Gemini CLI's Product Lead Just Dropped a Roadmap Preview, and It Hints at Gemini 3.1

Dmitry Lyalin, the Product Lead for Gemini CLI at Google, posted a thread tonight listing six features the team is working on. He was clear that none of this is confirmed for a specific release: “when it’s ready it’s ready, and everything is subject to change.” But taken together with the Gemini 3.1 hints from earlier tonight, the timing is hard to ignore.

Let’s go through each one.

1. Model Hinting

“Steer the model during execution.” Gemini CLI already has a configuration option for user hints, but it sounds like this will become a more prominent, interactive feature. The idea is that you can nudge the model mid-task without stopping the whole run and re-prompting. Claude Code has something similar with its interrupt-and-redirect flow, so this would bring Gemini CLI closer to parity there.

2. OS Notifications (macOS First)

“OS notifications when a task is completed or your attention is needed.” This is a small but genuinely useful quality-of-life feature. If you kick off a long task and switch to a browser tab, you’ll get a native macOS notification when it finishes or hits a point where it needs input. Claude Code doesn’t have this natively (you’d need a hook script), so Gemini CLI would be first to ship it as a built-in.

3. Input Bar Customization

“More customization of the elements above/below the input bar.” Not much detail here, but Gemini CLI’s input area already shows context like the current model, token count, and project path. Sounds like they’re making that configurable, maybe letting you add or remove status elements.

4. Clean UI Mode

“Clean UI mode vs. full mode to reduce noise when you want a cleaner interface.” Gemini CLI can be visually busy, especially with verbose tool output. A toggle between a stripped-down view and the full debug-style output makes sense. Claude Code handles this differently (it’s pretty minimal by default), but for Gemini CLI’s more information-dense interface, a clean mode is welcome.

5. Plan Mode Graduating from Preview

“Plan mode will exit preview soon, we also plan to add more advanced planning features.” Plan mode has been experimental in Gemini CLI since v0.29.0-preview.0 shipped on February 10. It’s a read-only mode where the model builds a plan without being able to write files. Taking it out of preview means it’s stable enough for daily use. “More advanced planning features” could mean anything from sub-agents for research to better plan iteration, but the detail is thin.

6. “Gemini 3.x” Performance Tuning

Here’s the line that stands out: “We’ve done a lot of work tuning the coding agent performance, the harness is getting much more optimized with Gemini 3.x.”

He didn’t say Gemini 3. He said Gemini 3.x. Combined with the “G3m1ni” tweet from Ammaar Reshi and the Gemini 3.1 Pro Preview model ID that leaked on a tracker last week, this is another data point suggesting 3.1 is the target model for the next round of Gemini CLI improvements.

The phrasing “the harness is getting much more optimized” is interesting too. It means they’re not just swapping in a better model. They’re tuning how the CLI agent loop interacts with the model: tool calling patterns, context management, how the agent decides what to do next. That kind of work is what separates a good model from a good coding agent.

The “House Is on Fire” Energy

Lyalin ended with: “the team is cooking so much the house is on fire.” The vibe across Google tonight is not subtle. Logan Kilpatrick posted “Gemini.” Ammaar Reshi posted “G3m1ni.” And now the Gemini CLI product lead is talking about 3.x optimizations and a bunch of features in various stages of readiness. Something is landing soon, probably Thursday.

What This Means for the CLI Wars

Gemini CLI, Claude Code, and GitHub Copilot CLI are all racing on the same set of features. These items are at different stages, so don’t expect them all at once, but here’s where Gemini CLI stands when they do ship:

Feature	Gemini CLI (upcoming)	Claude Code (now)
Model steering mid-task	Planned	Yes (interrupt)
OS notifications	Planned (macOS)	No (needs hooks)
Plan mode	Graduating from preview	Yes (stable)
Clean/minimal UI toggle	Planned	Default minimal
Agent performance tuning	Active (“3.x” optimized)	Active (Opus 4.6)

The notification feature and clean UI mode are nice differentiators. But the real question is whether “Gemini 3.x” closes the gap on coding performance. Right now Claude Code with Opus 4.6 is the benchmark for real-world software engineering tasks. If Gemini 3.1 Pro is the model behind “3.x” and it ships alongside a tuned agent harness, that’s the matchup to watch. Just don’t expect it all to drop at once.