Rewrites the JSON repair algorithm to address two correctness bugs
identified in code review:
1. Interior quoted word before comma: "say "yes", and go" was
misidentified as structural close because "," followed the quote.
2. JSON-shaped content in strings: {"key": "val"} inside a string
value was being parsed as actual JSON structure.
The new approach:
- Distinguishes keys from values (only values need repair)
- Uses first-valid-candidate scan with deep lookahead
- Verifies that after a candidate close, the continuation is not just
a structural char but a complete valid JSON pattern
- Validates keyword tokens (true/false/null) fully, not just first char
- Checks container closes recursively for valid continuation
Adds comprehensive tests for all reported edge cases plus a complex
combined scenario with nested JSON-like content, quoted words before
commas, and multiple failure modes in one string.
LLMs (especially Sonnet) sometimes emit JSON with unescaped double
quotes inside string values, e.g. (e.g. "28") instead of properly
escaping them. This caused parse failures in CI.
Add a repairJSON fallback that uses a character-by-character scanner
to identify interior quotes (those not followed by structural JSON
characters) and escape them before retrying the parse.
Fixes sonnet-review failures on gargoyle PR #551.
- CLI binary with flag/env var configuration
- Gitea API client (PR metadata, diff, CI status, post review)
- OpenAI-compatible LLM client
- Structured review prompt with conventions support
- JSON response parser with validation
- Markdown review formatter for Gitea
- CI failure auto-detection (REQUEST_CHANGES)
- Dry-run mode for testing
Rodin