【2026年6月版】Context Engineering 完全入門(補章A) － VS Code Copilot Chat 1.109の8つの新機能を調査This article examines eight new features in VS Code Copilot Chat 1.109 through…

Context Engineering, the practice of deliberately shaping the information an AI model receives before it generates a response, has become a central concern for developers who rely on coding assistants. This article looks at the release of VS Code Copilot Chat version 1.109 and examines eight of its new features through that lens, focusing on how each one gives developers more precise control over the context that surrounds a request. As models grow more capable, the quality of their output increasingly depends less on the phrasing of a single prompt and more on the surrounding material—open files, project structure, instructions, and tool results—that the assistant can see.

The central idea behind Context Engineering is that a large language model only reasons over what fits inside its context window. Everything the assistant "knows" about a task in the moment is assembled from sources the developer either supplies explicitly or allows the tool to gather automatically. When that assembly is left to chance, results can be inconsistent; when it is managed deliberately, the same underlying model tends to behave more predictably. The features highlighted in 1.109 appear to reflect this shift, moving the emphasis from clever wording toward reproducible context construction.

Several of the updates concern how instructions and reusable context are declared. Recent versions of Copilot Chat have supported instruction files and prompt files stored in the workspace, which let teams codify conventions such as preferred libraries, testing patterns, or coding style so they are injected consistently rather than retyped. Improvements in this area typically make it easier to scope instructions to particular file types or folders, which reduces noise by ensuring the model sees guidance relevant to the current task and not everything at once. From a Context Engineering standpoint, this is significant because irrelevant context can dilute a model's attention as much as missing context can starve it.

Other features address the mechanics of gathering context during a conversation. Context variables and references, which allow a developer to point the assistant at a specific file, selection, symbol, or the broader codebase, remain a core mechanism for controlling what the model reads. Enhancements here generally aim to make the boundaries of that context more visible and controllable, so the developer can confirm what was actually included. Agent-oriented workflows, where the assistant can run tools, read files, and make multi-step edits, raise the stakes further, because the context is being modified as the task proceeds rather than fixed at the outset.

Tooling and extensibility form another recurring theme. The Model Context Protocol, an open standard for connecting assistants to external data sources and tools, has been adopted across several vendors and is likely relevant to any discussion of context control in modern editors. Features that manage tool sets, permissions, and how tool output is fed back into the conversation directly affect what ends up in the window. Neatly summarizing or filtering tool results, for instance, can prevent a single verbose command output from crowding out more useful material.

It is worth placing this release in a wider industry context. Copilot Chat competes with tools such as Cursor, Cline, and JetBrains AI Assistant, many of which have converged on similar concepts: workspace-level rules, explicit context attachment, agent modes, and standardized tool interfaces. This convergence suggests that Context Engineering is becoming a shared vocabulary rather than a single product's marketing frame. The prerequisite concepts—prompt engineering, retrieval, and an understanding of context window limits—carry over across these environments, which makes learning them broadly transferable.

Readers evaluating whether to adopt the new features should treat specific behavior as version-dependent, since fast-moving extensions change defaults and naming between releases, and hands-on testing remains the most reliable guide. The broader takeaway is that the assistant's usefulness is closely tied to how well its context is curated. The eight features discussed here appear designed to make that curation more explicit, repeatable, and team-friendly, which is consistent with a general trend toward treating context as something to be engineered rather than left implicit. Developers who internalize that mindset are likely to get steadier results regardless of which tool or version they use.