ea-diagnostic

Logic: AI + AI Power Tools checks. AI Power Tools collects the facts; the AI writes and ranks the findings.

What it is for

Use it to get support when something has failed and the session cannot resolve it. You end with a reviewed Markdown report that a support engineer can act on, ready for you to email to help@novocircle.com.

ea-diagnostic produces a structured report that a support engineer can act on without reproducing your session. The AI collects system context, records every distinct problem from the session in a fixed format, ranks them, and ends with instructions for sending the report to help@novocircle.com.

It is a write-up skill. It does not fix anything and it does not send anything. The AI writes the report; you review it and email it.

It is separate from the server's automatic diagnostics (Diagnostics). That module writes a machine-generated issue file when EA_MCP_DIAGNOSTICS is on and an operation fails. This skill writes the narrative: what was observed, what was tried, what is suspected, what should change. A support case is strongest with both.

Capabilities it uses

When to use it, and when not

Use it when:

• you ask for help with an issue,
• an unexpected error cannot be resolved in the session,
• a workaround was applied but the underlying defect should still be reported, or
• before ending a session in which several issues were met.

ea-start-here escalates here after two failed attempts at the same thing, and when EA behaves inconsistently.

How it works

The steps run in order. The AI does not skip the automatic collection even if some calls fail, because a failed call is itself diagnostic data.

1. Collect system context. AI Power Tools The AI calls three read-only operations and records their output verbatim: get_repository_info, list_available_skills and assess_mdg_situation. An error is recorded with its exact message. Nothing is omitted. The header carries the session date, EA version, project file, server version, loaded technologies, whether diagnostics is active, and who prepared the report (the AI model, on behalf of the end user). A field whose call failed is written as UNAVAILABLE with the error message returned.
2. Write the report header. AI A fixed template with those fields.
3. Write one issue block per distinct problem. AI Numbered from 1. Each block has:
   • Severity and affected component, which must be specific: the exact operation name, a skill and section, or an EA interface element, never "the server".
   • Observed behavior: factual only, with exact error messages, returned values or interface behavior. No speculation. Dialog text and what EA's window shows are read directly. AI with computer use; otherwise you Without computer use, you paste a screenshot of the dialog into the conversation.
   • Root cause, in one of three forms: a confirmed cause stated plainly with how it was established; a suspected cause prefixed Suspected: with its reasoning; or Under investigation — insufficient data to determine root cause.
   • Workaround applied: the exact calls or numbered steps used, or None found.
   • Recommended fix: the specific file, operation, skill or EA setting and the exact change, not "fix the bug".
4. Apply the severity definitions. AI Critical blocks the session with no workaround and possible data loss. High significantly impairs progress, and the workaround is complex or fragile. Medium causes friction, and the workaround is straightforward. Low is minor or cosmetic.
5. Add a summary table. AI Ranked by severity, Critical first, then by issue number. Columns: number, severity, short title, affected component, whether a workaround was found, and where the fix belongs.
6. Add the reproduction and environment footer. AI Numbered reproduction steps with the calls and inputs, the model state when it happened (packages, elements, diagrams, technologies, with identifiers where known), relevant log output, and a transcript note. If server diagnostics was off, the footer tells you to turn it on before reproducing, to capture a structured log for the next report.
7. Add the submission section, verbatim. AI The email address, a subject format [Diagnostic] <short issue summary> — EA <version> — <date>, and what to attach: the report file, the EA project file if the issue is model-specific and you are comfortable sharing it, any MDG Technology files involved, and the server log if available.
8. Review and send. You You read the report, remove anything you do not want to share, and email it.

A partial report is better than none, so a failed collection call is recorded and the report continues. Error messages are pasted verbatim and never redacted.

What you get

• A Markdown report: header, one block per distinct problem, a ranked summary table, a reproduction footer and a ready-to-use submission section.
• Nothing is sent and nothing in the model changes.

Tools it relies on

• get_repository_info (Project): EA version, project path, repository type, and the diagnostics setting.
• list_available_skills (Skills (legacy)): the server version and the skills in the release bundle. It reads the release information over the network, so it needs network access.
• assess_mdg_situation (Discovering technologies): loaded technologies and how much of the model conforms.
• The diagnostics module (Diagnostics): the automatic report file the skill refers to.
• Shared reference: ea-ui-verification.md, for looking at EA before reporting a hang.

Sample prompts

1. "Run the diagnostic skill. The connector creation failed twice." One issue block, three collected calls, the submission section.
2. "Write this up for support before we close the session." A multi-issue report with a summary table.
3. "I worked around the missing tagged value with a direct database update. Report it so it gets fixed." The report records the exact update as the workaround and names the operation that should have worked.
4. "EA froze during the XMI import and I had to dismiss a dialog. Include that." The AI records the dialog text as the observed behavior and notes whether the cause is a blocked dialog or a defect.
5. "Turn on diagnostics, reproduce the problem, then write the report." The AI explains that EA_MCP_DIAGNOSTICS=1 is set in the server's environment or the extension's settings and needs a restart, reproduces the failure, and collects both the automatic file and the narrative.

Recommended models

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

Checking accuracy and completeness

1. Header against the server. Re-run get_repository_info and compare each header field. Correct: every value is the server's output or UNAVAILABLE with the error. Red flag: a build number or file path that no call returned.
2. Observed versus root cause. Read each "Observed behavior". Correct: only what was seen, namely messages, values and interface behavior, and each root cause in one of the three allowed forms. Red flag: an unqualified cause with no stated evidence.
3. Reproduction replay. Follow the numbered steps on a scratch model, never on a customer repository. Correct: the failure reproduces. Red flag: a step that cannot be run as written, or a missing input, which means the report is incomplete.
4. Completeness of issues. Count the distinct failures in the transcript and compare with the issue blocks. Complete: every failed call or surprising result that affected progress maps to an issue or is explained away, and the summary table has one row per block in severity order.
5. Cross-check with the automatic report. If diagnostics was on, the issue file's operation name, outcome and detail must match the issue it describes. Red flag: a mismatch, which suggests the AI blamed the wrong call.
6. Format. Correct: sequential numbers with none skipped or reused, only the four severities, specific affected components, and a submission section with the subject line filled in.
7. Sensitive content. Because error messages are pasted verbatim, the report can name packages and elements. Red flag: credentials, tokens or customer names in the text. You read it before it goes anywhere.

Limitations

• The report contains what the session contained. Element names, package paths and free text from the repository can appear. You must review it before sending. The server's automatic report truncates long strings; this skill's narrative does not.
• list_available_skills lists the skills in the release bundle, with a compatibility flag for each, and the server version. It is not a list of what is loaded, and it fails without network access.
• Some header fields can come back UNAVAILABLE where an operation does not return the field the template names. That is a gap in the template, not an EA problem.
• It does not send mail.
• It writes nothing to the model.