l6e: Enforce budgets for AI agents to optimize cost and performance

For too long, AI coding agents have operated with a 'blank check' mentality, leading to runaway token spend and inefficient context usage. l6e addresses this by introducing a budget primitive via the Model Context Protocol (MCP). Instead of simply reporting costs after the bill arrives, l6e inserts itself as a gatekeeper. Every expensive operation—be it a codebase search or a complex refactor—must be authorized by the budget server, which returns a status of allow, reroute, or halt based on the remaining session funds. The technical elegance here is that l6e doesn't just stop the agent; it changes the agent's internal reasoning. When a model is told it has a $3 budget, it naturally shifts toward targeted file reads and tighter planning to avoid hitting the hard cap. This 'budget pressure' effectively optimizes the agent's performance by discouraging lazy, exhaustive searches in favor of precision. The inclusion of a calibration engine that ingests actual billing CSVs to tune estimates is a practical touch that moves the tool from 'rough guess' to 'financial tool'. From a product standpoint, the friction is remarkably low. Because it leverages MCP, there is no need for a proxy or a complex SDK integration; it's a simple pip install and a config line. However, the system's effectiveness relies heavily on the model's ability to respect the MCP tool's output. While frontier models like Claude 3.5/4.6 are generally compliant, the 'behavioral' aspect of the constraint may vary across different LLMs. This is a must-have for developers using high-cost frontier models (like Opus) who are tired of unexpected API bills. It transforms the agent from a potential liability into a managed resource. If you are automating large-scale refactors or building custom agent pipelines via LangChain or CrewAI, l6e provides the missing circuit breaker for your wallet.