Staff Access to Your Sessions

Your session transcripts contain your source code and can contain your secrets. This page says who at PlatformSmith can read them and under what circumstances. It forms part of our privacy policy.

We have written it to describe what we actually do, not what sounds best. In particular, we are not going to tell you that no one ever reads a transcript — we sometimes need to, and a policy that says otherwise would be untrue.

Two tiers

Session data separates cleanly into two kinds, and we treat them differently.

Metadata — read routinely

Every session event carries structural information: the type of event, whether it errored, which phase of a run it belonged to, when it happened, and its position in the sequence.

Our engineers can query this as part of normal support and debugging. Almost every question we need to answer about a broken session — where did this run stop, what failed, did the agent ever start — is answerable from this alone, without seeing anything you wrote or the agent produced.

This tier requires no special approval. It is the ordinary business of keeping the Service working.

Content — requires a reason and your consent

The content of a transcript — your prompts, the agent’s replies, the file and command output it surfaced — is different.

We access it only when:

1. there is a specific, stated reason that metadata cannot resolve, typically a support issue you have raised;
2. access is limited to the affected session, not your account or history at large; and
3. you have consented to it for that incident.

We will ask. If you say no, we will keep debugging with the metadata tier and tell you if that means we cannot resolve the problem. The exceptions to this are set out under “Where we access content without asking” below.

What we do not do

• We do not read transcripts for product research, benchmarking, or curiosity.
• We do not use your code or prompts to train models, ours or anyone else’s.
• We do not share transcript content with third parties except as set out in the privacy policy or where legally compelled.
• We do not put your session content into our own AI tools.

That last point deserves spelling out, because we build this product with AI assistants and you would be right to wonder. Our engineers use them on our source code. Your session content does not go into them — not while debugging, not while investigating an incident. When we need to understand a failing run we work from the metadata tier, and when that is not enough we reproduce the problem with synthetic data.

This is what keeps two other statements true: that model providers are not our sub-processors (sub-processors), and that the data we hold about you stays in the region named in the privacy policy. If we ever changed this, those pages would change first.

Where we access content without asking

There are two situations where we may read transcript content without your consent, and we would rather name them than rely on you not noticing the gap:

• Suspected breach of our acceptable use policy — where we have specific reason to believe a session is being used for something that policy forbids. Asking permission first would defeat the purpose.
• A security incident affecting the Service, another customer, or your own account.

In both cases access stays limited to what the investigation requires, and we will tell you afterwards unless doing so would prejudice the investigation or we are legally prohibited.

This is not a general-purpose exception. It does not cover debugging, product research, or curiosity, all of which still follow the consent rule above.

Where the law overrides this

If we are legally compelled to produce data we will comply, and we will tell you unless we are prohibited from doing so.

Honesty about the control

This policy is currently enforced by process and by a small team, not by a technical control that makes content access impossible without your approval. We would rather state that than imply a stronger guarantee than we have built. Strengthening it — so the boundary is enforced by the system rather than by discipline — is planned work.

Contact

Questions, or to withdraw consent for an ongoing incident: privacy@platformsmith.com.