Google supercharges Gemini Code Assist with Gemini 2.5 and 1M context

If you write code in VS Code or JetBrains, you know the drill: coding assistants either forget the context three files in, or keep generating code that completely ignores your team's style guide.

Google just rolled out a major overhaul to Gemini Code Assist to fix exactly that. The assistant is now powered by the Gemini 2.5 model family, paired with new controls for repo context and project rules.

Individual developers can try the updates for free right now. Here is what is under the hood.

Gemini 2.5 and custom rules

The headline update is the switch to Gemini 2.5, which currently sits at the top of the LMArena leaderboard. Google says this delivers sharper reasoning for debugging, cleaner code transformations during refactoring, and a more thorough pull request reviewer for GitHub.

More interesting for day-to-day work, though, are the new personalization tools.

You can now set persistent "Rules" in your IDE settings. For instance, you can tell it to "Always generate unit tests when creating a new function," and Gemini will follow that constraint across every prompt without you having to retype it.

There is also a new Custom Commands feature. You can map complex scaffolding prompts or boilerplate patterns to custom shortcuts and trigger them right from the quick-pick menu.

Visual context management and 1M tokens

Feeding context into chat used to be hit-or-miss. Gemini Code Assist now lets you include entire folders or your full workspace in prompts using the "@" operator, backed by a massive 1M token context window.

To keep track of what the AI is actually reading, Google introduced a Context Drawer panel. It displays every active file and folder in the conversation, letting you add or drop files with a single click.

The chat interface also gained multi-session support, meaning you can start separate threads for different bugs and resume them later without losing your chat history.

Why it matters

For everyday developers, AI tools succeed or fail on predictability, not just benchmark scores. Adding persistent rules and a visual context manager directly tackles the frustration of AI ignoring codebase conventions or silently hallucinating on unread files. Since it is free for individual users in VS Code and JetBrains, trying it out on a medium-sized project is an easy call.

Across the industry, this update puts direct pressure on GitHub Copilot and Cursor. Cursor won over developers primarily through superior local codebase indexing and system prompts. Google is now bringing that exact playbook to its native extensions, backed by a 1M token context that few rivals can match on cost.

We think Google made the right trade-off by focusing on developer UX instead of pitching flashy, hands-off autonomous agents. The risk, however, is that massive context windows often tempt developers to dump entire repositories blindly into prompts, which can still dilute accuracy if the model fixates on the wrong files.