"It worked when I generated it" is the new "it works on my machine." If you cannot reproduce how a piece of software was built, you cannot audit it, you cannot debug the build, and you cannot trust it in production.
What replay pins
When vb commit captures a vibe, it records everything that influences the output, not just the output itself:
- The exact model identifier and version.
- Sampling config — temperature, top-p, and seed.
- The tool definitions the agent had access to.
- The ordered transcript of prompts and tool results.
Replay, then diff against reality
vb replay re-runs the captured vibe and compares the regenerated output against the stored diff. If they match, the vibe is reproducible. If they drift — a model was deprecated, a tool changed — you see exactly where, instead of discovering it in production six weeks later.
vb replay 4f3a9c
replaying with claude-opus-4-7 (temp 0.2, seed 7)...
3 files regenerated
diff vs. recorded output: identical
eval: 5/5 checks pass
✓ vibe is reproducibleUpgrading models, safely
Replay is also how you migrate. Point an old vibe at a newer model, run its eval suite, and you get an honest answer about whether the upgrade holds. The eval is the contract; the model is swappable underneath it.
“A build you cannot reproduce is a story, not a fact.”
