Your AI Might Be Lying to Your Boss (22 min read)

Windsurf's dashboard has a metric called PCW (percentage of code written by Windsurf), and the author felt this number was too far off from their actual sense of usage, so they ran their own experiment. According to the guidance, 85~95%+ is common, but the author's own session showed 98%.

At first, they tried sniffing the network with mitmproxy, but decoding was difficult since Windsurf makes heavy use of protobuf. Instead, through the web dashboard's GetAnalytics response and the message definitions bundled in the frontend, they identified user_bytes, codeium_bytes, total_bytes, and percent_code_written, and the calculation was effectively codeium_bytes / total_bytes. Unlike codeium_bytes, which retains the old name Codeium, this value reacted instantly to code additions/deletions during editing, not just at commit time, and restarting the session caused it to lose its generation history.

• Auto-inserted closing brackets or quotes were not properly captured as human-typed code.
• Code pasted in by the user was barely reflected in user_bytes.
• Cut/paste operations and move/copy operations performed by Cascade severely distorted the human vs. AI contribution counts.

As a result, even when the author had written all the code themselves, mixing in refactoring and move operations could cause the entire session to be tallied as 100% AI-generated. The author concluded that PCW fails to reflect actual work volume because of a calculation method that is strict toward user_bytes but lenient toward codeium_bytes.