[brynary]

Stacked on #45.

The grammar in #45 needed a locator rule for each kind of line and many lookaheads. The cause was the syntax: one line could split in more than one way, and the parser had to look ahead to choose. This PR changes the syntax so that the parser reads each line left to right, one token at a time. PRESS is the only line that still looks one token ahead.

This changes existing flows. Whirl has no users to migrate yet, so there is no compatibility mode. The sample flows, SPEC examples, eval caches, and tests use the new syntax.

Locators

Before	After
CLICK role:button "Sign in"	CLICK button:"Sign in"
CLICK role~:button Sign	CLICK button:~Sign
ASSERT role:dialog visible	ASSERT dialog:* visible
ASSERT text~:Added visible	ASSERT text:~Added visible

• Every ARIA role that Playwright accepts is its own prefix, except generic, none, and presentation: 79 roles. role: is gone.
• * matches any name. A quoted "*" is the literal name.
• The ~ of a substring match follows the colon, for roles and the five text prefixes alike.
• Prefixes are strict. In a locator, a bare colon always marks a prefix, and an unknown prefix is a parse error. Quote unprefixed text that holds a colon. >> is never a segment; quote it to match it as text.

Role options such as level are not in this PR. A later form could be heading[level=2]:"Welcome back".

Windows

Browser tabs are windows, so tab is free for the ARIA role:

Before	After
TAB payment	WINDOW payment
ASSERT tab:payment closed	ASSERT window:payment closed

The lint codes are duplicate-window and unknown-window. The shim protocol does not change; its wire command is still named tab.

Tokens

• A bare value cannot start with @. A bare @ token is always the step timeout and must end its line. Quote a value such as "@60s".
• A name: value line needs a space after the colon: in options, CAPTURE lines, headers, and snapshot settings.
• A quoted string joins a token only right after a prefix, as in label:"First name" or button:~"Sign in". This is also true for json: and xpath:, which had a rule of their own.
• A # starts a comment only at the start of a line or after white space. Before, a # inside a token started a comment. So VISIT /docs#install visited /docs, and ASSERT url == https://shop.example.com/#/cart compared against the URL without its fragment, and whirl check passed both. Now css:#submit and URL fragments work without quotes.
• After ACT, EXTRACT, or JUDGE, a first token with a prefix starts the scope. DRAG needs no quotes on a to that is text. After SCROLL, a bare direction or to is always the motion.

Removed

The [Asserts] and [Captures] sections are gone. They are unknown sections, whirl fmt no longer rewrites them, and the sections-removed and mixed-check-syntax errors are gone.

Result

The grammar in SPEC section 17 went from 107 rules and 46 lookaheads to 91 rules and 31 lookaheads, and that includes the new list of 79 roles. Its locator rules no longer depend on the kind of line. parse.rs is about 100 lines shorter.

Commits

1. Simplify the syntax so every line reads left to right.
2. Start a comment only at a # after white space.
3. Let json: and xpath: join a quoted string like other prefixes.

Validation

• mise run check.
• The routine grammar test compares the grammar with the parser on every sample and about 60,000 changed copies.
• The nightly grammar sweep (about 480,000 changed copies) passed locally on the last commit.
• For the first commit, I ran whirl check on 185 sample files and SPEC examples: the old build on the original files, and the new build on the converted files. 182 gave the same result, down to each diagnostic's code and line. The other 3 were tests of the removed sections, which this PR deletes.
• CI below.

🤖 Generated with Claude Code