August 23, 2026 · Joy
Handoff rediscovery — the measured claim
In this benchmark, JoyMux cut measured rediscovery during a coding handoff without changing the passing test outcome.
That sentence is allowed because a frozen comparison file still shows it. It is not allowed to say JoyMux was faster.
The work was not abstract. Agent A left a checkout API unfinished: Idempotency-Key is still a TODO, duplicate POSTs mint a second order, and two unittests fail.
The actual snapshot. Same remaining files in both arms. Scripted rediscovery on the receiving side — not a live model race.
What we ran
One checkout-API fixture. Agent A left idempotent replay unfinished. The receiving side then continued — twice:
- Baseline: start from the task only; rediscover git, files, tests, process.
- JoyMux: start from durable execution context, then the same remaining files.
The receiving side is scripted rediscovery, not a live coding model. Both arms applied the same remaining-work files and the same unittest suite.
Numbers (this run only)
| Metric | Baseline | JoyMux |
|---|---|---|
| Rediscovery tool calls | 15 | 1 |
| Files reread | 8 | 0 |
| Context tokens (density-4.0) | 2409 | 271 |
| Final tests | pass | pass |
| Time to first edit | 384 ms | 743 ms |
Baseline won the scripted clock. JoyMux won rediscovery volume.
Source of truth: demos/handoff-rediscovery/results/compare.json
SHA-256: 5d3f87ae2a3ffb9489efe6813ecae2488db4e894958f6708e9f485911f982fce
What this does not prove
- JoyMux is faster.
- Universal token savings on every project.
- Live LLM behavior.
- That JoyMux wrote the remaining implementation (both arms copied the same fixup).
If your agent is paying again to find work it already did, that is the loop this runtime is built to break.
Start a 14-day free trial — then $29/month. Cancel anytime.