analyze-skill-issues（分析技能问题）

Analyze Skill Issues

Queries the strdashboarddevveobvk Azure Storage account that stores all integration test results and retrieves error details for a given skill.

Quick Reference

TOOL USAGE — MANDATORY: Use mcp_azure_mcp_storage_blob_get to enumerate and identify the correct blob paths. Then use az storage blob download to download and read the content of those blobs. These tools serve different purposes and are both always required — they are not alternatives.

When to Use

• User asks why a skill's tests are failing
• User asks to debug or investigate a skill's recent test failures
• User wants to see errors from integration test runs for a specific skill
• User asks what errors or exceptions caused skill test failures

Skill Name Mapping

Resolve the user's skill name to its canonical directory name using the Skill Name Mapping.

MCP Tools

MCP tools are used exclusively for blob discovery — finding which blob paths exist. All calls require account: "strdashboarddevveobvk" and container: "integration-reports". Use az storage blob download (see Phase 2) to read blob content.

Output Requirements

Every response MUST include all of the following. Do not omit any item:

• Root cause explanation of why the skill has low confidence or failing tests
• Per-test breakdown (test name, pass/fail, confidence or error)
• Referenced Files section listing every downloaded file as a vscode://file/ link

⚠️ MANDATORY: The Referenced Files section is not optional. Always resolve $env:TEMP to the actual path (run echo $env:TEMP) and include a clickable link for every file downloaded. If this section is missing from your response, the output is incomplete.

Workflow

TOOL RESPONSIBILITIES: mcp_azure_mcp_storage_blob_get is used exclusively for blob discovery (listing available paths). az storage blob download is used exclusively for reading blob content. Both are always required.

Phase 1 — Enumerate Recent Blobs for the Skill

1. Resolve the skill name using the Skill Name Mapping.
2. List all blobs in the container to discover available date/run paths: mcp_azure_mcp_storage_blob_get({account: "strdashboarddevveobvk", container: "integration-reports"}) This returns a flat list of all blob names. Filter for entries that:

- Start with a date string matching yyyy-mm-dd within the last 7 days - Contain /{skill_name}/ in the path - Do not end with token-usage.json or agent-metadata.json

1. Group the matching blobs by date (descending) and by run ID (the second path segment).
2. Limit scope: If the user specified a date in their prompt, use only that date. Otherwise, take the 3 most recent dates from the filtered results and discard the rest. Do not process more than 3 dates.

Phase 2 — Read Test Result Files

For each matching blob path identified in Phase 1, download it to a local temp file and read its content:

az storage blob download --account-name strdashboarddevveobvk --container-name integration-reports `
  --name "<full-blob-path>" --file "$env:TEMP\<filename>" --auth-mode login --no-progress

Use just the last path segment as <filename>.

⚠️ REQUIRED: After each download, immediately append the resolved local path to your working list of referenced files. Run echo $env:TEMP once to resolve the actual temp directory. Every downloaded file must appear in the Referenced Files section of your final response — this is not optional.

Priority order for files to read (most useful first):

1. Any *.json files that are NOT token-usage.json or agent-metadata.json — these contain test pass/fail results and error messages
2. *-SKILL-REPORT.md or *report*.json files — contain skill invocation details and test narratives
3. junit.xml blobs if present — contain structured test failure messages

Phase 3 — Extract Failure Information

For each test result file, extract:

• Test name — the test case identifier
• Status — passed / failed / error / skipped
• Error message — the exception or assertion failure text
• Stack trace — if present in the file
• Date and run ID — for traceability

Look specifically for fields such as:

• "status": "failed" or "passed": false
• "error", "errorMessage", "failureMessage", "message" JSON keys
• <failure> or <error> XML elements in JUnit format
• Any Error: or FAIL prefixed lines in text reports

Phase 4 — Summarize Failures

Present the findings grouped by test name:

## Failures for <skill-name> — <date> (Run: <run_id>)

### <test-name>
- **Status:** failed
- **Error:** <error message>
- **Detail:** <stack trace or additional context, truncated to ~10 lines>

### <test-name-2>
...

At the end of the summary, include a Referenced Files section listing every file that was downloaded as a clickable VS Code link. Resolve $env:TEMP to the actual Windows temp path (e.g. C:\Users\<user>\AppData\Local\Temp) and format each entry as a markdown link using the vscode://file/ URI scheme with forward slashes:

## Referenced Files
- [<filename1>](vscode://file/C:/Users/<user>/AppData/Local/Temp/<filename1>)
- [<filename2>](vscode://file/C:/Users/<user>/AppData/Local/Temp/<filename2>)
- ...

To resolve the actual temp path, run echo $env:TEMP in the terminal before constructing the links.

Include:

• Total failing tests vs. total tests found
• Most recent run date analyzed
• Common error patterns across multiple tests (e.g., all failing due to auth, quota, timeout)

Date iteration rule: Process the most recent date first. Only proceed to the next date if the current date yields zero failures. Stop as soon as failures are found — do not process all 3 dates if the first one has results.

See Blob Path Layout for the full container tree structure.

Error Handling