Skip to content
last verified against v0.1.26 verified 2026-10-16

FAQ

Short answers, each linked to where it is measured or stated.

What decides whether a Cruise result is accepted?

The repository checks you approve before the run. Cruise pins the tests in an acceptance manifest and verifies each candidate against those copies. The current scope is bounded coding tasks with tests at the repository root or nested under it; it does not write the checks for you. See Cruise.

What is the verifier independent from?

The agent doing the work. The candidate is extracted and the agent’s machine stopped before the checks run on the host, using the pinned tests. A candidate’s own pass log is not accepted as proof (verifier record). This does not make the tests a guarantee against every defect.

What happens if the last model does not pass?

A job that exhausts its ladder waits in review. You can resume it from the same paused parent, on the same or another ladder, or cancel it and release the parent. Cruise does not promise to retry until it passes. The verdict and its signed receipt prove only that the approved tests passed (Cruise).

Is my data safe if the host dies?

Host loss has been survived in one stated test, in one region; cross-region recovery, continuous exercise of the failover and any recovery-time objective are not claimed (Limits, section 2). Process loss is a separate and older result: the machine’s hypervisor and the supervising daemon were terminated together, and the session came back with its work intact (gate 4).

How long does a resume take?

Seconds, not sub-second: 13.7 to 25.2 s across the bench’s recorded runs, of which the hypervisor’s own snapshot load is about 4 ms (resume times). “Zero downtime” is not established by these tests; see resume measurements.

Which agents run, and is KeepState tied to Claude Code?

KeepState is not tied to one agent. In an agent session the program that drives the model is the agent’s runner, and everything you do with an agent works the same whatever the runner is. Two runners are certified today: Claude Code, pinned at version 2.1.296 since 2026-10-14 (sessions created before keep 2.1.251, certified by its own run), and Codex CLI 0.162.1, which calls OpenAI on your own OpenAI key, and a paused session can switch between them; others are not built or certified for agent sessions yet (Runners). Plain sessions run any program you start; which ones have been tested is on the compatibility matrix.

Does my agent need changes to run on KeepState?

Claude Code ran unmodified (gate 3). aider needed one line in the machine’s image, export ANTHROPIC_API_BASE="${ANTHROPIC_BASE_URL:-}", and no change to KeepState (gate 7).

Can I do everything from the app?

Almost. Every ks command runs inside an agent’s window as a / command, except agent open, attach, login and logout; and ks key add, which reads your key from standard input, runs in a terminal (slash commands).

What gets billed on an interrupted model call?

At most once. A completed response caught by a save is kept and served from the gateway’s store when the program retries, with no new tokens bought (gate 3, re-run green on 13 October 2026). From 20 September to 6 October 2026 a retry after a resume could be billed twice; the fix is in production since 6 October 2026 (Limits, section 5). Whether a program retries is its own behavior; see the compatibility matrix.

What happens when a session hits its budget?

Further model calls are refused. A plain session is saved and paused; for an agent session, pausing at the budget has not been tested on its own. The defaults and what /usage shows: budgets.

Do old checkpoints go stale?

No. A checkpoint about 15.2 hours old woke in 14.3 s, reading its saved counter at 63 while its living twin had reached 4841. A checkpoint is a complete, self-contained object: a manifest plus content-addressed chunks.

What do I pay for?

KeepState charges for session time and checkpoint storage after deduplication. On your own key you pay model costs directly to your provider, with no KeepState markup. Where KeepState’s key is offered, a session can use it instead and pay the provider’s list price plus a published spread, from credit you have bought; Pricing says whether it is offered today. Cruise bills as the sessions it runs. The free tier runs one session at a time, and forking needs a paid plan. Rates are on the pricing page.

Can I self-host KeepState?

No. KeepState runs as a hosted product on KeepState-operated infrastructure only: there is no self-hosted, on-premises or customer-network deployment, and none is intended.

Is the source public?

No. The engine source and internal operating records are private. The public evidence summaries retain the tested outcomes, dates, and scope; each published artifact has its own digest. These summaries do not establish current deployment or permission to read internal records. The recorded demo shows the historical recovery walkthrough.

What can KeepState staff see?

Nobody at KeepState opens a customer session as a matter of course: access only with your explicit consent on a named session, logged in the same ledger you can read. What is stored at all, and its retention and deletion rules, is on Security.

Spotted an error on this page? Tell us, and we will correct it.

Report it
← previous Troubleshooting next → Compatibility matrix