Read `documented_names` against the failure the post opens with. It calls `public_functions(path)`, which walks the AST and keeps every `FunctionDef` and `AsyncFunctionDef` name that does not start with an underscore [7]. It then returns the subset of those names for which the file's own text contains the name wrapped in backticks [6]. The result is a subset of the code's public functions by construction, so `missing = funcs - documented` can only ever list names that exist [15]. The check runs in one direction only, flagging code that nothing mentions; it has no way to flag a mention that no code backs. The docstring describing a function renamed or deleted in the last commit, which is the exact failure the post uses to motivate the workflow [1], passes this gate untouched [15].
That leaves the doctest run holding the load. Of the three gates proposed, one executes the documented behavior and two test for the presence of text [16]: pydocstyle or a small AST script asserts a docstring exists [3], the coverage script asserts a name appears [6], and `python -m doctest module.py` is the only step that runs anything [4]. A docstring with a summary, an Args section and no `>>>` lines clears both presence gates and hands doctest nothing to execute [17]. It can also clear the name check by mentioning itself. The author is straight about this and calls the coverage script a heuristic rather than a proof, on the grounds that a docstring can name a function and still be wrong [8].
For the executing gate to transfer, your setup has to match a couple of conditions this one meets. Your examples need stable output, because the sample asserts that `format_bytes(1024)` renders exactly `'1.0 KiB'` [9], which is fine for a formatting helper and painful for anything returning a dict, a float, or an object with a default repr. And your language needs an example runner at all; doctest is what makes this cheap in Python [4], and I would not expect the shape to survive a move to a stack with no equivalent.
The economics arrive with a disclosure. The post says it was prepared as part of MonkeyCode's product outreach [14], and the nightly drafting job leans on that project's advertised free access, currently stated as 10 million tokens with the README named as the authority [13]. No per-function token figure appears anywhere, so the cost of scanning every new public function every night is yours to measure before you plan a budget around a free tier.
If I adopted this, I would keep the first two gates and rewrite the third to run the other way: pull the backticked names out of each docstring, resolve them against the module namespace, and fail on the ones that no longer resolve. That is a short loop over `getattr`, and it catches the case the article describes while the shipped script cannot [15]. The division of labor holds up regardless. Drafting is repetitive and low risk, but reviewing the draft still demands real scrutiny [18].
