test-docs

name: test-docs description: > Test ToolHive documentation by executing the steps in tutorials and guides against a real environment, verifying that examples are correct and ToolHive has not regressed. Use when the user asks to test, validate, or verify documentation such as: "test the vault integration tutorial", "verify the K8s quickstart", "check the CLI install docs", "test toolhive vault integration in kubernetes", or any request to run through a doc's instructions to confirm they work. Supports both Kubernetes-based docs (tutorials, K8s guides) and CLI-based docs.

Test Docs

Test ToolHive documentation by running each step against a live environment, reporting pass/fail with evidence, and recommending fixes for failures.

Workflow

1. Find the documentation - launch an Explore agent to locate the relevant doc file
2. Read and parse - extract testable steps (bash code blocks, YAML manifests, expected outputs)
3. Check prerequisites - run scripts/check-prereqs.sh; confirm the thv version with the user
4. Prepare the environment - create a dedicated test namespace unless the doc specifies otherwise
5. Execute steps - run each step sequentially, capture output
6. Report results - pass/fail per step with evidence
7. Clean up - delete all resources created during testing
8. Recommend fixes - classify failures as doc issues or ToolHive bugs

Step 1: Find the documentation

If the user provides a file path to the doc, use it directly. Assume you are running in the repository that contains the file.

Otherwise, use the Task tool with subagent_type=Explore to search the docs directory for the page matching the user's request. Documentation lives under:

• docs/toolhive/tutorials/ - step-by-step tutorials
• docs/toolhive/guides-cli/ - CLI how-to guides
• docs/toolhive/guides-k8s/ - Kubernetes how-to guides
• docs/toolhive/guides-mcp/ - MCP server guides
• docs/toolhive/guides-vmcp/ - Virtual MCP Server guides
• docs/toolhive/guides-registry/ - registry guides

All doc files use the .mdx extension. Search by keywords from the user's request. If multiple docs match, ask the user which one to test.

Step 2: Read and parse the doc

Read the full doc file. Extract an ordered list of testable steps:

• Bash code blocks (\``bash`) - commands to execute
• YAML code blocks with title="*.yaml" - manifests to apply with kubectl apply
• Expected outputs - text code blocks (\``text`) that follow a command, or prose describing expected behavior ("You should see...")

Build a test plan: a numbered list of steps, each with:

• The command(s) to run
• What constitutes a pass (expected output, exit code 0, resource created)
• Any dependencies on previous steps (variables, resources)

Present the test plan to the user for approval before executing.

Handling placeholders

Docs often contain placeholder values like ghp_your_github_token_here or your-org. Before executing, scan for obvious placeholders (ALL_CAPS patterns, "your-*", "example-*", "replace-*") and ask the user for real values. If a step requires a secret that the user cannot provide, mark it as skipped with a note explaining why.

Step 3: Check prerequisites

Run the prerequisites checker:

bash <skill-path>/scripts/check-prereqs.sh

Where <skill-path> is the absolute path to this skill's directory.

If the script reports errors, inform the user and stop. For K8s docs, the script checks for a kind cluster named "toolhive". For CLI docs, confirm with the user that the installed thv version matches what the doc expects.

Step 4: Prepare the environment

For Kubernetes docs, ask the user about existing infrastructure before deploying anything:

• Operator CRDs and operator: if the doc includes steps to install CRDs or the ToolHive operator, ask the user whether these are already installed. If so, skip those installation steps and proceed to the doc-specific content. Use AskUserQuestion with options like "Already installed", "Install fresh", "Reinstall (upgrade)".
• Namespaces: create a dedicated namespace test-docs-<timestamp> (e.g., test-docs-1706886400) unless the doc explicitly uses a specific namespace (like toolhive-system). If the doc requires toolhive-system, use it but track all resources created for cleanup.

For CLI docs:

• Use a temporary directory for any files created
• Note the state of running MCP servers before testing (thv list)
• Use unique server names: when running thv run <server>, always use --name <server>-test to avoid conflicts with existing servers. The default name is derived from the registry entry, which may already be running. Example: thv run --name osv-test osv
• Verify registry configuration: if a registry server isn't found, confirm with the user that they're using the default registry (not a custom registry configuration)

Step 5: Execute steps

Run each step sequentially. For each step:

1. Print the step number and a short description
2. Run the command(s)
3. Capture stdout, stderr, and exit code
4. Compare against expected output or success criteria
5. Record pass/fail with evidence

Execution rules

• Wait for readiness: when a step deploys resources, wait for them (e.g., kubectl wait --for=condition=ready) before proceeding. If the doc includes a wait command, use it. Otherwise add a reasonable wait (up to 120s for pods).
• Variable propagation: some steps set shell variables used by later steps (e.g., VAULT_POD=$(kubectl get pods ...)). Maintain these across steps by running in the same shell session.
• Skip non-testable steps: informational code blocks (showing file contents, expected output) are not commands to run. Use them as verification targets instead.
• Timeout: if any command hangs for more than 5 minutes, kill it and mark the step as failed with a timeout note.

Known pitfalls

• Server name conflicts: thv run <server> uses the registry entry name by default, which may already exist. Always use thv run --name <server>-test <server> to create a uniquely-named instance. This avoids "workload with name already exists" errors.
• Port-forward needs time: when testing via kubectl port-forward, wait at least 5 seconds before issuing curl. Start the port-forward in the background, sleep 5, then curl. Kill the port-forward PID after the check.
• MCPServer status field: the MCPServer CRD uses .status.phase (not .status.state) for the running state. When polling for readiness, use: kubectl get mcpservers <name> -n <ns> -o jsonpath='{.status.phase}' or simply poll kubectl get mcpservers and check the STATUS column.
• MCPServer readiness polling: after kubectl apply for an MCPServer, it may take 30-60 seconds to reach Running. Poll with 5-second intervals up to 120 seconds before declaring a timeout.

Step 6: Report results

After all steps complete, produce a summary table:

## Test Results: <doc title>

| Step | Description          | Result | Notes                    |
|------|----------------------|--------|--------------------------|
| 1    | Install Vault        | PASS   |                          |
| 2    | Configure auth       | PASS   |                          |
| 3    | Store secrets        | PASS   |                          |
| 4    | Deploy MCPServer     | FAIL   | Pod CrashLoopBackOff    |
| 5    | Verify integration   | SKIP   | Depends on step 4       |

For each FAIL, include:

• The command that failed
• Actual output vs expected output
• A classification:
  • Doc issue: the command or example in the doc is wrong or outdated (e.g., wrong image tag, deprecated flag, missing step)
  • ToolHive bug: the doc appears correct but ToolHive behaves unexpectedly (e.g., crash, wrong output from a correct command)
  • Environment issue: missing prerequisite, network problem, or transient error
• A specific recommendation (e.g., "Update image tag from v0.7.0 to v0.8.2" or "File a bug: MCPServer pods fail to start when Vault annotations are present")

Step 7: Clean up

Always clean up after testing:

• Delete test namespaces: kubectl delete namespace <test-namespace>
• Remove Helm releases installed during testing
• Delete any temporary files or directories
• If resources were created in toolhive-system, delete them individually by name

If cleanup fails, report what could not be cleaned up so the user can handle it manually.

Example session

User: "test the vault integration tutorial"

1. Explore agent finds docs/toolhive/tutorials/vault-integration.mdx
2. Parse the doc: 5 major steps with 12 bash commands and 1 YAML manifest
3. Run check-prereqs.sh - kind cluster "toolhive" exists, thv v0.8.2
4. Ask user: "The doc uses a placeholder GitHub token ghp_your_github_token_here. Provide a real token or skip step 2?"
5. Present test plan, user approves
6. Execute steps 1-5, report results
7. Clean up: delete vault namespace, remove MCPServer resource
8. Summary: 4/5 PASS, 1 FAIL with recommendation