#845 Replay simulation runs while-loop to the 10000 cap: agent_scheduled never terminates under mocked side effects
Description
EditEARS SPEC:
- The Highway replay service shall bound simulation of a while loop to the iteration count recorded in the original run's checkpoints.
- While executing in replay/simulation mode, when a while-loop exit condition depends on a variable that is only produced by a mocked side effect, the Highway replay service shall terminate the loop at the historical iteration count instead of re-evaluating until the runaway cap.
- If the original run's checkpoints contain no iteration count for a while loop, then the Highway replay service shall stop that loop at a configurable simulation cap (default well below 10000) and mark the simulation status "diverged", not "failed".
- If a while loop in replay is stopped by a cap, then the Highway replay service shall report a diagnostic naming the loop task_id, the condition text, and the variables that never resolved, instead of a bare RuntimeError.
- When a simulation is requested from the dashboard, the Highway API shall complete within a bounded time budget and shall not hold the request open while executing thousands of loop iterations.
EVIDENCE (reproduced 2026-08-03):
- Dashboard replay of workflow definition d0a0cb4c-1fa7-4d6a-a27b-bd19e573331c (workflow "agent_scheduled", name agent_sched_a1c703a7, tenant demo) -> RuntimeError: While loop exceeded maximum iterations (10000) at engine/interpreters/operators/loops.py:351.
- Definition: task agent_loop is operator_type=while with condition "{{agent_done}} == 0"; loop_body = single task agent_turn calling tools.agent.run_goal (result_key agent_step). agent_done is set by the tool as a side effect; under ReplayContext the side effect is mocked, so the condition is permanently true.
- Page load reported Finish: 50.64s - the request stays open for the whole runaway loop.
- Note the real (non-replay) runs of this definition complete normally (b4a5fb24-... completed in 12.4s), so this is a replay-fidelity defect, not a workflow defect.
Comments
Loading comments...
Context
Loading context...
Audit History
View AllLoading audit history...