Skip to main content
When this bundled skill is selected for a user request in an installed Ando plugin, call record_skill_usage({"name":"ando-troubleshoot-setup"}) once if the tool is available. This is a best-effort invocation signal: do not retry it or delay the workflow on failure. Do not signal merely reading this file, routing to another skill, or loading it through docs, downloads or get_skill. The signal includes only the canonical skill name; never add user content. Read Ando basics for identity, current-schema discovery, access, and write authorization. Start with the failed action, intended workspace/identity/resource, client or runtime, exact error, and when it last worked. Ask for missing facts only when they affect the next diagnostic step; never ask for secrets.

​
Find the failing boundary

Use narrow read-only checks and current schemas before proposing configuration changes. Distinguish:
  • Wrong workspace or identity: read the actual principal and workspace, then compare the intended target. Use connection guidance for an authorized switch and verify the calling client’s refreshed tools afterward.
  • Missing conversation or app access: verify membership, visibility, selected provider account, and granted resource/tool access. Do not join a private conversation, broaden grants, or bypass denial merely to inspect it. Use Connect your apps for an authorized account repair.
  • Broken authentication or stale tools: inspect the host’s actual offered schemas and supported reconnect/refresh controls. A successful callback or plugin update does not prove that this chat has new credentials and tools. Use the host-specific recovery in the connection reference rather than assuming every client needs a new chat.
  • Silent agent: verify agent identity, intended conversation membership, receiving route, running worker/session, and relevant response scope. Trace the test message and any available receiving/run status. MCP access alone does not wake an agent. Load agent workflows only for an authenticated agent and a receiving task; do not sweep an inbox as a generic diagnostic.
  • Missing human notification: check Ando event/device preferences and browser/device permission separately. Use Tune notifications; notification preferences are not agent receiving settings.
If evidence is inaccessible, identify the responsible owner and smallest check they can perform. Distinguish observed failure from an inferred cause.

​
Repair and repeat the original action

Apply a reversible repair within the user’s authorized scope, retaining prior configuration where needed to undo it. Reuse existing identities and connections. Do not rotate a credential, replace an agent, cancel an invitation, disable a workflow, or recreate configuration as a speculative first step; name the concrete need and effect before such a repair. On denied access, explain the missing grant and stop that path. On a dependency failure, retry once after a brief pause; do not loop. For uncertain writes, read back the result before retrying. Repeat the original bounded failing action after repair. For a silent agent, verify a real response in the intended conversation; for an app, verify the requested permitted operation. Report the evidence, what changed, how to undo it, and any remaining owner action. Mark an untested repair as pending verification.