ros-claw/rosclaw is a Physical AI runtime for embodied agents. PR #792, merged on 2026-11-06, added a path that separates the response wait time from the upper limit in this project's simulation (SIM) execution contract. Only one capability, coverage.program_finite_v1, designated as a trusted agent configuration, can register a 310-second response wait, a 310-second verification, and a 330-second upper limit.
This separation is a mechanism intended to keep a late-arriving response from being mistaken for a successful execution. The wait, verification, and upper values must satisfy 0 < response ≤ verification ≤ upper ≤ 330 seconds, and the response must always be shorter than the upper limit. Any request to change the upper limit, to invoke REAL/SHADOW mode, or to redefine the existing coverage.execute_slice contract is rejected before remote execution. If the response expires first, the result becomes ACTION_RESULT_UNKNOWN, and the original waiter and session, along with the upper deadline, are kept as they are until the original termination is settled. During that time, requests to introduce a substitute behavior are rejected, and release requires closing the channel. The repository description states that this runtime deals with controlled execution and verified experience, physical memory, and technological evolution.
The boundary is drawn at the line where merely storing an execution intent does not confer execution authority. The public coverage path re-checks the original identifier, the observation moment, Body/epochs, energy, native termination status, and the full footprint path immediately after persistence, and the registered budget check runs again before launch with the same upper limit and deadline. A budget adjustment can only shorten the front portion of work that has not yet been dispatched, and the requirement for consecutive strokes remains. A successful simulation query does not prove actual movement, nor does a termination receipt or cancellation confirmation alone return authority or establish a physical stop.
According to the verification record revealed in the commit message, the 12 newly registered boundary tests and the existing authorization, session, cancellation, receipt, and consent checks passed together, 78 cases in 4.39 seconds, and the Ruff check on the changed files is clean. Full SDK collection fails at tests/knowledge/test_contracts.py because the optional dependency rosclaw_know.contracts is absent (11264/11324 collected, 60 excluded). The full agentd attempt was aborted after 158.16 seconds, with 206 cases passing, but 4 renderer tests failed due to the absence of the ur5e and canonical renderer dependencies. It cannot be claimed that the whole agentd passes. Verification of the SDK in the installed image environment, the actual Native configuration, and the 310/330/310 timings on real hardware (common-tool) remain outstanding, and the new tests contain no model, provider, or RobotIO calls. The rules themselves are confirmed by code and tests, but no hardware-level results are presented. At present, this is closer to a change worth watching to see whether the rules are observed than to a point where operating conditions can be set.