Seamless Agent

Seamless Agent enhances GitHub Copilot by providing interactive user confirmation tools. It allows AI agents to ask for user approval before executing actions, ensuring you stay in control.

Features
A Language Model tool that enables Copilot to interactively prompt you for confirmation or additional input during chat sessions.
- User Confirmation — Get explicit approval before Copilot executes critical actions
- Interactive Input — Provide additional context or instructions mid-conversation
- Task Validation — Confirm whether a task was fulfilled your specs
- Seamless Integration — Works naturally within the Copilot Chat workflow
A Language Model tool that presents a multi-step plan in a dedicated review panel, so you can approve it or request changes with comments anchored to specific parts of the plan.
- Plan Review Panel — Review the plan in a focused editor-like view
- Targeted Feedback — Add comments tied to specific headings/paragraphs/list items
- Structured Output — Returns
{ approved, comments: [{ citation, comment }] } to the agent
- Safer Automation — Prevents execution before you approve the approach
Usage
Once installed, the ask_user and approve_plan tools are automatically available to GitHub Copilot Chat.
Automatic Usage
Copilot will automatically use this tool when it needs your confirmation. When triggered:
- A notification appears in VS Code
- Click "Open Console" to open the request panel
- Type your response
- Copilot continues based on your input
Reviewing a Plan with approve_plan
Copilot will use this tool when it wants your sign-off on a plan before proceeding. When triggered:
- A “Review Plan” editor panel opens
- Hover a heading/paragraph/list item and click the comment icon to add feedback
- Click Approve to proceed, or Request Changes to send feedback back to the agent
- Copilot continues based on
{ approved, comments }
Tips
Recommended System Prompt
To ensure the AI requests approval at the right times, add the following to your custom instructions or system prompt:
When the task requires multiple steps or non-trivial changes, present a detailed plan using #approvePlan and wait for approval before executing.
If the plan is rejected, incorporate the comments and submit an updated plan with #approvePlan.
Always use #askUser before completing any task to confirm the result matches what the user asked for.
You can add this into your .github/copilot-instructions.md file in your project
Quick tutorial: using approve_plan
If you want to explicitly trigger plan review from the start, ask Copilot something like:
Before you change anything, write a step-by-step plan and present it with #approvePlan.
Wait for my approval (or requested changes). Only then implement the plan.
Requirements
- VS Code 1.104.1 or higher
- GitHub Copilot Chat extension
Extension Settings
This extension works out of the box with no configuration required. You only need to instruct your agent to use it.
Known Issues
None at this time. Please report issues on GitHub.
Release Notes
0.1.8
Added
- Approve Plan Tool: Added
approve_plan (#approvePlan) to let users review/approve plans with inline comments before execution (VSCode)
0.1.7
Fixed
- Webview: Fixed issue where the fallback prompt was opened when the Webview didn't received focus (VSCode)
Changed
- Documentation: Added instruction to Antigravity users (Antigravity)
0.1.6
Fixed
- Antigravity: Fixed Antigravity integration
0.1.5
Added
- Antigravity: Added support to Antigravity
0.1.4
Fixed
- Badge Counter: Fixed badge not resetting to 0 after all requests are closed (was showing "1" incorrectly)
- Notification Behavior: Notifications now only appear when the Seamless Agent panel is not visible, reducing interruptions when the panel is already open
0.1.3
Added
- Multiple Concurrent Requests: Support for handling multiple requests with list view
- Attachments: File attachments support with VS Code Quick Pick file selector
Changed
- Layout: A few updates on layout to make the request panel looks like copilot chat
- Task List:
Improved task list UI with better visual hierarchy
- Disposal: The request will be dispose when the agent stops.
- Panel Icon:
Updated panel icon to match VS Code's design language
- Badge Counter: Visual badge showing the number of pending requests
0.1.2
Added
- Dedicated Panel View:
A new "Seamless Agent" panel is registered in the bottom panel area (alongside Terminal/Output), providing a non-intrusive workspace for agent interactions.
- Rich Markdown Rendering:
Full Markdown support including: - Headers, bold, italic, strikethrough - Code blocks with syntax highlighting for 10 languages: - JavaScript/TypeScript - Python - C# - Java - CSS - HTML/XML - JSON - Bash/Shell - SQL - Block quotes - Ordered and unordered lists - Links (auto-linkified) - Tables
- Multi-Line Input:
A proper
<textarea> element allows users to: - Write detailed multi-paragraph responses - Paste code snippets - Use Ctrl+Enter to submit
- Non-Intrusive Notifications:
- Badge indicator on the panel tab showing pending request count
- Information notification as supplementary alert with "Open Console" action
- Panel auto-reveals but preserves focus (
preserveFocus: true)
- Graceful Fallback:
If the webview panel is unavailable (e.g., not yet resolved), the tool automatically falls back to the original VS Code dialog approach.
Changed
- Moved user confirmation UI from VS Code popup dialogs to dedicated panel
- Updated esbuild configuration to compile webview scripts separately
- Improved localization system with support for EN, PT-BR and PT variants
Fixed
- Added
dist/ to .gitignore to avoid committing build artifacts
0.0.4
- Initial beta release
- Added
ask_user Language Model tool
- Multi-language support (English, Portuguese)
License
MIT