AI coding activity tracking

Elapsed editor time cannot distinguish a hand-typed function from a large accepted AI suggestion. Attribution keeps both signals visible.

A client posts explicit AI metadata with the heartbeat.

{
  "is_ai_generated": true,
  "ai_agent": "your-agent-name",
  "ai_lines_added": 42,
  "ai_lines_deleted": 3
}

The server aggregates line counts, human keystrokes, agent totals and daily series.

Bundled hooks cover Claude Code, Codex CLI and Gemini CLI. Cursor and Copilot need provider-specific adapters.

How fields are populated

• Verified bundled hooks post an event id and provider metadata
• Your own wrapper can post the AI fields directly
• WakaTime imports with an AI category remain AI-attributed

Every line change lands in one of three classes, and the third one is the point. A tool that only has "AI" and "human" has to guess when it cannot tell, and a guess in a metric you will use to make decisions is worse than an admission.

• AI: a hook or client named the agent that made the change
• Human: the editor observed the keystrokes and characters
• Unattributed: the change arrived without proof of either, and is reported as its own quantity

Unattributed is normal, not an error. Pasting from a browser, a refactor run by a tool without a hook, or a merge all produce real work that nothing can honestly attribute. Its share is visible so you can decide whether it is small enough to ignore.

A terminal agent has no editor events to hook. When Claude Code rewrites six files while VS Code is closed, an editor plugin sees nothing, so a per-user agent reads each tool’s own local transcript instead and turns new events into AI-attributed heartbeats.

• The agent runs as a per-user service and checks active transcripts on a short interval
• It reads local transcript files only, and sends identifiers and line counts from them
• Work done with the editor closed is still recorded, with the same attribution
• A saved cursor tracks what has been read, so restarting it does not re-send events

The VS Code setup flow detects which agents are installed and registers hooks for them. It never edits another tool’s configuration without asking first.

Command Palette: AtoAxis: Show AI Tracking Status
Command Palette: AtoAxis: Set Up AI Tracking
Command Palette: AtoAxis: Repair AI Integrations

The same state is readable, and the same work is doable, from the terminal.

wakaclock-cli ai status              # what each agent looks like right now
wakaclock-cli ai sync --dry-run      # parse without sending or advancing anything
wakaclock-cli agent status           # is the always-on transcript service running

Token usage and estimated cost are not invented. They appear only when provider-reported tokens and a versioned effective price are available; subscription activity without authoritative pricing is null with coverage metadata.

What coverage means here

• Token totals separate input, output and cache reads and writes when the provider reports them
• Estimated spend needs both the usage and a price that was in effect at the time
• Each of attribution, tokens and cost carries its own coverage percentage
• A missing value is null with its coverage, never a zero that reads like a measurement

• AI flag, agent name and AI lines added/deleted
• Human keystrokes and character counts
• Project, branch, language, file path and timestamps

What is never recorded

• Prompts or model responses
• Source file contents, patches or transcripts

AI attribution uses the same metadata-only pipeline as ordinary heartbeats. No conversation body is stored.