The session log

The session log is a file that is only ever appended to. What is already written is never modified.

Out of that come all the features you would expect and some you would not.

• Continuity Resume Pick a conversation up where it was left
• Branches Fork Split in two from an earlier point
• Audit Replay Walk through it step by step exactly as it happened
• Search Find Look for something inside the session
• Output Export Take it out of the system
• Accounting Measure Know what it consumed and where

It all derives from the same event stream.

And there is a nuance almost nobody explains. The model does not read the log. It reads a projection built from it. The raw fragments are stored separately, and they are what makes it possible to replay a session faithfully instead of only summarising it.

The rule

Why I trust that claim

Because I looked at how the check is built, and the design is serious.

Every package in the project publishes its own verifications under its exact name, so a failure points at whoever is responsible. And there is a script that automatically rejects packages declaring they have nothing to check without explaining why, with a specific hand written justification for that package.

That is a lot of discipline for a promise nobody was going to audit. That is what makes it credible.

What this is for in practice

If you have to answer to someone for what an agent does on your infrastructure, this matters more than any performance benchmark. It is traceability by design, not an activity log bolted on afterwards.