During and after a run

Failed or interrupted runs, rate limits, low scores, empty previews and exports.

Problems while a run is going, or once it's finished.

The run failed ("Run failed" / "Run error")#

A provider returned an error the swarm couldn't retry, so the run stopped. The phase it stopped in is marked on the phase track, and a Run error banner shows the provider's message. Tap it to open the Run Log, and filter by Errors.

401, 403, unauthorized, invalid key

Likely cause: The key was revoked or changed.
Fix: Paste a fresh key in Settings and tap Validate.

402, billing, balance, credit, quota, insufficient

Likely cause: The account is out of credit or over a spending limit.
Fix: Top up, or raise the limit, in the provider's console.

404, model not found, not available on your plan

Likely cause: Your account can't use that model.
Fix: Pick another model for that role.

400, context length, too many tokens

Likely cause: The prompt was too large for the model.
Fix: Use a model with a larger context window, or reduce the project's scope.

The message mentions	Likely cause	Fix
401, 403, unauthorized, invalid key	The key was revoked or changed.	Paste a fresh key in Settings and tap Validate.
402, billing, balance, credit, quota, insufficient	The account is out of credit or over a spending limit.	Top up, or raise the limit, in the provider's console.
404, model not found, not available on your plan	Your account can't use that model.	Pick another model for that role.
400, context length, too many tokens	The prompt was too large for the model.	Use a model with a larger context window, or reduce the project's scope.

Once it's fixed, you don't have to start over. Open the run and tap Run next iteration to continue from where it stopped. See Continue a finished run.

The run seems stuck#

Check the active worker rows. Generating response… N KB received rising means the model is still writing. Large files and reasoning models can take several minutes.
Tap the active phase node to see its live elapsed time and calls.
Providers sometimes answer "rate limited" (429) or "overloaded" (503). The swarm waits and retries these automatically, which can look like a pause. Retries show in the Leaderboard detail and in llm-calls.csv.
A connection that drops mid-answer, for example when switching from Wi-Fi to mobile data, is retried too.
A call that gets no response at all times out after 10 minutes and is retried once.

If nothing has moved for a long time, you can Cancel and then Run next iteration to pick it up again.

Lots of rate-limit retries#

Each developer and QA reviewer calls the provider at the same time, which can exceed a low rate limit, especially on free or new accounts.

Lower Developer agents and QA agents.
Spread roles across more than one provider.
Ask your provider for a higher usage tier.

The run was cancelled when I left the app#

Interrupted — the app closed while the run was active means the app was closed or killed while a run was going. This can happen if you swipe it away, the phone restarts, or the OS reclaims memory.

Android: the run continues in the background with an ongoing notification. Battery-saver or "app sleeping" settings on some phones can still kill it. Exclude rdn-swarm from battery optimization if this happens.
iPhone: iOS suspends background apps after a short grace period. Keep rdn-swarm in the foreground for long runs.

Either way, Run next iteration picks the run back up.

A QA review was "unreadable, not scored"#

The reviewer's reply was cut off at its output-token limit before it said anything usable, so it wasn't counted. That's safer than counting it as a clean review. Raise the QA role's Max output tokens, or use a model with a larger output limit. Output-cap hits on the Leaderboard detail shows how often this happens.

The score won't go up#

Read the issues. The Issues Aggregated list and the scorecard show which dimension is dragging and why. A single critical issue halves its dimension.
Check the acceptance criteria. Vague or contradictory criteria can't be met. Tighten them with a note, or reopen with a clearer brief.
Use a note. Tell the swarm exactly what's wrong: "Clicking a filled cell still changes it."
Try a stronger model on the Developer or Project Manager role.
Check the weights. A dimension that doesn't matter for your project, often Accessibility on a quick prototype, can be down-weighted under Scorecard.

"Run finished · not Successful · N critical"#

The Score reached your threshold, but critical issues are still open. Successful needs both. Reopen the run with a note that points at the critical issues.

The preview is blank or says "Nothing to preview"#

Nothing to preview — this run produced no HTML entry file: the build has no HTML page to open. Make sure the brief asks for a single-page HTML app.
Couldn't prepare the preview on this device: the files couldn't be prepared for the preview. Free up some storage and tap ↻ Reload.
The page loads but is blank or unresponsive: the generated JavaScript probably has an error. Add a note describing what you see, or continue the run. QA in Visual or Interactive mode catches this kind of problem automatically.

The zip export failed#

The export is refused if the phone doesn't have room for the archive, rather than writing a broken zip. Free up storage and try again. Runs with many screenshots are the largest.

I can't find a run#

History lists every run. Check that the filter is on All.
Deleted runs can't be recovered, and runs aren't backed up to the cloud. A reset or new phone starts empty. Export runs you want to keep.