jp.ParseString in OjG, a Go JSON library with JSONPath, rejected a filter with no space between an operator and a negative number, such as $[?@.a==-1]. I fixed it in PR #252, merged about 22 minutes after I opened it.

When it happens

_, err := jp.ParseString(`$[?@.a==-1]`)
// '==-' is not a valid operation at 9 in $[?@.a==-1]

It should parse as $[?(@.a == -1)]. RFC 9535 section 2.3.5.1 makes the blank space around a comparison operator optional and lets a number start with -. The other comparison operators failed the same way, as did +, -, * and / (@.x+-1, @.x--1, 2*-1, @.x/-1). The spaced form already parsed.

How the operator is read

readEqOp in jp/parse.go reads a run of operator characters, and raises when a - lands in a token longer than one character. So ==- is read as one token and rejected.

Why the gap existed

The check dates from the first version of readEqOp, in June 2020. At the base commit no test in jp had an operator directly before a minus sign and a digit.

The fix

The run now stops at a - that follows a complete operator and is followed by a digit. The whole jp/parse.go diff:

@@ -702,6 +702,9 @@ func (p *parser) readEqOp() (o *op) {
 			if len(token) == 0 || !partialOp(token, b) {
 				break
 			}
+		} else if b == '-' && opMap[string(token)] != nil && p.pos+1 < len(p.buf) && '0' <= p.buf[p.pos+1] && p.buf[p.pos+1] <= '9' {
+			// A minus directly after an operator and before a digit starts a negative number.
+			break
 		}
 		token = append(token, b)
 		if b == '-' && 1 < len(token) {

The token is then == and -1 is the next operand. The len(p.buf) test keeps a trailing ==- an error. - is itself a complete operator, so @.x--1 now parses as @.x - -1; it failed before.

The test

Fifteen rows joined TestParse: eleven that must parse and four that must stay errors (==- 1, ==-.1, a trailing ==-, =-1). Two:

{src: "$[?@.x==-1]", expect: "$[?(@.x == -1)]"},
{src: "[?(@.x==-.1)]", err: "'==-' is not a valid operation at 9 in [?(@.x==-.1)]"},

Without the fix TestParse stops at the first new row; the error rows pass either way. The * row is 2*-1 because @.a*-1 already parsed: @.a* reads as the path @.a.*.

What the maintainer did

@ohler55 approved it with no comment and squash-merged it. The merged files match my branch.

The lesson

  • Write the lookahead from the grammar and no wider: a digit, not just “not a space”. Give each near miss an error row.
  • Run each sibling on the unfixed code before it gets a row. A form can already work for another reason.

If a lexer of yours reads a run of operator characters, reply on X.