Limits · Saphan Studio documentation

limit	what it means for you
No retention policy	Nothing deletes run directories, journals, model transcripts or custody records automatically. Size your storage, and remove old runs by your own policy.
No model transcript where the recorder cannot run	Agents with no recorder channel (`cursor`, `proc`, `agent-loop`, `openai-compat`); `droid` without a bring-your-own-key model and a declared codec and upstream; `qwen-code` without a configured upstream, or signed in with its own OAuth; walled runs the recorder cannot be carried into (`qwen-code` on any walled run; `codex` and `droid` on macOS; container and vendor confinement); runs with `--no-recording` or `guardrail.record = false`. These runs still have a run record, a journal and a raw transcript, and the run record names the reason.
Console sessions are recorded for `claude-code` only	Console conversations and shells with other agents are not recorded through the guardrail.
No viewer for the signed transcript	Returned, signed segments are stored with their custody records, but no console page or command reads them back in this release.
No standalone verification command	`saphan return` verifies segment signatures when it takes them into custody; there is no command to check a chain or a signature on demand.
Local model transcripts are not taken into custody	Only remote runs' segments are copied home, checked and stored.
No running cost while a run works	The run's cost is known when it ends.
The recorder is given, not enforced	Outside sealed Linux walls, an agent that ignores the recorder's address is not recorded.
One console route shows the raw transcript	A finished run with no journal is replayed unredacted on its output stream.
Attached sessions are not copies	If the vendor's own session file is deleted or changed, so is what the attached session shows.
Runs started outside Saphan Studio are not recorded	Only runs the engine starts are recorded.