[{'Text': '```\n{\n  "body": "## Summary\\n\\nTransform the ontology from \\"one of six datasets an agent can load\\" into the agent\'s default way of understanding broadband data. Implement a layered context architecture where the agent starts every conversation already knowing the conceptual model.\\n\\n## The Problem\\n\\nCurrently the agent must call `assemble_dataset_context([\'ontology\'])` to load ~4,000 tokens of context before it can query ontology data. The ontology summary visible in tool descriptions is a capability listing (\\"you can ask about ISPs, networks, etc.\\") rather than a mental model. The agent knows *what topics exist* but not *how to think about them*.\\n\\n## The Architecture\\n\\n**Layer 0 — Always visible (tool descriptions):**\\nRewrite `get_dataset_summary()` for ontology to convey the mental model: entity chain (ISP → Network → Link), measurement pattern (`cto_measurement_record`), schema navigation, gotchas (varchar dates/magnitudes). ~500-800 chars, always present in the agent\'s view.\\n\\n**Layer 1 — Cheap discovery tools:**\\nNew lightweight tools that run single SQL calls: `get_aspect_names()`, `get_entity_names()`, `get_link_types()`, `get_isp_networks()`. Their descriptions add another thin layer of orientation.\\n\\n**Layer 2 — Full context (optional):**\\n`assemble_dataset_context` stays but becomes a fallback. `execute_query` no longer mandates it.\\n\\n## Scope\\n\\n- Rewrite ontology `get_dataset_summary()` as mental model\\n- Rewrite ontology `get_db_info()` with improved content\\n- Add lightweight discovery tools in `ontology_tools.py`\\n- Soften `execute_query` mandatory context loading\\n- Fix duplicate SQL example (example 4 = copy of example 3)\\n- Leave BI reports untouched\\n\\n## Context\\n\\nThis connects the ontology (under construction 1+ years) with the MCP server (working for months) — making the ontology the agent\'s native language rather than a translation layer.",\n  "createdAt": "2026-04-06T11:08:01Z",\n  "state": "OPEN",\n  "title": "Ontology as the agent\'s mental model — layered context architecture"\n}\n```'}]