Analyze Test Run

Downloads artifacts from a GitHub Actions integration test run, generates a summarized skill invocation report, and files GitHub issues for each test failure with root-cause analysis.

When to Use

• Summarize results of a GitHub Actions integration test run
• Calculate skill invocation rates for the skill under test
• For azure-deploy tests: track the full deployment chain (azure-prepare → azure-validate → azure-deploy)
• Compare skill invocation across two runs
• File issues for test failures with root-cause context

Input

Parameter	Required	Description
Run ID or URL	Yes	GitHub Actions run ID (e.g. 22373768875) or full URL
Comparison Run	No	Second run ID/URL for side-by-side comparison

MCP Tools

All tools use owner: "microsoft" and repo: "GitHub-Copilot-for-Azure" as fixed parameters. method selects the operation within the tool.

Tool	method	Key Parameter	Purpose
actions_get	get_workflow_run	resource_id: run ID	Fetch run status and metadata
actions_list	list_workflow_run_artifacts	resource_id: run ID	List all artifacts for a run
actions_get	download_workflow_run_artifact	resource_id: artifact ID	Get a temporary download URL for an artifact ZIP
get_job_logs	—	run_id + failed_only: true	Retrieve job logs when artifact content is inaccessible
search_issues	—	query: search string	Find existing open issues before creating new ones
create_issue	—	title, body, labels, assignees	File a new GitHub issue for a test failure

Workflow

Phase 1 — Download & Parse

1. Extract the numeric run ID from the input (strip URL prefix if needed)

2. Fetch run metadata using the MCP actions_get tool:

   actions_get({ method: "get_workflow_run", owner: "microsoft", repo: "GitHub-Copilot-for-Azure", resource_id: "<run-id>" })
   

3. List artifacts using the MCP actions_list tool, then download each relevant artifact:

   // List artifacts
   actions_list({ method: "list_workflow_run_artifacts", owner: "microsoft", repo: "GitHub-Copilot-for-Azure", resource_id: "<run-id>" })
   // Download individual artifacts by ID
   actions_get({ method: "download_workflow_run_artifact", owner: "microsoft", repo: "GitHub-Copilot-for-Azure", resource_id: "<artifact-id>" })
   

   The download returns a temporary URL. Fetch the ZIP archive from that URL and extract it locally. If the environment restricts outbound HTTP (e.g. AWF sandbox), record in the analysis report that artifact content was unavailable and fall back to job logs via the get_job_logs MCP tool.

4. Locate these files in the downloaded artifacts:

   • junit.xml — test pass/fail/skip/error results
   • *-SKILL-REPORT.md — generated skill report with per-test details
   • agent-metadata-*.md files — raw agent session logs per test

   ⚠️ Note: If artifact ZIP files cannot be downloaded due to network restrictions, or if downloaded files cannot be extracted, use the get_job_logs MCP tool to identify test failures and produce a best-effort analysis from whatever data is accessible.

Phase 2 — Build Summary Report

Produce a markdown report with four sections. See report-format.md for the exact template.

Section 1 — Test Results Overview

Parse junit.xml to build:

Metric	Value
Total tests	count from <testsuites tests=…>
Executed	total − skipped
Skipped	count of <skipped/> elements
Passed	executed − failures − errors
Failed	count of <failure> elements
Test Pass Rate	passed / executed as %

Include a per-test table with name, duration (from time attribute, convert seconds to Xm Ys), and Pass/Fail result.

Section 2 — Skill Invocation Rate

Read the SKILL-REPORT.md "Per-Test Case Results" sections. For each executed test determine whether the skill under test was invoked.

The skills to track depend on which integration test suite the run belongs to:

azure-deploy integration tests — track the full deployment chain:

Skill	How to detect
azure-prepare	Mentioned as invoked in the narrative or agent-metadata
azure-validate	Mentioned as invoked in the narrative or agent-metadata
azure-deploy	Mentioned as invoked in the narrative or agent-metadata

Build a per-test invocation matrix (Yes/No for each skill) and compute rates:

Skill	Invocation Rate
azure-deploy	X% (n/total)
azure-prepare	X% (n/total)
azure-validate	X% (n/total)
Full skill chain (P→V→D)	X% (n/total)

The azure-deploy integration tests exercise the full deployment workflow where the agent is expected to invoke azure-prepare, azure-validate, and azure-deploy in sequence. This three-skill chain tracking is specific to azure-deploy tests only.

All other integration tests — track only the skill under test:

Skill	Invocation Rate
{skill-under-test}	X% (n/total)

For non-deploy tests (e.g. azure-prepare, azure-ai, azure-kusto), only track whether the primary skill under test was invoked. Do not include azure-prepare/azure-validate/azure-deploy chain columns.

Section 3 — Report Confidence & Pass Rate

Extract from SKILL-REPORT.md:

• Skill Invocation Success Rate (from the report's statistics section)
• Overall Test Pass Rate (from the report's statistics section)
• Average Confidence (from the report's statistics section)

Section 4 — Comparison (only when a second run is provided)

Repeat Phase 1–3 for the second run, then produce a side-by-side delta table. See report-format.md § Comparison.

Phase 3 — File Issues for Failures

For every test with a <failure> element in junit.xml:

1. Read the failure message and file:line from the XML
2. Read the actual line of code from the test file at that location
3. Read the agent-metadata-*.md for that test from the artifacts
4. Read the corresponding section in the SKILL-REPORT.md for context on what the agent did
5. Determine root cause category:
   • Skill not invoked — agent bypassed skills and used manual commands
   • Deployment failure — infrastructure or RBAC error during deployment
   • Timeout — test exceeded time limit
   • Assertion mismatch — expected files/links not found
   • Quota exhaustion — Azure region quota prevented deployment
6. Search for existing open issue before creating a new one using the search_issues MCP tool:

   search_issues({
     owner: "microsoft", repo: "GitHub-Copilot-for-Azure",
     query: "Integration test failure: {skill} in:title is:open"
   })
   

   Match criteria: an open issue whose title and body describe a similar problem. If a match is found, skip issue creation for this failure and note the existing issue number(s) in the summary report.
7. If no existing issue was found, create a GitHub issue using the create_issue MCP tool, assign the label with the name of the skill, and assign it to the code owners listed in .github/CODEOWNERS file based on which skill it is for:

create_issue({
  owner: "microsoft", repo: "GitHub-Copilot-for-Azure",
  title: "Integration test failure: <skill> – <keywords> [<root-cause-category>]",
  labels: ["bug", "integration-test", "test-failure", "<skill>"],
  body: "<body>",
  assignees: ["<codeowners>"]
})

Title format: Integration test failure: {skill} – {keywords} [{root-cause-category}]

• {keywords}: 2-4 words from the test name — app type (function app, static web app) + IaC type (Terraform, Bicep) + trigger if relevant
• {root-cause-category}: one of the categories from step 5 in brackets

Issue body template — see issue-template.md.

⚠️ Note: Do NOT include the Error Details (JUnit XML) or Agent Metadata sections in the issue body. Keep issues concise with the diagnosis, prompt context, skill report context, and environment sections only. ⚠️ Note: Do NOT create issues for skill invocation test failures.

For azure-deploy integration tests, include an "azure-deploy Skill Invocation" section showing whether azure-deploy was invoked (Yes/No), with a note that the full chain is azure-prepare → azure-validate → azure-deploy. For all other integration tests, include a "{skill} Skill Invocation" section showing only whether the primary skill under test was invoked.

Error Handling

Error	Cause	Fix
no artifacts found	Run has no uploadable reports	Verify the run completed the "Export report" step
HTTP 404 on actions_get	Invalid run ID or no access	Check the run ID and ensure the MCP token has repo access
rate limit exceeded	Too many GitHub API calls	Wait and retry; reduce concurrent MCP tool calls
Artifact ZIP download blocked	AWF sandbox restricts outbound HTTP to blob storage	Use get_job_logs MCP tool to get failure details from job logs; produce best-effort analysis from metadata

References

• report-format.md — Output report template
• issue-template.md — GitHub issue body template