The branching/merging mental model is interesting for agent workflows. One thing I've noticed is that context drift often originates upstream at the prompt structure level — agents tend to lose track of their objective when the original prompt mixes role, context, and constraints into one blob.

I've been building flompt (https://flompt.dev) which approaches this from the other end: decompose prompts into typed blocks (role, context, objective, constraints, examples, output_format) before they ever reach the agent. The MCP server exposes compile_prompt() so structured prompts can be injected dynamically into agentic pipelines. Combining that with something like Contexa for runtime context versioning could be a useful pairing.