Snapshotting: More complete AI code attribution with auspex

Last updated: October 27, 2026

Snapshotting lets auspex capture agent-written code from every source, including shell commands, scripts, MCP tools, and skills, not just dedicated edit tools. Snapshotting is in closed beta on macOS. Your Span account team can tell you whether your organization can join.

Why it matters

Coding agents like Claude Code increasingly change files through shell commands and scripts rather than a dedicated edit tool. Without Snapshotting, Span attributes code only from those dedicated edit-tool calls, so it misses a growing share of agent-written code and undercounts AI attribution. Teams that rely heavily on agents notice this most.

Snapshotting closes that gap, so your AI Code Ratio reflects how your teams actually work.

How it works

Instead of watching only for specific edit-tool calls, auspex observes changes in your Git repository while an agent session is running. It takes a lightweight snapshot of the working tree every couple of minutes in the background, using its own scratch space, so it adds no latency to your agents and writes nothing into your repository.

This changes how attribution works: any change made in the repository during an active agent session is counted as AI-driven, including edits you make by hand. For the full rules and a worked example, see How AI Code Ratio is calculated.

Setup

Upgrade auspex to the latest version on your Macs. Find the latest version and what changed in the auspex release notes. Use the install guide for your device management tool, or the self-install guide, linked from Understanding Agent Traces in Span (auspex).
Ask about joining the beta. Snapshotting is in closed beta. Contact your Span account team to find out whether your organization can join. If it can, Span turns on Snapshotting for your organization once auspex is updated, and the data starts flowing.

Things to know

Closed beta. Snapshotting is in closed beta on macOS. Your Span account team can tell you whether your organization can join.
Hand edits during a session count as AI. Snapshotting credits every repository change made during an active agent session to the agent, so AI Code Ratio can overcount. For example, if an agent writes 80 lines and you type 20 more lines by hand in the same repository while the session is running, all 100 lines count as AI-written, and the PR shows 100% instead of 80%.
Most teams will see AI Code Ratio rise on new PRs. That's expected: Span now captures agent work it previously missed, plus the hand edits described above.
It's forward-looking. Snapshotting applies to PRs matched with traces collected after it is turned on. Past PRs aren't recalculated.
Expect a step change in trends. Because history isn't recalculated, AI Code Ratio trend charts may jump around your upgrade date.
macOS only. Snapshotting runs on macOS. Windows and Linux devices keep attributing code from edit-tool calls only.
Git repositories only. Changes outside Git repositories aren't captured.
Some files are skipped. Lockfiles and minified files aren't included in snapshots.
Cross-repo edits may be missed. If an agent edits a repository auspex isn't tracking, those changes may not be captured.
Concurrent sessions are less precise. Running several agent sessions in the same repository at once can make matching less precise.

Need help?

Contact your Span account team or reach out to support@span.app.