- What cost the most: research before the first edit, which used most of the turns before the context was summarised. I read winnow-grammar’s source to check four things: that a
state bound works with input bytes, how an extern rule resolves its function, the rt::Input/Slice types, and whether the skip_trivia memo keeps a recording WS correct when a parse backtracks. After that, two smaller costs: my first pass of edits to grammar.rs actually applied even though bash reported a quoting error, so the retry’s assertion failed and confused me; and the brief’s own-WS grammar would not compile at all, which took a few probes to trace to a separate defect.
- What the brief could have handed me:
- That rules in a grammar without a state are generic over
__WgS, so entries can give every grammar a LineBreaks state without changing anything else.
- The extern-rule function signature (
fn<B: rt::Slice, S: StateOf<T>>(i: &mut rt::Input<B,S>)), and that the text and bytes copies share one Rust module, so the functions need different names.
- Where
std’s program-side Nikaia lives (src/runtime/, lowered by lower-std), as opposed to tools/.
- That
rule WS needs an action (-> { }).
- That an inline choice whose alternatives give different values does not compile; this broke the brief’s own test grammar.
- Checks run: fmt, clippy, the
grammar:: unit tests, two build cycles, a fixpoint check, the old binary on both new tests, tests/compiler with both settings, it filtered on grammar, and pre-push. Wasted ones:
- The first old-binary run failed in
nikaia-std, not in the test, because runtime had not been lowered yet. It has to come after lower-std.
- A
nikaia run straight on a test file was refused (not the project’s entry point).
- An extra fixpoint build that changed nothing.
- The one change that would save most: have the brief check its own example code (test grammars, signatures) against the current binary before handing it over. Or builder.md could add a line: “probe each grammar in the brief on the old binary, without the new built-in, before the first edit.” That would have caught the choice defect and the missing
WS action at turn 5, not after the build. A short “winnow-grammar facts” section in builder-facts.md covering state, extern rule, rt::Input and the trivia memo would have saved most of the research.
This site is open source.
Improve this page.
