Foundation profile, which connectors did you use, what was the last error. A no, a wrong account or a connector error is you. The same error in a fresh session with Foundation read and the right connectors is us.

What should I send you when I ask for help?

The exact error text, the session link, and one sentence on what you expected. With those three we usually ship a fixed pack the same day. Without them we spend the day asking.

An agent is giving generic output. Is that a bug?

Almost never. It is Foundation being thin on voice, or the agent not reading it. Ask: did you read the Foundation profile and the voice file. The fix table has the full line.

Something sent that I did not approve.

Stop. Check what went out and clean up by hand first. Then Connectors, open the tool, set every write action to Always ask, and add ask me before you send to the automation. Then tell us, so we can see how it slipped.

Credits are climbing and I do not know why.

Settings, Analytics, look at the model split and the skills that ran. Then open your automations and check two things on each: the effort, and how old the thread is. A long thread on Ultra is almost always the answer.

Should I delete and reinstall when in doubt?

For an automation, yes, that is the fix for cost and for stalls. For an agent, no. Reinstalling the same pack changes nothing. Ask it the three questions first.

Is there a status page?

Yes, Perplexity has one. If everything is slow or odd at once, check it before you touch anything. Your Project is saved. Nothing is lost by waiting an hour.

Done when