The three channel-finder pipelines The three pipelines are not size tiers of one design — each takes a differently shaped channel database, resolves a query by a different mechanism, and is exposed through its own MCP server. In-context puts the whole database in one prompt; hierarchical selects recursively level by level; middle layer runs a React agent over query tools. SELECTION pipeline_mode set → that pipeline. unset → the first pipeline with a database configured, preferring middle_layer › hierarchical › in_context in_context hierarchical middle_layer DATABASE SHAPE flat JSON standalone entries plus template entries per family nested tree levels + naming_pattern {system}:{family}[{device}]… MML functional System → Family → Field → ChannelNames QUERY MECHANISM one LLM call the whole database sits in the system prompt no splitting, no retry recursive selection the model picks once per level: system → family → device → field → subfield React agent loop calls query tools — list_systems, inspect_fields, list_channels, run_sql MCP SERVER channel_finder_in_context channel_finder_hierarchical channel_finder_middle_layer