The latest commit of minervacomputing/harness, 82ef266 (October 27, 2026), changed things so that when the platform stalls one object, the tests that broke because of it are run one more time. This repository is an AI agent project in which agents do not hold credentials, the gateway compares per-resource permissions on every call, and each response is executed in a sealed container.
The core is the retry rule that went into vitest.faults.config.ts. The test is re-run only when there are traces of the platform having stalled — for example when the last error is a store operation on a Cloudflare Durable Object (a server-side object that holds state for a long time inside the platform) that exceeded 30 seconds and was reset, when world initialization did not answer for 1 minute, or when a call to a new FaultProvider took less than 1 second — and the log displays (retry x1). Tests that directly create and use their own world or sandbox mostly own these themselves, so they start over on a new object and do not inherit the state left behind by the previous run. In the sandbox cleanup stage, if the platform hands back an object reset, only a warning is left behind and the whole file is not treated as a failure. The same behavior is also written down in the The test Worker section of the README that describes the test Worker.
However, what this change touches is not the runtime but the reliability of the tests. This commit is on the scale of 3 files, 28 lines added and 13 lines deleted; rather than adding a new feature, it only tidied up the rules that decide failures and their explanation. Because the retry ends after a single attempt, tests can still be flaky in environments where the platform stall continues, and real operational performance or cases of application are not presented in this commit's information.