- SakeAnalysisResult.isFromCache flag added (not serialized to JSON)
- Both cache-hit paths return result.asCached() to signal caller
- camera_screen: EXP +10 only awarded on fresh API calls, not cache hits
- camera_screen: show '解析済みの結果を使用（経験値なし）' on cache hit
- camera_screen: clear _capturedImages after successful analysis
- camera_screen: catch(_) -> catch(e) with debugPrint logging
- SakeAnalysisResult.fromJson: auto-fill missing tasteStats keys with 3,
clamp all values to 1-5 range to prevent broken charts
- Bump version 1.0.37+44 -> 1.0.38+45
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- name, brand, prefecture: still OCR-strict (no completion/inference)
- alcoholContent, polishingRatio, tasteStats etc: restored from label + sake knowledge
- Overly strict 'no knowledge at all' rule was causing all detail specs to return null
- Sync fallback prompt with same policy
- Bump version 1.0.36+43 -> 1.0.37+44
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Rewrote prompt: 'ソムリエ' role -> pure OCR system role
- Explicit prohibition on brand name completion/correction
(e.g. label shows 東魁 -> output 東魁, never 東魁盛)
- prefecture: output null if not written on label; inference from brand knowledge forbidden
- tasteStats: default to mid-value 3 instead of 'reasonable estimate' (was encouraging hallucination)
- Added concrete examples for white-deer, kubota, tokai cases
- Bumped version 1.0.35+42 -> 1.0.36+43
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
secrets.dart: change AI_PROXY_URL defaultValue from 'http://100.76.7.3:8080'
to empty string. Consumer APKs use useProxy=false and never reach this code
path, so there is zero functional impact. Internal network topology is no
longer embedded in distributed binaries.
Made-with: Cursor
Analysis cache redesign:
- Remove Hive persistence from AnalysisCacheService entirely
- Use Map<String, SakeAnalysisResult> in-memory instead of Hive boxes
- Cache now lives only for the duration of an app session; restart always
produces a fresh analysis — no stale misidentifications can persist
- Remove Hive init(), TTL/cleanupExpired(), getCachedByBrand() dead code
- API surface unchanged: callers (gemini_service, dev_menu) need no edits
- main.dart: delete legacy Hive boxes (analysis_cache_v1, brand_index_v1)
from disk on startup for existing users
- dev_menu_screen: update cache description text to reflect new behavior
Rationale:
- Camera captures always produce unique files -> cache hit rate was ~0%
- Each user supplies their own Gemini API key -> no shared cost benefit
- Persistent wrong results (e.g. misrecognized brand names) could survive
up to 30 days under the old design
- Different sake editions photographed separately have different hashes
and were never affected by the cache in the first place
Made-with: Cursor
✨ Improvements:
- Replace print() with debugPrint() (7 instances)
- Migrate withOpacity to withValues (9 instances)
- Remove unused imports (2 files)
- Fix BuildContext async gaps with mounted checks
- Remove unused local variables
📊 Analysis Results:
- Flutter analyzer: 57 → 46 issues (-11)
- Security audit: Passed ✅
- Code quality: Production ready ✅🎯 Purpose:
Complete snapshot before QR feature removal.
This is the last version with full QR functionality.
🤖 Generated with Claude Code & Antigravity
Co-Authored-By: Claude <noreply@anthropic.com>
Ponshu Developer