Report a problem to support

Goal

Turn a failure into a structured, complete diagnostic report and send it to support. For any user who has hit a problem the session cannot resolve. You end with a reviewed Markdown report that a support engineer can act on without reproducing the session, and an email to help@novocircle.com that carries it.

Where it runs

This scenario needs the MCP server. Computer use covers its UI steps.

Before you start

• A problem. ea-diagnostic is for failures, not for general questions.
• For richer detail: the ability to restart the server with diagnostics switched on (the EA_MCP_DIAGNOSTICS setting), and to reproduce the failure once more. How you set it depends on how the server was installed. See Deployment packages.
• Nothing in the model is changed by this scenario.

A failure is only useful to support if it is specific. The skill collects the system context, writes one entry per problem in a fixed format, ranks them, and fills in the reproduction steps and the submission instructions. It does not send anything. You read the report, remove anything you do not want to share, and send it.

Two things have similar names, and the scenario uses both. The diagnostic report is the document the skill writes, and it exists to be emailed. Server diagnostics is an opt-in server mode that writes a Markdown file for each failing or slow call, with the arguments, the error trace, the EA build and the last 25 calls, which the AI cannot always see.

Steps

1. AI ea-start-here recognizes that it is time to report. The triggers: an operation that fails twice with the same error after you corrected the input; EA running but behaving inconsistently (an element that exists is not found, a tagged value does not stick, a diagram does not show what the model holds); a technology that passes every check yet whose stereotypes are unavailable; or two operations that contradict each other. Two failed attempts at the same thing is the signal to report instead of guessing a third time.
2. AI Power Tools ea-start-here rules out the three common causes. EA may be busy and not hung: (Not Responding) in the title bar means it is processing, so wait, and do not repeat the command or open a second instance. The technology may not be loaded (get_mdg_from_runtime). The write may have gone to a different model (get_repository_info). Looking at EA's title bar and dialogs is a UI step: AI with computer use; otherwise you.
3. You Optional: switch on server diagnostics, restart the server, and repeat the failure once.
4. AI Power Tools ea-diagnostic collects the system context: the repository information, the skills list, and the language state. A read that fails is recorded as UNAVAILABLE with its exact error, so a partial report is still produced. Offline, the server-version field can read UNAVAILABLE, because the skills list is read from GitHub.
5. AI ea-diagnostic writes one issue entry per distinct problem: a severity (Critical, High, Medium or Low), the affected component named exactly (create_connector, not "the server"), the observed behavior with the error text verbatim, the root cause (stated plainly if confirmed, prefixed "Suspected:" if not, or "Under investigation"), the workaround applied, and a recommended fix that names the file, tool or skill.
6. AI ea-diagnostic ranks the issues and writes the summary, the reproduction steps and the submission section.
7. You Review and redact the report. It holds the project file path, error messages, and element names and ids. Server diagnostic files also hold call arguments, with long strings truncated.
8. You Send the report yourself. The subject line is [Diagnostic] <short summary> — EA <version> — <YYYY-MM-DD>. Attach the report. Attach the model file only if the problem is model-specific and you are comfortable sharing it, plus any MDG XML files involved and the server log if you captured one.

With server diagnostics on, each failing or slow call writes a file named issue_<UTC-timestamp>_<tool>.md to %LOCALAPPDATA%\ea-mcp-server\diagnostics\. A setting changes the folder, and another sets how many seconds a call may take before it is flagged (the default is 30). When diagnostics is on and a call fails, the response carries the path of the file and the support address.

Prompts to use

Opening

Follow-ups

Adds a second numbered issue.

The review in step 7.

Recommended models

Ratings are guidance based on the kind of reasoning the skill needs, not measured benchmarks.

Checking the result

• Accurate. Observed behavior carries the exact error text, not a paraphrase. Anything not confirmed is labeled "Suspected:" or "Under investigation". Someone else can repeat the reproduction steps, which name the operations with their inputs.
• Complete. Every header field is filled or marked UNAVAILABLE with its error. Every distinct problem has its own numbered entry. The summary table ranks by severity. The submission section carries a subject line with the real summary, EA version and date.
• Nothing unintended changed. No model change is made. Check that you sent the report yourself and attached only what you chose to.

Variations

• Server report only. When a call fails with diagnostics already on, the file in the diagnostics folder is itself an attachment. You do not need the skill if the failure is a single call.
• A skill that gave wrong guidance. Name the skill and the section as the affected component, and give the recommended text change.
• End of session. Ask for the report before ending a long session that hit several issues, so none is lost with the conversation.

Limitations

• Nothing is sent automatically. AI Power Tools has no telemetry. A report reaches support only when you send it.
• Diagnostics is off by default. With it off, no files are written and a failing call leaves no record on disk. If the first failure was not captured, repeat it with the mode on.
• The AI can write a plausible root cause that is wrong. Distinguish "confirmed" from "Suspected:" before you send. Support relies on that distinction.
• The report is only as good as its collection. A read that fails leaves a gap. The skill records the gap instead of omitting it.