[tactino]

The follow-up to #15 and #16, now that plugrl-protocol#11 is merged.

• Pin: plugrl-protocol moves from extras: robomimic and LIBERO on MuJoCo 2.3.7, mujoco on MuJoCo 3, as declared conflicts; lock current #11's branch head (4c877bc) to main (0a76192), the merge of extras: robomimic and LIBERO on MuJoCo 2.3.7, mujoco on MuJoCo 3, as declared conflicts; lock current #11. The code is the same.

• CI probes this client against SPEC.md. The test job runs plugrl-conformance --probe --scenario all with this client running probe-v1, twice:

  • once as a version 1 server;
  • once offering reuse-feedback-obs, which the client then uses.

  The probe works out from each feedback what the client did with each action. It also drives a resync and a text frame. This is the checker that found the stale-chunk bug a reconnect voids every chunk in flight (ConnectionReplaced) - found by the probe checker; probe-v1 env #15 fixed, and from now on a change that brings it back fails CI.

Locally, both runs report no violations. In the second, 9 infers reused rows, all of them rows the server held, including after the resync.

[tactino]

Holding this as a draft: the new probe step hangs on GitHub's runners.

• Run 36945849606 (2026-10-01) sat in "Probe this client against SPEC.md" until the job's 6-hour limit, and kept no log of the step.
• Run 36975670026 (2026-10-02, with timeout-minutes: 10 on the step and timeout 240 around each probe) sat in the same step for more than an hour. Neither timeout fired, and a manual cancel did not take effect at first. That points at a runner too starved to act, not at a slow probe.
• The same commands on a Linux workstation (guangzhao), on a clone of this branch after uv sync, pass every check in 2-3 s per run, with no violations and no leftover processes.

Next: run the step with a watchdog that prints ps and free to the log every 30 s, and upload the probe's output as an artifact if: always(), so a hang leaves evidence. The pin to plugrl-protocol main could go in on its own meanwhile.