1. One hand-written parser for six dialects, under a contract with check-json5 1.4

Status: accepted (2026-10-13)

Context

jsonvet 2.0 replaced check-json5 1.4 in place. A linter needs what a deserializer skips: an exact position for every problem, and every duplicate key with both positions. 1.4's users also need the hook to pass and fail the same files.

Options

• Build on serde_json, json5 or jsonc-parser. On 2026-10-13 every one turned {a:1,a:2} into {"a": 2}, so none could report a duplicate. jsonc-parser is lax by design, and json5 1.3.1 went quadratic on hex numbers.
• One parser per dialect. Six grammars to keep in step.
• One byte scanner for JSON, I-JSON, JSONC, JWCC, JSON5 and JSON Lines, with the crates as fuzz oracles.

Decision

One scanner, built for linting. Verdicts, exit codes and duplicate-key reports stay equal to 1.4; syntax-error wording is ours, marked Breaking.

Consequences

• We own every edge case, so the parser has a corpus, json5-tests (114 of 114), proptest and fuzz targets.
• All 40,228 differential runs agreed with 1.4.1 at 2.0 (#34); the 385 intended differences since are listed in design.md.

Confirmation

The differential CI job on every pull request, and the weekly fuzz run against serde_json and json5.

Revisit if

A maintained crate offers strict dialect modes with offsets and duplicate reports and passes our corpus, or 4.0 drops check-json5 compatibility.