feat(mcp-get-started): fastmcp 4, the SEP-2663 tasks extension, and a task-augmented tool

[JonZeolla]

Change Description

Adds slow_shout to the MCP 101 rig — an async tool with task=True — and moves the rig to FastMCP 4 so it can drive the MCP task cycle on the 2026-07-28 protocol revision.

• server.py — slow_shout(text, seconds=10), plus mcp.add_extension(TasksExtension()). FastMCP 4 splits task support into the fastmcp-tasks distribution and requires registering the extension explicitly; the tasks extra alone is no longer enough.
• client.py — passes Path for a file target and a string for a URL. FastMCP 4 deprecates inferring a stdio transport from a bare string (removed in 5), so Client("server.py") printed a FastMCPDeprecationWarning on every learner's first run.
• pyproject.toml / Dockerfile — fastmcp[tasks]>=4,<5, lockfile re-pinned.
• evidence/ — every file recaptured from the e2e harness instance. New: discover.txt (the server/discover response) and tasks-capability-refusal.txt (the -32021).
• READMEs regenerated with export-lab-readme.js --all.

Background (Why we did this)

The lab took a server from stdio to a container to goose without ever showing what MCP does when a tool takes minutes. On 2026-07-28 that story is short enough to drive with curl.

What changed between SEP-1686 and SEP-2663 is worth knowing before reviewing:

	SEP-1686 (fastmcp 3)	SEP-2663 (fastmcp 4)
Enablement	fastmcp[tasks] extra	plus mcp.add_extension(TasksExtension())
Requesting a task	params.task = {ttl}	nothing on the request; declaring the extension in the envelope does it
Polling	tasks/get then tasks/result	tasks/get only, result inline once completed
Advertising	execution.taskSupport on the tool	absent from the tool descriptor

pydocket's default backend is an in-process queue, so the lab needs no Redis and no extra container.

Testing

The full lab ran on the e2e harness (Rocky Linux 8 EC2, provisioned from the sandbox image recipe): 17 blocks ran, 0 skipped, 0 failed. Evidence in this PR is that run's output, not a laptop's.

{"result":"THE SLOW WAY!"}
waited 10s for that
submitted d49tt8EqL5SD0h8g8DbRIFUXxoaMx2HGPyZOMOafufs
{"status":"working","pollIntervalMs":5000.0}
{"status":"completed","result":{"result":"TASKS WORK!"}}

client.py passes over both stdio and Streamable HTTP with no deprecation warning, and goose reached both tools on its first attempt.

Pull Request Checklist

• If you are adding a dependency, please explain how it was chosen — fastmcp[tasks] is the same fastmcp already pinned plus the extra carrying task support; fastmcp-tasks and pydocket arrive through it. No new vendor.
• Ensure all items left for future development have been appropriately documented

Hub-side content: Zenable-io/next-gen-governance#6752

🤖 Generated with Claude Code

https://claude.ai/code/session_01ReAbCfx4UvDZE3RKSGBDQM

[ai-coding-guardrails]

Nice work! 😎

I didn't find anything of concern

List of skipped files due to configuration

• labs/mcp-get-started/uv.lock

Risk: 🟢 Low

Reviewed with 🤟 by Zenable