test_assistant_service.py 69 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094
  1. """Tests for the assistant service (Phase 2) using synthetic event streams."""
  2. import itertools
  3. import json
  4. import re
  5. import threading
  6. import time
  7. from typing import Any
  8. from stt_daemon import events
  9. from stt_daemon.assistant.inference import InferenceJob, InferenceQueue
  10. from stt_daemon.assistant.service import EXECUTE_BUSY, AssistantService
  11. from stt_daemon.assistant.triage_dump import TriageDumper
  12. from stt_daemon.config import Config
  13. _sequence = itertools.count(1)
  14. class Clock:
  15. def __init__(self, start=1000.0):
  16. self.t = float(start)
  17. def __call__(self):
  18. return self.t
  19. def advance(self, dt):
  20. self.t += float(dt)
  21. class FakeSubscriber:
  22. def events(self):
  23. return iter(())
  24. def close(self):
  25. pass
  26. def _service():
  27. cfg = Config()
  28. clock = Clock()
  29. emitted: list[dict] = []
  30. def emit(type_, **fields):
  31. emitted.append({"type": type_, **fields})
  32. service = AssistantService(cfg, FakeSubscriber(), emit=emit, time_fn=clock)
  33. return service, emitted, clock
  34. def _ev(type_, clock, **fields):
  35. event = {"v": 1, "type": type_, "seq": next(_sequence), "ts": clock()}
  36. event.update(fields)
  37. return event
  38. def _transcript(clock, text, ts=None):
  39. return {
  40. "v": 1,
  41. "type": events.TRANSCRIPT,
  42. "seq": next(_sequence),
  43. "ts": clock() if ts is None else ts,
  44. "text": text,
  45. }
  46. def test_ready_requires_matching_segment_done():
  47. service, _, clock = _service()
  48. service.handle_event(_ev(events.SPEECH_START, clock))
  49. assert service.ready() is False
  50. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  51. assert service.ready() is False # awaiting segment 1
  52. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  53. assert service.ready() is True
  54. def test_stale_segment_done_does_not_satisfy_later_turn():
  55. service, _, clock = _service()
  56. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  57. service.handle_event(_ev(events.SPEECH_START, clock))
  58. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=2))
  59. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  60. assert service.ready() is False # still awaiting segment 2
  61. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=2, emitted=True))
  62. assert service.ready() is True
  63. def test_gated_segment_done_clears_readiness():
  64. service, _, clock = _service()
  65. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=3))
  66. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=3, emitted=False))
  67. assert service.ready() is True
  68. def test_speech_stop_without_seg_id_clears_readiness():
  69. """A discarded turn closes with speech_stop but no seg_id; readiness must recover."""
  70. service, _, clock = _service()
  71. service.handle_event(_ev(events.SPEECH_START, clock))
  72. assert service.ready() is False
  73. service.handle_event(_ev(events.SPEECH_STOP, clock)) # discarded: no seg_id
  74. assert service.speech_open is False
  75. assert service.awaited_seg_id is None
  76. assert service.ready() is True
  77. def test_status_resets_stale_speech_state_on_reconnect():
  78. service, _, clock = _service()
  79. service.handle_event(_ev(events.SPEECH_START, clock))
  80. assert service.speech_open is True
  81. service.handle_event(_ev(events.STATUS, clock, capture="x", segments_done=0))
  82. assert service.speech_open is False
  83. assert service.awaited_seg_id is None
  84. assert service.ready() is True
  85. def test_command_buffer_uses_arm_timestamp():
  86. service, _, clock = _service()
  87. clock.advance(10)
  88. assert service.arm() is True
  89. arm_ts = service.arm_ts
  90. assert arm_ts is not None
  91. # Replayed transcript from before arming is ignored.
  92. service.handle_event(_transcript(clock, "old", ts=arm_ts - 5))
  93. assert service.command_text == ""
  94. service.handle_event(_transcript(clock, "what time is it"))
  95. assert service.command_text == "what time is it"
  96. def test_mutual_exclusion_between_command_and_chat():
  97. service, _, _ = _service()
  98. assert service.toggle_chat() is True
  99. assert service.arm() is False # cannot arm in chat
  100. assert service.toggle_chat() is True # back to idle
  101. assert service.arm() is True
  102. assert service.toggle_chat() is False # cannot chat while armed
  103. def test_arm_timeout_disarms_and_emits_error():
  104. service, emitted, clock = _service()
  105. service.arm()
  106. clock.advance(service.cfg.assistant.command_arm_timeout_s + 0.1)
  107. service.tick()
  108. assert service.modes.mode.value == "idle"
  109. assert service.command_text == ""
  110. assert any(e["type"] == "error" and e["code"] == "arm_timeout" for e in emitted)
  111. def test_long_speech_does_not_disarm():
  112. service, emitted, clock = _service()
  113. service.arm()
  114. service.handle_event(_ev(events.SPEECH_START, clock))
  115. clock.advance(service.cfg.assistant.command_arm_timeout_s + 30.0)
  116. service.tick()
  117. assert service.modes.mode.value == "command_armed"
  118. assert not any(e["type"] == "error" for e in emitted)
  119. def test_captured_command_survives_arm_timeout():
  120. service, emitted, clock = _service()
  121. service.arm()
  122. service.handle_event(_transcript(clock, "what time is it"))
  123. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  124. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  125. clock.advance(service.cfg.assistant.command_arm_timeout_s + 30.0)
  126. service.tick()
  127. assert service.modes.mode.value == "command_armed"
  128. assert service.command_text == "what time is it"
  129. assert not any(e["type"] == "error" for e in emitted)
  130. def test_stt_stall_cancels_armed_command():
  131. service, emitted, clock = _service()
  132. service.handle_event(_ev(events.HEARTBEAT, clock, worker_stall_s=0.0))
  133. service.arm()
  134. clock.advance(service.cfg.assistant.stall_timeout_ms / 1000.0 + 0.5)
  135. service.tick()
  136. assert service.modes.mode.value == "idle"
  137. assert any(e["type"] == "error" and e["code"] == "stt_stall" for e in emitted)
  138. def test_worker_stall_marks_stt_degraded():
  139. service, _, clock = _service()
  140. long_stall = service.cfg.assistant.worker_stall_ms / 1000.0 + 1.0
  141. service.handle_event(_ev(events.HEARTBEAT, clock, worker_stall_s=long_stall))
  142. assert service.stt_ok() is False
  143. assert service.ready() is False
  144. def test_chat_idle_timeout_exits_chat():
  145. service, _, clock = _service()
  146. assert service.toggle_chat() is True
  147. clock.advance(service.cfg.assistant.chat_idle_timeout_s + 0.1)
  148. service.tick()
  149. assert service.modes.mode.value == "idle"
  150. def test_chat_exit_phrase_exits_chat():
  151. service, _, clock = _service()
  152. assert service.toggle_chat() is True
  153. service.handle_event(_transcript(clock, "That's all."))
  154. assert service.modes.mode.value == "idle"
  155. assert service.chat_history == []
  156. def test_viewer_detached_exits_chat_and_forgets_history():
  157. service, _, clock = _service()
  158. assert service.toggle_chat() is True
  159. service.handle_event(_transcript(clock, "hello"))
  160. assert service.viewer_detached() is True
  161. assert service.modes.mode.value == "idle"
  162. assert service.chat_history == []
  163. def test_viewer_detached_is_noop_outside_chat():
  164. service, _, _ = _service()
  165. assert service.viewer_detached() is False
  166. assert service.modes.mode.value == "idle"
  167. def test_exit_chat_invokes_chat_reset():
  168. cfg = Config()
  169. resets: list[int] = []
  170. service = AssistantService(
  171. cfg, FakeSubscriber(), time_fn=Clock(), chat_reset=lambda: resets.append(1)
  172. )
  173. service.toggle_chat()
  174. service.viewer_detached()
  175. assert resets == [1]
  176. def test_chat_accumulates_turns():
  177. service, _, clock = _service()
  178. service.toggle_chat()
  179. for text in ("hello there", "how are you"):
  180. service.handle_event(_transcript(clock, text))
  181. assert service.chat_history == ["Unknown: hello there", "Unknown: how are you"]
  182. def test_execute_requires_readiness():
  183. service, _, clock = _service()
  184. service.arm()
  185. service.handle_event(_transcript(clock, "do it"))
  186. service.handle_event(_ev(events.SPEECH_START, clock))
  187. assert service.execute() is None # speech still open
  188. assert service.modes.mode.value == "command_armed"
  189. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  190. assert service.execute() is None # awaiting segment 1
  191. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  192. assert service.execute() == "do it"
  193. assert service.modes.mode.value == "executing"
  194. assert service.finish() is True
  195. assert service.modes.mode.value == "idle"
  196. def test_execute_reports_busy_when_queue_rejects():
  197. service, emitted, clock = _service()
  198. # A zero-capacity lane rejects the priority command job on submit.
  199. service._inference = InferenceQueue(size=0)
  200. service.executor = object()
  201. service.arm()
  202. service.handle_event(_transcript(clock, "do it"))
  203. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  204. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  205. assert service.execute() is EXECUTE_BUSY
  206. assert service.modes.mode.value == "idle"
  207. assert not any(e.get("code") == "busy" for e in emitted)
  208. def test_turn_ended_requires_quiet_gap():
  209. service, _, clock = _service()
  210. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  211. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  212. assert service.ready() is True
  213. assert service.turn_ended() is False
  214. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  215. assert service.turn_ended() is True
  216. def test_emits_mode_and_command_text():
  217. service, emitted, clock = _service()
  218. service.arm()
  219. assert {"type": "mode", "mode": "command_armed"} in emitted
  220. service.handle_event(_transcript(clock, "hi"))
  221. assert any(e["type"] == "command_text" and e["text"] == "hi" for e in emitted)
  222. def test_cancel_clears_command_buffer():
  223. service, _, clock = _service()
  224. service.arm()
  225. service.handle_event(_transcript(clock, "hi"))
  226. assert service.cancel() is True
  227. assert service.modes.mode.value == "idle"
  228. assert service.command_text == ""
  229. assert service.arm_ts is None
  230. def test_pending_chat_is_not_lost_across_threads():
  231. """Transcripts appended on the reader thread while the ticker dispatches
  232. must all reach the model: the join-and-clear in _dispatch_chat is atomic,
  233. and the user bubble is emitted by the worker in _run_chat."""
  234. cfg = Config()
  235. cfg.assistant.turn_gap_s = 0.01
  236. dispatched: list[str] = []
  237. lock = threading.Lock()
  238. def emit(type_, **fields):
  239. if type_ == "chat_message" and fields.get("role") == "user":
  240. with lock:
  241. dispatched.append(fields["text"])
  242. service = AssistantService(cfg, FakeSubscriber(), emit=emit, chat_fn=lambda text: "ok")
  243. service.toggle_chat()
  244. service.start()
  245. try:
  246. total = 200
  247. producer = threading.Thread(
  248. target=lambda: [
  249. service.handle_event(
  250. {
  251. "v": 1,
  252. "type": events.TRANSCRIPT,
  253. "seq": i,
  254. "ts": time.time(),
  255. "text": f"w{i}",
  256. }
  257. )
  258. for i in range(total)
  259. ],
  260. daemon=True,
  261. )
  262. producer.start()
  263. deadline = time.time() + 10.0
  264. while time.time() < deadline:
  265. service.tick()
  266. with lock:
  267. words = sum(len(text.split()) for text in dispatched)
  268. done = words == 2 * total
  269. if not producer.is_alive() and done:
  270. break
  271. time.sleep(0.002)
  272. producer.join(timeout=5)
  273. with lock:
  274. words = sum(len(text.split()) for text in dispatched)
  275. assert words == 2 * total
  276. finally:
  277. service.stop()
  278. def _full_queue_service():
  279. service, emitted, clock = _service()
  280. service.toggle_chat()
  281. service._inference = InferenceQueue(size=1)
  282. service._inference.submit(
  283. InferenceJob(name="blocker", run=lambda: None, evictable=False)
  284. )
  285. return service, emitted, clock
  286. def test_chat_refusal_rebuffers_and_emits_queue_full_once():
  287. """A refused chat submit re-buffers under lock and emits queue_full on flip."""
  288. service, emitted, _ = _full_queue_service()
  289. service._pending_chat = ["hello world"]
  290. service._dispatch_chat()
  291. assert service._pending_chat == ["hello world"]
  292. assert service._chat_queue_full is True
  293. errors = [e for e in emitted if e["type"] == "error" and e.get("code") == "queue_full"]
  294. assert len(errors) == 1
  295. assert [e for e in emitted if e["type"] == "chat_message"] == []
  296. service._dispatch_chat()
  297. assert service._pending_chat == ["hello world"]
  298. errors = [e for e in emitted if e["type"] == "error" and e.get("code") == "queue_full"]
  299. assert len(errors) == 1
  300. def test_chat_dispatch_success_clears_latch_without_user_bubble():
  301. """Success clears the latch; the user bubble comes from _run_chat, not dispatch."""
  302. service, emitted, _ = _full_queue_service()
  303. service._pending_chat = ["hello world"]
  304. service._chat_queue_full = True
  305. service._inference.drop(lambda j: True)
  306. service._dispatch_chat()
  307. assert service._pending_chat == []
  308. assert service._chat_queue_full is False
  309. assert [e for e in emitted if e["type"] == "chat_message"] == []
  310. def test_run_chat_emits_user_first_then_reply():
  311. service, emitted, _ = _service()
  312. service.chat_fn = lambda text: "reply!"
  313. service._run_chat("hello")
  314. assert emitted == [
  315. {"type": "chat_message", "role": "user", "text": "hello"},
  316. {"type": "chat_message", "role": "assistant", "text": "reply!"},
  317. ]
  318. def test_chat_notify_emits_compaction_note():
  319. service, emitted, _ = _service()
  320. service.chat_notify("compacted", turns=4)
  321. assert {
  322. "type": "chat_message",
  323. "role": "note",
  324. "text": "Compacted 4 earlier turns into a summary.",
  325. } in emitted
  326. def test_chat_notify_emits_truncated_error():
  327. service, emitted, _ = _service()
  328. service.chat_notify("truncated", message="cut off")
  329. assert {"type": "error", "code": "truncated", "message": "cut off"} in emitted
  330. def test_clear_empties_chat_history():
  331. service, emitted, clock = _service()
  332. service.toggle_chat()
  333. for text in ("hello", "how are you"):
  334. service.handle_event(_transcript(clock, text))
  335. assert service.chat_history == ["Unknown: hello", "Unknown: how are you"]
  336. assert len(service._pending_chat) == 2
  337. assert service.clear() is True
  338. assert service.chat_history == []
  339. assert service._pending_chat == []
  340. assert {"type": "cleared"} in emitted
  341. def test_clear_invokes_chat_reset():
  342. cfg = Config()
  343. resets: list[int] = []
  344. service = AssistantService(
  345. cfg, FakeSubscriber(), time_fn=Clock(), chat_reset=lambda: resets.append(1)
  346. )
  347. service.toggle_chat()
  348. service.clear()
  349. assert resets == [1]
  350. def test_clear_does_not_change_mode():
  351. service, _, clock = _service()
  352. service.toggle_chat()
  353. assert service.modes.mode.value == "chat"
  354. service.clear()
  355. assert service.modes.mode.value == "chat"
  356. def test_clear_always_returns_true():
  357. service, _, _ = _service()
  358. assert service.clear() is True
  359. service.toggle_chat()
  360. service.handle_event(_ev(events.TRANSCRIPT, Clock(), text="test"))
  361. assert service.clear() is True
  362. def test_heartbeat_emits_inference_depth_and_in_flight():
  363. """Calling tick() emits heartbeat with inference_depth (int) and inference_in_flight (bool)."""
  364. service, emitted, clock = _service()
  365. service.tick()
  366. heartbeats = [e for e in emitted if e["type"] == "heartbeat"]
  367. assert len(heartbeats) == 1
  368. heartbeat = heartbeats[0]
  369. assert "inference_depth" in heartbeat
  370. assert isinstance(heartbeat["inference_depth"], int)
  371. assert "inference_in_flight" in heartbeat
  372. assert isinstance(heartbeat["inference_in_flight"], bool)
  373. def test_status_contains_inference_depth_and_in_flight():
  374. """status() snapshot contains inference_depth (int) and inference_in_flight (bool)."""
  375. service, _, _ = _service()
  376. status = service.status()
  377. assert "inference_depth" in status
  378. assert isinstance(status["inference_depth"], int)
  379. assert "inference_in_flight" in status
  380. assert isinstance(status["inference_in_flight"], bool)
  381. _AUTO = object()
  382. class FakeTriage:
  383. def __init__(
  384. self,
  385. questions=_AUTO,
  386. state_pair=("ok", ""),
  387. escalate=False,
  388. parsed=_AUTO,
  389. normalized=_AUTO,
  390. gate=True,
  391. ):
  392. if questions is _AUTO:
  393. questions = {"urgency": {"type": "noul", "instructions": "how urgent"}}
  394. self._questions = questions
  395. self._state_pair = state_pair
  396. self._escalate = escalate
  397. self._parsed = parsed
  398. self._normalized = normalized
  399. self._gate = gate
  400. self.build_state_calls: list[tuple] = []
  401. self.normalize_calls: list = []
  402. self.prompt_calls: list[tuple] = []
  403. self.gate_calls: list = []
  404. @property
  405. def questions(self):
  406. return self._questions
  407. def state(self):
  408. return self._state_pair
  409. def build_state(self, window):
  410. self.build_state_calls.append(tuple(window))
  411. return "\n".join("- " + str(item) for item in (window or []))
  412. def normalize(self, payload):
  413. self.normalize_calls.append(payload)
  414. if self._normalized is not _AUTO:
  415. return self._normalized
  416. if not isinstance(payload, dict):
  417. return {}
  418. signals = {"order": [], "items": {}, "escalated": False, "input_tokens": 0}
  419. signals.update(payload)
  420. if self._escalate:
  421. signals["escalated"] = True
  422. return signals
  423. def build_brief_prompt(self, deltas, signals):
  424. self.prompt_calls.append(
  425. (tuple(deltas), dict(signals) if isinstance(signals, dict) else signals)
  426. )
  427. return "BRIEF-PROMPT"
  428. def parse_brief(self, text):
  429. if self._parsed is _AUTO:
  430. return text if text else None
  431. return self._parsed
  432. def should_update_brief(self, signals):
  433. self.gate_calls.append(signals)
  434. return self._gate
  435. class FakeBriefFile:
  436. """Recording stand-in for BriefFile (write/clear never raise)."""
  437. def __init__(self):
  438. self.writes: list[str] = []
  439. self.clears = 0
  440. def write(self, text):
  441. self.writes.append(text)
  442. def clear(self):
  443. self.clears += 1
  444. class FakeLane:
  445. def __init__(self, ready=True):
  446. self._ready = ready
  447. self.submits: list[tuple] = []
  448. self.drops = 0
  449. self.predict_gate = threading.Event()
  450. def ready(self):
  451. return self._ready
  452. def state(self):
  453. return ("ok", "") if self._ready else ("loading", "")
  454. def submit(self, state, questions, lang, generation):
  455. self.submits.append((state, questions, lang, generation))
  456. return True
  457. def drop_pending(self):
  458. self.drops += 1
  459. def _ok_reply(text):
  460. return "ok"
  461. def _triage_service(
  462. triage=None,
  463. lane=None,
  464. brief_fn=None,
  465. chat_fn: Any = _AUTO,
  466. brief_file=None,
  467. brief_reset=None,
  468. chat_reset=None,
  469. ):
  470. cfg = Config()
  471. cfg.assistant.turn_gap_s = 0.5
  472. clock = Clock()
  473. emitted: list[dict] = []
  474. def emit(type_, **fields):
  475. emitted.append({"type": type_, **fields})
  476. if triage is None:
  477. triage = FakeTriage()
  478. if lane is None:
  479. lane = FakeLane()
  480. if chat_fn is _AUTO:
  481. chat_fn = _ok_reply
  482. service = AssistantService(
  483. cfg,
  484. FakeSubscriber(),
  485. emit=emit,
  486. time_fn=clock,
  487. triage=triage,
  488. laya_lane=lane,
  489. brief_fn=brief_fn,
  490. brief_reset=brief_reset,
  491. brief_file=brief_file,
  492. chat_fn=chat_fn,
  493. chat_reset=chat_reset,
  494. )
  495. return service, emitted, clock, triage, lane
  496. def _triage_events(emitted):
  497. return [e for e in emitted if e["type"] == "triage"]
  498. def _escalated_payload():
  499. return {"order": [], "items": {}, "escalated": True, "input_tokens": 3}
  500. def test_triage_status_disabled_when_no_analyzer():
  501. service, _, _ = _service()
  502. status = service.status()
  503. assert status["triage_state"] == "disabled"
  504. assert status["triage_reason"] == ""
  505. assert status["triage"] is None
  506. def test_triage_status_reflects_analyzer():
  507. service, _, _, _, _ = _triage_service(
  508. triage=FakeTriage(state_pair=("loading", "warming up"))
  509. )
  510. status = service.status()
  511. assert status["triage_state"] == "loading"
  512. assert status["triage_reason"] == "warming up"
  513. assert status["triage"] is None
  514. def test_status_contains_questions_from_analyzer():
  515. questions = {
  516. "urgency": {"type": "noul", "instructions": "how urgent"},
  517. "sentiment": {"type": "score", "instructions": "tone", "criteria": ["low", "high"]},
  518. }
  519. service, _, _, triage, _ = _triage_service(triage=FakeTriage(questions=questions))
  520. status = service.status()
  521. assert status["questions"] == questions
  522. assert status["questions"] is triage.questions
  523. def test_status_questions_none_without_analyzer():
  524. service, _, _ = _service()
  525. assert service.status()["questions"] is None
  526. def test_status_event_carries_questions():
  527. service, emitted, _, triage, _ = _triage_service()
  528. service.tick()
  529. statuses = [e for e in emitted if e["type"] == "status"]
  530. assert len(statuses) == 1
  531. assert statuses[0]["questions"] is triage.questions
  532. def test_triage_captures_transcript_lang():
  533. service, _, clock, _, _ = _triage_service()
  534. assert service._last_transcript_lang is None
  535. event = _transcript(clock, "hola")
  536. event["lang"] = "es"
  537. service.handle_event(event)
  538. assert service._last_transcript_lang == "es"
  539. service.handle_event(_transcript(clock, "hello"))
  540. assert service._last_transcript_lang is None
  541. def test_maybe_triage_passes_lang_to_lane():
  542. service, _, clock, _, lane = _triage_service()
  543. event = _transcript(clock, "hola")
  544. event["lang"] = "es"
  545. service.handle_event(event)
  546. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  547. service.tick()
  548. assert len(lane.submits) == 1
  549. _, _, lang, _ = lane.submits[0]
  550. assert lang == "es"
  551. def test_maybe_triage_runs_in_idle_mode():
  552. service, _, clock, triage, lane = _triage_service()
  553. service.handle_event(_transcript(clock, "hello"))
  554. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  555. service.tick()
  556. assert len(lane.submits) == 1
  557. state, questions, lang, gen = lane.submits[0]
  558. assert state == "- hello"
  559. assert questions is triage.questions
  560. assert lang is None
  561. assert gen == 0
  562. assert triage.build_state_calls == [("hello",)]
  563. assert service.triage_history == ["hello"]
  564. def test_maybe_triage_uses_transcript_buffer_not_chat_history():
  565. service, _, clock, triage, lane = _triage_service(chat_fn=None)
  566. service.chat_history = ["stale chat only"]
  567. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  568. service.tick()
  569. assert lane.submits == []
  570. assert triage.build_state_calls == []
  571. service.handle_event(_transcript(clock, "hello"))
  572. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  573. service.tick()
  574. assert len(lane.submits) == 1
  575. assert triage.build_state_calls == [("hello",)]
  576. def test_maybe_triage_requires_settled_turn():
  577. service, _, clock, _, lane = _triage_service()
  578. service.handle_event(_transcript(clock, "hello"))
  579. service.tick()
  580. assert lane.submits == []
  581. def test_maybe_triage_skips_unready_lane():
  582. service, _, clock, _, lane = _triage_service(lane=FakeLane(ready=False))
  583. service.handle_event(_transcript(clock, "hello"))
  584. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  585. service.tick()
  586. assert lane.submits == []
  587. def test_maybe_triage_skips_without_questions():
  588. service, _, clock, triage, lane = _triage_service(
  589. triage=FakeTriage(questions=None)
  590. )
  591. service.handle_event(_transcript(clock, "hello"))
  592. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  593. service.tick()
  594. assert lane.submits == []
  595. assert triage.build_state_calls == []
  596. def test_maybe_triage_dedupes_identical_window():
  597. service, _, clock, triage, lane = _triage_service()
  598. service.handle_event(_transcript(clock, "hello"))
  599. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  600. service.tick()
  601. assert len(lane.submits) == 1
  602. state, questions, lang, gen = lane.submits[0]
  603. assert state == "- hello"
  604. assert questions is triage.questions
  605. assert lang is None
  606. assert gen == 0
  607. service.tick()
  608. service.tick()
  609. assert len(lane.submits) == 1
  610. service.handle_event(_transcript(clock, "second"))
  611. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  612. service.tick()
  613. assert len(lane.submits) == 2
  614. assert lane.submits[1][0] == "- hello\n- second"
  615. def test_second_transcript_produces_second_submit():
  616. service, _, clock, _, lane = _triage_service()
  617. service.handle_event(_transcript(clock, "first"))
  618. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  619. service.tick()
  620. assert len(lane.submits) == 1
  621. service.handle_event(_transcript(clock, "second"))
  622. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  623. service.tick()
  624. assert len(lane.submits) == 2
  625. assert lane.submits[1][0] == "- first\n- second"
  626. def test_tick_under_50ms_with_blocked_lane():
  627. service, _, clock, _, lane = _triage_service()
  628. assert not lane.predict_gate.is_set()
  629. service.toggle_chat()
  630. service.handle_event(_transcript(clock, "hello"))
  631. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  632. start = time.perf_counter()
  633. service.tick()
  634. elapsed = time.perf_counter() - start
  635. assert len(lane.submits) == 1
  636. assert elapsed < 0.05
  637. def test_on_triage_result_emits_signals():
  638. service, emitted, clock, _, lane = _triage_service()
  639. service.toggle_chat()
  640. service.handle_event(_transcript(clock, "hello"))
  641. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  642. service.tick()
  643. _, _, _, gen = lane.submits[0]
  644. payload = {
  645. "order": ["urgency"],
  646. "items": {"urgency": {"kind": "noul"}},
  647. "escalated": False,
  648. "input_tokens": 12,
  649. }
  650. service._on_triage_result(gen, payload)
  651. staged = _triage_events(emitted)
  652. assert len(staged) == 1
  653. assert staged[0]["stage"] == "signals"
  654. assert staged[0]["triage"] == service.last_triage
  655. assert staged[0]["triage"]["escalated"] is False
  656. assert staged[0]["triage"]["input_tokens"] == 12
  657. assert service.status()["triage"] == service.last_triage
  658. def test_triage_events_carry_the_server_timestamp():
  659. service, emitted, clock, _, lane = _triage_service(brief_fn=lambda prompt: "RAW")
  660. service.toggle_chat()
  661. service.handle_event(_transcript(clock, "hello"))
  662. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  663. service.tick()
  664. _, _, _, gen = lane.submits[0]
  665. expected_ts = clock()
  666. service._on_triage_result(gen, _escalated_payload())
  667. signals = [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  668. assert len(signals) == 1
  669. assert isinstance(signals[0]["ts"], (int, float))
  670. assert signals[0]["ts"] == expected_ts
  671. job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  672. job.run()
  673. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  674. assert len(contents) == 1
  675. assert isinstance(contents[0]["ts"], (int, float))
  676. assert contents[0]["ts"] == expected_ts # content reuses the signals timestamp
  677. def test_on_triage_result_suppresses_stale_generation():
  678. service, emitted, clock, _, lane = _triage_service()
  679. service.toggle_chat()
  680. service.handle_event(_transcript(clock, "hello"))
  681. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  682. service.tick()
  683. _, _, _, gen = lane.submits[0]
  684. service.clear()
  685. service._on_triage_result(gen, _escalated_payload())
  686. assert service.last_triage is None
  687. assert [e for e in _triage_events(emitted) if e.get("stage") == "signals"] == []
  688. service.handle_event(_transcript(clock, "hello"))
  689. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  690. service.tick()
  691. assert len(lane.submits) == 2
  692. def test_pending_brief_capped_by_char_budget():
  693. """The buffer keeps the newest deltas within budget, dropping the oldest."""
  694. service, _, clock, _, _ = _triage_service(brief_fn=lambda prompt: "RAW")
  695. service.cfg.triage.brief_pending_char_budget = 30
  696. service.handle_event(_transcript(clock, "aaa"))
  697. service.handle_event(_transcript(clock, "bbbb"))
  698. assert service._pending_brief == ["Unknown: aaa", "Unknown: bbbb"]
  699. service.handle_event(_transcript(clock, "ccccc"))
  700. assert service._pending_brief == ["Unknown: bbbb", "Unknown: ccccc"]
  701. assert service.triage_history == ["aaa", "bbbb", "ccccc"]
  702. def test_pending_brief_untouched_without_triage():
  703. service, _, clock = _service()
  704. service.handle_event(_transcript(clock, "hello"))
  705. assert service._pending_brief == []
  706. assert service.triage_history == ["hello"]
  707. def test_on_triage_result_accepts_non_chat_mode():
  708. service, emitted, clock, _, lane = _triage_service()
  709. service.handle_event(_transcript(clock, "hello"))
  710. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  711. service.tick()
  712. _, _, _, gen = lane.submits[0]
  713. window = service._last_triage_window
  714. service._on_triage_result(gen, _escalated_payload())
  715. assert service.last_triage is not None
  716. assert service._last_triage_window == window
  717. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  718. def test_on_triage_result_ignores_none_payload():
  719. service, emitted, clock, _, lane = _triage_service()
  720. service.toggle_chat()
  721. service.handle_event(_transcript(clock, "hello"))
  722. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  723. service.tick()
  724. _, _, _, gen = lane.submits[0]
  725. service._on_triage_result(gen, None)
  726. assert service.last_triage is None
  727. assert _triage_events(emitted) == []
  728. def test_none_payload_dispatches_nothing_and_keeps_deltas():
  729. """A Laya predict failure never dispatches: no job, no events, deltas stay."""
  730. calls: list[str] = []
  731. service, emitted, clock, _, lane = _triage_service(
  732. brief_fn=lambda prompt: calls.append(prompt) or "RAW"
  733. )
  734. service.toggle_chat()
  735. service.handle_event(_transcript(clock, "hello"))
  736. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  737. service.tick()
  738. _, _, _, gen = lane.submits[0]
  739. assert service._pending_brief == ["Unknown: hello"]
  740. service._on_triage_result(gen, None)
  741. assert service.last_triage is None
  742. assert [e.get("stage") for e in _triage_events(emitted)] == []
  743. assert [j.name for j in service._inference._queue] == ["chat"]
  744. assert calls == []
  745. assert service._pending_brief == ["Unknown: hello"]
  746. assert service._brief_outstanding is False
  747. def test_none_payload_without_brief_fn_enqueues_nothing():
  748. service, emitted, clock, _, lane = _triage_service()
  749. service.toggle_chat()
  750. service.handle_event(_transcript(clock, "hello"))
  751. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  752. service.tick()
  753. _, _, _, gen = lane.submits[0]
  754. service._on_triage_result(gen, None)
  755. assert service.last_triage is None
  756. assert _triage_events(emitted) == []
  757. assert [j.name for j in service._inference._queue] == ["chat"]
  758. assert service._pending_brief == ["Unknown: hello"]
  759. def test_on_triage_result_ignores_empty_signals():
  760. service, emitted, clock, _, lane = _triage_service(
  761. triage=FakeTriage(normalized={})
  762. )
  763. service.toggle_chat()
  764. service.handle_event(_transcript(clock, "hello"))
  765. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  766. service.tick()
  767. _, _, _, gen = lane.submits[0]
  768. service._on_triage_result(gen, _escalated_payload())
  769. assert service.last_triage is None
  770. assert _triage_events(emitted) == []
  771. def test_substantive_turn_dispatches_one_brief_job():
  772. """A substantive settled turn calls the gate once and queues one brief job."""
  773. service, emitted, clock, triage, lane = _triage_service(
  774. triage=FakeTriage(escalate=True), brief_fn=lambda prompt: "RAW"
  775. )
  776. service.toggle_chat()
  777. service.handle_event(_transcript(clock, "we might churn"))
  778. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  779. service.tick()
  780. _, _, _, gen = lane.submits[0]
  781. service._on_triage_result(
  782. gen, {"order": [], "items": {}, "escalated": False, "input_tokens": 0}
  783. )
  784. assert service.last_triage is not None
  785. assert service.last_triage["escalated"] is True
  786. assert service._last_brief_signals == service.last_triage
  787. assert service._inference.depth() == 2
  788. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  789. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  790. assert len(triage.gate_calls) == 1
  791. assert service._brief_outstanding is True
  792. assert service._brief_token == 1
  793. def test_brief_dispatches_without_escalation():
  794. service, _, clock, _, lane = _triage_service(brief_fn=lambda prompt: "RAW")
  795. service.toggle_chat()
  796. service.handle_event(_transcript(clock, "all calm"))
  797. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  798. service.tick()
  799. _, _, _, gen = lane.submits[0]
  800. service._on_triage_result(
  801. gen, {"order": [], "items": {}, "escalated": False, "input_tokens": 0}
  802. )
  803. assert service.last_triage is not None
  804. assert service._inference.depth() == 2
  805. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  806. def test_gated_out_turn_skips_brief():
  807. """A gated-out turn stores signals but never calls brief_fn."""
  808. calls: list[str] = []
  809. service, emitted, clock, triage, lane = _triage_service(
  810. triage=FakeTriage(gate=False),
  811. brief_fn=lambda prompt: calls.append(prompt) or "RAW",
  812. )
  813. service.toggle_chat()
  814. service.handle_event(_transcript(clock, "small talk"))
  815. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  816. service.tick()
  817. _, _, _, gen = lane.submits[0]
  818. service._on_triage_result(gen, _escalated_payload())
  819. assert service.last_triage is not None
  820. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  821. assert calls == []
  822. assert [j.name for j in service._inference._queue] == ["chat"]
  823. assert service._brief_outstanding is False
  824. assert len(triage.gate_calls) == 1
  825. def test_no_brief_when_disabled_in_config():
  826. calls: list[str] = []
  827. service, _, clock, _, lane = _triage_service(
  828. triage=FakeTriage(escalate=True),
  829. brief_fn=lambda prompt: calls.append(prompt) or "RAW",
  830. )
  831. service.cfg.triage.extract_enabled = False
  832. service.toggle_chat()
  833. service.handle_event(_transcript(clock, "churn risk"))
  834. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  835. service.tick()
  836. _, _, _, gen = lane.submits[0]
  837. service._on_triage_result(gen, _escalated_payload())
  838. assert service.last_triage is not None
  839. assert service.last_triage["escalated"] is True
  840. assert calls == []
  841. assert [j.name for j in service._inference._queue] == ["chat"]
  842. def test_no_brief_without_brief_fn():
  843. service, _, clock, _, lane = _triage_service(triage=FakeTriage(escalate=True))
  844. service.toggle_chat()
  845. service.handle_event(_transcript(clock, "churn risk"))
  846. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  847. service.tick()
  848. _, _, _, gen = lane.submits[0]
  849. service._on_triage_result(gen, _escalated_payload())
  850. assert service.last_triage is not None
  851. assert [j.name for j in service._inference._queue] == ["chat"]
  852. assert service._brief_outstanding is False
  853. def test_brief_refused_when_queue_full():
  854. """A refused submit leaves deltas buffered and the outstanding flag clear."""
  855. service, emitted, clock, _, lane = _triage_service(
  856. triage=FakeTriage(escalate=True), brief_fn=lambda prompt: "RAW"
  857. )
  858. service._inference = InferenceQueue(size=2)
  859. service.toggle_chat()
  860. service.handle_event(_transcript(clock, "churn risk"))
  861. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  862. service.tick()
  863. assert service._inference.depth() == 1
  864. _, _, _, gen = lane.submits[0]
  865. service._on_triage_result(gen, _escalated_payload())
  866. assert service._inference.depth() == 1
  867. assert service._inference._queue[0].name == "chat"
  868. assert service.last_triage is not None
  869. assert service.last_triage["escalated"] is True
  870. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  871. assert service._pending_brief == ["Unknown: churn risk"]
  872. assert service._brief_outstanding is False
  873. def test_second_result_while_outstanding_is_refused():
  874. """At most one brief job is outstanding: the second turn only stores signals."""
  875. service, emitted, clock, _, lane = _triage_service(brief_fn=lambda prompt: "RAW")
  876. service.toggle_chat()
  877. service.handle_event(_transcript(clock, "first"))
  878. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  879. service.tick()
  880. _, _, _, gen0 = lane.submits[0]
  881. service._on_triage_result(gen0, _escalated_payload())
  882. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  883. service.handle_event(_transcript(clock, "second"))
  884. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  885. service.tick()
  886. _, _, _, gen1 = lane.submits[1]
  887. service._on_triage_result(gen1, _escalated_payload())
  888. jobs = [j for j in service._inference._queue if j.name == "brief-update"]
  889. assert len(jobs) == 1
  890. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals", "signals"]
  891. assert service._brief_outstanding is True
  892. assert service._pending_brief == ["Unknown: first", "Unknown: second"]
  893. def _seed_brief(service, *deltas):
  894. service._pending_brief = list(deltas)
  895. service._brief_outstanding = True
  896. return service._brief_token
  897. def test_brief_merges_content_and_writes_file():
  898. brief_file = FakeBriefFile()
  899. service, emitted, _, triage, _ = _triage_service(
  900. brief_fn=lambda prompt: "the brief", brief_file=brief_file
  901. )
  902. service.toggle_chat()
  903. signals = _escalated_payload()
  904. service._last_brief_signals = signals
  905. token = _seed_brief(service, "Unknown: hello")
  906. ts = service._now()
  907. service._run_brief_update(0, signals, ts, token)
  908. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  909. assert len(contents) == 1
  910. assert contents[0]["triage"] == {**signals, "brief": "the brief"}
  911. assert service.last_triage == contents[0]["triage"]
  912. assert triage.prompt_calls == [(("Unknown: hello",), signals)]
  913. assert brief_file.writes == ["the brief"]
  914. assert service._pending_brief == []
  915. assert service._brief_outstanding is False
  916. def test_empty_brief_requeues_deltas():
  917. """A `""` reply keeps signals, re-buffers the deltas, and clears the flag."""
  918. service, emitted, _, _, _ = _triage_service(brief_fn=lambda prompt: "")
  919. service.toggle_chat()
  920. signals = _escalated_payload()
  921. service.last_triage = signals
  922. service._last_brief_signals = signals
  923. token = _seed_brief(service, "Unknown: hello")
  924. ts = service._now()
  925. service._run_brief_update(0, signals, ts, token)
  926. assert [e for e in _triage_events(emitted) if e.get("stage") == "content"] == []
  927. assert service.last_triage == signals
  928. assert "brief" not in service.last_triage
  929. assert service._pending_brief == ["Unknown: hello"]
  930. assert service._brief_outstanding is False
  931. def test_brief_error_requeues_and_clears_flag():
  932. """A raising brief_fn leaves the flag clear and the deltas buffered."""
  933. def boom(prompt):
  934. raise RuntimeError("llm down")
  935. service, emitted, _, _, _ = _triage_service(brief_fn=boom)
  936. service.toggle_chat()
  937. token = _seed_brief(service, "Unknown: hello")
  938. ts = service._now()
  939. service._run_brief_update(0, _escalated_payload(), ts, token)
  940. assert _triage_events(emitted) == []
  941. assert service.last_triage is None
  942. assert service._pending_brief == ["Unknown: hello"]
  943. assert service._brief_outstanding is False
  944. def test_brief_empty_buffer_returns_early():
  945. """An empty buffer returns early: no call, no event, flag cleared."""
  946. calls: list[str] = []
  947. service, emitted, _, _, _ = _triage_service(
  948. brief_fn=lambda prompt: calls.append(prompt) or "RAW"
  949. )
  950. service.toggle_chat()
  951. service._brief_outstanding = True
  952. token = service._brief_token
  953. ts = service._now()
  954. assert service._pending_brief == []
  955. service._run_brief_update(0, _escalated_payload(), ts, token)
  956. assert calls == []
  957. assert _triage_events(emitted) == []
  958. assert service._brief_outstanding is False
  959. def test_brief_suppressed_after_reset():
  960. service, emitted, _, _, _ = _triage_service(brief_fn=lambda prompt: "RAW")
  961. service.toggle_chat()
  962. service._pending_brief = ["Unknown: hello"]
  963. service._reset_triage()
  964. emitted.clear()
  965. service._run_brief_update(0, _escalated_payload(), service._now(), 0)
  966. assert _triage_events(emitted) == []
  967. assert service.last_triage is None
  968. assert service._pending_brief == ["Unknown: hello"]
  969. def test_brief_works_after_leaving_chat():
  970. seen = []
  971. service, emitted, _, _, _ = _triage_service(
  972. brief_fn=lambda prompt: seen.append(prompt) or "RAW"
  973. )
  974. service.toggle_chat()
  975. assert service.modes.exit_chat() is True
  976. token = _seed_brief(service, "Unknown: hello")
  977. signals = _escalated_payload()
  978. service._last_brief_signals = signals
  979. service._run_brief_update(0, signals, service._now(), token)
  980. assert seen != []
  981. assert [e.get("stage") for e in _triage_events(emitted)] == ["content"]
  982. assert service.last_triage is not None
  983. assert service.last_triage["brief"] == "RAW"
  984. def test_deltas_arriving_during_call_survive_and_flush():
  985. """Deltas appended mid-call are not consumed; the finally flushes them."""
  986. service, emitted, clock, _, _ = _triage_service()
  987. service.toggle_chat()
  988. signals = _escalated_payload()
  989. service._last_brief_signals = signals
  990. token = _seed_brief(service, "Unknown: first")
  991. appended: list[int] = []
  992. def _brief(prompt):
  993. if not appended:
  994. appended.append(1)
  995. service.handle_event(_transcript(clock, "during call"))
  996. return "first brief"
  997. service._brief_fn = _brief
  998. service._run_brief_update(0, signals, clock(), token)
  999. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1000. assert len(contents) == 1
  1001. assert contents[0]["triage"]["brief"] == "first brief"
  1002. flushed = [j for j in service._inference._queue if j.name == "brief-update"]
  1003. assert len(flushed) == 1
  1004. assert service._brief_outstanding is True
  1005. assert service._pending_brief == ["Unknown: during call"]
  1006. flushed[0].run()
  1007. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1008. assert len(contents) == 2
  1009. assert contents[1]["triage"]["brief"] == "first brief"
  1010. assert service._pending_brief == []
  1011. assert service._brief_outstanding is False
  1012. def test_in_flight_brief_after_clear_emits_and_writes_nothing():
  1013. """Stale in-flight job: no event, no file write, no requeue, flag untouched."""
  1014. brief_file = FakeBriefFile()
  1015. service, emitted, clock, _, lane = _triage_service(
  1016. brief_fn=lambda prompt: "RAW", brief_file=brief_file
  1017. )
  1018. service.toggle_chat()
  1019. service.handle_event(_transcript(clock, "hello"))
  1020. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1021. service.tick()
  1022. _, _, _, gen = lane.submits[0]
  1023. service._on_triage_result(gen, _escalated_payload())
  1024. job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  1025. assert service._brief_outstanding is True
  1026. service.clear()
  1027. assert service._brief_outstanding is False
  1028. token_after_clear = service._brief_token
  1029. emitted.clear()
  1030. job.run()
  1031. assert [e for e in _triage_events(emitted)] == []
  1032. assert brief_file.writes == []
  1033. assert service._pending_brief == []
  1034. assert service.last_triage is None
  1035. assert service._brief_token == token_after_clear
  1036. assert service._brief_outstanding is False
  1037. def test_clear_resets_triage():
  1038. resets: list[int] = []
  1039. chats: list[int] = []
  1040. brief_file = FakeBriefFile()
  1041. service, emitted, clock, _, lane = _triage_service(
  1042. triage=FakeTriage(escalate=True),
  1043. brief_fn=lambda prompt: "RAW",
  1044. brief_file=brief_file,
  1045. brief_reset=lambda: resets.append(1),
  1046. chat_reset=lambda: chats.append(1),
  1047. )
  1048. service.toggle_chat()
  1049. service.handle_event(_transcript(clock, "churn"))
  1050. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1051. service.tick()
  1052. _, _, _, gen = lane.submits[0]
  1053. service._on_triage_result(gen, _escalated_payload())
  1054. assert service._inference.depth() == 2
  1055. assert service._pending_brief == ["Unknown: churn"]
  1056. assert service._brief_outstanding is True
  1057. token_before = service._brief_token
  1058. assert service.clear() is True
  1059. assert service._triage_generation == 1
  1060. assert service.last_triage is None
  1061. assert service._last_triage_window == ()
  1062. assert service.triage_history == []
  1063. assert service._pending_brief == []
  1064. assert service._brief_token == token_before + 1
  1065. assert service._brief_outstanding is False
  1066. assert service._last_brief_signals is None
  1067. assert resets == [1]
  1068. assert brief_file.clears == 1
  1069. assert chats == [1]
  1070. assert lane.drops == 1
  1071. assert service._inference.depth() == 0
  1072. assert list(service._inference._queue) == []
  1073. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1074. assert len(cleared) == 1
  1075. assert cleared[0]["triage"] is None
  1076. assert {"type": "cleared"} in emitted
  1077. signals_before = len(
  1078. [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  1079. )
  1080. service._on_triage_result(gen, _escalated_payload())
  1081. assert service.last_triage is None
  1082. assert (
  1083. len([e for e in _triage_events(emitted) if e.get("stage") == "signals"])
  1084. == signals_before
  1085. )
  1086. submits_before = len(lane.submits)
  1087. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1088. service.tick()
  1089. assert len(lane.submits) == submits_before
  1090. def test_exit_chat_does_not_reset_triage():
  1091. chats: list[int] = []
  1092. bresets: list[int] = []
  1093. brief_file = FakeBriefFile()
  1094. service, emitted, clock, _, lane = _triage_service(
  1095. brief_fn=lambda prompt: "RAW",
  1096. chat_reset=lambda: chats.append(1),
  1097. brief_reset=lambda: bresets.append(1),
  1098. brief_file=brief_file,
  1099. )
  1100. service.toggle_chat()
  1101. service.handle_event(_transcript(clock, "hello"))
  1102. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1103. service.tick()
  1104. assert len(lane.submits) == 1
  1105. _, _, _, gen = lane.submits[0]
  1106. service._on_triage_result(gen, _escalated_payload())
  1107. assert service.last_triage is not None
  1108. assert service._triage_generation == 0
  1109. assert service.toggle_chat() is True
  1110. assert service.modes.mode.value == "idle"
  1111. assert service._triage_generation == 0
  1112. assert service.last_triage is not None
  1113. assert service._last_triage_window == ("hello",)
  1114. assert service.triage_history == ["hello"]
  1115. assert service._pending_brief == ["Unknown: hello"]
  1116. assert service._last_brief_signals is not None
  1117. assert service._brief_outstanding is True
  1118. assert lane.drops == 0
  1119. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1120. assert cleared == []
  1121. assert chats == [1]
  1122. assert bresets == []
  1123. assert brief_file.clears == 0
  1124. def test_toggle_chat_enter_does_not_touch_triage():
  1125. """Entering chat resets only the chat buffers; the brief engine is independent."""
  1126. bresets: list[int] = []
  1127. brief_file = FakeBriefFile()
  1128. service, emitted, clock, _, lane = _triage_service(
  1129. brief_fn=lambda prompt: "RAW",
  1130. brief_reset=lambda: bresets.append(1),
  1131. brief_file=brief_file,
  1132. )
  1133. service.handle_event(_transcript(clock, "hello"))
  1134. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1135. service.tick()
  1136. _, _, _, gen = lane.submits[0]
  1137. service._on_triage_result(gen, _escalated_payload())
  1138. triage_before = service.last_triage
  1139. pending_before = list(service._pending_brief)
  1140. assert service.toggle_chat() is True
  1141. assert service.modes.mode.value == "chat"
  1142. assert service._triage_generation == 0
  1143. assert service.last_triage is triage_before
  1144. assert service._last_triage_window == ("hello",)
  1145. assert service.triage_history == ["hello"]
  1146. assert service._pending_brief == pending_before
  1147. assert service._last_brief_signals is not None
  1148. assert service._brief_outstanding is True
  1149. assert lane.drops == 0
  1150. assert bresets == []
  1151. assert brief_file.clears == 0
  1152. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1153. assert cleared == []
  1154. def test_clear_without_triage_emits_no_triage_event():
  1155. service, emitted, _ = _service()
  1156. brief_file = FakeBriefFile()
  1157. bresets: list[int] = []
  1158. service._brief_file = brief_file
  1159. service._brief_reset = lambda: bresets.append(1)
  1160. service.clear()
  1161. assert {"type": "cleared"} in emitted
  1162. assert _triage_events(emitted) == []
  1163. assert brief_file.clears == 1
  1164. assert bresets == []
  1165. def test_stale_brief_after_clear_does_not_clear_newer_dispatch_flag():
  1166. """A stale in-flight job must not clear a newer dispatch's outstanding flag."""
  1167. brief_file = FakeBriefFile()
  1168. service, emitted, clock, _, lane = _triage_service(
  1169. brief_fn=lambda prompt: "RAW", brief_file=brief_file
  1170. )
  1171. service.toggle_chat()
  1172. service.handle_event(_transcript(clock, "hello"))
  1173. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1174. service.tick()
  1175. _, _, _, gen = lane.submits[0]
  1176. service._on_triage_result(gen, _escalated_payload())
  1177. stale_job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  1178. assert service._brief_outstanding is True
  1179. service.clear()
  1180. # A fresh turn dispatches a new brief AFTER clear bumped the token.
  1181. service.handle_event(_transcript(clock, "fresh"))
  1182. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1183. service.tick()
  1184. _, _, _, gen2 = lane.submits[1]
  1185. service._on_triage_result(gen2, _escalated_payload())
  1186. assert service._brief_outstanding is True
  1187. newer_token = service._brief_token
  1188. pending_before = list(service._pending_brief)
  1189. triage_before = service.last_triage
  1190. emitted.clear()
  1191. stale_job.run()
  1192. assert _triage_events(emitted) == []
  1193. assert brief_file.writes == []
  1194. assert service._pending_brief == pending_before
  1195. assert service._brief_outstanding is True
  1196. assert service._brief_token == newer_token
  1197. assert service.last_triage is triage_before
  1198. def test_clear_resets_chat_latch():
  1199. service, emitted, _ = _service()
  1200. service.toggle_chat()
  1201. service._chat_queue_full = True
  1202. service._pending_chat = ["stuck text"]
  1203. service.clear()
  1204. assert service._chat_queue_full is False
  1205. assert service._pending_chat == []
  1206. assert {"type": "cleared"} in emitted
  1207. def test_triage_end_to_end_signals_content_cleared():
  1208. """Two settled turns collapse into ONE brief job (single-outstanding guard).
  1209. The second `_on_triage_result` is refused as `"outstanding"`, so the queue
  1210. holds one `brief-update` that consumes both deltas and emits one `content`.
  1211. The panel keeps the FIRST turn's signals until the next settled turn (lag).
  1212. """
  1213. prompts: list[str] = []
  1214. service, emitted, clock, triage, lane = _triage_service(
  1215. brief_fn=lambda prompt: prompts.append(prompt) or "the brief"
  1216. )
  1217. service.toggle_chat()
  1218. service.handle_event(_transcript(clock, "first thought"))
  1219. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1220. service.tick()
  1221. assert len(lane.submits) == 1
  1222. _, _, _, gen0 = lane.submits[0]
  1223. service._on_triage_result(
  1224. gen0, {"order": ["urgency"], "items": {}, "escalated": False, "input_tokens": 1}
  1225. )
  1226. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  1227. assert prompts == []
  1228. assert service.last_triage is not None
  1229. assert service.last_triage["escalated"] is False
  1230. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  1231. triage._escalate = True
  1232. service.handle_event(_transcript(clock, "we might churn"))
  1233. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1234. service.tick()
  1235. assert len(lane.submits) == 2
  1236. _, _, _, gen1 = lane.submits[1]
  1237. service._on_triage_result(gen1, _escalated_payload())
  1238. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals", "signals"]
  1239. assert [j.name for j in service._inference._queue] == [
  1240. "chat",
  1241. "brief-update",
  1242. "chat",
  1243. ]
  1244. for job in [j for j in service._inference._queue if j.name == "brief-update"]:
  1245. job.run()
  1246. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1247. assert len(contents) == 1
  1248. assert len(prompts) == 1
  1249. deltas, _ = triage.prompt_calls[0]
  1250. assert deltas == ("Unknown: first thought", "Unknown: we might churn")
  1251. assert contents[0]["triage"]["escalated"] is False
  1252. assert contents[0]["triage"]["brief"] == "the brief"
  1253. assert service.last_triage == contents[0]["triage"]
  1254. assert service.status()["triage"] == contents[0]["triage"]
  1255. service.clear()
  1256. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1257. assert len(cleared) == 1
  1258. assert cleared[0]["triage"] is None
  1259. assert service.last_triage is None
  1260. service._on_triage_result(gen1, _escalated_payload())
  1261. assert [e.get("stage") for e in _triage_events(emitted)] == [
  1262. "signals",
  1263. "signals",
  1264. "content",
  1265. "cleared",
  1266. ]
  1267. assert service.toggle_chat() is True
  1268. service.handle_event(_transcript(clock, "idle thought"))
  1269. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1270. service.tick()
  1271. assert len(lane.submits) == 3
  1272. assert lane.submits[2][0].endswith("- idle thought")
  1273. def test_tick_emits_status_on_triage_state_change():
  1274. service, emitted, _, triage, _ = _triage_service(
  1275. triage=FakeTriage(state_pair=("loading", "warming up"))
  1276. )
  1277. service.tick()
  1278. statuses = [e for e in emitted if e["type"] == "status"]
  1279. assert len(statuses) == 1
  1280. assert statuses[0]["triage_state"] == "loading"
  1281. assert statuses[0]["triage_reason"] == "warming up"
  1282. assert "triage" in statuses[0]
  1283. triage._state_pair = ("ok", "")
  1284. service.tick()
  1285. statuses = [e for e in emitted if e["type"] == "status"]
  1286. assert len(statuses) == 2
  1287. assert statuses[1]["triage_state"] == "ok"
  1288. assert statuses[1]["triage_reason"] == ""
  1289. assert "triage" in statuses[1]
  1290. service.tick()
  1291. assert len([e for e in emitted if e["type"] == "status"]) == 2
  1292. def test_no_status_emitted_when_triage_disabled():
  1293. service, emitted, _ = _service()
  1294. service.tick()
  1295. service.tick()
  1296. assert [e for e in emitted if e["type"] == "status"] == []
  1297. def test_tick_heals_triage_panel_after_lane_ready():
  1298. import time as _time
  1299. from pathlib import Path
  1300. from stt_daemon.assistant.laya_lane import LayaLane
  1301. from stt_daemon.assistant.triage import TriageAnalyzer
  1302. from stt_daemon.tui.state import ViewState
  1303. class _FakeRouter:
  1304. def preload(self, names):
  1305. return self
  1306. def predict(self, state, questions, model=None, lang=None, **kwargs):
  1307. if "_warmup" in questions:
  1308. return {"model": model, "answers": {}, "usage": {}, "routing": {}}
  1309. return {
  1310. "model": model,
  1311. "answers": {
  1312. "frustration": {
  1313. "noul": 0.1,
  1314. "confidence": 0.9,
  1315. "answer_confidence": 0.9,
  1316. },
  1317. "urgency": {
  1318. "choice": "low",
  1319. "confidence": 0.9,
  1320. "answer_confidence": 0.9,
  1321. },
  1322. },
  1323. "usage": {"input_tokens": 5},
  1324. }
  1325. def _factory(**kwargs):
  1326. return _FakeRouter()
  1327. cfg = Config()
  1328. cfg.assistant.turn_gap_s = 0.5
  1329. cfg.triage.questions_file = str(
  1330. Path(__file__).resolve().parent.parent / "questions.example.json"
  1331. )
  1332. clock = Clock()
  1333. emitted: list[dict] = []
  1334. def emit(type_, **fields):
  1335. emitted.append({"type": type_, **fields})
  1336. analyzer = TriageAnalyzer(cfg.triage)
  1337. assert analyzer.questions, "example questions must load"
  1338. holder: dict = {}
  1339. lane = LayaLane(
  1340. cfg.triage,
  1341. agent_factory=_factory,
  1342. on_result=lambda gen, payload: holder["svc"]._on_triage_result(gen, payload),
  1343. on_state=lambda state, reason: analyzer.refresh(state, reason),
  1344. )
  1345. service = AssistantService(
  1346. cfg,
  1347. FakeSubscriber(),
  1348. emit=emit,
  1349. time_fn=clock,
  1350. triage=analyzer,
  1351. laya_lane=lane,
  1352. chat_fn=_ok_reply,
  1353. )
  1354. holder["svc"] = service
  1355. try:
  1356. lane.start()
  1357. deadline = _time.monotonic() + 5.0
  1358. while not lane.ready() and _time.monotonic() < deadline:
  1359. _time.sleep(0.005)
  1360. assert lane.ready(), "lane did not become ready"
  1361. assert analyzer.state() == ("ok", "")
  1362. service.tick()
  1363. statuses = [e for e in emitted if e["type"] == "status"]
  1364. assert statuses, "tick emitted no status after lane ready"
  1365. assert statuses[-1]["triage_state"] == "ok"
  1366. service.handle_event(_transcript(clock, "hello"))
  1367. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1368. service.tick()
  1369. deadline = _time.monotonic() + 5.0
  1370. while service.last_triage is None and _time.monotonic() < deadline:
  1371. _time.sleep(0.005)
  1372. assert service.last_triage is not None, "no triage signals from lane"
  1373. view = ViewState()
  1374. for event in emitted:
  1375. view.ingest(event)
  1376. assert view.triage_state == "ok"
  1377. lines = view.triage_lines()
  1378. assert lines
  1379. assert not any("unavailable" in line for line in lines)
  1380. assert not any("loading" in line for line in lines)
  1381. finally:
  1382. lane.stop()
  1383. def _dump_brief(prompt, on_io=None):
  1384. """Brief fake that mimics LlmClient.chat's on_io phases."""
  1385. if on_io is not None:
  1386. on_io({"phase": "request", "url": "http://llm/v1", "body": {"prompt": prompt}})
  1387. on_io({"phase": "response", "payload": {"content": "parsed"}, "error": None})
  1388. return "BRIEF"
  1389. def _run_triage_flow(service, emitted, clock, lane, payload):
  1390. service.toggle_chat()
  1391. service.handle_event(_transcript(clock, "hello"))
  1392. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1393. service.tick()
  1394. _, _, _, gen = lane.submits[0]
  1395. service._on_triage_result(gen, payload)
  1396. return gen
  1397. def _dump_files(tmp_path):
  1398. return list(tmp_path.rglob("*.json"))
  1399. def test_dump_writes_full_laya_and_llm_io(tmp_path):
  1400. service, emitted, clock, triage, lane = _triage_service(brief_fn=_dump_brief)
  1401. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1402. service.cfg.triage.dump_io = True
  1403. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1404. jobs = [j for j in service._inference._queue if j.name == "brief-update"]
  1405. assert len(jobs) == 1
  1406. jobs[0].run()
  1407. files = _dump_files(tmp_path)
  1408. assert len(files) == 1
  1409. assert re.fullmatch(r"\d{6}-\d{3}\.json", files[0].name)
  1410. data = json.loads(files[0].read_text(encoding="utf-8"))
  1411. assert data["ts"] == clock()
  1412. assert isinstance(data["iso"], str) and data["iso"]
  1413. assert data["laya"]["request"]["state"] == "- hello"
  1414. assert data["laya"]["request"]["questions"] == triage.questions
  1415. assert data["laya"]["request"]["model"] == "multilingual"
  1416. assert data["laya"]["request"]["max_len"] == 1024
  1417. assert data["laya"]["response"] == _escalated_payload()
  1418. assert data["laya"]["error"] is None
  1419. assert data["llm"]["request"]["url"] == "http://llm/v1"
  1420. assert data["llm"]["request"]["body"] == {"prompt": "BRIEF-PROMPT"}
  1421. assert data["llm"]["response"] == {"content": "parsed"}
  1422. assert data["llm"]["error"] is None
  1423. assert service._triage_dump._pending == {}
  1424. def test_dump_laya_error_writes_laya_only_and_keeps_deltas(tmp_path):
  1425. """A Laya predict failure finalizes the Laya half with `llm: null`.
  1426. Nothing is dispatched, so no LLM half is ever recorded and the buffered
  1427. deltas survive for the next settled turn.
  1428. """
  1429. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1430. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1431. service.cfg.triage.dump_io = True
  1432. _run_triage_flow(service, emitted, clock, lane, None)
  1433. assert [j.name for j in service._inference._queue] == ["chat"]
  1434. assert service._pending_brief == ["Unknown: hello"]
  1435. files = _dump_files(tmp_path)
  1436. assert len(files) == 1
  1437. data = json.loads(files[0].read_text(encoding="utf-8"))
  1438. assert data["laya"]["response"] is None
  1439. assert data["laya"]["error"] == "predict failed"
  1440. assert data["llm"] is None
  1441. assert service._triage_dump._pending == {}
  1442. def test_dump_without_brief_has_laya_only(tmp_path):
  1443. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1444. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1445. service.cfg.triage.dump_io = True
  1446. service.cfg.triage.extract_enabled = False
  1447. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1448. files = _dump_files(tmp_path)
  1449. assert len(files) == 1
  1450. data = json.loads(files[0].read_text(encoding="utf-8"))
  1451. assert data["laya"]["response"] == _escalated_payload()
  1452. assert data["llm"] is None
  1453. assert service._triage_dump._pending == {}
  1454. def test_dump_queue_full_has_laya_only(tmp_path):
  1455. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1456. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1457. service.cfg.triage.dump_io = True
  1458. service._inference = InferenceQueue(size=2)
  1459. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1460. assert service._inference.depth() == 1
  1461. assert [j.name for j in service._inference._queue] == ["chat"]
  1462. assert service._pending_brief == ["Unknown: hello"]
  1463. files = _dump_files(tmp_path)
  1464. assert len(files) == 1
  1465. data = json.loads(files[0].read_text(encoding="utf-8"))
  1466. assert data["laya"]["response"] == _escalated_payload()
  1467. assert data["llm"] is None
  1468. assert service._triage_dump._pending == {}
  1469. def test_dump_disabled_writes_nothing(tmp_path):
  1470. service, emitted, clock, _, lane = _triage_service(
  1471. brief_fn=lambda prompt: "RAW"
  1472. )
  1473. assert service._triage_dump.enabled is False
  1474. service._triage_dump = TriageDumper(False, directory=str(tmp_path))
  1475. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1476. jobs = [j for j in service._inference._queue if j.name == "brief-update"]
  1477. assert len(jobs) == 1
  1478. jobs[0].run()
  1479. assert _dump_files(tmp_path) == []
  1480. def test_dump_stale_result_discards_without_file(tmp_path):
  1481. service, emitted, clock, triage, lane = _triage_service(brief_fn=_dump_brief)
  1482. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1483. service.cfg.triage.dump_io = True
  1484. service.toggle_chat()
  1485. service.handle_event(_transcript(clock, "hello"))
  1486. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1487. service.tick()
  1488. _, _, _, gen = lane.submits[0]
  1489. real_normalize = triage.normalize
  1490. def stale_normalize(payload):
  1491. # Simulate a clear landing between the first check/begin and the re-check.
  1492. service._triage_generation += 1
  1493. return real_normalize(payload)
  1494. triage.normalize = stale_normalize
  1495. service._on_triage_result(gen, _escalated_payload())
  1496. assert _dump_files(tmp_path) == []
  1497. assert service._triage_dump._pending == {}
  1498. def test_dump_stale_brief_discards_without_file(tmp_path):
  1499. service, _, clock, _, _ = _triage_service(brief_fn=_dump_brief)
  1500. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1501. service.cfg.triage.dump_io = True
  1502. ts = clock()
  1503. service._triage_dump.begin(ts, laya={"request": {}, "response": {}, "error": None})
  1504. service._pending_brief = ["Unknown: hello"]
  1505. service._triage_generation += 1 # stale before the brief even starts
  1506. service._run_brief_update(0, _escalated_payload(), ts, service._brief_token)
  1507. assert _dump_files(tmp_path) == []
  1508. assert service._triage_dump._pending == {}
  1509. assert service._pending_brief == ["Unknown: hello"]
  1510. def test_dump_brief_cleared_mid_flight_writes_nothing(tmp_path):
  1511. service, _, clock, _, _ = _triage_service(brief_fn=_dump_brief)
  1512. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1513. service.cfg.triage.dump_io = True
  1514. def stale_brief(prompt, on_io=None):
  1515. # Simulate a clear landing mid-LLM-call (after the request phase fired).
  1516. service._triage_generation += 1
  1517. if on_io is not None:
  1518. on_io({"phase": "request", "url": "http://llm/v1", "body": {}})
  1519. on_io({"phase": "response", "payload": {"c": 1}, "error": None})
  1520. return "RAW"
  1521. service._brief_fn = stale_brief
  1522. service._pending_brief = ["Unknown: hello"]
  1523. service._brief_outstanding = True
  1524. token = service._brief_token
  1525. ts = clock()
  1526. service._triage_dump.begin(ts, laya={"request": {}, "response": {}, "error": None})
  1527. service._run_brief_update(0, _escalated_payload(), ts, token)
  1528. assert _dump_files(tmp_path) == []
  1529. assert service._triage_dump._pending == {}
  1530. assert service._pending_brief == []
  1531. assert service._brief_outstanding is False
  1532. def test_dump_gated_out_writes_laya_only_and_no_pending(tmp_path):
  1533. """A turn the gate leaves out writes `llm: null` and leaks no pending record."""
  1534. service, emitted, clock, triage, lane = _triage_service(
  1535. triage=FakeTriage(escalate=True, gate=False), brief_fn=_dump_brief
  1536. )
  1537. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1538. service.cfg.triage.dump_io = True
  1539. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1540. assert [j.name for j in service._inference._queue] == ["chat"]
  1541. assert len(triage.gate_calls) == 1
  1542. files = _dump_files(tmp_path)
  1543. assert len(files) == 1
  1544. data = json.loads(files[0].read_text(encoding="utf-8"))
  1545. assert data["laya"]["response"] == _escalated_payload()
  1546. assert data["llm"] is None
  1547. assert service._triage_dump._pending == {}
  1548. def test_dump_coalesced_update_writes_laya_only_then_llm_only_flush(tmp_path):
  1549. """A coalesced turn writes a Laya-only record; the later flush a separate llm-only one."""
  1550. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1551. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1552. service.cfg.triage.dump_io = True
  1553. service.toggle_chat()
  1554. service.handle_event(_transcript(clock, "first"))
  1555. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1556. service.tick()
  1557. _, _, _, gen0 = lane.submits[0]
  1558. service._on_triage_result(gen0, _escalated_payload())
  1559. service.handle_event(_transcript(clock, "second"))
  1560. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1561. service.tick()
  1562. _, _, _, gen1 = lane.submits[1]
  1563. service._on_triage_result(gen1, _escalated_payload())
  1564. assert service._brief_outstanding is True
  1565. appended: list[int] = []
  1566. def _brief(prompt, on_io=None):
  1567. if not appended:
  1568. appended.append(1)
  1569. service.handle_event(_transcript(clock, "during call"))
  1570. if on_io is not None:
  1571. on_io({"phase": "request", "url": "http://llm/v1", "body": {"prompt": prompt}})
  1572. on_io({"phase": "response", "payload": {"content": "parsed"}, "error": None})
  1573. return "BRIEF"
  1574. service._brief_fn = _brief
  1575. clock.advance(1.0)
  1576. job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  1577. service._inference._queue.remove(job)
  1578. job.run()
  1579. flushed = [j for j in service._inference._queue if j.name == "brief-update"]
  1580. assert len(flushed) == 1
  1581. flushed[0].run()
  1582. assert service._triage_dump._pending == {}
  1583. records = [
  1584. json.loads(path.read_text(encoding="utf-8")) for path in _dump_files(tmp_path)
  1585. ]
  1586. assert len(records) == 3
  1587. coalesced = [r for r in records if r["laya"] is not None and r["llm"] is None]
  1588. flushed_records = [r for r in records if r["laya"] is None and r["llm"] is not None]
  1589. both = [r for r in records if r["laya"] is not None and r["llm"] is not None]
  1590. assert len(coalesced) == 1
  1591. assert len(flushed_records) == 1
  1592. assert len(both) == 1
  1593. assert coalesced[0]["ts"] != flushed_records[0]["ts"]
  1594. assert flushed_records[0]["llm"]["request"]["body"] == {"prompt": "BRIEF-PROMPT"}
  1595. def test_chat_client_failure_emits_llm_error_and_service_survives():
  1596. """Adversarial (a): a chat client raising on `chat` surfaces as an `llm_error`
  1597. event and the service keeps functioning afterwards."""
  1598. from stt_daemon.assistant.llm import ConversationEngine, LlmError
  1599. class BoomClient:
  1600. def chat(self, messages, max_tokens=None, timeout_s=None, **kwargs):
  1601. raise LlmError("boom")
  1602. engine = ConversationEngine(Config().llm, BoomClient(), system_prompt="SYS")
  1603. service, emitted, clock = _service()
  1604. service.chat_fn = engine.send
  1605. assert service.toggle_chat() is True
  1606. service._run_chat("hello")
  1607. errors = [e for e in emitted if e["type"] == "error" and e.get("code") == "llm_error"]
  1608. assert errors and "boom" in errors[-1]["message"]
  1609. assert service.toggle_chat() is True
  1610. assert service.modes.mode.value == "idle"