Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.
Python agent-orchestrationai-agentai-coding-assistantclaude-codecoding-agentcomputer-usedesktop-applocal-firstmcppersonal-jarvisvoice-assistantvoice-coding
"""Shared realtime postmortem verdicts (probe + diag script consume these).
The whole point of the module is that the live-call probe and the
flight-recorder audit judge a session identically — so these tests pin the
verdict boundaries, not the wording.
"""
from __future__ import annotations
from dodaios.diagnostics.realtime_forensics import (
SPAWN_FIRST_AUDIO_BUDGET_MS,
SPAWN_READY_BUDGET_MS,
evaluate_postmortem,
extract_postmortems,
)
def _healthy() -> dict[str, object]:
return {
"session_id": "healthy",
"provider": "codex-subscription-realtime",
"turns_completed": 4,
"ready_ms": 2400,
"first_audio_ms": 4100,
"close_clean": True,
}
def _kinds(pm: dict[str, object]) -> set[str]:
return {finding.kind for finding in evaluate_postmortem(pm)}
def test_a_healthy_session_yields_no_findings() -> None:
assert evaluate_postmortem(_healthy()) == []
def test_every_failure_class_maps_to_its_finding() -> None:
sick = _healthy() | {
"response_splices": 2,
"quiescence_boundary_turns": 4,
"ungrounded_captions_dropped": 3,
"ungrounded_responses_refused": 7,
"self_dialogue_rebuilds": 1,
"handoff_obligation_misses": 2,
"stale_generations_dropped": 1,
"mute_emergency_releases": 2,
"max_loop_stall_ms": 15_000,
"sender_pacing_resyncs": 1,
"sender_shed_frames": 40,
"recv_dropped_frames": 5,
"rebuilds": 2,
"stun_retries": 1,
"language_flips": 2,
"ready_ms": SPAWN_READY_BUDGET_MS + 1,
"first_audio_ms": SPAWN_FIRST_AUDIO_BUDGET_MS + 1,
"close_clean": False,
}
assert _kinds(sick) == {
"splice-seam",
"turn-without-boundary",
"ungrounded-caption-storm",
"refusal-storm",
"self-dialogue",
"handoff-obligation-miss",
"stale-generation-dropped",
"mute-emergency-release",
"event-loop-stall",
"sender-behind-wall-clock",
"recv-overflow",
"rebuild-loop",
"language-mismatch",
"spawn-over-budget",
"unclean-close",
}
def test_severity_boundaries_hold() -> None:
# One ungrounded caption is a warn; two are the storm.
one = evaluate_postmortem(_healthy() | {"ungrounded_captions_dropped": 1})
assert [(f.severity, f.kind) for f in one] == [("warn", "ungrounded-caption")]
# Refusals: quiet below 3, warn at 3-4, high at the adapter's storm verdict.
assert _kinds(_healthy() | {"ungrounded_responses_refused": 2}) == set()
assert _kinds(_healthy() | {"ungrounded_responses_refused": 3}) == {"refusal-run"}
assert _kinds(_healthy() | {"ungrounded_responses_refused": 5}) == {
"refusal-storm"
}
# A single rebuild is a warn; two (or any STUN retry) are the loop class.
assert _kinds(_healthy() | {"rebuilds": 1}) == {"transport-rebuild"}
assert _kinds(_healthy() | {"stun_retries": 1}) == {"rebuild-loop"}
# A sub-second scheduling wobble is not a stall.
assert _kinds(_healthy() | {"max_loop_stall_ms": 900}) == set()
# first_audio_ms == 0 means "no audio ever" — reported by other counters,
# never as a spawn overrun.
assert _kinds(_healthy() | {"first_audio_ms": 0}) == set()
# One language flip is normal conversation behaviour (de -> en on purpose).
assert _kinds(_healthy() | {"language_flips": 1}) == set()
one_handoff_miss = evaluate_postmortem(
_healthy() | {"handoff_obligation_misses": 1}
)
assert [(f.severity, f.kind) for f in one_handoff_miss] == [
("warn", "handoff-obligation-miss")
]
def test_extract_postmortems_reads_both_capture_shapes() -> None:
recorder_row = {
"event": "RealtimeSessionPostmortem",
"payload": {"session_id": "a", "turns_completed": 1},
}
probe_row = {"src": "postmortem", "data": {"session_id": "b"}}
noise = [
{"event": "VoiceTurnCompleted", "payload": {}},
{"src": "notif", "data": {"method": "thread/realtime/x"}},
{"src": "postmortem", "data": "not-a-mapping"},
]
payloads = extract_postmortems([recorder_row, probe_row, *noise])
assert [p.get("session_id") for p in payloads] == ["a", "b"]