Sizing, timeouts and limits

Resources. As Host requirements.2: take the declared shape as a floor.

Tool timeout. Each call has a wall-clock limit, after which the session is killed and restarted. The default is a minute; delivered environments raise it, because a build or a test run needs longer. A patch environment divides its own allowance between building, running tests, and re-checking the defect, so a patch task's effective build timeout is less than the whole.

Setup and grading have their own limits, generous relative to the work.

Input size. There is a ceiling on the input the agent may submit, and it is not a fixed number: it is derived per environment from the largest file in that task's own test suite, with a floor of 8 KiB and a ceiling of 100 MiB. The task text your agent receives states the actual figure. Do not hard-code one.

Episode limits are yours. The environment does not stop your agent; it only refuses work that exceeds a budget. Turn count, wall-clock time and token spend are limits your orchestrator imposes, and they should be part of how you report a result — an agent that ran out of turns did not fail the same way as one whose patch was rejected.