cache-cost-tracking

# Cache Cost Tracking

Monitor LLM costs and cache effectiveness.

## Langfuse Automatic Tracking

```python
from langfuse.decorators import observe, langfuse_context

@observe(as_type="generation")
async def call_llm_with_cache(
    prompt: str,
    agent_type: str,
    analysis_id: UUID
) -> str:
    """LLM call with automatic cost tracking."""

    # Link to parent trace
    langfuse_context.update_current_trace(
        name=f"{agent_type}_generation",
        session_id=str(analysis_id)
    )

    # Check caches
    if cache_key in lru_cache:
        langfuse_context.update_current_observation(
            metadata={"cache_layer": "L1", "cache_hit": True}
        )
        return lru_cache[cache_key]

    similar = await semantic_cache.get(prompt, agent_type)
    if similar:
        langfuse_context.update_current_observation(
            metadata={"cache_layer": "L2", "cache_hit": True}
        )
        return similar

    # LLM call - Langfuse tracks tokens/cost automatically
    response = await llm.generate(prompt)

    langfuse_context.update_current_observation(
        metadata={
            "cache_layer": "L4",
            "cache_hit": False,
            "prompt_cache_hit": response.usage.cache_read_input_tokens > 0
        }
    )

    return response.content
```

## Hierarchical Cost Rollup

```python
class AnalysisWorkflow:
    @observe(as_type="trace")
    async def run_analysis(self, url: str, analysis_id: UUID):
        """Parent trace aggregates child costs.

        Trace Hierarchy:
        run_analysis (trace)
        ├── security_agent (generation)
        ├── tech_agent (generation)
        └── synthesis (generation)
        """
        langfuse_context.update_current_trace(
            name="content_analysis",
            session_id=str(analysis_id),
            tags=["multi-agent"]
        )

        for agent in self.agents:
            await self.run_agent(agent, content, analysis_id)

    @observe(as_type="generation")
    async def run