agent-feedback: postfix-! parentheses and html-style body formatting

agent-feedback/items/2026-08-28-document-html-style-body-formatting.md

@@ -0,0 +1,12 @@
+---
+type: unclear
+impact: low
+effort: low
+site: README.md › # Options
+---
+
+# Say in the README that `<html-style>`/`<html-script>` bodies are formatted and that `embeddedLanguageFormatting: "off"` is the only (global) way to keep them as written
+
+`getTagParser` in `src/utils/get-parser-name.ts` returns `"css"` for `html-style` with no `type=` test, and `getScriptTagParser` returns `babel-ts`/`json` for `html-script` unless its `type=` falls outside module/text/javascript/application/javascript/importmap/speculationrules/application/json, so both bodies are reprinted by prettier's embedded printers. Marko writes those bodies out verbatim (`packages/runtime-tags/src/translator/core/html-style.ts` parses them with `text: true, preserveWhitespace: true`), so a `--write` changes the rendered bytes of every page that uses the tags, and a project holding a "formatting must not change the rendered output" gate has to read the plugin's source to learn that this is intended. It is intended: `src/__tests__/fixtures/html-style-element` and `src/__tests__/fixtures/script-with-type` pin it, and `getCompiledText` in `src/__tests__/index.test.ts` skips `script`, `html-script`, `style` and `html-style` before asserting the compiled text is unchanged. README.md's only option section is `markoSyntax` and never mentions any of it. Add a paragraph under `# Options` naming the tag bodies that get reformatted, saying that Prettier's `embeddedLanguageFormatting: "off"` is the only lever and that it is global (prettier 3.9.6 returns from `printEmbeddedLanguages` before running `embed` whenever the option is not `"auto"`, so it also stops formatting `<style>`, `<script>`, `static`, attribute values and `${}` placeholders), and that `<html-style>` ignores `type=` entirely while `<html-script>` with an unrecognised `type=` is left as written.
+
+Check: `grep -n -i 'embedded\|html-style\|html-script\|preserve\|verbatim' README.md` exits 1 with no output. Then `pnpm run build && printf '<html-style type="text/css">\n  body { margin: 0; padding: 0; }\n</html-style>\n<html-style>#incidents-placeholder { display: none; }</html-style>\n' > /tmp/hs.marko && pnpm exec prettier --no-config --plugin ./dist/index.mjs --parser marko /tmp/hs.marko` expands both bodies to one declaration per line, while adding `--embedded-language-formatting=off` returns the input byte-identical.

agent-feedback/items/2026-08-28-keep-parens-around-postfix-bang-attr-value.md

@@ -0,0 +1,12 @@
+---
+type: bug
+impact: high
+effort: med
+site: src/utils/to-valid-doc.ts › toValidAttrValue
+---
+
+# Keep the parentheses around an attribute value that ends in a postfix `!`
+
+`toValidAttrValue` asks htmljs-parser's `isValidAttrValue(code, concise)` whether the printed value has to be wrapped, and that check parses the bare expression to EOF without the `>` terminator, so a value ending in a TypeScript non-null `!` answers `Validity.enclosed` and the function returns the doc untouched, never building the `b.ifBreak("(")` group. htmljs-parser cannot parse that value unparenthesised (`lookBehindForOperator` in its `src/states/EXPRESSION.ts` treats char 33 as a pending operator, so `<const/tenant=input.tenant!>` runs past the `>` and the CompileError lands on the following tag), which makes the parentheses load-bearing rather than cosmetic: `--write` rewrites `<const/tenant=(input.tenant!)>` to `<const/tenant=input.tenant!>`, the template stops compiling, and `--check` then reports the broken file as already formatted. Dropping redundant parentheses is otherwise wanted, and `<div disabled=(1 > 2)>` in the same run keeps its own because that value validates as `invalid`, so the fix is narrow: treat a printed value whose last non-whitespace character is `!` as needing the group. Handing the checker the terminator is not enough on its own, since `isValidAttrValue("input.tenant!>", false)` is also `enclosed`, and a guard keyed on `Validity.valid` would miss the case entirely. No fixture under `src/__tests__/fixtures` puts a postfix `!` at the end of an attribute value, so one belongs with the fix.
+
+Check: `pnpm run build && printf '<const/tenant=(input.tenant!)>\n<const/item=(list[0]!)>\n<div disabled=(1 > 2)>x</div>\n' > /tmp/x.marko && pnpm exec prettier --no-config --plugin ./dist/index.mjs --parser marko /tmp/x.marko` prints `<const/tenant=input.tenant!>`, `<const/item=list[0]!>` and an unchanged `<div disabled=(1 > 2)>x</div>`; compiling that output with `@marko/compiler` at `output: "html"` fails with `Unexpected token, expected "</>/<=/>="` pointing at the next line, while the parenthesised input compiles. Root cause: `node -e 'const {isValidAttrValue,Validity}=require("htmljs-parser");console.log(Validity,isValidAttrValue("input.tenant!",false),isValidAttrValue("input.tenant!>",false),isValidAttrValue("1 > 2",false))'` prints `{ enclosed: 2, invalid: 0, valid: 1 } 2 2 0`.