Skip to content
| Marketplace
Sign in
Visual Studio Code>Data Science>TruthMeter AI - Evidence-Based AI Code MetricsNew to Visual Studio Code? Get it now.
TruthMeter AI - Evidence-Based AI Code Metrics

TruthMeter AI - Evidence-Based AI Code Metrics

Jeffrey Jose

|
40 installs
| (0) | Free
How much of your code is AI-written, and does it stick? Measured from git attribution, agent session logs and line survival, not guesses.
Installation
Launch VS Code Quick Open (Ctrl+P), paste the following command, and press enter.
Copied to clipboard
More Info

TruthMeter AI

How much of your code is AI-written, and does it stick? TruthMeter answers from evidence your repository already has (git attribution, git blame and, if you allow it, local agent session logs) instead of guessing from typing speed.

TruthMeter dashboard

What it measures

Metric How it's computed
AI-attributed commits and lines Commits that record AI involvement: a Co-authored-by trailer, an agent bot as author, or a tool footer like “Generated with Claude Code”.
Line survival Of the lines a commit added, the share git blame -w still attributes to it at HEAD. Only commits older than 14 days count. AI-attributed vs. everything else.
Fix within 72h Share of commits followed within 72 hours by a fix/revert-style commit touching the same files (manifests and READMEs ignored).
By tool Claude Code, GitHub Copilot, Cursor, OpenAI Codex, Gemini/Jules, Aider, Devin, Windsurf, Amazon Q/Kiro, plus your own rules.
Agent session evidence (opt-in) Files that Claude Code or Codex edited, read from their local session logs, matched to commits that have no AI trailer. This shows how much AI work your trailers miss.

Rates stay hidden until each group has at least 5 commits old enough to judge. A rate based on two commits is noise.

Why 2.0 is a rewrite

Version 1.x flagged edits as AI-generated using regular expressions and keystroke timing. In practice those patterns matched almost all code: one import line, items.map(, a typed parameter, or any three lines of text all came back as "AI". The percentages and dollar figures built on top of that weren't measuring anything.

No editor API reports when another extension's suggestion is accepted, so in-editor detection can't be made reliable. The evidence that does exist is in git history and agent session logs, and that's all 2.0 uses.

Getting started

  1. Install from the VS Code Marketplace or Open VSX (Cursor, Windsurf, VSCodium, Antigravity).
  2. Open a folder that is a git repository. The status bar shows the AI-attributed commit share.
  3. Click it, or run TruthMeter: Show Dashboard.
  4. Optional: TruthMeter: Turn On Agent Session Evidence.

Getting good data

TruthMeter can only count what your history records. To make AI work visible:

  • Claude Code adds Co-Authored-By: Claude … by default. Keep it on.
  • Cursor adds Co-authored-by: Cursor Agent for agent commits.
  • Codex CLI has a commit_attribution setting in config.toml.
  • VS Code + Copilot: git.addAICoAuthor. Because it can stamp commits where Copilot wrote little, Copilot trailers are shown as medium confidence. Turn off truthmeter.attribution.trustCopilotTrailer to exclude them.
  • Squash merges keep trailers only if your merge tool copies them into the squash message.

Settings

Setting Default
truthmeter.windowDays 90 Days of history to analyse.
truthmeter.survival.minAgeDays 14 Minimum commit age for survival and fix-rate stats.
truthmeter.survival.maxFiles 400 Cap on files blamed. Least-touched files are skipped first.
truthmeter.agentLogs.enabled false Read local Claude Code / Codex session logs.
truthmeter.attribution.trustCopilotTrailer true Count Copilot co-author trailers.
truthmeter.attribution.customRules [] [{ "tool": "Our Bot", "pattern": "\\[ai\\]" }], a regex tested against author, committer and message.
truthmeter.excludePatterns lock files, build output Path regexes left out of line counts.
truthmeter.statusBar.enabled true Show the status bar item.

Privacy

Everything runs locally: git log, git blame, and (only if you opt in) reading session logs under ~/.claude/projects and ~/.codex/sessions. From those logs TruthMeter extracts only file paths inside the open workspace and their timestamps. No telemetry, no network calls. The extension only runs in trusted workspaces, because it executes git there.

Limits worth knowing

  • These are correlations in your own history, not causal proof. AI commits are often different work (tests, scaffolding, refactors) from the rest.
  • Survival treats files renamed after a commit as "not surviving", for both groups equally.
  • Commits made without any trailer and without a local agent log are invisible to TruthMeter. That's the point: it reports only what it can show.

Development

npm ci
npm run test:unit      # unit + integration tests against back-dated fixture git repos
npm run test:e2e       # launches VS Code 1.90 with the extension; VSCODE_VERSION=stable for latest
npm run analyze -- /path/to/repo --days 180 --agent-logs   # same analysis from the command line

Publishing: push a vX.Y.Z tag matching package.json. GitHub Actions tests, then publishes to the VS Code Marketplace (OIDC trusted publishing, with a PAT fallback until 2026-12-01), Open VSX, and a GitHub release.

License

AGPL-3.0

  • Contact us
  • Jobs
  • Privacy
  • Manage cookies
  • Terms of use
  • Trademarks
  • Your Privacy Choices
  • Consumer Health Privacy
© 2026 Microsoft