Coding Agents Need Typed Context Surfaces, Not One Flat Context Window
Coding agents often struggle with complex tasks because they treat context as a flat combination of various data types like source files and chat history. This approach is insufficient for tasks requiring exact values, structured relationships, or reusable procedures. A more effective solution involves providing coding agents with typed context surfaces, which allows for better organization and utilization of information. This architecture is currently implemented in the open-source Codex MCP plugin OpenDCAI/DataMind.
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年10月9日 08:53 UTC
收录当时偏移：UTC+02026年10月9日 13:00 UTC
- 发布
- 2026年10月9日 08:53
- 收录
- 2026年10月9日 13:00
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 同步延迟
档位是按信源手工设定的编辑判断，不是逐条打分。
Most coding agents treat context as a combination of source files, chat history, configuration files, and retrieved text. That works for simple code completion, but becomes unreliable when a task requires exact values, structured relationships, reusable procedures, or durable project-specific facts.
The problem is that different information types have different access patterns:
- Documents require semantic and keyword retrieval.
- Structured data requires filtering, joins, aggregation, and SQL.
- Relationships require graph traversal rather than text similarity.
- Procedures need reusable, versioned instructions.
- Durable facts need scope, persistence, and controlled updates.
A practical architecture is to expose these as typed context surfaces instead of flattening them into a single vector index:
- KB for documents and notes, with hybrid retrieval.
- DB for structured records and SQL queries.
- Graph for entities, relationships, and bounded traversal.
- Skills for reusable operational procedures.
- Memory for scoped facts, preferences, and durable state.
The agent can route each request to the relevant surface, combine results when necessary, and preserve provenance for every piece of evidence. Read operations and write operations should also remain separate: retrieval returns normalized evidence, while updates return explicit receipts containing the affected source, profile, revision, and surface-level results.
This architecture is currently implemented in the open-source Codex MCP plugin OpenDCAI/DataMind .