ollama-autolisted

# ollama-auto — Smart Ollama Delegation + Audit Delegates cheap mechanical tasks to local Ollama instead of burning Claude tokens. Logs every call. Audit mode shows if it's actually saving tokens. --- ## Auto-Trigger Rules Claude MUST invoke Ollama automatically (without user asking) when ALL conditions are true: | Condition | Threshold | |-----------|-----------| | Task is mechanical (classify, sort, format, filter, describe) | Yes | | No codebase context needed | Yes | | Input is plain text / list / JSON | Yes | | Output doesn't need Claude reasoning | Yes | **Auto-trigger examples:** - Scan result with 30+ folders → classify each as `bloatware / app / game / dev / media / system` - winget list → tag each as `keeper / remove / update` - List of file paths → group by type/purpose - Raw command output → extract only relevant lines - Template generation with repetitive structure **Never auto-delegate:** - Tasks needing repo/file context - Debugging or architecture decisions - Image/video analysis - Anything requiring Claude reasoning or judgment --- ## How to Call Ollama ```bash # Simple prompt curl -s http://localhost:11434/api/generate \ -d "{\"model\":\"mistral\",\"prompt\":\"PROMPT_HERE\",\"stream\":false}" \ | python -c "import sys,json; print(json.load(sys.stdin)['response'])" ``` ```bash # Long prompt via temp file (avoids shell escaping hell) cat > /tmp/ollama_prompt.json << 'EOF' { "model": "mistral", "prompt": "YOUR LONG PROMPT HERE", "stream": f