You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fix Google provider for large specs and improve JSON robustness
- google.go: set ResponseMIMEType=application/json to suppress code fence
wrapping; create genai.Client per-call using caller context with defer Close()
- llm.go: handle orphaned opening fence (truncated fenced responses) in
stripMarkdownFences; add fixInvalidJSONEscapes sanitizer for \d/\w/\s
regex patterns LLMs embed unescaped in JSON strings
- main.go: update default Google model to gemini-2.5-flash (65k output
tokens vs 8k for gemini-2.0-flash, required for large spec/plan files)
- README.md: update default model reference for Google provider
Smoke tested all three providers against the full specs/SPEC.md + specs/PLAN.md:
anthropic DRIFT_DETECTED score=78 (134s)
openai VIOLATION score=60 ( 10s)
google VIOLATION score=73 (117s)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
0 commit comments