Recordings

What is recorded

While recorded, a machine runs on one core, and everything that comes in from outside is logged at the exact instruction where it arrives:

• network packets;
• disk contents;
• the clock and randomness;
• the commands of each step.

Because the machine is deterministic, that log plus the starting state reproduce the run exactly.

What is kept

• The recording of every failed job, for 30 days. That covers the job's recorded steps, and in a split job, each shard with a failing step. Recordings of passing jobs and passing shards are deleted when the job ends.
• A replay check of every kept recording, run automatically. The dashboard shows whether the replay matched; so far every one has.
• The job summary and logs, for 30 days.

Replaying

A kept recording replays the run exactly. A replay can stop at any point of it, down to a single instruction, and show the whole machine there: memory, registers, files and processes. It can jump to a later or an earlier point. During the beta, these replay and time-travel sessions run on our hosts (a recording needs the same CPU model it was made on). Ask on the job's page or at hello@hypervise.sh. A self-serve replay and debugging API, including one for AI agents, is next on the roadmap.

Recordings contain your secrets

A recording holds everything the job read, including GITHUB_TOKEN and any secrets the job used. We treat recordings like the secrets themselves: see security.