MCP Testing

Podman Desktop MCP Testing

Interactively test and explore Podman Desktop through automated browser commands, connected via Chrome DevTools Protocol (CDP).

When to use this skill

• Manually testing Podman Desktop UI changes during development
• Testing the production (installed) Podman Desktop app
• Debugging specific user workflows or UI issues
• Exploring new features interactively
• Verifying functionality without writing test code
• Quick acceptance testing

Prerequisites

• MCP server podman-desktop-mcp configured via .mcp.json at the repo root (already done)
• For dev mode: Node.js 24+ and pnpm 10+ installed
• For production mode: Podman Desktop installed

Workflow

Phase 1: Detect Environment

Always run detection yourself. Never ask the user to run scripts manually.

Detect the OS — run uname -s via the Bash tool:

• Returns Darwin → macOS
• Returns Linux → Linux
• Starts with MINGW, MSYS, or CYGWIN → Windows (Git Bash)
• Command unavailable or unrecognized → Windows (native PowerShell)

Run the probe script to detect what's running:

Linux / macOS:

bash .agents/skills/mcp-testing/probe.sh

Windows:

pwsh .agents/skills/mcp-testing/probe.ps1

Parse the key=value output. You will get eleven variables:

• PROD_RUNNING — whether a production (installed) Podman Desktop is running
• PROD_CDP_PORT — CDP port for production (none if CDP not available)
• PROD_APP_TITLE — window title of production app (none if not available)
• DEV_RUNNING — whether pnpm watch is running
• DEV_CDP_PORT — CDP port for dev (none if not available)
• DEV_APP_TITLE — window title of dev app (none if not available)
• STALE_SESSION — true if a previous session left a state file without cleaning up
• PORT_9222_NAME / PORT_9222_PID — process on port 9222 (none if free)
• PORT_9223_NAME / PORT_9223_PID — process on port 9223 (none if free)

Stale session: If STALE_SESSION=true, run the stop script immediately — no confirmation needed — then re-run the probe to get fresh state:

Linux / macOS:

bash .agents/skills/mcp-testing/stop.sh

Windows:

pwsh .agents/skills/mcp-testing/stop.ps1

Port conflict: For each port where PORT_<port>_NAME is not none, electron, node, or Podman Desktop, use AskUserQuestion:

• Question: Port {PORT} is in use by "{NAME}" (PID {PID}), which doesn't look like Podman Desktop. Kill it to free the port?
• Header: Port conflict
• Options:
  1. Kill it — Stop the process and free the port
  2. Skip — Leave it running (choose a different mode or abort)

If the user chooses "Kill it", kill the process using the PID from the probe output:

Linux / macOS: kill <PID> Windows: Stop-Process -Id <PID> -Force

If the user chooses Skip and neither port is usable, report the conflict and stop — do not run the start script.

Phase 2: Choose Mode

Always ask the user which mode to use. Use AskUserQuestion with options based on the probe results. Annotate each option with its current status.

Build the options based on probe output:

Production mode option:

• If PROD_RUNNING=true and PROD_CDP_PORT is not none: Label: Production mode Description: Connect to installed Podman Desktop on port {PROD_CDP_PORT} — changes affect your real environment

• If PROD_RUNNING=true and PROD_CDP_PORT=none: Label: Production mode Description: Will relaunch Podman Desktop with CDP enabled — changes affect your real environment

• If PROD_RUNNING=false: Label: Production mode Description: Will launch Podman Desktop with CDP enabled — changes affect your real environment

Dev mode option:

• If DEV_RUNNING=true and DEV_CDP_PORT is not none: Label: Dev mode Description: pnpm watch already running on port {DEV_CDP_PORT} — instant connect

• If DEV_RUNNING=true and DEV_CDP_PORT=none: Label: Dev mode Description: pnpm watch is running but CDP is not responding — will restart pnpm watch

• If DEV_RUNNING=false and PROD_RUNNING=false: Label: Dev mode Description: Start pnpm watch for isolated development testing (2-4 minutes first run)

• If DEV_RUNNING=false and PROD_RUNNING=true: Label: Dev mode Description: Will close production app first, then start pnpm watch (2-4 minutes)

Use the AskUserQuestion tool with header Mode and these two options. The question should be: Which mode would you like to use for testing?

Phase 3: Start and Connect

Based on the chosen mode and probe results, prepare the app and run the startup script.

Production mode

Run the startup script — it handles detecting the current state (including stopping any dev instance that holds the single-instance lock), relaunching with CDP if needed, waiting for the app, and verifying the connection:

Linux / macOS:

bash .agents/skills/mcp-testing/start.sh --mode prod

Windows:

pwsh .agents/skills/mcp-testing/start.ps1 -Mode prod

Dev mode

If DEV_RUNNING=true and DEV_CDP_PORT is not none (already running):

bash .agents/skills/mcp-testing/start.sh --mode dev-fast

pwsh .agents/skills/mcp-testing/start.ps1 -Mode dev-fast

If DEV_RUNNING=true and DEV_CDP_PORT=none (pnpm watch alive but CDP not responding — treat as full restart):

Use --mode dev (not dev-fast). The start script will kill the stale process and restart cleanly.

If DEV_RUNNING=false (full startup) — set Bash tool timeout to 600000 (10 minutes):

bash .agents/skills/mcp-testing/start.sh --mode dev

pwsh .agents/skills/mcp-testing/start.ps1 -Mode dev

Error handling

If the start script exits with a non-zero code, report the error output to the user and refer to the Troubleshooting section below. Common failures:

• Port still in use → suggest killing the process manually
• pnpm install timeout → check network connectivity
• CDP not available within 120s → check for single-instance lock (see Troubleshooting)

Do not retry the script automatically — let the user decide after seeing the error.

Extract the CDP port

When the start script prints Ready — call mcp__podman-desktop-mcp__connect(...), extract the port number from that line. You will pass it to the Haiku agent in the next phase.

Phase 4: Execute task (Haiku agent)

1. Read .agents/skills/mcp-testing/agent-prompt.md

2. Select relevant operation references. Using the task description and the table below, select which files from .agents/skills/mcp-testing/references/operations/ to include. Always include navigation.md. Use your judgment — if the task implies an area even without exact keywords, include its reference.

   File	Covers
   navigation.md	Always include. Sidebar nav, welcome/onboarding page, dashboard, breadcrumbs, back/forward, status bar
   images.md	Pull/build/push/save/delete/prune images, image details, edit/rename, Containerfile, platform selection, registry search
   containers.md	Create/run/start/stop/delete/export containers, container details, logs, terminal, port mapping, prune
   pods.md	Kube Play (YAML deploy), podify (create pod from containers), pod lifecycle, pod logs, prune
   volumes.md	Create/delete/prune volumes, volume details, gather sizes, usage status
   networks.md	Create/delete networks, subnet configuration, network details, bulk delete
   extensions.md	View/install/enable/disable/delete extensions, catalog, custom OCI install, extension status
   settings.md	Preferences, proxy configuration, resources, CLI tools section, Kubernetes settings
   registries.md	Add/remove/login to registries, credentials, authentication (under Settings)
   kubernetes.md	Kubernetes contexts, namespaces, resource browsing (nodes/pods/deployments/services), apply YAML
   compose.md	Docker Compose onboarding, CLI tool install/uninstall/version check
   search-palette.md	Command palette (F1), search commands, navigate via palette

   Cross-dependencies — when one area implies another, include both:

   • Container operations → also include images.md (creating a container requires an image)
   • Pod operations → also include containers.md (podify requires selecting containers)
   • Registry operations → also include settings.md (registries is under Settings)
   • Settings exploration → also include registries.md (registries is a Settings subsection)
   • Compose/CLI tools → also include settings.md (CLI tools is under Settings)
   • Settings exploration → also include compose.md when CLI tools may be relevant
   • Docker mentioned (docker, docker-compose, Docker Hub) → include compose.md + containers.md + images.md + extensions.md + settings.md
   • Kubernetes settings → also include kubernetes.md
   • Any image name in the task (httpd, nginx, alpine, redis, etc.) → include images.md
   • Ambiguous or broad scope (no single area clearly implied) → include navigation.md + containers.md + images.md + settings.md + extensions.md
   • General exploration or unknown scope → include all references

3. Read the selected reference files and concatenate their contents into an Operations Reference block

4. Spawn an agent with model: "haiku", passing: CDP port: {PORT}. Task: {TASK} followed by the full content of agent-prompt.md, then the operations reference block

The agent uses operation references directly when they cover the elements needed for the task. It falls back to compact snapshots only for elements not in the reference, and full snapshot() as a last resort.

The agent returns a summary of what it found or did, and optionally a structured workflow trace (for E2E test writing tasks). The agent handles connecting to CDP, initial setup (dialogs), and disconnecting — the orchestrator only needs to pass the port number.

After the agent returns, use AskUserQuestion:

• Question: What would you like to do next?
• Header: Next
• Options:
  1. Continue testing — Provide another task (user enters free text via "Other")
  2. Done — finish testing — Disconnect and clean up

If the user continues, run Phase 4 again with the new task. If done, proceed to Phase 5.

Phase 5: Cleanup

Use AskUserQuestion:

• Question: What should I do with the app?
• Header: Cleanup
• Options:
  1. Leave running — Keep the app alive so you can reconnect later
  2. Close it — Stop the app and free the port

If the user chooses "Leave running" — remove the session file so the next invocation finds the app cleanly via probe (no process killing):

Linux / macOS: bash .agents/skills/mcp-testing/stop.sh --session-only Windows: pwsh .agents/skills/mcp-testing/stop.ps1 -SessionOnly

If the user chooses "Close it" — run the stop script, which reads the session state and kills the right processes:

Linux / macOS:

bash .agents/skills/mcp-testing/stop.sh

Windows:

pwsh .agents/skills/mcp-testing/stop.ps1

Troubleshooting

pnpm: command not found

npm install -g pnpm

Then re-open your terminal and verify with pnpm --version.

Production App Without CDP

Symptom: Probe shows PROD_RUNNING=true but PROD_CDP_PORT=none.

Fix: Run start.sh --mode prod (or start.ps1 -Mode prod) — it detects this case automatically, closes the running app, and relaunches it with CDP enabled.

Production Blocks Dev Instance

start.sh --mode dev (and start.ps1 -Mode dev) automatically detects and closes a running production Podman Desktop before starting the dev instance — no manual action is needed.

If startup still times out after 120s: the production app may have restarted in the gap between detection and the kill, or a non-Podman-Desktop process holds port 9222. Run the stop script, verify port 9222 is free (lsof -ti :9222), then re-run the start script.

Quick Reference