A developer productivity tool that provides insights into how you interact with AI coding assistants.
How it Works
The extension automatically detects when you accept AI-generated code. It supports both Inline Suggestions (ghost text) with deterministic detection, and Agentic AI interactions (Copilot Chat, Inline Chat, Agent Mode) with confidence-based detection.
When an AI suggestion is confirmed:
- Detection: For inline suggestions, the extension intercepts the Tab key. For agentic AI, it uses a multi-signal confidence scoring system.
- Tracking: It sets a temporary flag in your local git repository.
- Git Integration: A custom
commit-msg hook is automatically installed in your repository. When you commit your changes, this hook checks for the flag and appends an "Impacted by AI" footer to your commit message.
Features
- Inline Suggestion Tracking: Deterministic, zero false positives. Tracks Tab, Ctrl+Right (word), and Ctrl+Shift+Right (line) acceptances.
- Agentic AI Detection: Confidence-based scoring system detects Agent Mode, Chat Apply, and autonomous code insertions.
- Configurable Threshold: Adjust the confidence threshold (50-100%) to balance detection vs false positives.
- Marker Consolidation: When both inline and agentic AI are used before a commit, the message shows "Impacted by AI (Inline + Agentic)".
- Git Integration: Automatically tags commits that contain AI-generated code with an "Impacted by AI" trailer.
- Privacy Focused: No code is sent to external servers. All processing happens locally within VS Code and your Git hooks.
Extension Settings
This extension contributes the following settings:
copilotInsightTracker.captureMode: Enable or disable all tracking functionality (default: true).copilotInsightTracker.trackAgenticEdits: Enable or disable tracking of agentic AI edits (default: true).copilotInsightTracker.agenticConfidenceThreshold: Minimum confidence percentage (50-100) to mark as agentic (default: 70).
Agentic Detection Signals
The confidence scoring system uses 14 signals to detect AI-generated code:
Negative Signals (Reduce Score)
| Signal |
Points |
Description |
| Recent paste operation |
-50 |
Paste detected via Ctrl+V within 500ms |
Positive Signals (Increase Score)
| Signal |
Points |
Description |
| VS Code window not focused |
+35 |
Changes made while window is unfocused |
| Sustained background activity |
+20/+30 |
3+ (or 6+) background changes in 5 seconds |
| Background document change |
+30 |
Edits to non-active document |
| Multi-file rapid changes |
+25 |
2+ files changed within 5 seconds |
| Large background deletion |
+25 |
50+ chars deleted in non-active document |
| No recent typing (large) |
+40 |
100+ chars with 3+ second typing gap |
| No recent typing (small) |
+20 |
20+ chars with 2+ second typing gap |
| Fast insertion rate |
+20 |
>100 chars inserted per second |
| Chat panel visible |
+20 |
Chat-related views detected |
| Background replacement |
+15 |
10+ chars inserted AND deleted in background |
| Large deletion (no typing) |
+15 |
50+ chars deleted without recent typing |
| Pure insertion |
+15 |
Insert-only change, no text removed |
| Large code block |
+15/+8 |
200+ (or 100+) characters changed |
| Markdown formatting |
+10 |
Headers, lists, or code blocks in .md files |
| Code structure patterns |
+10 |
Contains function/class definitions |
| Complete statements |
+5 |
Lines end with ; { } ) ] : |
Changes are only marked when confidence exceeds the threshold (default: 70%).
Key Detection Improvements
- Deletions tracked: Large deletions (50+ chars) now contribute to scoring
- Replacements caught: Background edit+delete operations detected
- Sustained activity: Multiple consecutive background changes aggregate score
Supported AI Interactions
Inline Suggestions (Deterministic)
- Full acceptance (Tab): Accepts the entire ghost text suggestion
- Word acceptance (Ctrl+Right): Accepts the next word of the suggestion
- Line acceptance (Ctrl+Shift+Right): Accepts the next line of the suggestion
Agentic AI (Confidence-Based)
- Agent Mode: Autonomous multi-file code changes
- Chat Apply in Editor: Applying code blocks from Chat
- Background edits: Changes to documents without active user focus
What is NOT Tracked (Intentionally)
- Copy from Chat + Manual Paste: You're in full control of the paste action.
- Terminal Commands: Code generated by AI-suggested terminal commands.
- Manual Typing/Editing: Regular code editing is never tracked.
Commit Message Examples
After accepting inline AI suggestions:
feat: add user authentication
Impacted by AI
After Agent Mode changes (with confidence score):
fix: resolve memory leak
Impacted by AI (Agentic - 85% confidence)
After using both inline and agentic in the same session:
refactor: improve performance
Impacted by AI (Inline + Agentic)
Threshold Guidelines
- 50%: More detections, higher false positive risk
- 70% (default): Balanced detection, recommended for most users
- 85%+: Conservative, only marks high-confidence changes
Known Limitations
- Threshold Edge Cases: Changes exactly at the threshold boundary may be inconsistently detected.
- Cannot Distinguish AI Sources: Cannot determine which AI tool made the change (Copilot vs Claude vs other).
- Fast Human Typing: Extremely fast typists making large changes may occasionally score above threshold.
- Multi-File Refactoring: VS Code's rename symbol or find-replace across files may score as agentic.
- Paste Penalty: The -50 paste penalty is deterministic and may occasionally penalize legitimate AI operations if you paste immediately before them.
How It Works (Technical)
Change Tracking
- totalChars: Tracks both insertions AND deletions (total activity)
- Background detection: Changes to non-active documents score higher
- Sustained activity: Consecutive background changes within 5 seconds aggregate
Score Preservation
- The highest confidence score in a session is preserved for commit tagging
- Sessions reset after 30 seconds of inactivity or after a commit
Custom Git Hooks Path
- Automatically detects
git config core.hooksPath setting
- Installs hook in the correct location (global or local)
Requirements
- VS Code 1.85 or later
- Git installed and repository initialized
Core Team
AI Contribution Tracker is a collaboration project by David Exterman and Yoav Lax.
