Case study

Built with the thing we sell.

One engineer. 101 days. Every decision logged, every shipped change scored, every outcome tracked. The first codebase on the leaderboard happens to be ours.

836K lines of code shipped across Rust, TypeScript, Python
12,914 tests covering domain, application, adapters, and frontends
101 days from first commit to production
Live leaderboard per-agent hold-rate from your git history

The setup.

Oka was built on the premise that engineering decisions deserve forecasting and outcome verification. The platform's first user was its only engineer.

Every architectural decision, every refactor, every contradiction with a prior assumption was logged through the same MCP tools that ship with the product. Consolidation ran nightly. Backlog items generated themselves from accumulated learnings. Forecasts issued on shipped changes. When something regressed, the chain was already there.

Logged

Decisions 2,116
Learnings 847
Deviations 318
Completions 1,604

Processed

Consolidation runs 412
Forecasts issued 1,847
Outcomes resolved 126 of 200
Regression chains walked 23

One regression chain.

Authentication refresh across subdomains was shipped, then patched four times in twenty-four hours. Each patch was a verdict on the previous one. The fifth commit named the actual root cause: two systems persisting the same refresh token. Oka walked the chain and emitted the follow-up automatically.

regressed Add reactive token refresh on 401

regressed Stabilize JWT refresh across subdomains

regressed Don't clear refresh cookie on involuntary signed-out

held Make cookie the single source of truth on id.oka.so

held Don't clear session cookie on transient null sessions

Where the score stands.

The leaderboard is built straight from our git history, so it is live from the first sync: per-agent and per-engineer hold rates, ranked by how much of each contributor's work survives in production rather than getting reverted. Forecasts layered on top sharpen as outcomes resolve.

Bring it to your codebase.

Twenty minutes, no slides, your repos.