Setup coverage (windows)

[bmerkle]

This PR:
fixed setup for uv coverage under Windows

[Copilot]

Pull request overview

Adds Windows support for running test coverage via uv, aligning the developer workflow with the existing Unix make coverage target.

Changes:

• Added a coverage command to make.bat that runs coverage.py via uv run.
• Updated .coveragerc to measure coverage from src instead of test.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
make.bat	Adds a coverage target and updates the usage string to include it.
.coveragerc	Adjusts coverage source configuration to the src tree.

[Copilot]

The Windows coverage target doesn't set COVERAGE_PROCESS_START. Some tests spawn the MCP server in a subprocess and only propagate COVERAGE_PROCESS_START if it’s already present, so coverage for subprocesses won’t be collected on Windows (unlike the Makefile target). Set COVERAGE_PROCESS_START=.coveragerc before coverage run (and optionally unset it afterwards) to match the Unix behavior.

Suggested change

	uv run coverage erase
	uv run coverage run -m pytest
	uv run coverage combine
	uv run coverage report
	uv run coverage erase
	set "COVERAGE_PROCESS_START=.coveragerc"
	uv run coverage run -m pytest
	uv run coverage combine
	uv run coverage report
	set "COVERAGE_PROCESS_START="

[KRRT7]

A couple of things I noticed looking at this:

1. source = src should probably be source = src,test — the previous source = test was wrong in the other direction, but this just swapped which half we're missing. Measuring test coverage catches dead test code from copy-paste naming mistakes, obsolete helpers, and tests that silently never run. Ned Batchelder has a good writeup on why: https://nedbatchelder.com/blog/202008/you_should_include_your_tests_in_coverage.html

Important

This matters even more now that AI agents are writing and modifying tests. They're great at producing code that looks correct — proper structure, reasonable assertions, good names — but they regularly introduce subtle bugs: tests that pass because they assert the wrong thing, helper functions that never get called, branches in test logic that are dead on arrival. The failure mode is silent: CI is green, coverage of src looks fine, nobody notices the test isn't actually exercising anything. Covering your test code is a cheap defensive check against that. If an agent generates a test function but misspells the name so pytest doesn't collect it, source = src won't catch it — source = src,test will show it as uncovered immediately.

---

2. make.bat coverage target is missing COVERAGE_PROCESS_START — the Makefile sets it; make.bat doesn't. Without it, you get zero coverage from the MCP server code on Windows. Should be:

   set "COVERAGE_PROCESS_START=.coveragerc"
   uv run coverage run -m pytest
   set "COVERAGE_PROCESS_START="

   Why this matters

   The MCP server tests spawn the server as a subprocess (test_mcp_server.py launches python -m typeagent.mcp.server via StdioServerParameters). coverage run -m pytest only measures the main pytest process — it can't see into child processes. COVERAGE_PROCESS_START is the mechanism that bridges that gap: the test fixture propagates it to the subprocess env, server.py calls coverage.process_startup() which sees the var and starts collecting, and coverage combine merges everything at the end. Without it, process_startup() is a no-op.

[bmerkle]

1. regarding the inclusion of the test directory, I would disagree.
   The coverage should report the coverage of the production code or production app. Mixing up the src and test code dilutes the number or metric.
   I would also not necessarily measure coverage on infrastructure code which was only generated via a template or template based code generator.
   The argument has not changed with AI IMO, but I also asked on a mailing list where I know some TDD experts, about their opinion....
   Also please look at the comments too the article you cited...

2. this is a bug for subprocesses. Without it, subprocess coverage (spawned by pytest) isn't collected.
   Thanks for finding it. Created a fix set COVERAGE_PROCESS_START, which the Makefile does. #253