fix: resolve screenshot, Ollama image/audio analysis, and build failures

[dafandikri]

Summary

Fixes 7 open issues by addressing 5 root causes across the screenshot pipeline, LLM integration, and build system.

Issues Resolved

Fix	Issues Closed	Root Cause
Screenshot ENOENT on macOS	#53, #10	screenshot-desktop strips spaces from filenames — "Application Support" → "ApplicationSupport". Switched to app.getPath("temp")
Image analysis crash (Ollama)	#41, #43	When USE_OLLAMA=true, this.model (Gemini) stays null. All vision methods called this.model.generateContent() without an Ollama fallback
Audio analysis crash	#6	Same null model issue. Added explicit error since Ollama doesn't support audio natively
Build fails	#35	Build script deleted dist-electron/ then never recompiled electron code. Added tsc -p electron/tsconfig.json to build
Ollama improvements	#25	Added vision support via images param, robust JSON parsing, default to llava model

Changes

electron/ScreenshotHelper.ts

• Use app.getPath("temp") instead of app.getPath("userData") to avoid spaces in path
• Add { recursive: true } to mkdirSync for safety
• Add file existence validation after capture with descriptive permission error

electron/LLMHelper.ts

• Add Ollama fallback to all 6 methods that used this.model.generateContent():
  • extractProblemFromImages — uses Ollama vision API with images param
  • generateSolution — uses callOllama()
  • debugSolutionWithImages — uses Ollama vision API
  • analyzeImageFile — uses Ollama vision API
  • analyzeAudioFile / analyzeAudioFromBase64 — clear error (Ollama can't do audio)
• Extended callOllama() to accept optional images parameter for vision models
• Added parseJsonSafe() for robust JSON extraction from LLM responses
• Updated default fallback model from gemma:latest to llava:latest (vision-capable)

package.json

• Fixed build script: "build": "npm run clean && tsc -p electron/tsconfig.json && tsc && vite build"

Test plan

• tsc -p electron/tsconfig.json — compiles with 0 errors
• tsc --noEmit — frontend type-check passes
• vite build — builds successfully, dist-electron/main.js exists
• Screenshots save to temp directory (verified: files created at /var/folders/.../T/)
• llava vision model correctly analyzes screenshots via Ollama API
• screencapture verified working from CLI (4.9MB file) and Node.js (3.3MB file)

Notes for reviewers

• Vision model required: Users with Ollama need a vision-capable model (llava, llama3.2-vision) for image analysis. llama3.2 (text-only) won't work.
• Audio still requires Gemini: Ollama doesn't support audio analysis. Users get a clear error message directing them to configure GEMINI_API_KEY.
• The screenshot-desktop library bug (space stripping in filenames) is upstream — using temp dir is the workaround.

🤖 Generated with Claude Code