Mobb AI Tracer

VS Code/Cursor extension that monitors and tracks AI-assisted coding activities.

What it does

• Cursor: Polls the Cursor database to capture AI conversations, tool calls, and code changes.
• VS Code/Copilot: Watches Copilot events, file edits, and log files to track AI-generated code.
• Uploads: Sends collected data (prompts, AI responses, code diffs) to Mobb backend for analysis.

Features

• Auto-detects whether you're using Cursor or VS Code.
• Tracks AI conversations, tool executions, and code changes.
• Monitors both chat completions and inline edits.
• Uploads data securely to Mobb's backend.

Installation

Install from VSIX package or build from source.

Requirements

• VS Code 1.99.0+ or Cursor
• Node.js 22.x

Cursor Debug instructions

1. Update the .env values with these values for local development:

 WEB_LOGIN_URL="http://localhost:5173/cli-login"
 API_URL="http://localhost:8080/v1/graphql"
 WEB_APP_URL="http://localhost:5173"

2. Run (this will overwrite your .env file, so, check which values you need or want to use)

package.sh

3. Open Cursor

If the extension is already installed,

• Go to Extensions and uninstall
• Restart extensions

4. Go to the file clients/cursor_ext/mobb-ai-tracer-VERSION.vsix

5. Right click, install extension

6. Open the Output panel and pick the channel you need:

   • mobb-ai-tracer → extension output channel

Human Code Tracking Notes

What the Human tracker does

• Listens to text document changes in the active editor only.
• Groups contiguous human edits into "segments" (line ranges).
• Closes a segment when:
  • You pause typing for ~segmentIdleMs, or
  • You make an edit outside the current contiguous line window (beyond the adjacency gap), or
  • The segment grows beyond segmentMaxChars (safety cap), or
  • The time since the last edit in the segment exceeds maxSegmentDurationMs, or
  • A non-human edit group occurs (multi-change event or large single insert). In this case we force-close any open human segment and ignore the non-human group itself.
• For each closed segment we:
  • Compute basic metrics (duration, edit cadence).
  • Classify it using conservative heuristics and treat only clearly human segments as HUMAN.
  • If classified as HUMAN, we read the full content of every line in the segment's line span into a single changedLines string.
  • If the non-whitespace character count of changedLines is above the configured threshold, we either emit a DRY-RUN upload log entry in the shared output channel or, when uploadEnabled is true, upload the segment via the AI blame pipeline.

Settings

• The human tracker uses internal defaults; there are no user-facing settings for this feature.

Internal configuration (for developers)

• Human tracker configuration

  • All tracker and classifier thresholds live in src/human/config.ts (HUMAN_TRACKING_CONFIG). Key fields:
    • segmentIdleMs — idle gap to close a segment (ms).
    • segmentMaxChars — max characters buffered per segment.
    • maxSegmentDurationMs — max time the segment can live before it is force-closed (ms).
    • adjacencyGapLines — maximum gap (in lines) to merge edits into a single segment.
    • minSegmentCharsWithNoWhitespace — minimum total length of changed lines required to upload (excluding whitespace).
    • uploadEnabled — whether to actually upload segments or just log artifacts.
  • Classifier thresholds (cadence, paste-like detection, etc.) are under HUMAN_TRACKING_CONFIG.classifier and used by:
    • src/human/eventClassifier.ts (per-change event classification: multi-change vs large insert vs single small change).
    • src/human/segmentClassifier.ts (future preparation for more complex classification logic per segment - we currently only use the event classifier).

• All thresholds above are internal constants; changing them requires a rebuild.