Troubleshooting

Each entry starts from what you see. Messages the app shows while you set up a project, edit a plan or start a run are listed with their causes on the pages for those screens: From a URL to a plan, Editing the plan, Test accounts, Starting a run and Your model key.

“That address is on a private network”

Setup in the app reads only public pages, and refuses any address that resolves to a private, loopback or reserved range — including through a redirect. To test localhost, a VPN or a private staging server, use the local runner, which reads and runs from your own machine.

The proposed people and goals are generic

Setup reads the HTML your server sends for that one page, without running its JavaScript. A single-page app that fills itself in after loading gives setup little text to work from. Rewrite the people and goals on the plan page — they are yours to edit — or paste the address of a page whose content is in the HTML.

“Hosted runs are in private beta”

Your email address is not on the private beta list for hosted runs. Signing in, setting up projects and editing plans work regardless. Write to contact@usetrawler.com to get access, or use the local runner.

Start run is disabled

The Start panel says why next to the button:

• Save or discard your changes to the plan first. — the plan has unsaved changes. Use Save plan or Discard in the bar at the bottom.
• Ask an owner of this workspace to add a model key. — the workspace has no key, and only an owner or admin can add one.

The people stop at the sign-in page

• No test account. A person without one is told to sign up if the product allows it, but cannot fill in a password field, so a sign-up form that asks for a password stops them. Add a test account and pick it under Signs in as.
• Username and password on separate screens. Trawler types both into the sign-in form in one go, so a sign-in that asks for the password on a second screen may stop the person.
• Sign-in on another domain. If your product sends people to a sign-in page on a different origin — a separate identity provider, say — the browser blocks it. The app does not let you add allowed origins yet; the local runner’s project file does.
• Staging behind basic auth or a secret header. The app cannot pass either yet. The local runner can.

The run has said “Waiting for a runner.” for a while

Runs wait until one of Trawler’s runners is free, and each run is worked one session at a time. The run starts on its own; you can close the tab.

A person could not finish

The card says Could not finish and gives the reason in red:

Reason	What happened
the runner stopped answering	The runner carrying out the session went silent for 10 minutes.
the job could not be prepared from the project; it may have changed since the run started	The project changed in a way the run could not start the session from.
the model stopped calling tools 3 turns in a row	The model kept answering in plain text instead of acting. Try another model.
the model’s replies were cut off 3 times in a row	The model’s answers kept running over its length limit. Try another model.
the browser failed 3 times in a row; last error: …	Three browser actions in a row failed; the last error follows.

The other people’s sessions go ahead regardless. A run where every session failed ends as Failed.

A defect could not be judged

The judge gave no verdict. Open the defect to see Why it was not judged: — for example Model error: the model gave no verdict (2 tries), or The run’s cap ran out before the judge answered. The replay is kept: once the run has ended, Judge again asks the judge once more, paid from what is left of the cap. See Judging again.

The run stopped at the cap

The run spent its cap before everything was done: whatever was still waiting was cancelled, so some people show Stopped, some defects are Not judged — The run ended before it was replayed. or …before it was judged., and The fresh agent could not follow the steps far enough to report. for a replay the cap cut off — and a defect whose judge was cut off is under Could not be judged. Start another run with a higher cap, fewer people, or a less expensive model — the estimate shows what a cap is likely to cover.

People finish early, with open goals and money left under the cap

If, during a run, the provider refuses the key (it was revoked, say), reports the account out of credit with a payment-required answer — as OpenRouter does — or the workspace key is replaced with one from another provider, each remaining session ends as though the cap had been reached, and the app does not yet say why. Check the key and your provider’s billing, then start another run.

A provider that reports an empty account some other way — OpenAI, for one, answers as if you were sending too many requests — shows up instead as Could not finish on the people’s cards, with the provider’s message. If the key stops working only once the replays have begun, the people’s cards are unaffected: a defect whose replay it cut off shows The fresh agent could not follow the steps far enough to report., and one whose judge it cut off goes under Could not be judged.

The price says “unknown”

The Start panel has no price for the name in the model field — Trawler knows none for that model, the models could not be listed, or the name is not in the list. Trawler looks again when the run starts: if it finds a price, the run is held to the default $2 cap; if not, it stops after 3 million tokens, and only through OpenRouter, which reports each call’s cost, is it also held to the $2. Watch your provider’s billing, or pick a model the panel shows a price for. See When the price is unknown.

“Lost contact with Trawler. Retrying…”

The page could not refresh. The run itself carries on; the page catches up when the connection returns.

I cannot find an earlier run

The app has no list of projects or runs yet. A run’s page stays at its address — keep it.