Meanwhile

Get paid while your AI coding agent works

September 7, 2026 589 "installs" turned out to be ~0 real ones -- here's what I learned checking

Meanwhile's live /network-stats page showed 589 installs and I nearly celebrated that at face value. Pulled every single record from production KV instead of trusting the counter: 93% matched an exact synthetic fingerprint (fixed cost, fixed tokens, sequential fake session IDs) from my own earlier testing, and most of the rest were test-labeled names hitting from different simulated clients. Real, verified external human installs: as close to 0 as makes no difference.

The counter itself wasn't buggy -- it counts every install: key exactly as designed. The actual mistake was upstream: real curl/script calls hit production instead of wrangler dev --local during earlier testing, and nobody audited the data until now.

Posting this because it felt more useful to say plainly than to quietly reset the number and move on. If you're checking your own growth metrics and haven't pulled the raw records at least once, it's worth doing before you trust the dashboard.

nirmeet011011

2 Likes

2 Comments

Say something nice…

Post Comment

1

Good question, and I had to actually define it precisely rather than eyeball it. The signal I'm treating as real: total_calls > 0 where the session evidence (cost/token deltas, or at minimum a cwd/session hash) genuinely progressed between calls -- not just a line getting served, since that fires on every request including a dead script pinging on a timer. A single genuinely-progressed call is a real person's editor or terminal actually doing something between two requests, which a synthetic script replaying fixed values can't fake without deliberately incrementing fake numbers. Since posting this I also found and fixed a real bug: the server's install-state reads were hitting Cloudflare's free-tier KV daily quota and silently 500ing for hours -- so some of that "0 real installs" number may have partly been real people hitting a broken endpoint, not just bad test data pollution. Moved the hot path onto a Durable Object so that specific failure mode can't recur. Watching the same raw records now to see if anything crosses that bar for real.

nirmeet011011

·
9 hours ago
·
Reply
1

The raw-record audit is more useful than the 589 headline number. Now that you have a clean baseline, what downstream action will you treat as the first reliable signal that a real install is actually becoming a real user?

Aryan Sinh

·
2 days ago
·
Reply

Meanwhile

About

People