DEV Community

Cover image for VS Code 1.147: Send Later, Drafts, Model Routers, and Visual Chat Responses Explained
Muhammad Hamid Raza
Muhammad Hamid Raza

Posted on

VS Code 1.147: Send Later, Drafts, Model Routers, and Visual Chat Responses Explained

You have a task in mind, but you're not ready to send it to an agent yet. Maybe the build is still running. Maybe you want to sleep on the wording first. So you leave the prompt sitting in the box and hope you don't lose it. šŸ˜…

VS Code 1.147 (released November 18, 2026) tackles that situation directly. This update is mostly about the Agents window, and it makes the "not yet, but soon" part of working with agents much smoother.

Which of these would save you the most time? Let's go through them.


What Is VS Code 1.147?

VS Code 1.147 is the latest release of Visual Studio Code. Think of it like a new school term: same school, a few new tools in your bag.

The headline themes are:

  • Send later: schedule a request for a later time
  • Drafts: save unsent requests and come back to them
  • Model routers: let a decision model pick which of your models handles a task
  • Visual responses: see charts, reports, and mockups inline in chat
  • Customizable assisted permissions: tell the reviewer what you consider safe

Many of these are Experimental or Preview. They may change, and some are off by default. Always check the setting name before you go looking for a feature.


Why VS Code 1.147 Matters

Working with agents isn't always "type, send, done." Real work has gaps. You wait for builds, you think before sending, and you follow up the next morning.

This release helps with that rhythm. It also gives you more control over how the agent works:

  • Which model handles what
  • Which actions need your approval
  • How answers are shown to you

Less retyping, less guessing, and fewer "where did my prompt go?" moments. šŸš€


Benefits with Real-Life Examples

1. Send Later (Experimental)

Setting: sessions.sendLater.enabled (Agents window only)

You can postpone a chat request until a specific time. Use the Send Later menu on the send button, or pick Choose Date and Time for an exact slot. You can also just ask the agent in plain language, like "Check the build in 10 minutes."

Example: Your CI build takes a while. Schedule a follow-up request to check on it instead of keeping a reminder in your head.

Good to know:

  • The local agent host must be running at the scheduled time.
  • Scheduling doesn't wake a sleeping computer.
  • A missed request is processed once the host is available again.
  • It works only for sessions on This Computer and requires chat.automations.enabled.

2. Drafts (Experimental)

Setting: sessions.drafts.enabled

Before, the new-session input held one unsent request at a time. Now you can keep several. Each draft remembers its text, attachments, inline references, workspace, agent, model, and other configuration. Drafts survive reloads and restarts.

Example: You're halfway through writing a refactor request when a bug report arrives. Leave the new session, and your draft is saved in the Drafts section. Nothing is sent.

A draft is removed only after the request is accepted, so a failed send doesn't cost you your work. šŸ‘

3. Model Routers (Experimental)

Setting: chat.copilot.modelRouter.enabled

A model router is a YAML file in ~/.agents/model-routers/. You describe which model you want for which kind of work. A decision model, gpt-6-luna-decisions, matches each request to one of your routes, and VS Code runs the chat on that route's model.

Here's the example from the release notes:

version: 1
description: Opus to build, GPT-6 Luna for quick questions.
routes:
  - model: opus
    when: Building or changing code over many steps, such as a feature, a bug fix, or a refactor.
  - model: gpt-6-luna
    when: Quick answers and small chores, such as where something is or what a setting does.
Enter fullscreen mode Exit fullscreen mode

Example: A big refactor goes to one model, and "what does this setting do?" goes to another.

To author one, type /model-router in a Copilot harness chat and describe how you want requests routed. Then pick the router in the model picker. The team says routers are at an early stage and wants feedback.

4. Visual Responses in Chat

Settings: chat.htmlPreview.enabled, chat.nativeComponents.enabled (Agents window only, both off by default)

Some answers are easier to see than read. There are two options:

  • HTML previews: for reports, diagrams, and mockups. Try /html-preview compare the two caching approaches. Scripts, forms, external links, and external resources aren't supported.
  • Native visuals (Experimental): for data such as metrics, badges, tables, and bar charts, rendered by VS Code so they match your theme. Try /vscode-ui compare the full and incremental build times.

Example: Instead of a wall of numbers about build times, you get a small chart right in the response.

5. Customizable Assisted Permissions

Setting: chat.permissions.assisted.instructions

With assisted permissions, an LLM reviewer checks each tool call in Copilot harness sessions and approves low-risk ones. Now you can tell it your own rules:

  • Always ask before installing packages or pushing to a remote.
  • Running tests and read-only Git commands doesn't need my approval.

Your instructions take priority over the reviewer's default guidance. They can't override permission rules, organization policy, or the sandbox, and they aren't sent to the agent.


Comparison: Send Later vs. Drafts

These two sound similar, but they solve different problems.

Drafts Send Later
What happens Request is saved, not sent Request is sent at a time you choose
Best for "I'm not ready yet" "Not now, but at this time"
Where it lives Drafts section of the sessions list Drafts (new session) or the chat's Scheduled area (existing chat)
Needs the host running? No sending happens Yes, at the scheduled time

Use a draft when you're still thinking. Use Send Later when you already know when.


Best Tips: Do's and Don'ts

Do āœ…

  • Check the exact setting name before searching for a feature. Many are off by default.
  • Start model routers with two simple routes, like the example above.
  • Write short, clear permission instructions in plain language.
  • Use Scheduled Only in the sessions filter to find pending requests.
  • Send feedback on experimental features. The team asks for it.

Don't āŒ

  • Don't expect Send Later to wake a sleeping computer.
  • Don't expect HTML previews to run scripts or submit forms.
  • Don't assume your permission instructions can override organization policy or the sandbox. They can't.

Common Mistakes People Make

1. Expecting a feature that is switched off. Several features here are experimental and off by default. If you don't see them, check the setting first.

2. Scheduling a request and closing the laptop. The local agent host has to be running. Scheduled requests work only for sessions on This Computer and need chat.automations.enabled.

3. Writing vague routes. A route like "hard stuff" gives the decision model little to work with. Describe the kind of work, like "building or changing code over many steps."

4. Forgetting that automation changes approvals. Requests started by an automation or another agent are no longer treated as if you wrote them. Tool calls there might ask for approval where they didn't before.

5. Ignoring the deprecations. VS Code desktop on macOS Ventura and Sonoma, and on 32-bit ARM Linux, is deprecated. The Local harness is deprecated and will be removed in February 2027. The experimental AI usage statistics in the Status Bar (editor.aiStats.enabled) were removed.


A Few More Things Worth a Look šŸ‘€

  • Browser Use (Preview): lets agents work in your default browser (Chrome or Edge). It needs the GitHub Copilot browser extension, a local Copilot session, and Node.js 22 or later.
  • Remote HTML preview (Preview): open HTML files from SSH, WSL, or Dev Container workspaces in the integrated browser with workbench.browser.enableRemoteProxy.
  • Panel corners and spacing: customize them with workbench.sizeCustomizations.
  • Markdown editor upgrades: edit tables cell by cell, and see <details> sections as collapsible blocks.

Conclusion

VS Code 1.147 makes agent work feel less rushed. You can save drafts, schedule requests, route tasks to preferred models, see visual answers, and tell the permission reviewer how you work. Many of these are still experimental, so try them gently and share feedback.

If this helped, read more developer posts on hamidrazadev.com, and share this with a teammate who still keeps prompts in a sticky note. 😊

Which feature will you try first: Send Later, Drafts, or Model Routers? Let me know in the comments!

Top comments (0)