What it is for
#Each job below, done without shadowlap and with it:
| Job | Without shadowlap | With shadowlap |
|---|---|---|
| Reproduce a failing run | Re-run and hope; pay for the whole run again | shadowlap replay <run>: same run, zero upstream requests |
| Find where a failed run went wrong | Read the trace and guess | shadowlap blame <run>: the model call that lost the run, found with forks that replay everything before it for free |
| Fix a prompt the agent keeps misreading | Edit it by hand, rerun, repeat | shadowlap repair: a model you choose proposes rewrites, and the smallest one verify accepts comes back as a patch with its evidence |
| Know a fix works before it ships | Rerun the failing case once and hope | shadowlap verify: the failing runs fixed, runs it was not written against too, and the good runs kept, several attempts each |
| Fix a harness bug (parsing, dispatch, state, stop condition) | Re-run live; the model may behave differently this time | Full replay against the recorded model behavior; the bug reproduces every time |
| Verify a fix that changes later requests | Re-run everything live, side effects included | Unchanged prefix replays; only the changed suffix runs live; writes are blocked or stubbed |
| Ask "what if call 9 had returned X?" | Hand-edit code and re-run | --override [email protected], and only the calls after it run live |
| Catch unintended prompt changes in CI | String snapshot tests | shadowlap test fails the build on the first changed request |
| Know what a prompt edit or a model switch decides differently | Run every golden live and read the transcripts | shadowlap decisions: each recorded model call asked again with only the change, ranked by consequence, writes first; shadowlap migrate for a model switch, verified and priced |
Three sentences to keep in mind
#A replay forks at the first request that changes, not where the bug is.
Fixes to the code around the model change no request and replay in full. Editing the system prompt changes every request and forks at call 1: you get the diff and a fenced fork, but no savings. See How it works.
Nothing is matched fuzzily.
Two requests match only if they are identical after the normalizers you wrote. A false replay hit would be the worst bug this tool could have, so it reports a false divergence instead and shows you how to fix it.
Side effects that don't go through shadowlap will run again.
Wrap tools with the SDK or a framework adapter. A tool the model called that the run never recorded is named, and a replay won't run one that writes until you wrap it, declare it a read or allow it. Anything else, such as a direct database client or an unwrapped HTTP call, is invisible to replay.