Repository navigation
TEST 128
Run the same seeded synthetic task through on-demand, periodic, and continuous fake capture streams using a pinned local model identity. Inject slow/partial token streams, queue saturation, repeated events, missing sequences, time-budget exhaustion, cancellation before and during inference, transport failure after partial output, and a late response. Observe every emitted request/token/action and capture-state timestamp.
All modes obey one finite session deadline, request/queue/resource bounds, and zero implicit retries. Evidence order and capture identity remain attached to every request/token/selection/action event. Cancellation settles or leaves an explicit unresolved/incomplete state; partial output never becomes a final assessment or grants new UI authority. No remote fallback occurs. scripts/tests/test_assisted_benchmark_session.py exercises the local candidate, failure injection, and one existing sealed UI-action grant in each mode; scripts/tests/test_assisted_benchmark_live_policy.py checks exact candidates, active-state/run/frame/VM/model binding, contradictory error acknowledgements, immutable contract state, and uncertain-call refusal. The scheduler records bounded timing/identity events and enum choices, not raw model prose. Incomplete sessions return no further choices; results retain actionApplied or actionOutcomeUnknown without pretending an attempted irreversible action was undone. Every result has authorizesAction: false. Run .venv/bin/python -m pytest scripts/tests/test_assisted_benchmark_session.py scripts/tests/test_assisted_benchmark_live_policy.py -q. Setting PROJECTBUDDY_ACTOR_FIXTURE_SERVER to a compiled Actor stdio entry enables SDK reads from active temporary capture fixtures through all three modes into one synthetic workload action, followed by fixture finalization. No real capture, VM, model inference, Windows execution, latency/quality calibration, or production pin is claimed.