Arness Infra Report

Report an Arness Infra workflow issue by running a diagnostic and filing a GitHub issue on the Arness plugin repository. The arn-infra-doctor agent analyzes Infra configuration and behavior — it never reads project source code or business logic.

Workflow

Step 0: Smart Routing

Before proceeding, check whether the user's issue actually belongs to a different Arness plugin:

1. Read the ## Arness section from the project's CLAUDE.md (if it exists).
2. Check the user's description for keyword signals:
   • Code keywords: "plan", "spec", "execute", "taskify", "swift", "standard", "batch", "ship", "PR", "review-pr", "assess", "catch-up", "document-project", "save-plan"
   • Spark keywords: "discover", "prototype", "greenfield", "stress test", "naming", "brainstorming", "feature extract", "visual sketch", "use cases"
3. If Code signals detected: "This sounds like an Arness Code issue. Run /arn-code-report instead."
4. If Spark signals detected: "This sounds like an Arness Spark issue. Run /arn-spark-report instead."
5. If the user confirms it is actually an Infra issue, proceed.

---

Step 1: Explain the Process

Inform the user:

"I'll help you report an Arness Infra workflow issue. Here's how this works:

1. You describe what happened — which skill you used and what went wrong
2. I'll run a diagnostic that checks Infra's configuration and expected behavior (I won't read your project code)
3. You'll review the issue report before it's submitted
4. The issue gets filed on the Arness plugin repository for the maintainers

Your project code and business logic are never included in the report."

---

Step 2: Gather User Description

Ask the user using AskUserQuestion:

• "What happened? Which Arness Infra skill were you using and what went wrong?"

Let the user type a free-form description. This becomes the user_description for the diagnostic.

---

Step 3: Check Prerequisites

1. Detect the plugin's GitHub repository:

   • Read the repository field from ${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json. Parse owner/repo from the URL.
   • Fallback: git -C ${CLAUDE_PLUGIN_ROOT} remote get-url origin. Extract owner/repo from the URL (strip .git suffix and https://github.com/ or git@github.com: prefix).

2. Check gh auth status — user must be authenticated to file issues.

3. Read plugin version from ${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json.

4. Read ## Arness config from the project's CLAUDE.md (if it exists).

If GitHub access is not available, offer an alternative: generate the diagnostic report as a local file (arness-infra-report-<YYYY-MM-DD>.md in the project root) the user can manually submit.

---

Step 4: Invoke arn-infra-doctor

Spawn the arn-infra-doctor agent via the Task tool with:

• The user's description of the issue
• Project root path
• ## Arness config content (or "not configured")
• Plugin version
• Instruction to read the knowledge base at ${CLAUDE_PLUGIN_ROOT}/skills/arn-infra-report/references/infra-knowledge-base.md

Wait for the agent to complete and collect the diagnostic report.

---

Step 5: Compose and Review Issue

Assemble the GitHub issue using the template from ${CLAUDE_PLUGIN_ROOT}/skills/arn-infra-report/references/issue-template.md:

• Include the user's original description in the "User Report" section
• Include the doctor's diagnostic findings (ISSUE items only, not OK items)
• Include the doctor's assessment
• Include plugin version and relevant config state
• Include environment info (OS, Git, GitHub, gh CLI, Docker, IaC tool, Cloud CLI status)
• Exclude any project code or business logic

Present the complete draft to the user, then request explicit consent.

Ask (using AskUserQuestion):

This report will be filed as a public GitHub issue on the Arness repository. It contains only Arness configuration state and diagnostic findings — no project source code or business logic. Please review the report above carefully.

1. Submit — I've reviewed it and consent to filing this publicly
2. Save locally — save as a file instead of submitting
3. Edit first — let me modify the report before deciding

• If Submit: proceed to Step 6.
• If Save locally: save as arness-infra-report-<YYYY-MM-DD>.md in the project root. Inform the user. Do not submit.
• If Edit first: let the user modify the draft, then ask again.

---

Step 6: Submit Issue

The arn-infra-report label must already exist on the plugin repository (maintained by the plugin maintainers, not created by this skill).

1. Create the issue with the arn-infra-report label:

   gh issue create --repo <owner/repo> --title "<title>" --body "<body>" --label "arn-infra-report"
   

2. If the command fails because the label does not exist, retry without --label and inform the user that the issue was filed without a label.

3. Report the issue URL to the user.

If submission fails for other reasons (permissions, network), save the report as arness-infra-report-<YYYY-MM-DD>.md in the project root and inform the user with instructions to submit manually.

Error Handling

• Plugin not installed from GitHub — no git remote found in plugin root and no repository field in plugin.json. Generate report as arness-infra-report-<YYYY-MM-DD>.md in the project root instead.
• gh not authenticated — suggest gh auth login, offer local file fallback (arness-infra-report-<YYYY-MM-DD>.md).
• User lacks permission to file issues on plugin repo — save report as arness-infra-report-<YYYY-MM-DD>.md in the project root, suggest opening manually.
• arn-infra-doctor agent fails — report the failure, offer to file a minimal issue with just the user's description.
• ## Arness config missing — proceed anyway, note "not configured" in the report.
• arn-infra-report label missing from plugin repo — retry without --label, note this in the output.

Constraints

• Allowed tools: Read, Glob, Grep, Bash, Agent (for arn-infra-doctor)
• NEVER include project source code, business logic, or sensitive data in the report
• NEVER file issues on the user's own project repository — this skill only files on the Arness plugin repository
• ALWAYS present the full report draft and obtain explicit user consent before submitting
• ALWAYS offer a local-file alternative — the user must never be forced to submit publicly