v0.6.0 — Lenient JSON parsing for LLM-emitted JS-literal style

Added

• Lenient JSON parsing fallback for every db positional JSON argument. When strict JSON.parse rejects the input, the plugin retries with a permissive parser (relaxJson) that handles three idioms LLMs commonly emit:

  • Bareword keys: {$gt: 1950} → {"$gt": 1950} (the recurring benchmark trap)
  • Single-quoted strings: {'name': 'Alice'} → {"name": "Alice"}
  • Trailing commas: {a:1,} and [1,2,] survive

  Hand-rolled char-by-char tokenizer (~75 LOC, zero dependency). Only transforms tokens outside string literals — {"x": "key: value"} and {"q": "$gt: 5"} are unchanged through both passes.

• 21 new unit tests in tests/lib/relax-json.test.ts + 7 new integration tests in tests/db.test.ts. Total suite: 167/167.

Why

The 8-model agent benchmark identified Llama 3.2 3B and Llama 3.1 8B FP as consistent emitters of {$gt: 1950} JS-literal style. v0.6.0 closes that gap without a JSON5 dependency or any change to strict-JSON behavior.

Compatibility

Zero breaking changes. The relaxer never runs on input that JSON.parse accepts directly. True garbage still produces exit 2.

Caveats

• Narrow relaxer; no JSON5 comments, hex numbers, Infinity, etc. are supported.
• Bareword true / false / null are recognized as JSON literals (not quoted as keys).