test_assistant_service.py 99 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963
  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 import TriageAnalyzer
  12. from stt_daemon.assistant.triage_dump import TriageDumper
  13. from stt_daemon.config import Config
  14. _sequence = itertools.count(1)
  15. # A real analyzer (pure, stdlib-only helpers) backing the fake's retention
  16. # guard so guard behaviour is exercised end to end.
  17. _real_analyzer = TriageAnalyzer(Config().triage)
  18. class Clock:
  19. def __init__(self, start=1000.0):
  20. self.t = float(start)
  21. def __call__(self):
  22. return self.t
  23. def advance(self, dt):
  24. self.t += float(dt)
  25. class FakeSubscriber:
  26. def events(self):
  27. return iter(())
  28. def close(self):
  29. pass
  30. def _service():
  31. cfg = Config()
  32. clock = Clock()
  33. emitted: list[dict] = []
  34. def emit(type_, **fields):
  35. emitted.append({"type": type_, **fields})
  36. service = AssistantService(cfg, FakeSubscriber(), emit=emit, time_fn=clock)
  37. return service, emitted, clock
  38. def _ev(type_, clock, **fields):
  39. event = {"v": 1, "type": type_, "seq": next(_sequence), "ts": clock()}
  40. event.update(fields)
  41. return event
  42. def _transcript(clock, text, ts=None, session_dir=None):
  43. event = {
  44. "v": 1,
  45. "type": events.TRANSCRIPT,
  46. "seq": next(_sequence),
  47. "ts": clock() if ts is None else ts,
  48. "text": text,
  49. }
  50. if session_dir is not None:
  51. event["session_dir"] = session_dir
  52. return event
  53. def test_ready_requires_matching_segment_done():
  54. service, _, clock = _service()
  55. service.handle_event(_ev(events.SPEECH_START, clock))
  56. assert service.ready() is False
  57. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  58. assert service.ready() is False # awaiting segment 1
  59. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  60. assert service.ready() is True
  61. def test_stale_segment_done_does_not_satisfy_later_turn():
  62. service, _, clock = _service()
  63. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  64. service.handle_event(_ev(events.SPEECH_START, clock))
  65. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=2))
  66. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  67. assert service.ready() is False # still awaiting segment 2
  68. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=2, emitted=True))
  69. assert service.ready() is True
  70. def test_gated_segment_done_clears_readiness():
  71. service, _, clock = _service()
  72. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=3))
  73. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=3, emitted=False))
  74. assert service.ready() is True
  75. def test_speech_stop_without_seg_id_clears_readiness():
  76. """A discarded turn closes with speech_stop but no seg_id; readiness must recover."""
  77. service, _, clock = _service()
  78. service.handle_event(_ev(events.SPEECH_START, clock))
  79. assert service.ready() is False
  80. service.handle_event(_ev(events.SPEECH_STOP, clock)) # discarded: no seg_id
  81. assert service.speech_open is False
  82. assert service.awaited_seg_id is None
  83. assert service.ready() is True
  84. def test_status_resets_stale_speech_state_on_reconnect():
  85. service, _, clock = _service()
  86. service.handle_event(_ev(events.SPEECH_START, clock))
  87. assert service.speech_open is True
  88. service.handle_event(_ev(events.STATUS, clock, capture="x", segments_done=0))
  89. assert service.speech_open is False
  90. assert service.awaited_seg_id is None
  91. assert service.ready() is True
  92. def test_command_buffer_uses_arm_timestamp():
  93. service, _, clock = _service()
  94. clock.advance(10)
  95. assert service.arm() is True
  96. arm_ts = service.arm_ts
  97. assert arm_ts is not None
  98. # Replayed transcript from before arming is ignored.
  99. service.handle_event(_transcript(clock, "old", ts=arm_ts - 5))
  100. assert service.command_text == ""
  101. service.handle_event(_transcript(clock, "what time is it"))
  102. assert service.command_text == "what time is it"
  103. def test_mutual_exclusion_between_command_and_chat():
  104. service, _, _ = _service()
  105. assert service.toggle_chat() is True
  106. assert service.arm() is False # cannot arm in chat
  107. assert service.toggle_chat() is True # back to idle
  108. assert service.arm() is True
  109. assert service.toggle_chat() is False # cannot chat while armed
  110. def test_arm_timeout_disarms_and_emits_error():
  111. service, emitted, clock = _service()
  112. service.arm()
  113. clock.advance(service.cfg.assistant.command_arm_timeout_s + 0.1)
  114. service.tick()
  115. assert service.modes.mode.value == "idle"
  116. assert service.command_text == ""
  117. assert any(e["type"] == "error" and e["code"] == "arm_timeout" for e in emitted)
  118. def test_long_speech_does_not_disarm():
  119. service, emitted, clock = _service()
  120. service.arm()
  121. service.handle_event(_ev(events.SPEECH_START, clock))
  122. clock.advance(service.cfg.assistant.command_arm_timeout_s + 30.0)
  123. service.tick()
  124. assert service.modes.mode.value == "command_armed"
  125. assert not any(e["type"] == "error" for e in emitted)
  126. def test_captured_command_survives_arm_timeout():
  127. service, emitted, clock = _service()
  128. service.arm()
  129. service.handle_event(_transcript(clock, "what time is it"))
  130. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  131. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  132. clock.advance(service.cfg.assistant.command_arm_timeout_s + 30.0)
  133. service.tick()
  134. assert service.modes.mode.value == "command_armed"
  135. assert service.command_text == "what time is it"
  136. assert not any(e["type"] == "error" for e in emitted)
  137. def test_stt_stall_cancels_armed_command():
  138. service, emitted, clock = _service()
  139. service.handle_event(_ev(events.HEARTBEAT, clock, worker_stall_s=0.0))
  140. service.arm()
  141. clock.advance(service.cfg.assistant.stall_timeout_ms / 1000.0 + 0.5)
  142. service.tick()
  143. assert service.modes.mode.value == "idle"
  144. assert any(e["type"] == "error" and e["code"] == "stt_stall" for e in emitted)
  145. def test_worker_stall_marks_stt_degraded():
  146. service, _, clock = _service()
  147. long_stall = service.cfg.assistant.worker_stall_ms / 1000.0 + 1.0
  148. service.handle_event(_ev(events.HEARTBEAT, clock, worker_stall_s=long_stall))
  149. assert service.stt_ok() is False
  150. assert service.ready() is False
  151. def test_chat_idle_timeout_exits_chat():
  152. service, _, clock = _service()
  153. assert service.toggle_chat() is True
  154. clock.advance(service.cfg.assistant.chat_idle_timeout_s + 0.1)
  155. service.tick()
  156. assert service.modes.mode.value == "idle"
  157. def test_chat_exit_phrase_exits_chat():
  158. service, _, clock = _service()
  159. assert service.toggle_chat() is True
  160. service.set_chat_capture(True)
  161. service.handle_event(_transcript(clock, "That's all."))
  162. assert service.modes.mode.value == "idle"
  163. def test_exit_phrase_preserves_chat_history():
  164. service, _, clock = _service()
  165. service.toggle_chat()
  166. service.set_chat_capture(True)
  167. service.handle_event(_transcript(clock, "hello"))
  168. service.handle_event(_transcript(clock, "That's all."))
  169. assert service.modes.mode.value == "idle"
  170. assert service.chat_history == ["Unknown: hello", "Unknown: That's all."]
  171. def test_viewer_detached_exits_chat_and_keeps_history():
  172. service, _, clock = _service()
  173. assert service.toggle_chat() is True
  174. service.set_chat_capture(True)
  175. service.handle_event(_transcript(clock, "hello"))
  176. assert service.viewer_detached() is True
  177. assert service.modes.mode.value == "idle"
  178. assert service.chat_history == ["Unknown: hello"]
  179. def test_viewer_detached_is_noop_outside_chat():
  180. service, _, _ = _service()
  181. assert service.viewer_detached() is False
  182. assert service.modes.mode.value == "idle"
  183. def test_exit_chat_preserves_history_and_does_not_reset():
  184. cfg = Config()
  185. resets: list[int] = []
  186. service = AssistantService(
  187. cfg, FakeSubscriber(), time_fn=Clock(), chat_reset=lambda: resets.append(1)
  188. )
  189. service.toggle_chat()
  190. service.set_chat_capture(True)
  191. service.handle_event(_transcript(Clock(), "hello"))
  192. service.viewer_detached()
  193. service.toggle_chat()
  194. assert resets == []
  195. assert service.chat_history == ["Unknown: hello"]
  196. def test_chat_accumulates_turns():
  197. service, _, clock = _service()
  198. service.toggle_chat()
  199. service.set_chat_capture(True)
  200. for text in ("hello there", "how are you"):
  201. service.handle_event(_transcript(clock, text))
  202. assert service.chat_history == ["Unknown: hello there", "Unknown: how are you"]
  203. def test_chat_capture_off_skips_spoken_transcripts_and_exit_phrase():
  204. service, _, clock = _service()
  205. service.toggle_chat()
  206. assert service.set_chat_capture(False) is True
  207. service.handle_event(_transcript(clock, "hello"))
  208. service.handle_event(_transcript(clock, "That's all."))
  209. assert service.chat_history == []
  210. assert service._pending_chat == []
  211. assert service.modes.mode.value == "chat"
  212. def test_chat_capture_off_still_allows_typed_turns_and_triage():
  213. service, _, clock = _service()
  214. service.toggle_chat()
  215. service.set_chat_capture(False)
  216. service.handle_event(_transcript(clock, "spoken"))
  217. assert service.submit_chat_text("typed") is True
  218. assert service.chat_history == ["typed"]
  219. assert service.triage_history == ["spoken"]
  220. def test_set_chat_capture_emits_once_and_reports_status():
  221. service, emitted, _ = _service()
  222. assert service.status()["chat_capture"] is False
  223. assert service.set_chat_capture(True) is True
  224. assert [e for e in emitted if e["type"] == "chat_capture"] == [
  225. {"type": "chat_capture", "on": True}
  226. ]
  227. assert service.status()["chat_capture"] is True
  228. emitted.clear()
  229. service.set_chat_capture(True)
  230. assert emitted == []
  231. def test_triage_chat_off_ignores_chat_turns():
  232. service, _, _, _, lane = _triage_service()
  233. service._run_chat("typed")
  234. assert lane.submits == []
  235. assert service.triage_history == []
  236. def test_triage_chat_sends_reply_only_when_capture_on():
  237. service, _, _, _, lane = _triage_service()
  238. service.toggle_chat()
  239. service.set_chat_capture(True)
  240. service.set_triage_chat(True)
  241. service._run_chat("spoken, already in triage")
  242. assert service.triage_history == ["Assistant: ok"]
  243. assert lane.submits[0][0] == "- Assistant: ok"
  244. def test_triage_chat_sends_user_and_reply_when_capture_off():
  245. service, _, _, _, lane = _triage_service()
  246. service.toggle_chat()
  247. service.set_chat_capture(False)
  248. service.set_triage_chat(True)
  249. service._run_chat("typed")
  250. assert service.triage_history == ["You: typed", "Assistant: ok"]
  251. assert lane.submits[0][0] == "- You: typed\n- Assistant: ok"
  252. def test_set_triage_chat_emits_once_and_reports_status():
  253. service, emitted, _, _, _ = _triage_service()
  254. assert service.status()["triage_chat"] is False
  255. assert service.set_triage_chat(True) is True
  256. assert [e for e in emitted if e["type"] == "triage_chat"] == [
  257. {"type": "triage_chat", "on": True}
  258. ]
  259. assert service.status()["triage_chat"] is True
  260. emitted.clear()
  261. service.set_triage_chat(True)
  262. assert emitted == []
  263. def test_submit_chat_text_requires_chat_and_dispatches_immediately():
  264. cfg = Config()
  265. cfg.assistant.turn_gap_s = 999.0
  266. replies: list[str] = []
  267. emitted: list[dict] = []
  268. def emit(type_, **fields):
  269. emitted.append({"type": type_, **fields})
  270. service = AssistantService(
  271. cfg,
  272. FakeSubscriber(),
  273. emit=emit,
  274. chat_fn=lambda text: replies.append(text) or "ok",
  275. time_fn=Clock(),
  276. )
  277. assert service.submit_chat_text("hi") is False
  278. assert service.submit_chat_text(" ") is False
  279. service.toggle_chat()
  280. service.start()
  281. try:
  282. assert service.submit_chat_text("hi") is True
  283. deadline = time.time() + 2.0
  284. while time.time() < deadline and not replies:
  285. time.sleep(0.01)
  286. finally:
  287. service.stop()
  288. assert replies == ["hi"]
  289. assert service.chat_history == ["hi"]
  290. assert service._pending_chat == []
  291. assert any(
  292. e["type"] == "chat_message" and e.get("role") == "user" and e.get("text") == "hi"
  293. for e in emitted
  294. )
  295. def test_execute_clears_command_buffer_immediately():
  296. service, emitted, clock = _service()
  297. service.arm()
  298. service.handle_event(_transcript(clock, "do it"))
  299. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  300. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  301. assert service.execute() == "do it"
  302. assert service.modes.mode.value == "executing"
  303. assert service.command_text == ""
  304. assert service.command_edited is False
  305. assert service.arm_ts is None
  306. assert any(e["type"] == "command_text" and e["text"] == "" for e in emitted)
  307. def test_execute_requires_readiness():
  308. service, _, clock = _service()
  309. service.arm()
  310. service.handle_event(_transcript(clock, "do it"))
  311. service.handle_event(_ev(events.SPEECH_START, clock))
  312. assert service.execute() is None # speech still open
  313. assert service.modes.mode.value == "command_armed"
  314. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  315. assert service.execute() is None # awaiting segment 1
  316. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  317. assert service.execute() == "do it"
  318. assert service.modes.mode.value == "executing"
  319. assert service.finish() is True
  320. assert service.modes.mode.value == "idle"
  321. def test_execute_reports_busy_when_queue_rejects():
  322. service, emitted, clock = _service()
  323. # A zero-capacity lane rejects the priority command job on submit.
  324. service._inference = InferenceQueue(size=0)
  325. service.executor = object()
  326. service.arm()
  327. service.handle_event(_transcript(clock, "do it"))
  328. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  329. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  330. assert service.execute() is EXECUTE_BUSY
  331. assert service.modes.mode.value == "idle"
  332. assert not any(e.get("code") == "busy" for e in emitted)
  333. def test_turn_ended_requires_quiet_gap():
  334. service, _, clock = _service()
  335. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  336. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  337. assert service.ready() is True
  338. assert service.turn_ended() is False
  339. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  340. assert service.turn_ended() is True
  341. def test_emits_mode_and_command_text():
  342. service, emitted, clock = _service()
  343. service.arm()
  344. assert {"type": "mode", "mode": "command_armed"} in emitted
  345. service.handle_event(_transcript(clock, "hi"))
  346. assert any(e["type"] == "command_text" and e["text"] == "hi" for e in emitted)
  347. def test_cancel_clears_command_buffer():
  348. service, _, clock = _service()
  349. service.arm()
  350. service.handle_event(_transcript(clock, "hi"))
  351. assert service.cancel() is True
  352. assert service.modes.mode.value == "idle"
  353. assert service.command_text == ""
  354. assert service.arm_ts is None
  355. def test_pending_chat_is_not_lost_across_threads():
  356. """Transcripts appended on the reader thread while the ticker dispatches
  357. must all reach the model: the join-and-clear in _dispatch_chat is atomic,
  358. and the user bubble is emitted by the worker in _run_chat."""
  359. cfg = Config()
  360. cfg.assistant.turn_gap_s = 0.01
  361. dispatched: list[str] = []
  362. lock = threading.Lock()
  363. def emit(type_, **fields):
  364. if type_ == "chat_message" and fields.get("role") == "user":
  365. with lock:
  366. dispatched.append(fields["text"])
  367. service = AssistantService(cfg, FakeSubscriber(), emit=emit, chat_fn=lambda text: "ok")
  368. service.toggle_chat()
  369. service.set_chat_capture(True)
  370. service.start()
  371. try:
  372. total = 200
  373. producer = threading.Thread(
  374. target=lambda: [
  375. service.handle_event(
  376. {
  377. "v": 1,
  378. "type": events.TRANSCRIPT,
  379. "seq": i,
  380. "ts": time.time(),
  381. "text": f"w{i}",
  382. }
  383. )
  384. for i in range(total)
  385. ],
  386. daemon=True,
  387. )
  388. producer.start()
  389. deadline = time.time() + 10.0
  390. while time.time() < deadline:
  391. service.tick()
  392. with lock:
  393. words = sum(len(text.split()) for text in dispatched)
  394. done = words == 2 * total
  395. if not producer.is_alive() and done:
  396. break
  397. time.sleep(0.002)
  398. producer.join(timeout=5)
  399. with lock:
  400. words = sum(len(text.split()) for text in dispatched)
  401. assert words == 2 * total
  402. finally:
  403. service.stop()
  404. def _full_queue_service():
  405. service, emitted, clock = _service()
  406. service.toggle_chat()
  407. service._inference = InferenceQueue(size=1)
  408. service._inference.submit(
  409. InferenceJob(name="blocker", run=lambda: None, evictable=False)
  410. )
  411. return service, emitted, clock
  412. def test_chat_refusal_rebuffers_and_emits_queue_full_once():
  413. """A refused chat submit re-buffers under lock and emits queue_full on flip."""
  414. service, emitted, _ = _full_queue_service()
  415. service._pending_chat = ["hello world"]
  416. service._dispatch_chat()
  417. assert service._pending_chat == ["hello world"]
  418. assert service._chat_queue_full is True
  419. errors = [e for e in emitted if e["type"] == "error" and e.get("code") == "queue_full"]
  420. assert len(errors) == 1
  421. assert [e for e in emitted if e["type"] == "chat_message"] == []
  422. service._dispatch_chat()
  423. assert service._pending_chat == ["hello world"]
  424. errors = [e for e in emitted if e["type"] == "error" and e.get("code") == "queue_full"]
  425. assert len(errors) == 1
  426. def test_chat_dispatch_success_clears_latch_without_user_bubble():
  427. """Success clears the latch; the user bubble comes from _run_chat, not dispatch."""
  428. service, emitted, _ = _full_queue_service()
  429. service._pending_chat = ["hello world"]
  430. service._chat_queue_full = True
  431. service._inference.drop(lambda j: True)
  432. service._dispatch_chat()
  433. assert service._pending_chat == []
  434. assert service._chat_queue_full is False
  435. assert [e for e in emitted if e["type"] == "chat_message"] == []
  436. def test_run_chat_emits_user_first_then_reply():
  437. service, emitted, _ = _service()
  438. service.chat_fn = lambda text: "reply!"
  439. service._run_chat("hello")
  440. assert emitted == [
  441. {"type": "chat_message", "role": "user", "text": "hello"},
  442. {"type": "chat_message", "role": "assistant", "text": "reply!"},
  443. ]
  444. def test_chat_notify_emits_compaction_note():
  445. service, emitted, _ = _service()
  446. service.chat_notify("compacted", turns=4)
  447. assert {
  448. "type": "chat_message",
  449. "role": "note",
  450. "text": "Compacted 4 earlier turns into a summary.",
  451. } in emitted
  452. def test_chat_notify_emits_truncated_error():
  453. service, emitted, _ = _service()
  454. service.chat_notify("truncated", message="cut off")
  455. assert {"type": "error", "code": "truncated", "message": "cut off"} in emitted
  456. def test_clear_empties_chat_history():
  457. service, emitted, clock = _service()
  458. service.toggle_chat()
  459. service.set_chat_capture(True)
  460. for text in ("hello", "how are you"):
  461. service.handle_event(_transcript(clock, text))
  462. assert service.chat_history == ["Unknown: hello", "Unknown: how are you"]
  463. assert len(service._pending_chat) == 2
  464. assert service.clear() is True
  465. assert service.chat_history == []
  466. assert service._pending_chat == []
  467. assert {"type": "cleared"} in emitted
  468. def test_clear_invokes_chat_reset():
  469. cfg = Config()
  470. resets: list[int] = []
  471. service = AssistantService(
  472. cfg, FakeSubscriber(), time_fn=Clock(), chat_reset=lambda: resets.append(1)
  473. )
  474. service.toggle_chat()
  475. service.clear()
  476. assert resets == [1]
  477. def test_clear_does_not_change_mode():
  478. service, _, clock = _service()
  479. service.toggle_chat()
  480. assert service.modes.mode.value == "chat"
  481. service.clear()
  482. assert service.modes.mode.value == "chat"
  483. def test_clear_always_returns_true():
  484. service, _, _ = _service()
  485. assert service.clear() is True
  486. service.toggle_chat()
  487. service.handle_event(_ev(events.TRANSCRIPT, Clock(), text="test"))
  488. assert service.clear() is True
  489. def test_heartbeat_emits_inference_depth_and_in_flight():
  490. """Calling tick() emits heartbeat with inference_depth (int) and inference_in_flight (bool)."""
  491. service, emitted, clock = _service()
  492. service.tick()
  493. heartbeats = [e for e in emitted if e["type"] == "heartbeat"]
  494. assert len(heartbeats) == 1
  495. heartbeat = heartbeats[0]
  496. assert "inference_depth" in heartbeat
  497. assert isinstance(heartbeat["inference_depth"], int)
  498. assert "inference_in_flight" in heartbeat
  499. assert isinstance(heartbeat["inference_in_flight"], bool)
  500. def test_status_contains_inference_depth_and_in_flight():
  501. """status() snapshot contains inference_depth (int) and inference_in_flight (bool)."""
  502. service, _, _ = _service()
  503. status = service.status()
  504. assert "inference_depth" in status
  505. assert isinstance(status["inference_depth"], int)
  506. assert "inference_in_flight" in status
  507. assert isinstance(status["inference_in_flight"], bool)
  508. _AUTO = object()
  509. class FakeTriage:
  510. def __init__(
  511. self,
  512. questions=_AUTO,
  513. state_pair=("ok", ""),
  514. escalate=False,
  515. parsed=_AUTO,
  516. normalized=_AUTO,
  517. gate=True,
  518. ):
  519. if questions is _AUTO:
  520. questions = {"urgency": {"type": "noul", "instructions": "how urgent"}}
  521. self._questions = questions
  522. self._state_pair = state_pair
  523. self._escalate = escalate
  524. self._parsed = parsed
  525. self._normalized = normalized
  526. self._gate = gate
  527. self.build_state_calls: list[tuple] = []
  528. self.normalize_calls: list = []
  529. self.prompt_calls: list[tuple] = []
  530. self.gate_calls: list = []
  531. @property
  532. def questions(self):
  533. return self._questions
  534. def state(self):
  535. return self._state_pair
  536. def build_state(self, window):
  537. self.build_state_calls.append(tuple(window))
  538. return "\n".join("- " + str(item) for item in (window or []))
  539. def normalize(self, payload):
  540. self.normalize_calls.append(payload)
  541. if self._normalized is not _AUTO:
  542. return self._normalized
  543. if not isinstance(payload, dict):
  544. return {}
  545. signals = {"order": [], "items": {}, "escalated": False, "input_tokens": 0}
  546. signals.update(payload)
  547. if self._escalate:
  548. signals["escalated"] = True
  549. return signals
  550. def build_brief_prompt(self, deltas, signals):
  551. self.prompt_calls.append(
  552. (tuple(deltas), dict(signals) if isinstance(signals, dict) else signals)
  553. )
  554. return "BRIEF-PROMPT"
  555. def build_ledger_prompt(self, narrative, ledger, deltas, signals):
  556. self.prompt_calls.append(
  557. (tuple(deltas), dict(signals) if isinstance(signals, dict) else signals)
  558. )
  559. ledger_text = "\n".join("- " + f for f in ledger) or "(none yet)"
  560. body = "\n".join("- " + d for d in deltas)
  561. return (
  562. f"CURRENT LEDGER:\n{ledger_text}\n\n"
  563. f"CURRENT BRIEF:\n{narrative or '(none yet)'}\n\n"
  564. f"NEW UTTERANCES:\n{body}"
  565. )
  566. def parse_ledger_reply(self, text, truncated=False):
  567. return _real_analyzer.parse_ledger_reply(text, truncated=truncated)
  568. def is_regression(self, prior_brief, new_brief):
  569. return _real_analyzer.is_regression(prior_brief, new_brief)
  570. def parse_facts(self, block):
  571. return _real_analyzer.parse_facts(block)
  572. def parse_brief(self, text):
  573. if self._parsed is _AUTO:
  574. return text if text else None
  575. return self._parsed
  576. def should_update_brief(self, signals):
  577. self.gate_calls.append(signals)
  578. return self._gate
  579. class FakeBriefFile:
  580. """Recording stand-in for BriefFile (write/clear/set_directory never raise)."""
  581. def __init__(self):
  582. self.writes: list[str] = []
  583. self.clears = 0
  584. self.directories: list[str] = []
  585. def set_directory(self, directory):
  586. self.directories.append(directory)
  587. def write(self, text):
  588. self.writes.append(text)
  589. def clear(self):
  590. self.clears += 1
  591. class FakeLane:
  592. def __init__(self, ready=True):
  593. self._ready = ready
  594. self.submits: list[tuple] = []
  595. self.drops = 0
  596. self.predict_gate = threading.Event()
  597. def ready(self):
  598. return self._ready
  599. def state(self):
  600. return ("ok", "") if self._ready else ("loading", "")
  601. def submit(self, state, questions, lang, generation):
  602. self.submits.append((state, questions, lang, generation))
  603. return True
  604. def drop_pending(self):
  605. self.drops += 1
  606. class FakeDebugSink:
  607. def __init__(self):
  608. self.events: list[tuple] = []
  609. def emit(self, type_, **fields):
  610. self.events.append((type_, fields))
  611. def _ok_reply(text):
  612. return "ok"
  613. def _brief_reply(narrative="", facts=None):
  614. """Build a ledger-format reply the analyzer's parser can consume."""
  615. body = "\n".join("- " + fact for fact in (facts or [])) or "none"
  616. return f"NEW FACTS:\n{body}\n\nBRIEF:\n{narrative}"
  617. def _triage_service(
  618. triage=None,
  619. lane=None,
  620. brief_fn=None,
  621. chat_fn: Any = _AUTO,
  622. brief_file=None,
  623. brief_reset=None,
  624. chat_reset=None,
  625. debug_sink=None,
  626. brief_finish: Any = "stop",
  627. ):
  628. cfg = Config()
  629. cfg.assistant.turn_gap_s = 0.5
  630. clock = Clock()
  631. emitted: list[dict] = []
  632. def emit(type_, **fields):
  633. emitted.append({"type": type_, **fields})
  634. if triage is None:
  635. triage = FakeTriage()
  636. if lane is None:
  637. lane = FakeLane()
  638. if chat_fn is _AUTO:
  639. chat_fn = _ok_reply
  640. factory = (lambda: debug_sink) if debug_sink is not None else None
  641. service = AssistantService(
  642. cfg,
  643. FakeSubscriber(),
  644. emit=emit,
  645. time_fn=clock,
  646. triage=triage,
  647. laya_lane=lane,
  648. brief_fn=_wrap_brief_fn(brief_fn, brief_finish),
  649. brief_reset=brief_reset,
  650. brief_file=brief_file,
  651. chat_fn=chat_fn,
  652. chat_reset=chat_reset,
  653. debug_sink_factory=factory,
  654. )
  655. return service, emitted, clock, triage, lane
  656. def _wrap_brief_fn(brief_fn, finish_reason="stop"):
  657. """Adapt a simple `prompt -> text` stub to the engine's call signature.
  658. A plain string return is treated as the volatile narrative half of a
  659. ledger reply, so existing stubs keep reading naturally. ``finish_reason``
  660. may be a callable taking the reply so a test can mark one reply truncated.
  661. """
  662. if brief_fn is None:
  663. return None
  664. def wrapped(prompt, on_io=None, finalize=None):
  665. kwargs = {}
  666. if on_io is not None:
  667. kwargs["on_io"] = on_io
  668. try:
  669. result = brief_fn(prompt, **kwargs)
  670. except TypeError:
  671. result = brief_fn(prompt)
  672. if finalize is not None:
  673. reason = finish_reason(result) if callable(finish_reason) else finish_reason
  674. result = finalize(result, reason)
  675. is_ledger = "BRIEF:" in result or "NEW FACTS:" in result
  676. if isinstance(result, str) and result and not is_ledger:
  677. result = _brief_reply(narrative=result)
  678. return result
  679. return wrapped
  680. def _triage_events(emitted):
  681. return [e for e in emitted if e["type"] == "triage"]
  682. def _escalated_payload():
  683. return {"order": [], "items": {}, "escalated": True, "input_tokens": 3}
  684. def test_triage_status_disabled_when_no_analyzer():
  685. service, _, _ = _service()
  686. status = service.status()
  687. assert status["triage_state"] == "disabled"
  688. assert status["triage_reason"] == ""
  689. assert status["triage"] is None
  690. def test_triage_status_reflects_analyzer():
  691. service, _, _, _, _ = _triage_service(
  692. triage=FakeTriage(state_pair=("loading", "warming up"))
  693. )
  694. status = service.status()
  695. assert status["triage_state"] == "loading"
  696. assert status["triage_reason"] == "warming up"
  697. assert status["triage"] is None
  698. def test_status_contains_questions_from_analyzer():
  699. questions = {
  700. "urgency": {"type": "noul", "instructions": "how urgent"},
  701. "sentiment": {"type": "score", "instructions": "tone", "criteria": ["low", "high"]},
  702. }
  703. service, _, _, triage, _ = _triage_service(triage=FakeTriage(questions=questions))
  704. status = service.status()
  705. assert status["questions"] == questions
  706. assert status["questions"] is triage.questions
  707. def test_status_questions_none_without_analyzer():
  708. service, _, _ = _service()
  709. assert service.status()["questions"] is None
  710. def test_status_event_carries_questions():
  711. service, emitted, _, triage, _ = _triage_service()
  712. service.tick()
  713. statuses = [e for e in emitted if e["type"] == "status"]
  714. assert len(statuses) == 1
  715. assert statuses[0]["questions"] is triage.questions
  716. def test_triage_captures_transcript_lang():
  717. service, _, clock, _, _ = _triage_service()
  718. assert service._last_transcript_lang is None
  719. event = _transcript(clock, "hola")
  720. event["lang"] = "es"
  721. service.handle_event(event)
  722. assert service._last_transcript_lang == "es"
  723. service.handle_event(_transcript(clock, "hello"))
  724. assert service._last_transcript_lang is None
  725. def test_maybe_triage_passes_lang_to_lane():
  726. service, _, clock, _, lane = _triage_service()
  727. event = _transcript(clock, "hola")
  728. event["lang"] = "es"
  729. service.handle_event(event)
  730. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  731. service.tick()
  732. assert len(lane.submits) == 1
  733. _, _, lang, _ = lane.submits[0]
  734. assert lang == "es"
  735. def test_maybe_triage_runs_in_idle_mode():
  736. service, _, clock, triage, lane = _triage_service()
  737. service.handle_event(_transcript(clock, "hello"))
  738. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  739. service.tick()
  740. assert len(lane.submits) == 1
  741. state, questions, lang, gen = lane.submits[0]
  742. assert state == "- hello"
  743. assert questions is triage.questions
  744. assert lang is None
  745. assert gen == 0
  746. assert triage.build_state_calls == [("hello",)]
  747. assert service.triage_history == ["hello"]
  748. def test_maybe_triage_uses_transcript_buffer_not_chat_history():
  749. service, _, clock, triage, lane = _triage_service(chat_fn=None)
  750. service.chat_history = ["stale chat only"]
  751. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  752. service.tick()
  753. assert lane.submits == []
  754. assert triage.build_state_calls == []
  755. service.handle_event(_transcript(clock, "hello"))
  756. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  757. service.tick()
  758. assert len(lane.submits) == 1
  759. assert triage.build_state_calls == [("hello",)]
  760. def test_maybe_triage_requires_settled_turn():
  761. service, _, clock, _, lane = _triage_service()
  762. service.handle_event(_transcript(clock, "hello"))
  763. service.tick()
  764. assert lane.submits == []
  765. def test_maybe_triage_periodic_fires_before_turn_settles():
  766. service, _, clock, _, lane = _triage_service()
  767. service.cfg.triage.periodic_s = 0.2
  768. service.handle_event(_transcript(clock, "hello"))
  769. assert service.turn_ended() is False
  770. clock.advance(0.25) # due (0.2) yet still quieter than turn_gap_s (0.5)
  771. service.tick()
  772. assert service.turn_ended() is False
  773. assert len(lane.submits) == 1
  774. assert lane.submits[0][0] == "- hello"
  775. def test_maybe_triage_periodic_disabled_ignores_accumulating_speech():
  776. service, _, clock, _, lane = _triage_service()
  777. service.cfg.triage.periodic_s = 0.0
  778. service.handle_event(_transcript(clock, "hello"))
  779. clock.advance(0.25)
  780. service.tick()
  781. assert service.turn_ended() is False
  782. assert lane.submits == []
  783. def test_maybe_triage_periodic_rearms_after_submit():
  784. service, _, clock, _, lane = _triage_service()
  785. service.cfg.triage.periodic_s = 0.2
  786. service.handle_event(_transcript(clock, "first"))
  787. clock.advance(0.25)
  788. service.tick()
  789. assert len(lane.submits) == 1
  790. service.tick()
  791. service.tick()
  792. assert len(lane.submits) == 1 # identical window, no new speech
  793. service.handle_event(_transcript(clock, "second"))
  794. clock.advance(0.25)
  795. service.tick()
  796. assert len(lane.submits) == 2
  797. assert lane.submits[1][0] == "- first\n- second"
  798. def test_triage_debug_emits_submit_and_result():
  799. sink = FakeDebugSink()
  800. service, _, clock, _, lane = _triage_service(debug_sink=sink)
  801. service.handle_event(_transcript(clock, "hello"))
  802. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  803. service.tick()
  804. submits = [event for event in sink.events if event[0] == "triage-submit"]
  805. assert submits == [
  806. ("triage-submit", {"trigger": "settled", "window": 1, "generation": 0})
  807. ]
  808. _, _, _, gen = lane.submits[0]
  809. service._on_triage_result(gen, _escalated_payload())
  810. results = [event for event in sink.events if event[0] == "triage-result"]
  811. assert len(results) == 1
  812. assert results[0][1]["escalated"] is True
  813. assert results[0][1]["generation"] == 0
  814. def test_triage_debug_marks_periodic_trigger():
  815. sink = FakeDebugSink()
  816. service, _, clock, _, lane = _triage_service(debug_sink=sink)
  817. service.cfg.triage.periodic_s = 0.2
  818. service.handle_event(_transcript(clock, "hello"))
  819. clock.advance(0.25)
  820. service.tick()
  821. submits = [event for event in sink.events if event[0] == "triage-submit"]
  822. assert submits[0][1]["trigger"] == "periodic"
  823. def test_triage_debug_emits_brief_dispatch():
  824. sink = FakeDebugSink()
  825. service, _, clock, _, lane = _triage_service(
  826. brief_fn=lambda prompt: "RAW", debug_sink=sink
  827. )
  828. service.handle_event(_transcript(clock, "hello"))
  829. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  830. service.tick()
  831. _, _, _, gen = lane.submits[0]
  832. service._on_triage_result(gen, _escalated_payload())
  833. brief = [event for event in sink.events if event[0] == "triage-brief"]
  834. assert brief == [("triage-brief", {"action": "dispatch", "generation": 0})]
  835. def test_triage_debug_emits_brief_gate_skip():
  836. sink = FakeDebugSink()
  837. service, _, clock, _, lane = _triage_service(
  838. triage=FakeTriage(gate=False), brief_fn=lambda prompt: "RAW", debug_sink=sink
  839. )
  840. service.handle_event(_transcript(clock, "hello"))
  841. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  842. service.tick()
  843. _, _, _, gen = lane.submits[0]
  844. service._on_triage_result(gen, _escalated_payload())
  845. brief = [event for event in sink.events if event[0] == "triage-brief"]
  846. assert brief == [
  847. ("triage-brief", {"action": "skip", "reason": "gate", "generation": 0})
  848. ]
  849. def test_maybe_triage_skips_unready_lane():
  850. service, _, clock, _, lane = _triage_service(lane=FakeLane(ready=False))
  851. service.handle_event(_transcript(clock, "hello"))
  852. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  853. service.tick()
  854. assert lane.submits == []
  855. def test_maybe_triage_skips_without_questions():
  856. service, _, clock, triage, lane = _triage_service(
  857. triage=FakeTriage(questions=None)
  858. )
  859. service.handle_event(_transcript(clock, "hello"))
  860. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  861. service.tick()
  862. assert lane.submits == []
  863. assert triage.build_state_calls == []
  864. def test_maybe_triage_dedupes_identical_window():
  865. service, _, clock, triage, lane = _triage_service()
  866. service.handle_event(_transcript(clock, "hello"))
  867. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  868. service.tick()
  869. assert len(lane.submits) == 1
  870. state, questions, lang, gen = lane.submits[0]
  871. assert state == "- hello"
  872. assert questions is triage.questions
  873. assert lang is None
  874. assert gen == 0
  875. service.tick()
  876. service.tick()
  877. assert len(lane.submits) == 1
  878. service.handle_event(_transcript(clock, "second"))
  879. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  880. service.tick()
  881. assert len(lane.submits) == 2
  882. assert lane.submits[1][0] == "- hello\n- second"
  883. def test_second_transcript_produces_second_submit():
  884. service, _, clock, _, lane = _triage_service()
  885. service.handle_event(_transcript(clock, "first"))
  886. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  887. service.tick()
  888. assert len(lane.submits) == 1
  889. service.handle_event(_transcript(clock, "second"))
  890. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  891. service.tick()
  892. assert len(lane.submits) == 2
  893. assert lane.submits[1][0] == "- first\n- second"
  894. def test_tick_under_50ms_with_blocked_lane():
  895. service, _, clock, _, lane = _triage_service()
  896. assert not lane.predict_gate.is_set()
  897. service.toggle_chat()
  898. service.handle_event(_transcript(clock, "hello"))
  899. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  900. start = time.perf_counter()
  901. service.tick()
  902. elapsed = time.perf_counter() - start
  903. assert len(lane.submits) == 1
  904. assert elapsed < 0.05
  905. def test_on_triage_result_emits_signals():
  906. service, emitted, clock, _, lane = _triage_service()
  907. service.toggle_chat()
  908. service.handle_event(_transcript(clock, "hello"))
  909. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  910. service.tick()
  911. _, _, _, gen = lane.submits[0]
  912. payload = {
  913. "order": ["urgency"],
  914. "items": {"urgency": {"kind": "noul"}},
  915. "escalated": False,
  916. "input_tokens": 12,
  917. }
  918. service._on_triage_result(gen, payload)
  919. staged = _triage_events(emitted)
  920. assert len(staged) == 1
  921. assert staged[0]["stage"] == "signals"
  922. assert staged[0]["triage"] == service.last_triage
  923. assert staged[0]["triage"]["escalated"] is False
  924. assert staged[0]["triage"]["input_tokens"] == 12
  925. assert service.status()["triage"] == service.last_triage
  926. def test_triage_events_carry_the_server_timestamp():
  927. service, emitted, clock, _, lane = _triage_service(brief_fn=lambda prompt: "RAW")
  928. service.toggle_chat()
  929. service.handle_event(_transcript(clock, "hello"))
  930. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  931. service.tick()
  932. _, _, _, gen = lane.submits[0]
  933. expected_ts = clock()
  934. service._on_triage_result(gen, _escalated_payload())
  935. signals = [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  936. assert len(signals) == 1
  937. assert isinstance(signals[0]["ts"], (int, float))
  938. assert signals[0]["ts"] == expected_ts
  939. job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  940. job.run()
  941. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  942. assert len(contents) == 1
  943. assert isinstance(contents[0]["ts"], (int, float))
  944. assert contents[0]["ts"] == expected_ts # content reuses the signals timestamp
  945. def test_on_triage_result_suppresses_stale_generation():
  946. service, emitted, clock, _, lane = _triage_service()
  947. service.toggle_chat()
  948. service.handle_event(_transcript(clock, "hello"))
  949. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  950. service.tick()
  951. _, _, _, gen = lane.submits[0]
  952. service.clear()
  953. service._on_triage_result(gen, _escalated_payload())
  954. assert service.last_triage is None
  955. assert [e for e in _triage_events(emitted) if e.get("stage") == "signals"] == []
  956. service.handle_event(_transcript(clock, "hello"))
  957. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  958. service.tick()
  959. assert len(lane.submits) == 2
  960. def test_pending_brief_capped_by_char_budget():
  961. """The buffer keeps the newest deltas within budget, dropping the oldest."""
  962. service, _, clock, _, _ = _triage_service(brief_fn=lambda prompt: "RAW")
  963. service.cfg.triage.brief_pending_char_budget = 30
  964. service.handle_event(_transcript(clock, "aaa"))
  965. service.handle_event(_transcript(clock, "bbbb"))
  966. assert service._pending_brief == ["Unknown: aaa", "Unknown: bbbb"]
  967. service.handle_event(_transcript(clock, "ccccc"))
  968. assert service._pending_brief == ["Unknown: bbbb", "Unknown: ccccc"]
  969. assert service.triage_history == ["aaa", "bbbb", "ccccc"]
  970. def test_pending_brief_untouched_without_triage():
  971. service, _, clock = _service()
  972. service.handle_event(_transcript(clock, "hello"))
  973. assert service._pending_brief == []
  974. assert service.triage_history == ["hello"]
  975. def test_on_triage_result_accepts_non_chat_mode():
  976. service, emitted, clock, _, lane = _triage_service()
  977. service.handle_event(_transcript(clock, "hello"))
  978. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  979. service.tick()
  980. _, _, _, gen = lane.submits[0]
  981. window = service._last_triage_window
  982. service._on_triage_result(gen, _escalated_payload())
  983. assert service.last_triage is not None
  984. assert service._last_triage_window == window
  985. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  986. def test_on_triage_result_ignores_none_payload():
  987. service, emitted, clock, _, lane = _triage_service()
  988. service.toggle_chat()
  989. service.handle_event(_transcript(clock, "hello"))
  990. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  991. service.tick()
  992. _, _, _, gen = lane.submits[0]
  993. service._on_triage_result(gen, None)
  994. assert service.last_triage is None
  995. assert _triage_events(emitted) == []
  996. def test_none_payload_dispatches_nothing_and_keeps_deltas():
  997. """A Laya predict failure never dispatches: no job, no events, deltas stay."""
  998. calls: list[str] = []
  999. service, emitted, clock, _, lane = _triage_service(
  1000. brief_fn=lambda prompt: calls.append(prompt) or "RAW"
  1001. )
  1002. service.toggle_chat()
  1003. service.set_chat_capture(True)
  1004. service.handle_event(_transcript(clock, "hello"))
  1005. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1006. service.tick()
  1007. _, _, _, gen = lane.submits[0]
  1008. assert service._pending_brief == ["Unknown: hello"]
  1009. service._on_triage_result(gen, None)
  1010. assert service.last_triage is None
  1011. assert [e.get("stage") for e in _triage_events(emitted)] == []
  1012. assert [j.name for j in service._inference._queue] == ["chat"]
  1013. assert calls == []
  1014. assert service._pending_brief == ["Unknown: hello"]
  1015. assert service._brief_outstanding is False
  1016. def test_none_payload_without_brief_fn_enqueues_nothing():
  1017. service, emitted, clock, _, lane = _triage_service()
  1018. service.toggle_chat()
  1019. service.set_chat_capture(True)
  1020. service.handle_event(_transcript(clock, "hello"))
  1021. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1022. service.tick()
  1023. _, _, _, gen = lane.submits[0]
  1024. service._on_triage_result(gen, None)
  1025. assert service.last_triage is None
  1026. assert _triage_events(emitted) == []
  1027. assert [j.name for j in service._inference._queue] == ["chat"]
  1028. assert service._pending_brief == ["Unknown: hello"]
  1029. def test_on_triage_result_ignores_empty_signals():
  1030. service, emitted, clock, _, lane = _triage_service(
  1031. triage=FakeTriage(normalized={})
  1032. )
  1033. service.toggle_chat()
  1034. service.handle_event(_transcript(clock, "hello"))
  1035. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1036. service.tick()
  1037. _, _, _, gen = lane.submits[0]
  1038. service._on_triage_result(gen, _escalated_payload())
  1039. assert service.last_triage is None
  1040. assert _triage_events(emitted) == []
  1041. def test_substantive_turn_dispatches_one_brief_job():
  1042. """A substantive settled turn calls the gate once and queues one brief job."""
  1043. service, emitted, clock, triage, lane = _triage_service(
  1044. triage=FakeTriage(escalate=True), brief_fn=lambda prompt: "RAW"
  1045. )
  1046. service.toggle_chat()
  1047. service.set_chat_capture(True)
  1048. service.handle_event(_transcript(clock, "we might churn"))
  1049. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1050. service.tick()
  1051. _, _, _, gen = lane.submits[0]
  1052. service._on_triage_result(
  1053. gen, {"order": [], "items": {}, "escalated": False, "input_tokens": 0}
  1054. )
  1055. assert service.last_triage is not None
  1056. assert service.last_triage["escalated"] is True
  1057. assert service._last_brief_signals == service.last_triage
  1058. assert service._inference.depth() == 2
  1059. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  1060. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  1061. assert len(triage.gate_calls) == 1
  1062. assert service._brief_outstanding is True
  1063. assert service._brief_token == 1
  1064. def test_brief_dispatches_without_escalation():
  1065. service, _, clock, _, lane = _triage_service(brief_fn=lambda prompt: "RAW")
  1066. service.toggle_chat()
  1067. service.set_chat_capture(True)
  1068. service.handle_event(_transcript(clock, "all calm"))
  1069. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1070. service.tick()
  1071. _, _, _, gen = lane.submits[0]
  1072. service._on_triage_result(
  1073. gen, {"order": [], "items": {}, "escalated": False, "input_tokens": 0}
  1074. )
  1075. assert service.last_triage is not None
  1076. assert service._inference.depth() == 2
  1077. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  1078. def test_gated_out_turn_skips_brief():
  1079. """A gated-out turn stores signals but never calls brief_fn."""
  1080. calls: list[str] = []
  1081. service, emitted, clock, triage, lane = _triage_service(
  1082. triage=FakeTriage(gate=False),
  1083. brief_fn=lambda prompt: calls.append(prompt) or "RAW",
  1084. )
  1085. service.toggle_chat()
  1086. service.set_chat_capture(True)
  1087. service.handle_event(_transcript(clock, "small talk"))
  1088. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1089. service.tick()
  1090. _, _, _, gen = lane.submits[0]
  1091. service._on_triage_result(gen, _escalated_payload())
  1092. assert service.last_triage is not None
  1093. assert "brief" not in service.last_triage # no prior brief to retain
  1094. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  1095. assert calls == []
  1096. assert [j.name for j in service._inference._queue] == ["chat"]
  1097. assert service._brief_outstanding is False
  1098. assert len(triage.gate_calls) == 1
  1099. def test_gated_out_turn_retains_previous_brief():
  1100. """A gated-out settled turn keeps the last accepted brief in the panel."""
  1101. service, emitted, clock, triage, lane = _triage_service(
  1102. triage=FakeTriage(gate=False), brief_fn=lambda prompt: "RAW"
  1103. )
  1104. service.toggle_chat()
  1105. signals = _escalated_payload()
  1106. service._last_brief_signals = signals
  1107. token = _seed_brief(service, "Unknown: first")
  1108. service._run_brief_update(0, signals, service._now(), token)
  1109. assert service.last_triage is not None
  1110. assert service.last_triage["brief"] == "RAW"
  1111. emitted.clear()
  1112. service.handle_event(_transcript(clock, "small talk"))
  1113. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1114. service.tick()
  1115. _, _, _, gen = lane.submits[0]
  1116. service._on_triage_result(gen, _escalated_payload())
  1117. assert service.last_triage is not None
  1118. assert service.last_triage["brief"] == "RAW"
  1119. assert "brief" not in service._last_brief_signals # raw signals stay raw
  1120. staged = [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  1121. assert len(staged) == 1
  1122. assert staged[0]["triage"] == service.last_triage
  1123. assert staged[0]["triage"]["brief"] == "RAW"
  1124. assert len(triage.gate_calls) == 1
  1125. def test_gated_out_turn_retains_brief_when_extract_disabled():
  1126. """`extract_enabled = false` skips the dispatch but still retains the brief."""
  1127. service, emitted, clock, _, lane = _triage_service(
  1128. triage=FakeTriage(gate=False), brief_fn=lambda prompt: "RAW"
  1129. )
  1130. service.toggle_chat()
  1131. service.set_chat_capture(True)
  1132. signals = _escalated_payload()
  1133. service._last_brief_signals = signals
  1134. token = _seed_brief(service, "Unknown: first")
  1135. service._run_brief_update(0, signals, service._now(), token)
  1136. assert service.last_triage is not None
  1137. assert service.last_triage["brief"] == "RAW"
  1138. service.cfg.triage.extract_enabled = False
  1139. emitted.clear()
  1140. service.handle_event(_transcript(clock, "small talk"))
  1141. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1142. service.tick()
  1143. _, _, _, gen = lane.submits[0]
  1144. service._on_triage_result(gen, _escalated_payload())
  1145. assert service.last_triage is not None
  1146. assert service.last_triage["brief"] == "RAW"
  1147. assert [j.name for j in service._inference._queue] == ["chat"]
  1148. staged = [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  1149. assert staged[0]["triage"]["brief"] == "RAW"
  1150. def test_gated_out_turn_retains_brief_without_brief_fn():
  1151. """No brief_fn means no dispatch, but a retained brief is still carried."""
  1152. service, emitted, clock, _, lane = _triage_service(triage=FakeTriage(gate=False))
  1153. service.toggle_chat()
  1154. service.last_triage = {**_escalated_payload(), "brief": "KEEP"}
  1155. emitted.clear()
  1156. service.handle_event(_transcript(clock, "small talk"))
  1157. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1158. service.tick()
  1159. _, _, _, gen = lane.submits[0]
  1160. service._on_triage_result(gen, _escalated_payload())
  1161. assert service.last_triage["brief"] == "KEEP"
  1162. staged = [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  1163. assert staged[0]["triage"]["brief"] == "KEEP"
  1164. def test_later_accepted_update_replaces_retained_brief():
  1165. """A later accepted update overwrites the retained brief text."""
  1166. texts = iter(["OLD", "NEW"])
  1167. service, emitted, _, _, _ = _triage_service(brief_fn=lambda prompt: next(texts))
  1168. service.toggle_chat()
  1169. signals = _escalated_payload()
  1170. service._last_brief_signals = signals
  1171. token = _seed_brief(service, "Unknown: first")
  1172. service._run_brief_update(0, signals, service._now(), token)
  1173. assert service.last_triage is not None
  1174. assert service.last_triage["brief"] == "OLD"
  1175. token = _seed_brief(service, "Unknown: second")
  1176. service._run_brief_update(0, signals, service._now(), token)
  1177. assert service.last_triage is not None
  1178. assert service.last_triage["brief"] == "NEW"
  1179. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1180. assert contents[-1]["triage"]["brief"] == "NEW"
  1181. service.clear()
  1182. assert service.last_triage is None
  1183. def test_no_brief_when_disabled_in_config():
  1184. calls: list[str] = []
  1185. service, _, clock, _, lane = _triage_service(
  1186. triage=FakeTriage(escalate=True),
  1187. brief_fn=lambda prompt: calls.append(prompt) or "RAW",
  1188. )
  1189. service.cfg.triage.extract_enabled = False
  1190. service.toggle_chat()
  1191. service.set_chat_capture(True)
  1192. service.handle_event(_transcript(clock, "churn risk"))
  1193. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1194. service.tick()
  1195. _, _, _, gen = lane.submits[0]
  1196. service._on_triage_result(gen, _escalated_payload())
  1197. assert service.last_triage is not None
  1198. assert service.last_triage["escalated"] is True
  1199. assert calls == []
  1200. assert [j.name for j in service._inference._queue] == ["chat"]
  1201. def test_no_brief_without_brief_fn():
  1202. service, _, clock, _, lane = _triage_service(triage=FakeTriage(escalate=True))
  1203. service.toggle_chat()
  1204. service.set_chat_capture(True)
  1205. service.handle_event(_transcript(clock, "churn risk"))
  1206. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1207. service.tick()
  1208. _, _, _, gen = lane.submits[0]
  1209. service._on_triage_result(gen, _escalated_payload())
  1210. assert service.last_triage is not None
  1211. assert [j.name for j in service._inference._queue] == ["chat"]
  1212. assert service._brief_outstanding is False
  1213. def test_brief_refused_when_queue_full():
  1214. """A refused submit leaves deltas buffered and the outstanding flag clear."""
  1215. service, emitted, clock, _, lane = _triage_service(
  1216. triage=FakeTriage(escalate=True), brief_fn=lambda prompt: "RAW"
  1217. )
  1218. service._inference = InferenceQueue(size=2)
  1219. service.toggle_chat()
  1220. service.set_chat_capture(True)
  1221. service.handle_event(_transcript(clock, "churn risk"))
  1222. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1223. service.tick()
  1224. assert service._inference.depth() == 1
  1225. _, _, _, gen = lane.submits[0]
  1226. service._on_triage_result(gen, _escalated_payload())
  1227. assert service._inference.depth() == 1
  1228. assert service._inference._queue[0].name == "chat"
  1229. assert service.last_triage is not None
  1230. assert service.last_triage["escalated"] is True
  1231. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  1232. assert service._pending_brief == ["Unknown: churn risk"]
  1233. assert service._brief_outstanding is False
  1234. def test_second_result_while_outstanding_is_refused():
  1235. """At most one brief job is outstanding: the second turn only stores signals."""
  1236. service, emitted, clock, _, lane = _triage_service(brief_fn=lambda prompt: "RAW")
  1237. service.toggle_chat()
  1238. service.set_chat_capture(True)
  1239. service.handle_event(_transcript(clock, "first"))
  1240. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1241. service.tick()
  1242. _, _, _, gen0 = lane.submits[0]
  1243. service._on_triage_result(gen0, _escalated_payload())
  1244. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  1245. service.handle_event(_transcript(clock, "second"))
  1246. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1247. service.tick()
  1248. _, _, _, gen1 = lane.submits[1]
  1249. service._on_triage_result(gen1, _escalated_payload())
  1250. jobs = [j for j in service._inference._queue if j.name == "brief-update"]
  1251. assert len(jobs) == 1
  1252. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals", "signals"]
  1253. assert service._brief_outstanding is True
  1254. assert service._pending_brief == ["Unknown: first", "Unknown: second"]
  1255. def _seed_brief(service, *deltas):
  1256. service._pending_brief = list(deltas)
  1257. service._brief_outstanding = True
  1258. return service._brief_token
  1259. def test_ledger_records_fact_and_survives_model_reemit():
  1260. """A recorded fact persists even when a later reply omits it."""
  1261. first = _brief_reply("Summary: rolling", ["The budget was $305 million."])
  1262. second = _brief_reply("Summary: newer", [])
  1263. replies = iter([first, second])
  1264. service, _, _, _, _ = _triage_service(brief_fn=lambda prompt: next(replies))
  1265. service._pending_brief = ["Unknown: one"]
  1266. service._brief_outstanding = True
  1267. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1268. service._pending_brief = ["Unknown: two"]
  1269. service._brief_outstanding = True
  1270. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1271. assert service.last_triage is not None
  1272. brief = service.last_triage["brief"]
  1273. assert "The budget was $305 million." in brief
  1274. assert service._fact_ledger.facts() == ["The budget was $305 million."]
  1275. def test_ledger_dedupes_repeated_fact():
  1276. """The same fact proposed twice is recorded once."""
  1277. reply = _brief_reply("Summary: s", ["Sense could find prey.", "sense could find prey"])
  1278. service, _, _, _, _ = _triage_service(brief_fn=lambda prompt: reply)
  1279. service._pending_brief = ["Unknown: one"]
  1280. service._brief_outstanding = True
  1281. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1282. assert service._fact_ledger.facts() == ["Sense could find prey."]
  1283. def test_truncated_brief_reply_keeps_prior_narrative_and_facts():
  1284. """A reply cut at the token limit must not replace the accepted narrative."""
  1285. sink = FakeDebugSink()
  1286. replies = iter(
  1287. [
  1288. _brief_reply("Summary: good\n\nOpen Questions:\n- who owns it?"),
  1289. _brief_reply("Summary: cut mid-wor", ["fact one durable", "fact two"]),
  1290. ]
  1291. )
  1292. service, _, _, _, _ = _triage_service(
  1293. brief_fn=lambda prompt: next(replies),
  1294. brief_finish=lambda text: "length" if "cut mid-wor" in text else "stop",
  1295. debug_sink=sink,
  1296. )
  1297. service._pending_brief = ["Unknown: one"]
  1298. service._brief_outstanding = True
  1299. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1300. service._pending_brief = ["Unknown: two"]
  1301. service._brief_outstanding = True
  1302. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1303. assert service._last_brief_narrative.startswith("Summary: good")
  1304. assert "cut mid-wor" not in service.last_triage["brief"]
  1305. assert service._fact_ledger.facts() == ["fact one durable"]
  1306. rejects = [
  1307. event
  1308. for event in sink.events
  1309. if event[0] == "triage-brief" and event[1].get("action") == "reject"
  1310. ]
  1311. assert rejects
  1312. assert rejects[-1][1]["reason"] == "truncated"
  1313. def test_brief_rewrite_that_drops_a_section_is_rejected():
  1314. """Dropping a canonical section keeps the prior, more complete brief."""
  1315. replies = iter(
  1316. [
  1317. _brief_reply("Summary: full\n\nOpen Questions:\n- who owns it?"),
  1318. _brief_reply("Summary: only"),
  1319. ]
  1320. )
  1321. service, _, _, _, _ = _triage_service(brief_fn=lambda prompt: next(replies))
  1322. service._pending_brief = ["Unknown: one"]
  1323. service._brief_outstanding = True
  1324. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1325. service._pending_brief = ["Unknown: two"]
  1326. service._brief_outstanding = True
  1327. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1328. assert "Open Questions:" in service.last_triage["brief"]
  1329. assert "Summary: only" not in service.last_triage["brief"]
  1330. def test_ledger_prompt_carries_ledger_and_narrative():
  1331. """The brief request hands the model the ledger, narrative, and new deltas."""
  1332. seen: list[str] = []
  1333. service, _, _, _, _ = _triage_service(
  1334. brief_fn=lambda prompt: seen.append(prompt) or _brief_reply("UPDATED")
  1335. )
  1336. service._fact_ledger.merge(["PRIOR FACT"])
  1337. service._last_brief_narrative = "PRIOR NARRATIVE"
  1338. service._pending_brief = ["Unknown: hello there"]
  1339. service._brief_outstanding = True
  1340. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1341. assert len(seen) == 1
  1342. assert "CURRENT LEDGER:" in seen[0]
  1343. assert "- PRIOR FACT" in seen[0]
  1344. assert "CURRENT BRIEF:\nPRIOR NARRATIVE" in seen[0]
  1345. assert "NEW UTTERANCES:\n- Unknown: hello there" in seen[0]
  1346. def test_brief_merges_content_and_writes_file():
  1347. brief_file = FakeBriefFile()
  1348. service, emitted, _, triage, _ = _triage_service(
  1349. brief_fn=lambda prompt: "the brief", brief_file=brief_file
  1350. )
  1351. service.toggle_chat()
  1352. signals = _escalated_payload()
  1353. service._last_brief_signals = signals
  1354. token = _seed_brief(service, "Unknown: hello")
  1355. ts = service._now()
  1356. service._run_brief_update(0, signals, ts, token)
  1357. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1358. assert len(contents) == 1
  1359. assert contents[0]["triage"] == {**signals, "brief": "the brief"}
  1360. assert service.last_triage == contents[0]["triage"]
  1361. assert triage.prompt_calls == [(("Unknown: hello",), signals)]
  1362. assert brief_file.writes == ["the brief"]
  1363. assert service._pending_brief == []
  1364. assert service._brief_outstanding is False
  1365. def test_empty_brief_requeues_deltas():
  1366. """A `""` reply keeps signals, re-buffers the deltas, and clears the flag."""
  1367. service, emitted, _, _, _ = _triage_service(brief_fn=lambda prompt: "")
  1368. service.toggle_chat()
  1369. signals = _escalated_payload()
  1370. service.last_triage = signals
  1371. service._last_brief_signals = signals
  1372. token = _seed_brief(service, "Unknown: hello")
  1373. ts = service._now()
  1374. service._run_brief_update(0, signals, ts, token)
  1375. assert [e for e in _triage_events(emitted) if e.get("stage") == "content"] == []
  1376. assert service.last_triage == signals
  1377. assert "brief" not in service.last_triage
  1378. assert service._pending_brief == ["Unknown: hello"]
  1379. assert service._brief_outstanding is False
  1380. def test_brief_error_requeues_and_clears_flag():
  1381. """A raising brief_fn leaves the flag clear and the deltas buffered."""
  1382. def boom(prompt):
  1383. raise RuntimeError("llm down")
  1384. service, emitted, _, _, _ = _triage_service(brief_fn=boom)
  1385. service.toggle_chat()
  1386. token = _seed_brief(service, "Unknown: hello")
  1387. ts = service._now()
  1388. service._run_brief_update(0, _escalated_payload(), ts, token)
  1389. assert _triage_events(emitted) == []
  1390. assert service.last_triage is None
  1391. assert service._pending_brief == ["Unknown: hello"]
  1392. assert service._brief_outstanding is False
  1393. def test_brief_empty_buffer_returns_early():
  1394. """An empty buffer returns early: no call, no event, flag cleared."""
  1395. calls: list[str] = []
  1396. service, emitted, _, _, _ = _triage_service(
  1397. brief_fn=lambda prompt: calls.append(prompt) or "RAW"
  1398. )
  1399. service.toggle_chat()
  1400. service._brief_outstanding = True
  1401. token = service._brief_token
  1402. ts = service._now()
  1403. assert service._pending_brief == []
  1404. service._run_brief_update(0, _escalated_payload(), ts, token)
  1405. assert calls == []
  1406. assert _triage_events(emitted) == []
  1407. assert service._brief_outstanding is False
  1408. def test_brief_suppressed_after_reset():
  1409. service, emitted, _, _, _ = _triage_service(brief_fn=lambda prompt: "RAW")
  1410. service.toggle_chat()
  1411. service._pending_brief = ["Unknown: hello"]
  1412. service._reset_triage()
  1413. emitted.clear()
  1414. service._run_brief_update(0, _escalated_payload(), service._now(), 0)
  1415. assert _triage_events(emitted) == []
  1416. assert service.last_triage is None
  1417. assert service._pending_brief == ["Unknown: hello"]
  1418. def test_brief_works_after_leaving_chat():
  1419. seen = []
  1420. service, emitted, _, _, _ = _triage_service(
  1421. brief_fn=lambda prompt: seen.append(prompt) or "RAW"
  1422. )
  1423. service.toggle_chat()
  1424. assert service.modes.exit_chat() is True
  1425. token = _seed_brief(service, "Unknown: hello")
  1426. signals = _escalated_payload()
  1427. service._last_brief_signals = signals
  1428. service._run_brief_update(0, signals, service._now(), token)
  1429. assert seen != []
  1430. assert [e.get("stage") for e in _triage_events(emitted)] == ["content"]
  1431. assert service.last_triage is not None
  1432. assert service.last_triage["brief"] == "RAW"
  1433. def test_deltas_arriving_during_call_survive_and_flush():
  1434. """Deltas appended mid-call are not consumed; the finally flushes them."""
  1435. service, emitted, clock, _, _ = _triage_service()
  1436. service.toggle_chat()
  1437. signals = _escalated_payload()
  1438. service._last_brief_signals = signals
  1439. token = _seed_brief(service, "Unknown: first")
  1440. appended: list[int] = []
  1441. def _brief(prompt, on_io=None, finalize=None):
  1442. if not appended:
  1443. appended.append(1)
  1444. service.handle_event(_transcript(clock, "during call"))
  1445. return _brief_reply("first brief")
  1446. service._brief_fn = _brief
  1447. service._run_brief_update(0, signals, clock(), token)
  1448. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1449. assert len(contents) == 1
  1450. assert contents[0]["triage"]["brief"] == "first brief"
  1451. flushed = [j for j in service._inference._queue if j.name == "brief-update"]
  1452. assert len(flushed) == 1
  1453. assert service._brief_outstanding is True
  1454. assert service._pending_brief == ["Unknown: during call"]
  1455. flushed[0].run()
  1456. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1457. assert len(contents) == 2
  1458. assert contents[1]["triage"]["brief"] == "first brief"
  1459. assert service._pending_brief == []
  1460. assert service._brief_outstanding is False
  1461. def test_in_flight_brief_after_clear_emits_and_writes_nothing():
  1462. """Stale in-flight job: no event, no file write, no requeue, flag untouched."""
  1463. brief_file = FakeBriefFile()
  1464. service, emitted, clock, _, lane = _triage_service(
  1465. brief_fn=lambda prompt: "RAW", brief_file=brief_file
  1466. )
  1467. service.toggle_chat()
  1468. service.handle_event(_transcript(clock, "hello"))
  1469. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1470. service.tick()
  1471. _, _, _, gen = lane.submits[0]
  1472. service._on_triage_result(gen, _escalated_payload())
  1473. job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  1474. assert service._brief_outstanding is True
  1475. service.clear()
  1476. assert service._brief_outstanding is False
  1477. token_after_clear = service._brief_token
  1478. emitted.clear()
  1479. job.run()
  1480. assert [e for e in _triage_events(emitted)] == []
  1481. assert brief_file.writes == []
  1482. assert service._pending_brief == []
  1483. assert service.last_triage is None
  1484. assert service._brief_token == token_after_clear
  1485. assert service._brief_outstanding is False
  1486. def test_clear_resets_triage():
  1487. resets: list[int] = []
  1488. chats: list[int] = []
  1489. brief_file = FakeBriefFile()
  1490. service, emitted, clock, _, lane = _triage_service(
  1491. triage=FakeTriage(escalate=True),
  1492. brief_fn=lambda prompt: "RAW",
  1493. brief_file=brief_file,
  1494. brief_reset=lambda: resets.append(1),
  1495. chat_reset=lambda: chats.append(1),
  1496. )
  1497. service.toggle_chat()
  1498. service.set_chat_capture(True)
  1499. service.handle_event(_transcript(clock, "churn"))
  1500. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1501. service.tick()
  1502. _, _, _, gen = lane.submits[0]
  1503. service._on_triage_result(gen, _escalated_payload())
  1504. assert service._inference.depth() == 2
  1505. assert service._pending_brief == ["Unknown: churn"]
  1506. assert service._brief_outstanding is True
  1507. token_before = service._brief_token
  1508. assert service.clear() is True
  1509. assert service._triage_generation == 1
  1510. assert service.last_triage is None
  1511. assert service._last_triage_window == ()
  1512. assert service.triage_history == []
  1513. assert service._pending_brief == []
  1514. assert service._brief_token == token_before + 1
  1515. assert service._brief_outstanding is False
  1516. assert service._last_brief_signals is None
  1517. assert resets == [1]
  1518. assert brief_file.clears == 1
  1519. assert chats == [1]
  1520. assert lane.drops == 1
  1521. assert service._inference.depth() == 0
  1522. assert list(service._inference._queue) == []
  1523. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1524. assert len(cleared) == 1
  1525. assert cleared[0]["triage"] is None
  1526. assert {"type": "cleared"} in emitted
  1527. signals_before = len(
  1528. [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  1529. )
  1530. service._on_triage_result(gen, _escalated_payload())
  1531. assert service.last_triage is None
  1532. assert (
  1533. len([e for e in _triage_events(emitted) if e.get("stage") == "signals"])
  1534. == signals_before
  1535. )
  1536. submits_before = len(lane.submits)
  1537. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1538. service.tick()
  1539. assert len(lane.submits) == submits_before
  1540. def test_exit_chat_does_not_reset_triage():
  1541. chats: list[int] = []
  1542. bresets: list[int] = []
  1543. brief_file = FakeBriefFile()
  1544. service, emitted, clock, _, lane = _triage_service(
  1545. brief_fn=lambda prompt: "RAW",
  1546. chat_reset=lambda: chats.append(1),
  1547. brief_reset=lambda: bresets.append(1),
  1548. brief_file=brief_file,
  1549. )
  1550. service.toggle_chat()
  1551. service.handle_event(_transcript(clock, "hello"))
  1552. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1553. service.tick()
  1554. assert len(lane.submits) == 1
  1555. _, _, _, gen = lane.submits[0]
  1556. service._on_triage_result(gen, _escalated_payload())
  1557. assert service.last_triage is not None
  1558. assert service._triage_generation == 0
  1559. assert service.toggle_chat() is True
  1560. assert service.modes.mode.value == "idle"
  1561. assert service._triage_generation == 0
  1562. assert service.last_triage is not None
  1563. assert service._last_triage_window == ("hello",)
  1564. assert service.triage_history == ["hello"]
  1565. assert service._pending_brief == ["Unknown: hello"]
  1566. assert service._last_brief_signals is not None
  1567. assert service._brief_outstanding is True
  1568. assert lane.drops == 0
  1569. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1570. assert cleared == []
  1571. assert chats == []
  1572. assert bresets == []
  1573. assert brief_file.clears == 0
  1574. def test_toggle_chat_enter_does_not_touch_triage():
  1575. """Entering chat resets only the chat buffers; the brief engine is independent."""
  1576. bresets: list[int] = []
  1577. brief_file = FakeBriefFile()
  1578. service, emitted, clock, _, lane = _triage_service(
  1579. brief_fn=lambda prompt: "RAW",
  1580. brief_reset=lambda: bresets.append(1),
  1581. brief_file=brief_file,
  1582. )
  1583. service.handle_event(_transcript(clock, "hello"))
  1584. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1585. service.tick()
  1586. _, _, _, gen = lane.submits[0]
  1587. service._on_triage_result(gen, _escalated_payload())
  1588. triage_before = service.last_triage
  1589. pending_before = list(service._pending_brief)
  1590. assert service.toggle_chat() is True
  1591. assert service.modes.mode.value == "chat"
  1592. assert service._triage_generation == 0
  1593. assert service.last_triage is triage_before
  1594. assert service._last_triage_window == ("hello",)
  1595. assert service.triage_history == ["hello"]
  1596. assert service._pending_brief == pending_before
  1597. assert service._last_brief_signals is not None
  1598. assert service._brief_outstanding is True
  1599. assert lane.drops == 0
  1600. assert bresets == []
  1601. assert brief_file.clears == 0
  1602. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1603. assert cleared == []
  1604. def test_clear_without_triage_emits_no_triage_event():
  1605. service, emitted, _ = _service()
  1606. brief_file = FakeBriefFile()
  1607. bresets: list[int] = []
  1608. service._brief_file = brief_file
  1609. service._brief_reset = lambda: bresets.append(1)
  1610. service.clear()
  1611. assert {"type": "cleared"} in emitted
  1612. assert _triage_events(emitted) == []
  1613. assert brief_file.clears == 1
  1614. assert bresets == []
  1615. def test_stale_brief_after_clear_does_not_clear_newer_dispatch_flag():
  1616. """A stale in-flight job must not clear a newer dispatch's outstanding flag."""
  1617. brief_file = FakeBriefFile()
  1618. service, emitted, clock, _, lane = _triage_service(
  1619. brief_fn=lambda prompt: "RAW", brief_file=brief_file
  1620. )
  1621. service.toggle_chat()
  1622. service.set_chat_capture(True)
  1623. service.handle_event(_transcript(clock, "hello"))
  1624. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1625. service.tick()
  1626. _, _, _, gen = lane.submits[0]
  1627. service._on_triage_result(gen, _escalated_payload())
  1628. stale_job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  1629. assert service._brief_outstanding is True
  1630. service.clear()
  1631. # A fresh turn dispatches a new brief AFTER clear bumped the token.
  1632. service.handle_event(_transcript(clock, "fresh"))
  1633. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1634. service.tick()
  1635. _, _, _, gen2 = lane.submits[1]
  1636. service._on_triage_result(gen2, _escalated_payload())
  1637. assert service._brief_outstanding is True
  1638. newer_token = service._brief_token
  1639. pending_before = list(service._pending_brief)
  1640. triage_before = service.last_triage
  1641. emitted.clear()
  1642. stale_job.run()
  1643. assert _triage_events(emitted) == []
  1644. assert brief_file.writes == []
  1645. assert service._pending_brief == pending_before
  1646. assert service._brief_outstanding is True
  1647. assert service._brief_token == newer_token
  1648. assert service.last_triage is triage_before
  1649. def test_clear_resets_chat_latch():
  1650. service, emitted, _ = _service()
  1651. service.toggle_chat()
  1652. service._chat_queue_full = True
  1653. service._pending_chat = ["stuck text"]
  1654. service.clear()
  1655. assert service._chat_queue_full is False
  1656. assert service._pending_chat == []
  1657. assert {"type": "cleared"} in emitted
  1658. def test_triage_end_to_end_signals_content_cleared():
  1659. """Two settled turns collapse into ONE brief job (single-outstanding guard).
  1660. The second `_on_triage_result` is refused as `"outstanding"`, so the queue
  1661. holds one `brief-update` that consumes both deltas and emits one `content`.
  1662. The panel keeps the FIRST turn's signals until the next settled turn (lag).
  1663. """
  1664. prompts: list[str] = []
  1665. service, emitted, clock, triage, lane = _triage_service(
  1666. brief_fn=lambda prompt: prompts.append(prompt) or "the brief"
  1667. )
  1668. service.toggle_chat()
  1669. service.set_chat_capture(True)
  1670. service.handle_event(_transcript(clock, "first thought"))
  1671. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1672. service.tick()
  1673. assert len(lane.submits) == 1
  1674. _, _, _, gen0 = lane.submits[0]
  1675. service._on_triage_result(
  1676. gen0, {"order": ["urgency"], "items": {}, "escalated": False, "input_tokens": 1}
  1677. )
  1678. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  1679. assert prompts == []
  1680. assert service.last_triage is not None
  1681. assert service.last_triage["escalated"] is False
  1682. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  1683. triage._escalate = True
  1684. service.handle_event(_transcript(clock, "we might churn"))
  1685. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1686. service.tick()
  1687. assert len(lane.submits) == 2
  1688. _, _, _, gen1 = lane.submits[1]
  1689. service._on_triage_result(gen1, _escalated_payload())
  1690. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals", "signals"]
  1691. assert [j.name for j in service._inference._queue] == [
  1692. "chat",
  1693. "brief-update",
  1694. "chat",
  1695. ]
  1696. for job in [j for j in service._inference._queue if j.name == "brief-update"]:
  1697. job.run()
  1698. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1699. assert len(contents) == 1
  1700. assert len(prompts) == 1
  1701. deltas, _ = triage.prompt_calls[0]
  1702. assert deltas == ("Unknown: first thought", "Unknown: we might churn")
  1703. assert contents[0]["triage"]["escalated"] is False
  1704. assert contents[0]["triage"]["brief"] == "the brief"
  1705. assert service.last_triage == contents[0]["triage"]
  1706. assert service.status()["triage"] == contents[0]["triage"]
  1707. service.clear()
  1708. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1709. assert len(cleared) == 1
  1710. assert cleared[0]["triage"] is None
  1711. assert service.last_triage is None
  1712. service._on_triage_result(gen1, _escalated_payload())
  1713. assert [e.get("stage") for e in _triage_events(emitted)] == [
  1714. "signals",
  1715. "signals",
  1716. "content",
  1717. "cleared",
  1718. ]
  1719. assert service.toggle_chat() is True
  1720. service.handle_event(_transcript(clock, "idle thought"))
  1721. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1722. service.tick()
  1723. assert len(lane.submits) == 3
  1724. assert lane.submits[2][0].endswith("- idle thought")
  1725. def test_tick_emits_status_on_triage_state_change():
  1726. service, emitted, _, triage, _ = _triage_service(
  1727. triage=FakeTriage(state_pair=("loading", "warming up"))
  1728. )
  1729. service.tick()
  1730. statuses = [e for e in emitted if e["type"] == "status"]
  1731. assert len(statuses) == 1
  1732. assert statuses[0]["triage_state"] == "loading"
  1733. assert statuses[0]["triage_reason"] == "warming up"
  1734. assert "triage" in statuses[0]
  1735. triage._state_pair = ("ok", "")
  1736. service.tick()
  1737. statuses = [e for e in emitted if e["type"] == "status"]
  1738. assert len(statuses) == 2
  1739. assert statuses[1]["triage_state"] == "ok"
  1740. assert statuses[1]["triage_reason"] == ""
  1741. assert "triage" in statuses[1]
  1742. service.tick()
  1743. assert len([e for e in emitted if e["type"] == "status"]) == 2
  1744. def test_no_status_emitted_when_triage_disabled():
  1745. service, emitted, _ = _service()
  1746. service.tick()
  1747. service.tick()
  1748. assert [e for e in emitted if e["type"] == "status"] == []
  1749. def test_tick_heals_triage_panel_after_lane_ready():
  1750. import time as _time
  1751. from pathlib import Path
  1752. from stt_daemon.assistant.laya_lane import LayaLane
  1753. from stt_daemon.assistant.triage import TriageAnalyzer
  1754. from stt_daemon.tui.state import ViewState
  1755. class _FakeRouter:
  1756. def preload(self, names):
  1757. return self
  1758. def predict(self, state, questions, model=None, lang=None, **kwargs):
  1759. if "_warmup" in questions:
  1760. return {"model": model, "answers": {}, "usage": {}, "routing": {}}
  1761. return {
  1762. "model": model,
  1763. "answers": {
  1764. "frustration": {
  1765. "noul": 0.1,
  1766. "confidence": 0.9,
  1767. "answer_confidence": 0.9,
  1768. },
  1769. "urgency": {
  1770. "choice": "low",
  1771. "confidence": 0.9,
  1772. "answer_confidence": 0.9,
  1773. },
  1774. },
  1775. "usage": {"input_tokens": 5},
  1776. }
  1777. def _factory(**kwargs):
  1778. return _FakeRouter()
  1779. cfg = Config()
  1780. cfg.assistant.turn_gap_s = 0.5
  1781. cfg.triage.questions_file = str(
  1782. Path(__file__).resolve().parent.parent / "questions.example.json"
  1783. )
  1784. clock = Clock()
  1785. emitted: list[dict] = []
  1786. def emit(type_, **fields):
  1787. emitted.append({"type": type_, **fields})
  1788. analyzer = TriageAnalyzer(cfg.triage)
  1789. assert analyzer.questions, "example questions must load"
  1790. holder: dict = {}
  1791. lane = LayaLane(
  1792. cfg.triage,
  1793. agent_factory=_factory,
  1794. on_result=lambda gen, payload: holder["svc"]._on_triage_result(gen, payload),
  1795. on_state=lambda state, reason: analyzer.refresh(state, reason),
  1796. )
  1797. service = AssistantService(
  1798. cfg,
  1799. FakeSubscriber(),
  1800. emit=emit,
  1801. time_fn=clock,
  1802. triage=analyzer,
  1803. laya_lane=lane,
  1804. chat_fn=_ok_reply,
  1805. )
  1806. holder["svc"] = service
  1807. try:
  1808. lane.start()
  1809. deadline = _time.monotonic() + 5.0
  1810. while not lane.ready() and _time.monotonic() < deadline:
  1811. _time.sleep(0.005)
  1812. assert lane.ready(), "lane did not become ready"
  1813. assert analyzer.state() == ("ok", "")
  1814. service.tick()
  1815. statuses = [e for e in emitted if e["type"] == "status"]
  1816. assert statuses, "tick emitted no status after lane ready"
  1817. assert statuses[-1]["triage_state"] == "ok"
  1818. service.handle_event(_transcript(clock, "hello"))
  1819. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1820. service.tick()
  1821. deadline = _time.monotonic() + 5.0
  1822. while service.last_triage is None and _time.monotonic() < deadline:
  1823. _time.sleep(0.005)
  1824. assert service.last_triage is not None, "no triage signals from lane"
  1825. view = ViewState()
  1826. for event in emitted:
  1827. view.ingest(event)
  1828. assert view.triage_state == "ok"
  1829. lines = view.triage_lines()
  1830. assert lines
  1831. assert not any("unavailable" in line for line in lines)
  1832. assert not any("loading" in line for line in lines)
  1833. finally:
  1834. lane.stop()
  1835. def _dump_brief(prompt, on_io=None, finalize=None):
  1836. """Brief fake that mimics LlmClient.chat's on_io phases."""
  1837. if on_io is not None:
  1838. on_io({"phase": "request", "url": "http://llm/v1", "body": {"prompt": prompt}})
  1839. on_io(
  1840. {
  1841. "phase": "response",
  1842. "payload": {"content": "parsed"},
  1843. "error": None,
  1844. "duration_ms": 250.0,
  1845. }
  1846. )
  1847. return _brief_reply("BRIEF")
  1848. def _run_triage_flow(service, emitted, clock, lane, payload):
  1849. service.toggle_chat()
  1850. service.set_chat_capture(True)
  1851. service.handle_event(_transcript(clock, "hello"))
  1852. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1853. service.tick()
  1854. _, _, _, gen = lane.submits[0]
  1855. service._on_triage_result(gen, payload)
  1856. return gen
  1857. def _dump_files(tmp_path):
  1858. return list(tmp_path.rglob("*.json"))
  1859. def test_dump_writes_full_laya_and_llm_io(tmp_path):
  1860. service, emitted, clock, triage, lane = _triage_service(brief_fn=_dump_brief)
  1861. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1862. service.cfg.triage.dump_io = True
  1863. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1864. jobs = [j for j in service._inference._queue if j.name == "brief-update"]
  1865. assert len(jobs) == 1
  1866. jobs[0].run()
  1867. files = _dump_files(tmp_path)
  1868. assert len(files) == 1
  1869. assert re.fullmatch(r"\d{6}-\d{3}\.json", files[0].name)
  1870. data = json.loads(files[0].read_text(encoding="utf-8"))
  1871. assert data["ts"] == clock()
  1872. assert isinstance(data["iso"], str) and data["iso"]
  1873. assert data["laya"]["request"]["state"] == "- hello"
  1874. assert data["laya"]["request"]["questions"] == triage.questions
  1875. assert data["laya"]["request"]["model"] == "multilingual"
  1876. assert data["laya"]["request"]["max_len"] == 1024
  1877. assert data["laya"]["response"] == _escalated_payload()
  1878. assert data["laya"]["error"] is None
  1879. assert data["llm"]["request"]["url"] == "http://llm/v1"
  1880. assert data["llm"]["request"]["body"] == {
  1881. "prompt": "CURRENT LEDGER:\n(none yet)\n\nCURRENT BRIEF:\n(none yet)\n\n"
  1882. "NEW UTTERANCES:\n- Unknown: hello"
  1883. }
  1884. assert data["llm"]["response"] == {"content": "parsed"}
  1885. assert data["llm"]["error"] is None
  1886. assert data["llm"]["duration_ms"] == 250.0
  1887. assert service._triage_dump._pending == {}
  1888. def test_dump_laya_error_writes_laya_only_and_keeps_deltas(tmp_path):
  1889. """A Laya predict failure finalizes the Laya half with `llm: null`.
  1890. Nothing is dispatched, so no LLM half is ever recorded and the buffered
  1891. deltas survive for the next settled turn.
  1892. """
  1893. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1894. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1895. service.cfg.triage.dump_io = True
  1896. _run_triage_flow(service, emitted, clock, lane, None)
  1897. assert [j.name for j in service._inference._queue] == ["chat"]
  1898. assert service._pending_brief == ["Unknown: hello"]
  1899. files = _dump_files(tmp_path)
  1900. assert len(files) == 1
  1901. data = json.loads(files[0].read_text(encoding="utf-8"))
  1902. assert data["laya"]["response"] is None
  1903. assert data["laya"]["error"] == "predict failed"
  1904. assert data["llm"] is None
  1905. assert service._triage_dump._pending == {}
  1906. def test_dump_without_brief_has_laya_only(tmp_path):
  1907. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1908. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1909. service.cfg.triage.dump_io = True
  1910. service.cfg.triage.extract_enabled = False
  1911. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1912. files = _dump_files(tmp_path)
  1913. assert len(files) == 1
  1914. data = json.loads(files[0].read_text(encoding="utf-8"))
  1915. assert data["laya"]["response"] == _escalated_payload()
  1916. assert data["llm"] is None
  1917. assert service._triage_dump._pending == {}
  1918. def test_dump_queue_full_has_laya_only(tmp_path):
  1919. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1920. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1921. service.cfg.triage.dump_io = True
  1922. service._inference = InferenceQueue(size=2)
  1923. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1924. assert service._inference.depth() == 1
  1925. assert [j.name for j in service._inference._queue] == ["chat"]
  1926. assert service._pending_brief == ["Unknown: hello"]
  1927. files = _dump_files(tmp_path)
  1928. assert len(files) == 1
  1929. data = json.loads(files[0].read_text(encoding="utf-8"))
  1930. assert data["laya"]["response"] == _escalated_payload()
  1931. assert data["llm"] is None
  1932. assert service._triage_dump._pending == {}
  1933. def test_dump_disabled_writes_nothing(tmp_path):
  1934. service, emitted, clock, _, lane = _triage_service(
  1935. brief_fn=lambda prompt: "RAW"
  1936. )
  1937. assert service._triage_dump.enabled is False
  1938. service._triage_dump = TriageDumper(False, directory=str(tmp_path))
  1939. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1940. jobs = [j for j in service._inference._queue if j.name == "brief-update"]
  1941. assert len(jobs) == 1
  1942. jobs[0].run()
  1943. assert _dump_files(tmp_path) == []
  1944. def test_dump_stale_result_discards_without_file(tmp_path):
  1945. service, emitted, clock, triage, lane = _triage_service(brief_fn=_dump_brief)
  1946. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1947. service.cfg.triage.dump_io = True
  1948. service.toggle_chat()
  1949. service.handle_event(_transcript(clock, "hello"))
  1950. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1951. service.tick()
  1952. _, _, _, gen = lane.submits[0]
  1953. real_normalize = triage.normalize
  1954. def stale_normalize(payload):
  1955. # Simulate a clear landing between the first check/begin and the re-check.
  1956. service._triage_generation += 1
  1957. return real_normalize(payload)
  1958. triage.normalize = stale_normalize
  1959. service._on_triage_result(gen, _escalated_payload())
  1960. assert _dump_files(tmp_path) == []
  1961. assert service._triage_dump._pending == {}
  1962. def test_dump_stale_brief_discards_without_file(tmp_path):
  1963. service, _, clock, _, _ = _triage_service(brief_fn=_dump_brief)
  1964. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1965. service.cfg.triage.dump_io = True
  1966. ts = clock()
  1967. service._triage_dump.begin(ts, laya={"request": {}, "response": {}, "error": None})
  1968. service._pending_brief = ["Unknown: hello"]
  1969. service._triage_generation += 1 # stale before the brief even starts
  1970. service._run_brief_update(0, _escalated_payload(), ts, service._brief_token)
  1971. assert _dump_files(tmp_path) == []
  1972. assert service._triage_dump._pending == {}
  1973. assert service._pending_brief == ["Unknown: hello"]
  1974. def test_dump_brief_cleared_mid_flight_writes_nothing(tmp_path):
  1975. service, _, clock, _, _ = _triage_service(brief_fn=_dump_brief)
  1976. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1977. service.cfg.triage.dump_io = True
  1978. def stale_brief(prompt, on_io=None, finalize=None):
  1979. # Simulate a clear landing mid-LLM-call (after the request phase fired).
  1980. service._triage_generation += 1
  1981. if on_io is not None:
  1982. on_io({"phase": "request", "url": "http://llm/v1", "body": {}})
  1983. on_io({"phase": "response", "payload": {"c": 1}, "error": None})
  1984. return _brief_reply("RAW")
  1985. service._brief_fn = stale_brief
  1986. service._pending_brief = ["Unknown: hello"]
  1987. service._brief_outstanding = True
  1988. token = service._brief_token
  1989. ts = clock()
  1990. service._triage_dump.begin(ts, laya={"request": {}, "response": {}, "error": None})
  1991. service._run_brief_update(0, _escalated_payload(), ts, token)
  1992. assert _dump_files(tmp_path) == []
  1993. assert service._triage_dump._pending == {}
  1994. assert service._pending_brief == []
  1995. assert service._brief_outstanding is False
  1996. def test_dump_gated_out_writes_laya_only_and_no_pending(tmp_path):
  1997. """A turn the gate leaves out writes `llm: null` and leaks no pending record."""
  1998. service, emitted, clock, triage, lane = _triage_service(
  1999. triage=FakeTriage(escalate=True, gate=False), brief_fn=_dump_brief
  2000. )
  2001. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  2002. service.cfg.triage.dump_io = True
  2003. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  2004. assert [j.name for j in service._inference._queue] == ["chat"]
  2005. assert len(triage.gate_calls) == 1
  2006. files = _dump_files(tmp_path)
  2007. assert len(files) == 1
  2008. data = json.loads(files[0].read_text(encoding="utf-8"))
  2009. assert data["laya"]["response"] == _escalated_payload()
  2010. assert data["llm"] is None
  2011. assert service._triage_dump._pending == {}
  2012. def test_dump_coalesced_update_writes_laya_only_then_llm_only_flush(tmp_path):
  2013. """A coalesced turn writes a Laya-only record; the later flush a separate llm-only one."""
  2014. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  2015. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  2016. service.cfg.triage.dump_io = True
  2017. service.toggle_chat()
  2018. service.handle_event(_transcript(clock, "first"))
  2019. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  2020. service.tick()
  2021. _, _, _, gen0 = lane.submits[0]
  2022. service._on_triage_result(gen0, _escalated_payload())
  2023. service.handle_event(_transcript(clock, "second"))
  2024. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  2025. service.tick()
  2026. _, _, _, gen1 = lane.submits[1]
  2027. service._on_triage_result(gen1, _escalated_payload())
  2028. assert service._brief_outstanding is True
  2029. appended: list[int] = []
  2030. def _brief(prompt, on_io=None, finalize=None):
  2031. if not appended:
  2032. appended.append(1)
  2033. service.handle_event(_transcript(clock, "during call"))
  2034. if on_io is not None:
  2035. on_io({"phase": "request", "url": "http://llm/v1", "body": {"prompt": prompt}})
  2036. on_io({"phase": "response", "payload": {"content": "parsed"}, "error": None})
  2037. return _brief_reply("BRIEF")
  2038. service._brief_fn = _brief
  2039. clock.advance(1.0)
  2040. job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  2041. service._inference._queue.remove(job)
  2042. job.run()
  2043. flushed = [j for j in service._inference._queue if j.name == "brief-update"]
  2044. assert len(flushed) == 1
  2045. flushed[0].run()
  2046. assert service._triage_dump._pending == {}
  2047. records = [
  2048. json.loads(path.read_text(encoding="utf-8")) for path in _dump_files(tmp_path)
  2049. ]
  2050. assert len(records) == 3
  2051. coalesced = [r for r in records if r["laya"] is not None and r["llm"] is None]
  2052. flushed_records = [r for r in records if r["laya"] is None and r["llm"] is not None]
  2053. both = [r for r in records if r["laya"] is not None and r["llm"] is not None]
  2054. assert len(coalesced) == 1
  2055. assert len(flushed_records) == 1
  2056. assert len(both) == 1
  2057. assert coalesced[0]["ts"] != flushed_records[0]["ts"]
  2058. assert flushed_records[0]["llm"]["request"]["body"] == {
  2059. "prompt": "CURRENT LEDGER:\n(none yet)\n\nCURRENT BRIEF:\nBRIEF\n\n"
  2060. "NEW UTTERANCES:\n- Unknown: during call"
  2061. }
  2062. def test_chat_client_failure_emits_llm_error_and_service_survives():
  2063. """Adversarial (a): a chat client raising on `chat` surfaces as an `llm_error`
  2064. event and the service keeps functioning afterwards."""
  2065. from stt_daemon.assistant.llm import ConversationEngine, LlmError
  2066. class BoomClient:
  2067. def chat(self, messages, max_tokens=None, timeout_s=None, **kwargs):
  2068. raise LlmError("boom")
  2069. engine = ConversationEngine(Config().llm, BoomClient(), system_prompt="SYS")
  2070. service, emitted, clock = _service()
  2071. service.chat_fn = engine.send
  2072. assert service.toggle_chat() is True
  2073. service._run_chat("hello")
  2074. errors = [e for e in emitted if e["type"] == "error" and e.get("code") == "llm_error"]
  2075. assert errors and "boom" in errors[-1]["message"]
  2076. assert service.toggle_chat() is True
  2077. assert service.modes.mode.value == "idle"
  2078. # ---------------------------------------------------------------------------
  2079. # Conversation session adoption (SESSION/TRANSCRIPT session_dir).
  2080. # ---------------------------------------------------------------------------
  2081. def _session_event(clock, session_dir):
  2082. event = _ev(events.SESSION, clock)
  2083. event["session_id"] = session_dir.rsplit("/", 1)[-1]
  2084. event["session_dir"] = session_dir
  2085. return event
  2086. def _ring(service, clock, *items):
  2087. """Feed a daemon-seeded replay window: status -> items -> ring_end."""
  2088. service.handle_event(_ev(events.STATUS, clock))
  2089. for item in items:
  2090. service.handle_event(item)
  2091. service.handle_event(_ev(events.RING_END, clock))
  2092. def _cleared(emitted):
  2093. return [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  2094. def test_status_includes_session_dir_without_brief_file():
  2095. service, _, clock = _service()
  2096. assert service.status()["session_dir"] is None
  2097. service.handle_event(_session_event(clock, "transcripts/2026-09-30-12-47-29"))
  2098. assert service.status()["session_dir"] == "transcripts/2026-09-30-12-47-29"
  2099. def test_live_session_event_sets_session_and_records_directory():
  2100. brief_file = FakeBriefFile()
  2101. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2102. session = "transcripts/2026-09-30-12-47-29"
  2103. service.handle_event(_session_event(clock, session))
  2104. assert service.status()["session_dir"] == session
  2105. assert brief_file.directories == [session]
  2106. assert service._triage_generation == 0
  2107. def test_live_transcript_session_sets_session_without_wiping_utterance():
  2108. brief_file = FakeBriefFile()
  2109. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2110. session = "transcripts/s1"
  2111. service.handle_event(_transcript(clock, "hello", session_dir=session))
  2112. assert service.status()["session_dir"] == session
  2113. assert brief_file.directories == [session]
  2114. assert service.triage_history == ["hello"]
  2115. assert service._pending_brief == ["Unknown: hello"]
  2116. assert service._triage_generation == 0
  2117. def test_brief_write_lands_under_session_dir(tmp_path):
  2118. from stt_daemon.assistant.brief_file import BriefFile
  2119. brief_file = BriefFile(True)
  2120. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2121. session = str(tmp_path / "transcripts" / "s1")
  2122. service.handle_event(_transcript(clock, "hello", session_dir=session))
  2123. brief_file.write("the brief")
  2124. path = tmp_path / "transcripts" / "s1" / "brief.md"
  2125. assert path.read_text(encoding="utf-8") == "the brief"
  2126. def test_live_session_change_resets_triage_and_keeps_chat():
  2127. brief_file = FakeBriefFile()
  2128. resets: list[int] = []
  2129. service, emitted, clock, _, lane = _triage_service(
  2130. brief_file=brief_file, brief_reset=lambda: resets.append(1)
  2131. )
  2132. service.toggle_chat()
  2133. service.set_chat_capture(True)
  2134. service.handle_event(_transcript(clock, "hello"))
  2135. service.handle_event(_session_event(clock, "transcripts/a"))
  2136. service.last_triage = {"signals": 1}
  2137. service._pending_brief = ["stale"]
  2138. emitted.clear()
  2139. service.handle_event(_session_event(clock, "transcripts/b"))
  2140. assert service.status()["session_dir"] == "transcripts/b"
  2141. assert service.last_triage is None
  2142. assert service._pending_brief == []
  2143. assert service._triage_generation == 1
  2144. assert service.chat_history == ["Unknown: hello"]
  2145. assert resets == [1]
  2146. assert lane.drops == 1
  2147. assert brief_file.directories == ["transcripts/a", "transcripts/b"]
  2148. assert len(_cleared(emitted)) == 1
  2149. def test_sessionless_transcript_does_not_reset():
  2150. service, _, clock, _, _ = _triage_service()
  2151. service.handle_event(_session_event(clock, "transcripts/a"))
  2152. service.last_triage = {"signals": 1}
  2153. service.handle_event(_transcript(clock, "hello"))
  2154. assert service._triage_generation == 0
  2155. assert service.last_triage == {"signals": 1}
  2156. assert service.status()["session_dir"] == "transcripts/a"
  2157. def test_same_dir_readoption_does_not_reset():
  2158. brief_file = FakeBriefFile()
  2159. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2160. session = "transcripts/a"
  2161. service.handle_event(_session_event(clock, session))
  2162. service.last_triage = {"signals": 1}
  2163. service.handle_event(_transcript(clock, "again", session_dir=session))
  2164. assert service._triage_generation == 0
  2165. assert service.last_triage == {"signals": 1}
  2166. assert brief_file.directories == [session]
  2167. def test_session_change_without_brief_file_does_not_raise():
  2168. service, _, clock, _, _ = _triage_service(brief_file=None)
  2169. service.handle_event(_session_event(clock, "transcripts/a"))
  2170. service.handle_event(_session_event(clock, "transcripts/b"))
  2171. assert service.status()["session_dir"] == "transcripts/b"
  2172. assert service._triage_generation == 1
  2173. def test_replay_single_session_adopts_without_reset():
  2174. brief_file = FakeBriefFile()
  2175. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2176. session = "transcripts/a"
  2177. _ring(service, clock, _transcript(clock, "hello", session_dir=session))
  2178. assert service.status()["session_dir"] == session
  2179. assert brief_file.directories == [session]
  2180. assert service._triage_generation == 0
  2181. assert _cleared(emitted) == []
  2182. assert service.triage_history == ["hello"]
  2183. def test_replay_two_sessions_commits_newest_and_resets_once():
  2184. brief_file = FakeBriefFile()
  2185. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2186. a, b = "transcripts/a", "transcripts/b"
  2187. _ring(
  2188. service,
  2189. clock,
  2190. _session_event(clock, a),
  2191. _transcript(clock, "old turn", session_dir=a),
  2192. _session_event(clock, b),
  2193. _transcript(clock, "new turn", session_dir=b),
  2194. )
  2195. assert service.status()["session_dir"] == b
  2196. assert brief_file.directories == [b]
  2197. assert service._triage_generation == 1
  2198. assert service.last_triage is None
  2199. assert service._pending_brief == []
  2200. assert len(_cleared(emitted)) == 1
  2201. def test_reconnect_tailing_current_session_is_noop():
  2202. brief_file = FakeBriefFile()
  2203. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2204. session = "transcripts/a"
  2205. service.handle_event(_session_event(clock, session))
  2206. service.last_triage = {"signals": 1}
  2207. emitted.clear()
  2208. _ring(service, clock, _transcript(clock, "replayed", session_dir=session))
  2209. assert service.status()["session_dir"] == session
  2210. assert brief_file.directories == [session]
  2211. assert service._triage_generation == 0
  2212. assert service.last_triage == {"signals": 1}
  2213. assert _cleared(emitted) == []
  2214. def test_reconnect_older_to_current_resets_once_without_rewind():
  2215. brief_file = FakeBriefFile()
  2216. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2217. a, b = "transcripts/a", "transcripts/b"
  2218. service.handle_event(_session_event(clock, a))
  2219. service.last_triage = {"signals": 1}
  2220. emitted.clear()
  2221. _ring(
  2222. service,
  2223. clock,
  2224. _session_event(clock, a),
  2225. _transcript(clock, "old", session_dir=a),
  2226. _session_event(clock, b),
  2227. _transcript(clock, "new", session_dir=b),
  2228. )
  2229. assert service.status()["session_dir"] == b
  2230. assert brief_file.directories == [a, b]
  2231. assert service._triage_generation == 1
  2232. assert service.last_triage is None
  2233. assert len(_cleared(emitted)) == 1
  2234. def test_reconnect_only_newer_session_resets_once():
  2235. brief_file = FakeBriefFile()
  2236. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2237. a, b = "transcripts/a", "transcripts/b"
  2238. service.handle_event(_session_event(clock, a))
  2239. service.last_triage = {"signals": 1}
  2240. _ring(service, clock, _transcript(clock, "new", session_dir=b))
  2241. assert service.status()["session_dir"] == b
  2242. assert service._triage_generation == 1
  2243. assert service.last_triage is None
  2244. assert brief_file.directories == [a, b]
  2245. def test_reconnect_span_older_current_clears_buffered_older_turns():
  2246. """A RING_END whose ring spanned an older session resets even when the
  2247. newest session already matches the adopted directory (no re-set_directory)."""
  2248. brief_file = FakeBriefFile()
  2249. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2250. a, b = "transcripts/a", "transcripts/b"
  2251. _ring(service, clock, _session_event(clock, a))
  2252. assert service._triage_generation == 0
  2253. service.handle_event(_transcript(clock, "old turn", session_dir=a))
  2254. assert service._pending_brief == ["Unknown: old turn"]
  2255. service.handle_event(_session_event(clock, b))
  2256. assert service._triage_generation == 1
  2257. service.handle_event(_transcript(clock, "new turn", session_dir=b))
  2258. assert service._pending_brief == ["Unknown: new turn"]
  2259. emitted.clear()
  2260. _ring(
  2261. service,
  2262. clock,
  2263. _session_event(clock, a),
  2264. _transcript(clock, "old turn", session_dir=a),
  2265. _session_event(clock, b),
  2266. _transcript(clock, "new turn", session_dir=b),
  2267. )
  2268. assert service.status()["session_dir"] == b
  2269. assert brief_file.directories == [a, b]
  2270. assert service._pending_brief == []
  2271. assert service._triage_generation == 2
  2272. assert len(_cleared(emitted)) == 1
  2273. def test_daemon_restart_new_session_resets_once():
  2274. brief_file = FakeBriefFile()
  2275. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2276. service.handle_event(_session_event(clock, "transcripts/a"))
  2277. service.last_triage = {"signals": 1}
  2278. service.handle_event(_session_event(clock, "transcripts/b"))
  2279. assert service.status()["session_dir"] == "transcripts/b"
  2280. assert service._triage_generation == 1
  2281. assert service.last_triage is None
  2282. assert brief_file.directories == ["transcripts/a", "transcripts/b"]