Troubleshooting Swarmbot automation: find the failed step before retrying

An online device is not proof that a workflow is ready, and a timeout does not prove nothing happened. Separate connection failures from workflow failures, preserve task evidence and prepare a useful support report.

简体中文

When an automation fails, resist repeated retries. Preserve the first error, identify the last confirmed successful step, and distinguish connection trouble from a workflow problem. For sending, submitting, installing or changing settings, check the result before repeating the action.

This is a troubleshooting guide, not a root-cause report for an observed incident. Relevant implementation and a read-only sample of app resources were reviewed on September 9, 2026. We did not operate devices, execute scripts or reproduce a failure. Diagnostic fields can differ by client version.

Connection failure or workflow failure?

“No response” can describe several situations:

Symptom Evidence to retain Question to answer
Device missing or no image Connection type, time and diagnostic report Is the device or image channel failing?
Image available but a step stalls Current screen, last successful step and first error Has the starting state or target changed?
Request timed out but the action appears complete Original task identifier, execution state and business result Did the response fail rather than the action?

A working stream does not prove that a script can complete. A failed script does not prove that a phone disconnected. A connection check covers the checks it actually performs at that moment, not the entire business task.

What the device diagnostic report tells you

In the device wall's context menu, use the diagnostics option under debugging and logs to inspect connection checks. After it finishes, copy the report rather than keeping only a screenshot of an error badge.

The current report implementation includes timestamps, individual checks and durations, and device/connection context. Runtime or startup details are included when the backend has them. An empty log section does not prove that an action never occurred. If diagnostics pass while the workflow fails, inspect its required screen and execution evidence next.

The copy function redacts common tokens and passwords, but that does not make the report safe for public posting. Identifiers, addresses, paths and business text may remain sensitive. Review it before sending it to trusted support; post only necessary sanitized excerpts publicly.

Check the starting state

A workflow may assume an existing login, a specific screen, no obstructing dialog and an available supporting script. The app skill sample inspected for this guide contained script dependency links. A dependency marked ready describes a resource state; it is not evidence that the phone satisfies the workflow's prerequisites.

Record four things:

  • Expected start: where should the workflow begin?
  • Actual start: what was on screen?
  • Last confirmed action: which step has observable evidence of success?
  • First divergence: what target or state change was missing?

A fixed three-second pause is not automatically safer than waiting for a meaningful condition with a limit. Selenium's official waiting guide explains timing races and the shortcomings of fixed sleeps. The transferable idea is condition-based waiting—not using Selenium APIs as Swarmbot script APIs. Selenium waiting strategies

Use the current Swarmbot script documentation for available checks. If no reliable condition is available, add a human checkpoint instead of an unlimited wait.

Read the task result, not just the AI's final message

When an interface returns an asynchronous task identifier, retain the original taskId and query its state and result using the client documentation. The current script-task query implementation distinguishes state, completion, errors, results and logs, and indicates truncated logs. A short log tail is not the entire execution history.

If an external AI or business system is calling Swarmbot, supply current documentation and explicitly ask it to inspect the existing task without submitting a new one. See the Chinese guide to external AI integration.

Be particularly careful when an uncertain action has side effects. AWS's engineering guidance explains that timeouts do not establish that nothing happened, while retries can add pressure to an overloaded service. This is a troubleshooting principle, not a claim that every Swarmbot operation provides idempotency. Timeouts and retries

A useful support report

Use this illustrative template rather than “it failed again”:

Client version: the version actually in use.
Time: date, time and timezone.
Scope: one or several devices, and connection type.
Goal: the intended result.
Starting state: screen, permissions and login prerequisites.
Last successful step: include the evidence.
First error: retain its wording after sanitizing sensitive content.
Actual outcome: did a submission, message or settings change occur?
Attachments: diagnostics, relevant task records and necessary screenshots.
Attempts so far: what changed and whether the symptom changed.

This lets support locate a layer, step and time before investigating a cause. If reproduction is needed, use one authorized device and one low-risk step, changing one condition at a time. For turning that validation into a reusable script, see our Chinese guide to small-step script development.

参考来源