Cloud|

Audit an application against the conventions

Check built-in and third-party applications and report findings by root cause.

2 min read Updated 2026-10-15 #conventions#audit#review#checklist

Use the Conventions for Cloud applications catalog to review one application. Follow the canonical link whenever a row needs more context.

1. Name the application and its surfaces

Record the app ID, the version or revision you inspect, and its public surfaces: routes and prefixes, pages, API, CLI commands, capabilities, Assistant Skills, and Help. Note which surfaces the application does not have.

2. Run the checks that apply

In the Cloud monorepo, run the named bun scripts/check.ts <rule> checks and focused tests from the relevant rows. For a third-party application, use their manual equivalents. Record failures and checks you could not run; capture the reproduction for every finding (the check command, a grep, or a click path).

3. Walk the catalog

Review each applicable area against the code. Where a rule names the running application, check it at desktop and phone width in light mode. Verify the canonical page's expected behavior, including denied requests and retries where relevant. Capture the reproduction for every finding (the check command, a grep, or a click path).

4. Record findings

Use this format. The filled row uses demo data:

text
ID | Location (file:line or URL) | Severity | Observed vs. expected | Reproduce | Fix
ACC-02 | src/service/items.ts:42 | blocker | Demo item is returned without checking read permission; the service must authorize the resource. | curl -i "$DEMO_APP_URL/api/items/demo-item" | Resolve the subject's permission before reading the item.
Severity Meaning
blocker Security, authorization, data loss, or an irreversible action
major Broken public contract or user-visible defect
minor Consistency or polish

5. Group findings by root cause

One root cause can violate several convention IDs. Group its findings and propose one fix per root cause instead of separate fixes for each symptom.

6. Report findings

Search the tracker for open issues before filing a report. Add evidence to an existing issue when it already describes the same problem.

In k2b-dev/cloud, file one GitHub issue per root cause using the Agent report issue form, with labels via-agent and needs-triage. Use its four sections exactly:

  • What is wrong: IDs, severity, and observed vs. expected behavior.
  • How to reproduce: the reproduction recorded for the finding.
  • Suspected cause: the location (file:line) with one sentence explaining why.
  • Proposal: the fix, or a decision needed from the maintainer.

Use demo data only. Include no personal data or screenshots with personal content; redact secrets and machine-specific paths. A third-party application uses its own tracker.

Leave out inapplicable rules and duplicates

Do not report a rule that does not apply: for example, a Built-in apps: row for a third-party application, or a rule about a surface the app does not have. Do not create duplicate issues; add new evidence to the existing report.

Connect a coding agent

The Cloud skill provides compact working instructions. MCP supplies exact current documentation when details matter.

CLI command
bunx skills add https://docs.example.com

This command installs the working instructions published by this website in the selected coding agent.

For exact, current details, also connect the MCP server through one of the agent tabs.

Installation uses the open-source Vercel Skills CLI.

The skill and MCP complement each other: the skill describes workflows, while MCP supplies current documentation.