Migrate to mcp 2.0 and raise test coverage to 99%

[FWDavide]

Summary

Unblocks uvx basecamp-cli-mcp, which had been dead on arrival since the mcp 2.0 release, then migrates properly to the 2.0 API and backfills the test suite that failed to catch the break.

Four commits, each self-contained:

1. Pin mcp<2 (v0.7.2) — the unbounded mcp>=1.2 range meant every fresh uvx install resolved mcp 2.0, which removed the @server.list_tools() / @server.call_tool() decorators, and died with AttributeError before serving a request. Pin so the published package installs again.
2. Migrate to the mcp 2.0 low-level Server API — port the two handlers to the on_list_tools / on_call_tool callbacks passed to Server(), and require mcp>=2,<3. Behaviour is preserved: tools still advertise the raw JSON Schema from tools.json (types.Tool now takes input_schema rather than inputSchema), and a failing call is still reported as an error result carrying the CLI message plus stderr. That last part is now explicit: 1.x's decorator wrapper turned a raised exception into isError: true for us, whereas a raw on_call_tool handler would surface it as a JSON-RPC error, so the handler returns CallToolResult(is_error=True) itself.
3. Raise test coverage from 50% to 99% — see below.
4. Ignore coverage artifacts — .coverage is regenerated on every run.

Why the test work

The mcp 2.0 break exposed how thin the suite was: it stayed green while the server could not be constructed at all, because three modules had no tests and the two that did stubbed out their own boundaries.

• Runner._invoke (tests/test_runner_invoke.py) — the seam neither existing suite touched. Runner tests overrode _invoke; server tests faked the runner. Nothing exercised argv reaching a process, or the {ok, data} / {ok: false, error} envelope becoming a value or a BasecampError. Driven by a real shell stub, so subprocess, stderr, and exit codes are real.
• setup_cmd (tests/test_setup_cmd.py) — the only code that writes outside the repo, and untested through the last three releases. Covers the config merge, the abort paths that must leave a malformed file untouched, backup creation, both tool-set choices, and the installer flow.
• generator (tests/test_generator.py) — the tree walk that produces the shipped contract, via a subclass replaying canned CLI output. Worth having before the pending regeneration against basecamp CLI 0.9.0.
• tools.json (tests/test_tools_contract.py) — the committed contract now has to load, satisfy the runner's expectations, pass SEP-986 tool naming, and build a server across all 255 entries.
• cli (tests/test_cli.py) — flag plumbing through to server.run.
• End-to-end protocol tests over in-memory streams (tests/test_server_protocol.py) — real handshake, tools/list, tools/call.

Two tests deliberately pin surprising behaviour rather than assert what it ought to be, so a future change surfaces it: alias dedupe drops sibling actions that share a short description, and the todos-update workaround cannot unassign everyone. Both are documented in place.

Also adds pytest-cov and coverage config.

Test plan

uv run pytest — 151 passed.

Name                                  Stmts   Miss Branch BrPart  Cover
-----------------------------------------------------------------------
src/basecamp_cli_mcp/__init__.py          5      2      0      0    60%
src/basecamp_cli_mcp/cli.py              35      0      6      0   100%
src/basecamp_cli_mcp/generator.py        90      0     36      0   100%
src/basecamp_cli_mcp/help_parser.py     115      0     54      1    99%
src/basecamp_cli_mcp/runner.py          120      0     60      0   100%
src/basecamp_cli_mcp/server.py           59      0     12      0   100%
src/basecamp_cli_mcp/setup_cmd.py       101      0     32      0   100%
-----------------------------------------------------------------------
TOTAL                                   525      2    200      1    99%

🤖 Generated with Claude Code