Journal

Progress

• Python 3.14.7 and uv 0.12.14 pinned in the course’s own project, code/python-for-csharp-java, with mypy 2.3.1 and pytest 9.1.1 in uv.lock
• check.sh: mypy on the project and on each error snippet, every example, script and solution run, pytest, the C# and Java comparisons compiled and run, all compared with expected/ on Windows
• CI on Linux, Windows and macOS (to verify: the workflow is written, not yet pushed)
• Lesson 1: installing and running Python
• Lesson 2: the execution model
• Lesson 3: built-in types and collections
• Lesson 4: functions
• Lesson 5: classes

2026-09-15 — Installing Python without touching the machine

• The machine already had a system Python and uv 0.10.4 on its PATH, and other projects use their own virtual environments. The course had to leave all of them alone. I downloaded the uv 0.12.14 binary into a scratch folder and pointed UV_PYTHON_INSTALL_DIR, UV_CACHE_DIR and UV_TOOL_DIR there, with UV_PYTHON_INSTALL_REGISTRY=0 so that the Windows registry doesn’t learn about the interpreter, and UV_NO_MODIFY_PATH=1. uv python install 3.14.7 then downloaded a build from python-build-standalone into that folder only. Lesson 1 shows the ordinary installation, which is what a reader wants.
• Python 3.14.7 was released on 2026-08-05. On Windows, python.org now recommends the Python install manager; the traditional installer and the py launcher that came with it are deprecated since 3.14. Lesson 1 installs with uv on the three OSes, and mentions the install manager.
• uv 0.12.0 changed uv init. It now creates a packaged project by default: a src/hello/__init__.py, a [project.scripts] entry and the uv_build backend, where 0.11 created a single main.py. uv init --no-package gives the old layout. Tutorials written before 2026 show the old files; lesson 1 shows the new ones, from check.sh.
• The course’s pyproject.toml sets [tool.uv] package = false: the course is a folder of examples, not a package to install.

2026-09-15 — Surprises while writing lessons 1-4

uv init inside a project adds a workspace member. check.sh creates lesson 1’s hello project in out/, under the course’s project, and the first run of uv init added members = ["out/uv-demo/hello"] to the course’s pyproject.toml and locked it again. --no-workspace stops that. A reader who runs uv init in a subfolder of a uv project gets the same edit.

uv tree filters by platform. On Windows, uv tree lists colorama under pytest; on Linux it wouldn’t, since the lock file records it with marker = "sys_platform == 'win32'". Lesson 1 uses uv tree --universal, which prints the same tree on the three OSes. The lock file itself is the same everywhere, which is the point of it.

Resolutions change every day. uv add httpx==0.28.1 pins httpx, not its dependencies: anyio, certifi and idna resolve to whatever is newest. check.sh sets UV_EXCLUDE_NEWER=2026-09-15T00:00:00Z for lesson 1’s demo project, so that the tree in the lesson doesn’t change when certifi publishes a release.

A .python-version that contradicts requires-python is a hard error, exit code 2, not a warning; lesson 1’s third exercise shows the message.

Python 3.14’s messages are better than the tutorials show. A list used as a dict key now says cannot use 'list' as a dict key (unhashable type: 'list'), where 3.13 said unhashable type: 'list'; the traceback puts carets under the failing part of the expression; is with a literal and an invalid escape sequence are SyntaxWarnings, printed when the file is compiled. The disassembly of lesson 1 shows 3.14’s LOAD_FAST_BORROW_LOAD_FAST_BORROW, a superinstruction that older articles about dis don’t have.

mypy 2.3.1 misses more than I expected in strict mode. It reports no error for lesson 2’s UnboundLocalError, for a list used as a dict key, for is with a literal, for an invalid escape sequence, or for a mutable default argument. Nor for ordered = tuning.sort(), which assigns None. It does catch every wrong call of lesson 4, and the subscript of a dict.get result that can be None (lesson 2). It refused the i=i lambda (Cannot infer type of lambda), which moved from examples/ to errors/. Lessons 2 to 4 say, for each error snippet, whether mypy finds it.

round isn’t the bug in lesson 1. I first wrote that f"{25.875:.2f}" prints 25.87; it prints 25.88. The total of lesson 1 prints 25.87 because the float computation gives 25.874999999999996. Lesson 3 now uses lesson 1’s own formula, and shows the Decimal version.

A set’s order depends on PYTHONHASHSEED. With seed 1 and 2, {'E', 'A', 'D', 'G', 'B'} prints in two different orders; check.sh fixes the seed for those two lines, and every other example sorts its sets before printing them.

.NET and Java disagree about changing a collection while iterating it. Dictionary.Remove inside a foreach doesn’t throw since .NET Core 3.0, List.Remove does, and Java’s LinkedHashMap throws ConcurrentModificationException. Lesson 3 prints all three.

pytest 9 reads [tool.pytest] in pyproject.toml, with native TOML types; older versions only read [tool.pytest.ini_options]. The course uses the new table.

dotnet run file.cs caches the build. As in the TypeScript course, a second run of an unchanged file doesn’t call the compiler, so check.sh runs dotnet clean then dotnet run --no-cache on each C# comparison. compare/global.json pins the SDK to 10.0.100 with latestFeature, since the machine also has a .NET 11 preview.

The CI can’t be pushed yet. The GitHub token used for pushing has no workflow scope, and GitHub refuses a push that adds a file under .github/workflows/. The code and the lessons are pushed; .github/workflows/python-examples.yml waits for gh auth refresh -s workflow. Until it runs, the Linux and macOS parts of lesson 1 are to verify.

2026-09-15 — What the lessons found in GuitarAlchemist/ga (cc42d21)

I cloned GuitarAlchemist/ga at cc42d21 into a scratch folder, installed each Python project with the course’s uv, and scanned the Python files with a small script based on the ast module (mutable defaults, bare except, annotations).

• A bug: HTTP 400 becomes 500 in the knowledge-graph service. In Apps/ga-graphiti-service/main.py, each endpoint raises HTTPException(status_code=400, ...) inside a try whose except Exception as e raises HTTPException(status_code=500, detail=str(e)). HTTPException is an Exception, so the 400 is caught and the client receives a 500 whose detail is 400: episode has no user. The same shape is at lines 111, 127, 143 and 159. Reproduction, reduced to FastAPI 0.121 with a TestClient: the script prints 500 {'detail': '400: episode has no user'}. The fix is an except HTTPException: raise before the general clause. Lesson 7 (errors) will use it; I haven’t opened an issue.
• The governance agent has no lock file. Apps/demerzel-agent/pyproject.toml declares lower bounds only (acp-sdk>=0.2.0), and the README says uv sync, which resolves today’s versions: acp-sdk 1.0.3 on 2026-09-15, a major version above the bound. The imports still work.
• Its entry point is never installed. [project.scripts] demerzel = "src.server:main" needs a build system, and the file has none: uv sync prints a warning that it skips the entry points, and uv run demerzel then found an unrelated demerzel.exe elsewhere on my PATH. The top-level package is also named src. Lesson 1 explains packaged projects; lesson 8 will come back to the layout.
• Importing the governance service can fail. src/services/governance.py, lines 13-28, calls find_governance_root() at module level, which raises FileNotFoundError when no governance/demerzel folder is found, so any import of the module, a test’s included, fails outside the repository. The environment variable DEMERZEL_ROOT is only read after the walk up the parent folders, so it can’t override a folder that exists. Lesson 2 reduces it, and its third exercise reads the root lazily.
• Errors swallowed. The same file has except Exception: continue or equivalent at lines 46-56, 83 and 130, and src/services/github.py at lines 33 and 53; epistemic_agent.py, lines 145-148, has except ValueError: pass. For lesson 7.
• Collections by hand. The tensor summary and the where filter of epistemic_agent.py are exercises 1 and 3 of lesson 3.
• The service’s image uses python:3.11-slim and uv 0.10.4 with a requirements.txt of lower bounds, so each image build can resolve different versions.
• Scripts with undeclared dependencies. Scripts/train-router-head.py imports NumPy and scikit-learn, and nothing in the repository declares them; a PEP 723 block would (lesson 1, exercise 2). It also reads JSON with json.load(open(path)), without with or an encoding (lines 63, 66 and 71). Scripts/validate_optick_sae_artifacts.py catches the ImportError of jsonschema at lines 34-39. Two scripts start with a UTF-8 BOM, which Python accepts.
• No mutable default argument in any Python file of the three repositories, according to the ast scan.

2026-09-15 — GuitarAlchemist/ix (2190c68) and tars (26c5f67)

• IX trains its sparse autoencoder from crates/ix-optick-sae/python, with a requirements.txt of lower bounds and no lock; the code imports Dict, List and Optional from typing (lesson 3), and its tests use unittest with a sys.path edit instead of pytest (lesson 9).
• TARS commits a virtual environment. Scripts/tts-venv holds 2,691 files: a Windows python.exe, pip, Flask and NumPy 2.2.4’s compiled wheel for CPython 3.12, added in commit 2402545 on 2025-04-01. It works on one OS and one Python version, and weighs on every clone. A pyproject.toml and a lock file would replace it (lesson 1).
• TARS’s Python tools declare nothing. Seven of the 23 scripts in tools/python import requests, aiohttp or websockets, and the folder has no dependency file. 272 of their 288 functions have no return annotation.
• Invalid escape sequences in implementation-starter.py, line 184, and meta-programming-demo.py, line 497, reported by Python 3.14.7 (lesson 3).
• Bare or swallowing except in autonomous-quality-loop.py (lines 388, 408, 462, 482 and 491), tars-mcp-client.py (89), tars-evolve-cli.py (443 and 462) and real-multimedia-processor.py (142). For lesson 7.

2026-09-15 — This site

• The machine learning course installs NumPy and scikit-learn with pip install into whatever Python is on the PATH. The versions are pinned, which is good, but the packages land in the system or user Python, and they can conflict with another project’s. A PEP 723 block and uv run --script, or a small uv project in code/, would isolate them. I haven’t changed that course.

To verify

• Lesson 1’s installation on Linux and macOS: the output of uv python install and the build names, from the CI’s logs once the workflow runs.
• check.sh on the three OSes: the outputs were captured on Windows, and path separators, line endings and the C# culture could differ.
• The behaviour of the Python install manager on Windows, which I didn’t install on this machine.