test_assistant_service.py 96 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877
  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.handle_event(_transcript(clock, "That's all."))
  161. assert service.modes.mode.value == "idle"
  162. def test_exit_phrase_preserves_chat_history():
  163. service, _, clock = _service()
  164. service.toggle_chat()
  165. service.handle_event(_transcript(clock, "hello"))
  166. service.handle_event(_transcript(clock, "That's all."))
  167. assert service.modes.mode.value == "idle"
  168. assert service.chat_history == ["Unknown: hello", "Unknown: That's all."]
  169. def test_viewer_detached_exits_chat_and_keeps_history():
  170. service, _, clock = _service()
  171. assert service.toggle_chat() is True
  172. service.handle_event(_transcript(clock, "hello"))
  173. assert service.viewer_detached() is True
  174. assert service.modes.mode.value == "idle"
  175. assert service.chat_history == ["Unknown: hello"]
  176. def test_viewer_detached_is_noop_outside_chat():
  177. service, _, _ = _service()
  178. assert service.viewer_detached() is False
  179. assert service.modes.mode.value == "idle"
  180. def test_exit_chat_preserves_history_and_does_not_reset():
  181. cfg = Config()
  182. resets: list[int] = []
  183. service = AssistantService(
  184. cfg, FakeSubscriber(), time_fn=Clock(), chat_reset=lambda: resets.append(1)
  185. )
  186. service.toggle_chat()
  187. service.handle_event(_transcript(Clock(), "hello"))
  188. service.viewer_detached()
  189. service.toggle_chat()
  190. assert resets == []
  191. assert service.chat_history == ["Unknown: hello"]
  192. def test_chat_accumulates_turns():
  193. service, _, clock = _service()
  194. service.toggle_chat()
  195. for text in ("hello there", "how are you"):
  196. service.handle_event(_transcript(clock, text))
  197. assert service.chat_history == ["Unknown: hello there", "Unknown: how are you"]
  198. def test_chat_capture_off_skips_spoken_transcripts_and_exit_phrase():
  199. service, _, clock = _service()
  200. service.toggle_chat()
  201. assert service.set_chat_capture(False) is True
  202. service.handle_event(_transcript(clock, "hello"))
  203. service.handle_event(_transcript(clock, "That's all."))
  204. assert service.chat_history == []
  205. assert service._pending_chat == []
  206. assert service.modes.mode.value == "chat"
  207. def test_chat_capture_off_still_allows_typed_turns_and_triage():
  208. service, _, clock = _service()
  209. service.toggle_chat()
  210. service.set_chat_capture(False)
  211. service.handle_event(_transcript(clock, "spoken"))
  212. assert service.submit_chat_text("typed") is True
  213. assert service.chat_history == ["typed"]
  214. assert service.triage_history == ["spoken"]
  215. def test_set_chat_capture_emits_once_and_reports_status():
  216. service, emitted, _ = _service()
  217. assert service.status()["chat_capture"] is True
  218. assert service.set_chat_capture(False) is True
  219. assert [e for e in emitted if e["type"] == "chat_capture"] == [
  220. {"type": "chat_capture", "on": False}
  221. ]
  222. assert service.status()["chat_capture"] is False
  223. emitted.clear()
  224. service.set_chat_capture(False)
  225. assert emitted == []
  226. def test_triage_chat_off_ignores_chat_turns():
  227. service, _, _, _, lane = _triage_service()
  228. service._run_chat("typed")
  229. assert lane.submits == []
  230. assert service.triage_history == []
  231. def test_triage_chat_sends_reply_only_when_capture_on():
  232. service, _, _, _, lane = _triage_service()
  233. service.toggle_chat()
  234. service.set_triage_chat(True)
  235. service._run_chat("spoken, already in triage")
  236. assert service.triage_history == ["Assistant: ok"]
  237. assert lane.submits[0][0] == "- Assistant: ok"
  238. def test_triage_chat_sends_user_and_reply_when_capture_off():
  239. service, _, _, _, lane = _triage_service()
  240. service.toggle_chat()
  241. service.set_chat_capture(False)
  242. service.set_triage_chat(True)
  243. service._run_chat("typed")
  244. assert service.triage_history == ["You: typed", "Assistant: ok"]
  245. assert lane.submits[0][0] == "- You: typed\n- Assistant: ok"
  246. def test_set_triage_chat_emits_once_and_reports_status():
  247. service, emitted, _, _, _ = _triage_service()
  248. assert service.status()["triage_chat"] is False
  249. assert service.set_triage_chat(True) is True
  250. assert [e for e in emitted if e["type"] == "triage_chat"] == [
  251. {"type": "triage_chat", "on": True}
  252. ]
  253. assert service.status()["triage_chat"] is True
  254. emitted.clear()
  255. service.set_triage_chat(True)
  256. assert emitted == []
  257. def test_submit_chat_text_requires_chat_and_dispatches_immediately():
  258. cfg = Config()
  259. cfg.assistant.turn_gap_s = 999.0
  260. replies: list[str] = []
  261. emitted: list[dict] = []
  262. def emit(type_, **fields):
  263. emitted.append({"type": type_, **fields})
  264. service = AssistantService(
  265. cfg,
  266. FakeSubscriber(),
  267. emit=emit,
  268. chat_fn=lambda text: replies.append(text) or "ok",
  269. time_fn=Clock(),
  270. )
  271. assert service.submit_chat_text("hi") is False
  272. assert service.submit_chat_text(" ") is False
  273. service.toggle_chat()
  274. service.start()
  275. try:
  276. assert service.submit_chat_text("hi") is True
  277. deadline = time.time() + 2.0
  278. while time.time() < deadline and not replies:
  279. time.sleep(0.01)
  280. finally:
  281. service.stop()
  282. assert replies == ["hi"]
  283. assert service.chat_history == ["hi"]
  284. assert service._pending_chat == []
  285. assert any(
  286. e["type"] == "chat_message" and e.get("role") == "user" and e.get("text") == "hi"
  287. for e in emitted
  288. )
  289. def test_execute_clears_command_buffer_immediately():
  290. service, emitted, clock = _service()
  291. service.arm()
  292. service.handle_event(_transcript(clock, "do it"))
  293. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  294. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  295. assert service.execute() == "do it"
  296. assert service.modes.mode.value == "executing"
  297. assert service.command_text == ""
  298. assert service.command_edited is False
  299. assert service.arm_ts is None
  300. assert any(e["type"] == "command_text" and e["text"] == "" for e in emitted)
  301. def test_execute_requires_readiness():
  302. service, _, clock = _service()
  303. service.arm()
  304. service.handle_event(_transcript(clock, "do it"))
  305. service.handle_event(_ev(events.SPEECH_START, clock))
  306. assert service.execute() is None # speech still open
  307. assert service.modes.mode.value == "command_armed"
  308. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  309. assert service.execute() is None # awaiting segment 1
  310. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  311. assert service.execute() == "do it"
  312. assert service.modes.mode.value == "executing"
  313. assert service.finish() is True
  314. assert service.modes.mode.value == "idle"
  315. def test_execute_reports_busy_when_queue_rejects():
  316. service, emitted, clock = _service()
  317. # A zero-capacity lane rejects the priority command job on submit.
  318. service._inference = InferenceQueue(size=0)
  319. service.executor = object()
  320. service.arm()
  321. service.handle_event(_transcript(clock, "do it"))
  322. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  323. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  324. assert service.execute() is EXECUTE_BUSY
  325. assert service.modes.mode.value == "idle"
  326. assert not any(e.get("code") == "busy" for e in emitted)
  327. def test_turn_ended_requires_quiet_gap():
  328. service, _, clock = _service()
  329. service.handle_event(_ev(events.SPEECH_STOP, clock, seg_id=1))
  330. service.handle_event(_ev(events.SEGMENT_DONE, clock, seg_id=1, emitted=True))
  331. assert service.ready() is True
  332. assert service.turn_ended() is False
  333. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  334. assert service.turn_ended() is True
  335. def test_emits_mode_and_command_text():
  336. service, emitted, clock = _service()
  337. service.arm()
  338. assert {"type": "mode", "mode": "command_armed"} in emitted
  339. service.handle_event(_transcript(clock, "hi"))
  340. assert any(e["type"] == "command_text" and e["text"] == "hi" for e in emitted)
  341. def test_cancel_clears_command_buffer():
  342. service, _, clock = _service()
  343. service.arm()
  344. service.handle_event(_transcript(clock, "hi"))
  345. assert service.cancel() is True
  346. assert service.modes.mode.value == "idle"
  347. assert service.command_text == ""
  348. assert service.arm_ts is None
  349. def test_pending_chat_is_not_lost_across_threads():
  350. """Transcripts appended on the reader thread while the ticker dispatches
  351. must all reach the model: the join-and-clear in _dispatch_chat is atomic,
  352. and the user bubble is emitted by the worker in _run_chat."""
  353. cfg = Config()
  354. cfg.assistant.turn_gap_s = 0.01
  355. dispatched: list[str] = []
  356. lock = threading.Lock()
  357. def emit(type_, **fields):
  358. if type_ == "chat_message" and fields.get("role") == "user":
  359. with lock:
  360. dispatched.append(fields["text"])
  361. service = AssistantService(cfg, FakeSubscriber(), emit=emit, chat_fn=lambda text: "ok")
  362. service.toggle_chat()
  363. service.start()
  364. try:
  365. total = 200
  366. producer = threading.Thread(
  367. target=lambda: [
  368. service.handle_event(
  369. {
  370. "v": 1,
  371. "type": events.TRANSCRIPT,
  372. "seq": i,
  373. "ts": time.time(),
  374. "text": f"w{i}",
  375. }
  376. )
  377. for i in range(total)
  378. ],
  379. daemon=True,
  380. )
  381. producer.start()
  382. deadline = time.time() + 10.0
  383. while time.time() < deadline:
  384. service.tick()
  385. with lock:
  386. words = sum(len(text.split()) for text in dispatched)
  387. done = words == 2 * total
  388. if not producer.is_alive() and done:
  389. break
  390. time.sleep(0.002)
  391. producer.join(timeout=5)
  392. with lock:
  393. words = sum(len(text.split()) for text in dispatched)
  394. assert words == 2 * total
  395. finally:
  396. service.stop()
  397. def _full_queue_service():
  398. service, emitted, clock = _service()
  399. service.toggle_chat()
  400. service._inference = InferenceQueue(size=1)
  401. service._inference.submit(
  402. InferenceJob(name="blocker", run=lambda: None, evictable=False)
  403. )
  404. return service, emitted, clock
  405. def test_chat_refusal_rebuffers_and_emits_queue_full_once():
  406. """A refused chat submit re-buffers under lock and emits queue_full on flip."""
  407. service, emitted, _ = _full_queue_service()
  408. service._pending_chat = ["hello world"]
  409. service._dispatch_chat()
  410. assert service._pending_chat == ["hello world"]
  411. assert service._chat_queue_full is True
  412. errors = [e for e in emitted if e["type"] == "error" and e.get("code") == "queue_full"]
  413. assert len(errors) == 1
  414. assert [e for e in emitted if e["type"] == "chat_message"] == []
  415. service._dispatch_chat()
  416. assert service._pending_chat == ["hello world"]
  417. errors = [e for e in emitted if e["type"] == "error" and e.get("code") == "queue_full"]
  418. assert len(errors) == 1
  419. def test_chat_dispatch_success_clears_latch_without_user_bubble():
  420. """Success clears the latch; the user bubble comes from _run_chat, not dispatch."""
  421. service, emitted, _ = _full_queue_service()
  422. service._pending_chat = ["hello world"]
  423. service._chat_queue_full = True
  424. service._inference.drop(lambda j: True)
  425. service._dispatch_chat()
  426. assert service._pending_chat == []
  427. assert service._chat_queue_full is False
  428. assert [e for e in emitted if e["type"] == "chat_message"] == []
  429. def test_run_chat_emits_user_first_then_reply():
  430. service, emitted, _ = _service()
  431. service.chat_fn = lambda text: "reply!"
  432. service._run_chat("hello")
  433. assert emitted == [
  434. {"type": "chat_message", "role": "user", "text": "hello"},
  435. {"type": "chat_message", "role": "assistant", "text": "reply!"},
  436. ]
  437. def test_chat_notify_emits_compaction_note():
  438. service, emitted, _ = _service()
  439. service.chat_notify("compacted", turns=4)
  440. assert {
  441. "type": "chat_message",
  442. "role": "note",
  443. "text": "Compacted 4 earlier turns into a summary.",
  444. } in emitted
  445. def test_chat_notify_emits_truncated_error():
  446. service, emitted, _ = _service()
  447. service.chat_notify("truncated", message="cut off")
  448. assert {"type": "error", "code": "truncated", "message": "cut off"} in emitted
  449. def test_clear_empties_chat_history():
  450. service, emitted, clock = _service()
  451. service.toggle_chat()
  452. for text in ("hello", "how are you"):
  453. service.handle_event(_transcript(clock, text))
  454. assert service.chat_history == ["Unknown: hello", "Unknown: how are you"]
  455. assert len(service._pending_chat) == 2
  456. assert service.clear() is True
  457. assert service.chat_history == []
  458. assert service._pending_chat == []
  459. assert {"type": "cleared"} in emitted
  460. def test_clear_invokes_chat_reset():
  461. cfg = Config()
  462. resets: list[int] = []
  463. service = AssistantService(
  464. cfg, FakeSubscriber(), time_fn=Clock(), chat_reset=lambda: resets.append(1)
  465. )
  466. service.toggle_chat()
  467. service.clear()
  468. assert resets == [1]
  469. def test_clear_does_not_change_mode():
  470. service, _, clock = _service()
  471. service.toggle_chat()
  472. assert service.modes.mode.value == "chat"
  473. service.clear()
  474. assert service.modes.mode.value == "chat"
  475. def test_clear_always_returns_true():
  476. service, _, _ = _service()
  477. assert service.clear() is True
  478. service.toggle_chat()
  479. service.handle_event(_ev(events.TRANSCRIPT, Clock(), text="test"))
  480. assert service.clear() is True
  481. def test_heartbeat_emits_inference_depth_and_in_flight():
  482. """Calling tick() emits heartbeat with inference_depth (int) and inference_in_flight (bool)."""
  483. service, emitted, clock = _service()
  484. service.tick()
  485. heartbeats = [e for e in emitted if e["type"] == "heartbeat"]
  486. assert len(heartbeats) == 1
  487. heartbeat = heartbeats[0]
  488. assert "inference_depth" in heartbeat
  489. assert isinstance(heartbeat["inference_depth"], int)
  490. assert "inference_in_flight" in heartbeat
  491. assert isinstance(heartbeat["inference_in_flight"], bool)
  492. def test_status_contains_inference_depth_and_in_flight():
  493. """status() snapshot contains inference_depth (int) and inference_in_flight (bool)."""
  494. service, _, _ = _service()
  495. status = service.status()
  496. assert "inference_depth" in status
  497. assert isinstance(status["inference_depth"], int)
  498. assert "inference_in_flight" in status
  499. assert isinstance(status["inference_in_flight"], bool)
  500. _AUTO = object()
  501. class FakeTriage:
  502. def __init__(
  503. self,
  504. questions=_AUTO,
  505. state_pair=("ok", ""),
  506. escalate=False,
  507. parsed=_AUTO,
  508. normalized=_AUTO,
  509. gate=True,
  510. ):
  511. if questions is _AUTO:
  512. questions = {"urgency": {"type": "noul", "instructions": "how urgent"}}
  513. self._questions = questions
  514. self._state_pair = state_pair
  515. self._escalate = escalate
  516. self._parsed = parsed
  517. self._normalized = normalized
  518. self._gate = gate
  519. self.build_state_calls: list[tuple] = []
  520. self.normalize_calls: list = []
  521. self.prompt_calls: list[tuple] = []
  522. self.gate_calls: list = []
  523. @property
  524. def questions(self):
  525. return self._questions
  526. def state(self):
  527. return self._state_pair
  528. def build_state(self, window):
  529. self.build_state_calls.append(tuple(window))
  530. return "\n".join("- " + str(item) for item in (window or []))
  531. def normalize(self, payload):
  532. self.normalize_calls.append(payload)
  533. if self._normalized is not _AUTO:
  534. return self._normalized
  535. if not isinstance(payload, dict):
  536. return {}
  537. signals = {"order": [], "items": {}, "escalated": False, "input_tokens": 0}
  538. signals.update(payload)
  539. if self._escalate:
  540. signals["escalated"] = True
  541. return signals
  542. def build_brief_prompt(self, deltas, signals):
  543. self.prompt_calls.append(
  544. (tuple(deltas), dict(signals) if isinstance(signals, dict) else signals)
  545. )
  546. return "BRIEF-PROMPT"
  547. def build_ledger_prompt(self, narrative, ledger, deltas, signals):
  548. self.prompt_calls.append(
  549. (tuple(deltas), dict(signals) if isinstance(signals, dict) else signals)
  550. )
  551. ledger_text = "\n".join("- " + f for f in ledger) or "(none yet)"
  552. body = "\n".join("- " + d for d in deltas)
  553. return (
  554. f"CURRENT LEDGER:\n{ledger_text}\n\n"
  555. f"CURRENT BRIEF:\n{narrative or '(none yet)'}\n\n"
  556. f"NEW UTTERANCES:\n{body}"
  557. )
  558. def parse_ledger_reply(self, text, truncated=False):
  559. return _real_analyzer.parse_ledger_reply(text, truncated=truncated)
  560. def parse_facts(self, block):
  561. return _real_analyzer.parse_facts(block)
  562. def parse_brief(self, text):
  563. if self._parsed is _AUTO:
  564. return text if text else None
  565. return self._parsed
  566. def should_update_brief(self, signals):
  567. self.gate_calls.append(signals)
  568. return self._gate
  569. class FakeBriefFile:
  570. """Recording stand-in for BriefFile (write/clear/set_directory never raise)."""
  571. def __init__(self):
  572. self.writes: list[str] = []
  573. self.clears = 0
  574. self.directories: list[str] = []
  575. def set_directory(self, directory):
  576. self.directories.append(directory)
  577. def write(self, text):
  578. self.writes.append(text)
  579. def clear(self):
  580. self.clears += 1
  581. class FakeLane:
  582. def __init__(self, ready=True):
  583. self._ready = ready
  584. self.submits: list[tuple] = []
  585. self.drops = 0
  586. self.predict_gate = threading.Event()
  587. def ready(self):
  588. return self._ready
  589. def state(self):
  590. return ("ok", "") if self._ready else ("loading", "")
  591. def submit(self, state, questions, lang, generation):
  592. self.submits.append((state, questions, lang, generation))
  593. return True
  594. def drop_pending(self):
  595. self.drops += 1
  596. class FakeDebugSink:
  597. def __init__(self):
  598. self.events: list[tuple] = []
  599. def emit(self, type_, **fields):
  600. self.events.append((type_, fields))
  601. def _ok_reply(text):
  602. return "ok"
  603. def _brief_reply(narrative="", facts=None):
  604. """Build a ledger-format reply the analyzer's parser can consume."""
  605. body = "\n".join("- " + fact for fact in (facts or [])) or "none"
  606. return f"NEW FACTS:\n{body}\n\nBRIEF:\n{narrative}"
  607. def _triage_service(
  608. triage=None,
  609. lane=None,
  610. brief_fn=None,
  611. chat_fn: Any = _AUTO,
  612. brief_file=None,
  613. brief_reset=None,
  614. chat_reset=None,
  615. debug_sink=None,
  616. ):
  617. cfg = Config()
  618. cfg.assistant.turn_gap_s = 0.5
  619. clock = Clock()
  620. emitted: list[dict] = []
  621. def emit(type_, **fields):
  622. emitted.append({"type": type_, **fields})
  623. if triage is None:
  624. triage = FakeTriage()
  625. if lane is None:
  626. lane = FakeLane()
  627. if chat_fn is _AUTO:
  628. chat_fn = _ok_reply
  629. factory = (lambda: debug_sink) if debug_sink is not None else None
  630. service = AssistantService(
  631. cfg,
  632. FakeSubscriber(),
  633. emit=emit,
  634. time_fn=clock,
  635. triage=triage,
  636. laya_lane=lane,
  637. brief_fn=_wrap_brief_fn(brief_fn),
  638. brief_reset=brief_reset,
  639. brief_file=brief_file,
  640. chat_fn=chat_fn,
  641. chat_reset=chat_reset,
  642. debug_sink_factory=factory,
  643. )
  644. return service, emitted, clock, triage, lane
  645. def _wrap_brief_fn(brief_fn):
  646. """Adapt a simple `prompt -> text` stub to the engine's call signature.
  647. A plain string return is treated as the volatile narrative half of a
  648. ledger reply, so existing stubs keep reading naturally.
  649. """
  650. if brief_fn is None:
  651. return None
  652. def wrapped(prompt, on_io=None, finalize=None):
  653. kwargs = {}
  654. if on_io is not None:
  655. kwargs["on_io"] = on_io
  656. try:
  657. result = brief_fn(prompt, **kwargs)
  658. except TypeError:
  659. result = brief_fn(prompt)
  660. if finalize is not None:
  661. result = finalize(result, "stop")
  662. is_ledger = "BRIEF:" in result or "NEW FACTS:" in result
  663. if isinstance(result, str) and result and not is_ledger:
  664. result = _brief_reply(narrative=result)
  665. return result
  666. return wrapped
  667. def _triage_events(emitted):
  668. return [e for e in emitted if e["type"] == "triage"]
  669. def _escalated_payload():
  670. return {"order": [], "items": {}, "escalated": True, "input_tokens": 3}
  671. def test_triage_status_disabled_when_no_analyzer():
  672. service, _, _ = _service()
  673. status = service.status()
  674. assert status["triage_state"] == "disabled"
  675. assert status["triage_reason"] == ""
  676. assert status["triage"] is None
  677. def test_triage_status_reflects_analyzer():
  678. service, _, _, _, _ = _triage_service(
  679. triage=FakeTriage(state_pair=("loading", "warming up"))
  680. )
  681. status = service.status()
  682. assert status["triage_state"] == "loading"
  683. assert status["triage_reason"] == "warming up"
  684. assert status["triage"] is None
  685. def test_status_contains_questions_from_analyzer():
  686. questions = {
  687. "urgency": {"type": "noul", "instructions": "how urgent"},
  688. "sentiment": {"type": "score", "instructions": "tone", "criteria": ["low", "high"]},
  689. }
  690. service, _, _, triage, _ = _triage_service(triage=FakeTriage(questions=questions))
  691. status = service.status()
  692. assert status["questions"] == questions
  693. assert status["questions"] is triage.questions
  694. def test_status_questions_none_without_analyzer():
  695. service, _, _ = _service()
  696. assert service.status()["questions"] is None
  697. def test_status_event_carries_questions():
  698. service, emitted, _, triage, _ = _triage_service()
  699. service.tick()
  700. statuses = [e for e in emitted if e["type"] == "status"]
  701. assert len(statuses) == 1
  702. assert statuses[0]["questions"] is triage.questions
  703. def test_triage_captures_transcript_lang():
  704. service, _, clock, _, _ = _triage_service()
  705. assert service._last_transcript_lang is None
  706. event = _transcript(clock, "hola")
  707. event["lang"] = "es"
  708. service.handle_event(event)
  709. assert service._last_transcript_lang == "es"
  710. service.handle_event(_transcript(clock, "hello"))
  711. assert service._last_transcript_lang is None
  712. def test_maybe_triage_passes_lang_to_lane():
  713. service, _, clock, _, lane = _triage_service()
  714. event = _transcript(clock, "hola")
  715. event["lang"] = "es"
  716. service.handle_event(event)
  717. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  718. service.tick()
  719. assert len(lane.submits) == 1
  720. _, _, lang, _ = lane.submits[0]
  721. assert lang == "es"
  722. def test_maybe_triage_runs_in_idle_mode():
  723. service, _, clock, triage, lane = _triage_service()
  724. service.handle_event(_transcript(clock, "hello"))
  725. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  726. service.tick()
  727. assert len(lane.submits) == 1
  728. state, questions, lang, gen = lane.submits[0]
  729. assert state == "- hello"
  730. assert questions is triage.questions
  731. assert lang is None
  732. assert gen == 0
  733. assert triage.build_state_calls == [("hello",)]
  734. assert service.triage_history == ["hello"]
  735. def test_maybe_triage_uses_transcript_buffer_not_chat_history():
  736. service, _, clock, triage, lane = _triage_service(chat_fn=None)
  737. service.chat_history = ["stale chat only"]
  738. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  739. service.tick()
  740. assert lane.submits == []
  741. assert triage.build_state_calls == []
  742. service.handle_event(_transcript(clock, "hello"))
  743. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  744. service.tick()
  745. assert len(lane.submits) == 1
  746. assert triage.build_state_calls == [("hello",)]
  747. def test_maybe_triage_requires_settled_turn():
  748. service, _, clock, _, lane = _triage_service()
  749. service.handle_event(_transcript(clock, "hello"))
  750. service.tick()
  751. assert lane.submits == []
  752. def test_maybe_triage_periodic_fires_before_turn_settles():
  753. service, _, clock, _, lane = _triage_service()
  754. service.cfg.triage.periodic_s = 0.2
  755. service.handle_event(_transcript(clock, "hello"))
  756. assert service.turn_ended() is False
  757. clock.advance(0.25) # due (0.2) yet still quieter than turn_gap_s (0.5)
  758. service.tick()
  759. assert service.turn_ended() is False
  760. assert len(lane.submits) == 1
  761. assert lane.submits[0][0] == "- hello"
  762. def test_maybe_triage_periodic_disabled_ignores_accumulating_speech():
  763. service, _, clock, _, lane = _triage_service()
  764. service.cfg.triage.periodic_s = 0.0
  765. service.handle_event(_transcript(clock, "hello"))
  766. clock.advance(0.25)
  767. service.tick()
  768. assert service.turn_ended() is False
  769. assert lane.submits == []
  770. def test_maybe_triage_periodic_rearms_after_submit():
  771. service, _, clock, _, lane = _triage_service()
  772. service.cfg.triage.periodic_s = 0.2
  773. service.handle_event(_transcript(clock, "first"))
  774. clock.advance(0.25)
  775. service.tick()
  776. assert len(lane.submits) == 1
  777. service.tick()
  778. service.tick()
  779. assert len(lane.submits) == 1 # identical window, no new speech
  780. service.handle_event(_transcript(clock, "second"))
  781. clock.advance(0.25)
  782. service.tick()
  783. assert len(lane.submits) == 2
  784. assert lane.submits[1][0] == "- first\n- second"
  785. def test_triage_debug_emits_submit_and_result():
  786. sink = FakeDebugSink()
  787. service, _, clock, _, lane = _triage_service(debug_sink=sink)
  788. service.handle_event(_transcript(clock, "hello"))
  789. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  790. service.tick()
  791. submits = [event for event in sink.events if event[0] == "triage-submit"]
  792. assert submits == [
  793. ("triage-submit", {"trigger": "settled", "window": 1, "generation": 0})
  794. ]
  795. _, _, _, gen = lane.submits[0]
  796. service._on_triage_result(gen, _escalated_payload())
  797. results = [event for event in sink.events if event[0] == "triage-result"]
  798. assert len(results) == 1
  799. assert results[0][1]["escalated"] is True
  800. assert results[0][1]["generation"] == 0
  801. def test_triage_debug_marks_periodic_trigger():
  802. sink = FakeDebugSink()
  803. service, _, clock, _, lane = _triage_service(debug_sink=sink)
  804. service.cfg.triage.periodic_s = 0.2
  805. service.handle_event(_transcript(clock, "hello"))
  806. clock.advance(0.25)
  807. service.tick()
  808. submits = [event for event in sink.events if event[0] == "triage-submit"]
  809. assert submits[0][1]["trigger"] == "periodic"
  810. def test_triage_debug_emits_brief_dispatch():
  811. sink = FakeDebugSink()
  812. service, _, clock, _, lane = _triage_service(
  813. brief_fn=lambda prompt: "RAW", debug_sink=sink
  814. )
  815. service.handle_event(_transcript(clock, "hello"))
  816. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  817. service.tick()
  818. _, _, _, gen = lane.submits[0]
  819. service._on_triage_result(gen, _escalated_payload())
  820. brief = [event for event in sink.events if event[0] == "triage-brief"]
  821. assert brief == [("triage-brief", {"action": "dispatch", "generation": 0})]
  822. def test_triage_debug_emits_brief_gate_skip():
  823. sink = FakeDebugSink()
  824. service, _, clock, _, lane = _triage_service(
  825. triage=FakeTriage(gate=False), brief_fn=lambda prompt: "RAW", debug_sink=sink
  826. )
  827. service.handle_event(_transcript(clock, "hello"))
  828. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  829. service.tick()
  830. _, _, _, gen = lane.submits[0]
  831. service._on_triage_result(gen, _escalated_payload())
  832. brief = [event for event in sink.events if event[0] == "triage-brief"]
  833. assert brief == [
  834. ("triage-brief", {"action": "skip", "reason": "gate", "generation": 0})
  835. ]
  836. def test_maybe_triage_skips_unready_lane():
  837. service, _, clock, _, lane = _triage_service(lane=FakeLane(ready=False))
  838. service.handle_event(_transcript(clock, "hello"))
  839. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  840. service.tick()
  841. assert lane.submits == []
  842. def test_maybe_triage_skips_without_questions():
  843. service, _, clock, triage, lane = _triage_service(
  844. triage=FakeTriage(questions=None)
  845. )
  846. service.handle_event(_transcript(clock, "hello"))
  847. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  848. service.tick()
  849. assert lane.submits == []
  850. assert triage.build_state_calls == []
  851. def test_maybe_triage_dedupes_identical_window():
  852. service, _, clock, triage, lane = _triage_service()
  853. service.handle_event(_transcript(clock, "hello"))
  854. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  855. service.tick()
  856. assert len(lane.submits) == 1
  857. state, questions, lang, gen = lane.submits[0]
  858. assert state == "- hello"
  859. assert questions is triage.questions
  860. assert lang is None
  861. assert gen == 0
  862. service.tick()
  863. service.tick()
  864. assert len(lane.submits) == 1
  865. service.handle_event(_transcript(clock, "second"))
  866. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  867. service.tick()
  868. assert len(lane.submits) == 2
  869. assert lane.submits[1][0] == "- hello\n- second"
  870. def test_second_transcript_produces_second_submit():
  871. service, _, clock, _, lane = _triage_service()
  872. service.handle_event(_transcript(clock, "first"))
  873. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  874. service.tick()
  875. assert len(lane.submits) == 1
  876. service.handle_event(_transcript(clock, "second"))
  877. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  878. service.tick()
  879. assert len(lane.submits) == 2
  880. assert lane.submits[1][0] == "- first\n- second"
  881. def test_tick_under_50ms_with_blocked_lane():
  882. service, _, clock, _, lane = _triage_service()
  883. assert not lane.predict_gate.is_set()
  884. service.toggle_chat()
  885. service.handle_event(_transcript(clock, "hello"))
  886. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  887. start = time.perf_counter()
  888. service.tick()
  889. elapsed = time.perf_counter() - start
  890. assert len(lane.submits) == 1
  891. assert elapsed < 0.05
  892. def test_on_triage_result_emits_signals():
  893. service, emitted, clock, _, lane = _triage_service()
  894. service.toggle_chat()
  895. service.handle_event(_transcript(clock, "hello"))
  896. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  897. service.tick()
  898. _, _, _, gen = lane.submits[0]
  899. payload = {
  900. "order": ["urgency"],
  901. "items": {"urgency": {"kind": "noul"}},
  902. "escalated": False,
  903. "input_tokens": 12,
  904. }
  905. service._on_triage_result(gen, payload)
  906. staged = _triage_events(emitted)
  907. assert len(staged) == 1
  908. assert staged[0]["stage"] == "signals"
  909. assert staged[0]["triage"] == service.last_triage
  910. assert staged[0]["triage"]["escalated"] is False
  911. assert staged[0]["triage"]["input_tokens"] == 12
  912. assert service.status()["triage"] == service.last_triage
  913. def test_triage_events_carry_the_server_timestamp():
  914. service, emitted, clock, _, lane = _triage_service(brief_fn=lambda prompt: "RAW")
  915. service.toggle_chat()
  916. service.handle_event(_transcript(clock, "hello"))
  917. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  918. service.tick()
  919. _, _, _, gen = lane.submits[0]
  920. expected_ts = clock()
  921. service._on_triage_result(gen, _escalated_payload())
  922. signals = [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  923. assert len(signals) == 1
  924. assert isinstance(signals[0]["ts"], (int, float))
  925. assert signals[0]["ts"] == expected_ts
  926. job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  927. job.run()
  928. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  929. assert len(contents) == 1
  930. assert isinstance(contents[0]["ts"], (int, float))
  931. assert contents[0]["ts"] == expected_ts # content reuses the signals timestamp
  932. def test_on_triage_result_suppresses_stale_generation():
  933. service, emitted, clock, _, lane = _triage_service()
  934. service.toggle_chat()
  935. service.handle_event(_transcript(clock, "hello"))
  936. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  937. service.tick()
  938. _, _, _, gen = lane.submits[0]
  939. service.clear()
  940. service._on_triage_result(gen, _escalated_payload())
  941. assert service.last_triage is None
  942. assert [e for e in _triage_events(emitted) if e.get("stage") == "signals"] == []
  943. service.handle_event(_transcript(clock, "hello"))
  944. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  945. service.tick()
  946. assert len(lane.submits) == 2
  947. def test_pending_brief_capped_by_char_budget():
  948. """The buffer keeps the newest deltas within budget, dropping the oldest."""
  949. service, _, clock, _, _ = _triage_service(brief_fn=lambda prompt: "RAW")
  950. service.cfg.triage.brief_pending_char_budget = 30
  951. service.handle_event(_transcript(clock, "aaa"))
  952. service.handle_event(_transcript(clock, "bbbb"))
  953. assert service._pending_brief == ["Unknown: aaa", "Unknown: bbbb"]
  954. service.handle_event(_transcript(clock, "ccccc"))
  955. assert service._pending_brief == ["Unknown: bbbb", "Unknown: ccccc"]
  956. assert service.triage_history == ["aaa", "bbbb", "ccccc"]
  957. def test_pending_brief_untouched_without_triage():
  958. service, _, clock = _service()
  959. service.handle_event(_transcript(clock, "hello"))
  960. assert service._pending_brief == []
  961. assert service.triage_history == ["hello"]
  962. def test_on_triage_result_accepts_non_chat_mode():
  963. service, emitted, clock, _, lane = _triage_service()
  964. service.handle_event(_transcript(clock, "hello"))
  965. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  966. service.tick()
  967. _, _, _, gen = lane.submits[0]
  968. window = service._last_triage_window
  969. service._on_triage_result(gen, _escalated_payload())
  970. assert service.last_triage is not None
  971. assert service._last_triage_window == window
  972. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  973. def test_on_triage_result_ignores_none_payload():
  974. service, emitted, clock, _, lane = _triage_service()
  975. service.toggle_chat()
  976. service.handle_event(_transcript(clock, "hello"))
  977. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  978. service.tick()
  979. _, _, _, gen = lane.submits[0]
  980. service._on_triage_result(gen, None)
  981. assert service.last_triage is None
  982. assert _triage_events(emitted) == []
  983. def test_none_payload_dispatches_nothing_and_keeps_deltas():
  984. """A Laya predict failure never dispatches: no job, no events, deltas stay."""
  985. calls: list[str] = []
  986. service, emitted, clock, _, lane = _triage_service(
  987. brief_fn=lambda prompt: calls.append(prompt) or "RAW"
  988. )
  989. service.toggle_chat()
  990. service.handle_event(_transcript(clock, "hello"))
  991. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  992. service.tick()
  993. _, _, _, gen = lane.submits[0]
  994. assert service._pending_brief == ["Unknown: hello"]
  995. service._on_triage_result(gen, None)
  996. assert service.last_triage is None
  997. assert [e.get("stage") for e in _triage_events(emitted)] == []
  998. assert [j.name for j in service._inference._queue] == ["chat"]
  999. assert calls == []
  1000. assert service._pending_brief == ["Unknown: hello"]
  1001. assert service._brief_outstanding is False
  1002. def test_none_payload_without_brief_fn_enqueues_nothing():
  1003. service, emitted, clock, _, lane = _triage_service()
  1004. service.toggle_chat()
  1005. service.handle_event(_transcript(clock, "hello"))
  1006. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1007. service.tick()
  1008. _, _, _, gen = lane.submits[0]
  1009. service._on_triage_result(gen, None)
  1010. assert service.last_triage is None
  1011. assert _triage_events(emitted) == []
  1012. assert [j.name for j in service._inference._queue] == ["chat"]
  1013. assert service._pending_brief == ["Unknown: hello"]
  1014. def test_on_triage_result_ignores_empty_signals():
  1015. service, emitted, clock, _, lane = _triage_service(
  1016. triage=FakeTriage(normalized={})
  1017. )
  1018. service.toggle_chat()
  1019. service.handle_event(_transcript(clock, "hello"))
  1020. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1021. service.tick()
  1022. _, _, _, gen = lane.submits[0]
  1023. service._on_triage_result(gen, _escalated_payload())
  1024. assert service.last_triage is None
  1025. assert _triage_events(emitted) == []
  1026. def test_substantive_turn_dispatches_one_brief_job():
  1027. """A substantive settled turn calls the gate once and queues one brief job."""
  1028. service, emitted, clock, triage, lane = _triage_service(
  1029. triage=FakeTriage(escalate=True), brief_fn=lambda prompt: "RAW"
  1030. )
  1031. service.toggle_chat()
  1032. service.handle_event(_transcript(clock, "we might churn"))
  1033. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1034. service.tick()
  1035. _, _, _, gen = lane.submits[0]
  1036. service._on_triage_result(
  1037. gen, {"order": [], "items": {}, "escalated": False, "input_tokens": 0}
  1038. )
  1039. assert service.last_triage is not None
  1040. assert service.last_triage["escalated"] is True
  1041. assert service._last_brief_signals == service.last_triage
  1042. assert service._inference.depth() == 2
  1043. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  1044. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  1045. assert len(triage.gate_calls) == 1
  1046. assert service._brief_outstanding is True
  1047. assert service._brief_token == 1
  1048. def test_brief_dispatches_without_escalation():
  1049. service, _, clock, _, lane = _triage_service(brief_fn=lambda prompt: "RAW")
  1050. service.toggle_chat()
  1051. service.handle_event(_transcript(clock, "all calm"))
  1052. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1053. service.tick()
  1054. _, _, _, gen = lane.submits[0]
  1055. service._on_triage_result(
  1056. gen, {"order": [], "items": {}, "escalated": False, "input_tokens": 0}
  1057. )
  1058. assert service.last_triage is not None
  1059. assert service._inference.depth() == 2
  1060. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  1061. def test_gated_out_turn_skips_brief():
  1062. """A gated-out turn stores signals but never calls brief_fn."""
  1063. calls: list[str] = []
  1064. service, emitted, clock, triage, lane = _triage_service(
  1065. triage=FakeTriage(gate=False),
  1066. brief_fn=lambda prompt: calls.append(prompt) or "RAW",
  1067. )
  1068. service.toggle_chat()
  1069. service.handle_event(_transcript(clock, "small talk"))
  1070. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1071. service.tick()
  1072. _, _, _, gen = lane.submits[0]
  1073. service._on_triage_result(gen, _escalated_payload())
  1074. assert service.last_triage is not None
  1075. assert "brief" not in service.last_triage # no prior brief to retain
  1076. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  1077. assert calls == []
  1078. assert [j.name for j in service._inference._queue] == ["chat"]
  1079. assert service._brief_outstanding is False
  1080. assert len(triage.gate_calls) == 1
  1081. def test_gated_out_turn_retains_previous_brief():
  1082. """A gated-out settled turn keeps the last accepted brief in the panel."""
  1083. service, emitted, clock, triage, lane = _triage_service(
  1084. triage=FakeTriage(gate=False), brief_fn=lambda prompt: "RAW"
  1085. )
  1086. service.toggle_chat()
  1087. signals = _escalated_payload()
  1088. service._last_brief_signals = signals
  1089. token = _seed_brief(service, "Unknown: first")
  1090. service._run_brief_update(0, signals, service._now(), token)
  1091. assert service.last_triage is not None
  1092. assert service.last_triage["brief"] == "RAW"
  1093. emitted.clear()
  1094. service.handle_event(_transcript(clock, "small talk"))
  1095. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1096. service.tick()
  1097. _, _, _, gen = lane.submits[0]
  1098. service._on_triage_result(gen, _escalated_payload())
  1099. assert service.last_triage is not None
  1100. assert service.last_triage["brief"] == "RAW"
  1101. assert "brief" not in service._last_brief_signals # raw signals stay raw
  1102. staged = [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  1103. assert len(staged) == 1
  1104. assert staged[0]["triage"] == service.last_triage
  1105. assert staged[0]["triage"]["brief"] == "RAW"
  1106. assert len(triage.gate_calls) == 1
  1107. def test_gated_out_turn_retains_brief_when_extract_disabled():
  1108. """`extract_enabled = false` skips the dispatch but still retains the brief."""
  1109. service, emitted, clock, _, lane = _triage_service(
  1110. triage=FakeTriage(gate=False), brief_fn=lambda prompt: "RAW"
  1111. )
  1112. service.toggle_chat()
  1113. signals = _escalated_payload()
  1114. service._last_brief_signals = signals
  1115. token = _seed_brief(service, "Unknown: first")
  1116. service._run_brief_update(0, signals, service._now(), token)
  1117. assert service.last_triage is not None
  1118. assert service.last_triage["brief"] == "RAW"
  1119. service.cfg.triage.extract_enabled = False
  1120. emitted.clear()
  1121. service.handle_event(_transcript(clock, "small talk"))
  1122. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1123. service.tick()
  1124. _, _, _, gen = lane.submits[0]
  1125. service._on_triage_result(gen, _escalated_payload())
  1126. assert service.last_triage is not None
  1127. assert service.last_triage["brief"] == "RAW"
  1128. assert [j.name for j in service._inference._queue] == ["chat"]
  1129. staged = [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  1130. assert staged[0]["triage"]["brief"] == "RAW"
  1131. def test_gated_out_turn_retains_brief_without_brief_fn():
  1132. """No brief_fn means no dispatch, but a retained brief is still carried."""
  1133. service, emitted, clock, _, lane = _triage_service(triage=FakeTriage(gate=False))
  1134. service.toggle_chat()
  1135. service.last_triage = {**_escalated_payload(), "brief": "KEEP"}
  1136. emitted.clear()
  1137. service.handle_event(_transcript(clock, "small talk"))
  1138. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1139. service.tick()
  1140. _, _, _, gen = lane.submits[0]
  1141. service._on_triage_result(gen, _escalated_payload())
  1142. assert service.last_triage["brief"] == "KEEP"
  1143. staged = [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  1144. assert staged[0]["triage"]["brief"] == "KEEP"
  1145. def test_later_accepted_update_replaces_retained_brief():
  1146. """A later accepted update overwrites the retained brief text."""
  1147. texts = iter(["OLD", "NEW"])
  1148. service, emitted, _, _, _ = _triage_service(brief_fn=lambda prompt: next(texts))
  1149. service.toggle_chat()
  1150. signals = _escalated_payload()
  1151. service._last_brief_signals = signals
  1152. token = _seed_brief(service, "Unknown: first")
  1153. service._run_brief_update(0, signals, service._now(), token)
  1154. assert service.last_triage is not None
  1155. assert service.last_triage["brief"] == "OLD"
  1156. token = _seed_brief(service, "Unknown: second")
  1157. service._run_brief_update(0, signals, service._now(), token)
  1158. assert service.last_triage is not None
  1159. assert service.last_triage["brief"] == "NEW"
  1160. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1161. assert contents[-1]["triage"]["brief"] == "NEW"
  1162. service.clear()
  1163. assert service.last_triage is None
  1164. def test_no_brief_when_disabled_in_config():
  1165. calls: list[str] = []
  1166. service, _, clock, _, lane = _triage_service(
  1167. triage=FakeTriage(escalate=True),
  1168. brief_fn=lambda prompt: calls.append(prompt) or "RAW",
  1169. )
  1170. service.cfg.triage.extract_enabled = False
  1171. service.toggle_chat()
  1172. service.handle_event(_transcript(clock, "churn risk"))
  1173. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1174. service.tick()
  1175. _, _, _, gen = lane.submits[0]
  1176. service._on_triage_result(gen, _escalated_payload())
  1177. assert service.last_triage is not None
  1178. assert service.last_triage["escalated"] is True
  1179. assert calls == []
  1180. assert [j.name for j in service._inference._queue] == ["chat"]
  1181. def test_no_brief_without_brief_fn():
  1182. service, _, clock, _, lane = _triage_service(triage=FakeTriage(escalate=True))
  1183. service.toggle_chat()
  1184. service.handle_event(_transcript(clock, "churn risk"))
  1185. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1186. service.tick()
  1187. _, _, _, gen = lane.submits[0]
  1188. service._on_triage_result(gen, _escalated_payload())
  1189. assert service.last_triage is not None
  1190. assert [j.name for j in service._inference._queue] == ["chat"]
  1191. assert service._brief_outstanding is False
  1192. def test_brief_refused_when_queue_full():
  1193. """A refused submit leaves deltas buffered and the outstanding flag clear."""
  1194. service, emitted, clock, _, lane = _triage_service(
  1195. triage=FakeTriage(escalate=True), brief_fn=lambda prompt: "RAW"
  1196. )
  1197. service._inference = InferenceQueue(size=2)
  1198. service.toggle_chat()
  1199. service.handle_event(_transcript(clock, "churn risk"))
  1200. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1201. service.tick()
  1202. assert service._inference.depth() == 1
  1203. _, _, _, gen = lane.submits[0]
  1204. service._on_triage_result(gen, _escalated_payload())
  1205. assert service._inference.depth() == 1
  1206. assert service._inference._queue[0].name == "chat"
  1207. assert service.last_triage is not None
  1208. assert service.last_triage["escalated"] is True
  1209. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  1210. assert service._pending_brief == ["Unknown: churn risk"]
  1211. assert service._brief_outstanding is False
  1212. def test_second_result_while_outstanding_is_refused():
  1213. """At most one brief job is outstanding: the second turn only stores signals."""
  1214. service, emitted, clock, _, lane = _triage_service(brief_fn=lambda prompt: "RAW")
  1215. service.toggle_chat()
  1216. service.handle_event(_transcript(clock, "first"))
  1217. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1218. service.tick()
  1219. _, _, _, gen0 = lane.submits[0]
  1220. service._on_triage_result(gen0, _escalated_payload())
  1221. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  1222. service.handle_event(_transcript(clock, "second"))
  1223. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1224. service.tick()
  1225. _, _, _, gen1 = lane.submits[1]
  1226. service._on_triage_result(gen1, _escalated_payload())
  1227. jobs = [j for j in service._inference._queue if j.name == "brief-update"]
  1228. assert len(jobs) == 1
  1229. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals", "signals"]
  1230. assert service._brief_outstanding is True
  1231. assert service._pending_brief == ["Unknown: first", "Unknown: second"]
  1232. def _seed_brief(service, *deltas):
  1233. service._pending_brief = list(deltas)
  1234. service._brief_outstanding = True
  1235. return service._brief_token
  1236. def test_ledger_records_fact_and_survives_model_reemit():
  1237. """A recorded fact persists even when a later reply omits it."""
  1238. first = _brief_reply("Summary: rolling", ["The budget was $305 million."])
  1239. second = _brief_reply("Summary: newer", [])
  1240. replies = iter([first, second])
  1241. service, _, _, _, _ = _triage_service(brief_fn=lambda prompt: next(replies))
  1242. service._pending_brief = ["Unknown: one"]
  1243. service._brief_outstanding = True
  1244. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1245. service._pending_brief = ["Unknown: two"]
  1246. service._brief_outstanding = True
  1247. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1248. assert service.last_triage is not None
  1249. brief = service.last_triage["brief"]
  1250. assert "The budget was $305 million." in brief
  1251. assert service._fact_ledger.facts() == ["The budget was $305 million."]
  1252. def test_ledger_dedupes_repeated_fact():
  1253. """The same fact proposed twice is recorded once."""
  1254. reply = _brief_reply("Summary: s", ["Sense could find prey.", "sense could find prey"])
  1255. service, _, _, _, _ = _triage_service(brief_fn=lambda prompt: reply)
  1256. service._pending_brief = ["Unknown: one"]
  1257. service._brief_outstanding = True
  1258. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1259. assert service._fact_ledger.facts() == ["Sense could find prey."]
  1260. def test_ledger_prompt_carries_ledger_and_narrative():
  1261. """The brief request hands the model the ledger, narrative, and new deltas."""
  1262. seen: list[str] = []
  1263. service, _, _, _, _ = _triage_service(
  1264. brief_fn=lambda prompt: seen.append(prompt) or _brief_reply("UPDATED")
  1265. )
  1266. service._fact_ledger.merge(["PRIOR FACT"])
  1267. service._last_brief_narrative = "PRIOR NARRATIVE"
  1268. service._pending_brief = ["Unknown: hello there"]
  1269. service._brief_outstanding = True
  1270. service._run_brief_update(0, _escalated_payload(), service._now(), service._brief_token)
  1271. assert len(seen) == 1
  1272. assert "CURRENT LEDGER:" in seen[0]
  1273. assert "- PRIOR FACT" in seen[0]
  1274. assert "CURRENT BRIEF:\nPRIOR NARRATIVE" in seen[0]
  1275. assert "NEW UTTERANCES:\n- Unknown: hello there" in seen[0]
  1276. def test_brief_merges_content_and_writes_file():
  1277. brief_file = FakeBriefFile()
  1278. service, emitted, _, triage, _ = _triage_service(
  1279. brief_fn=lambda prompt: "the brief", brief_file=brief_file
  1280. )
  1281. service.toggle_chat()
  1282. signals = _escalated_payload()
  1283. service._last_brief_signals = signals
  1284. token = _seed_brief(service, "Unknown: hello")
  1285. ts = service._now()
  1286. service._run_brief_update(0, signals, ts, token)
  1287. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1288. assert len(contents) == 1
  1289. assert contents[0]["triage"] == {**signals, "brief": "the brief"}
  1290. assert service.last_triage == contents[0]["triage"]
  1291. assert triage.prompt_calls == [(("Unknown: hello",), signals)]
  1292. assert brief_file.writes == ["the brief"]
  1293. assert service._pending_brief == []
  1294. assert service._brief_outstanding is False
  1295. def test_empty_brief_requeues_deltas():
  1296. """A `""` reply keeps signals, re-buffers the deltas, and clears the flag."""
  1297. service, emitted, _, _, _ = _triage_service(brief_fn=lambda prompt: "")
  1298. service.toggle_chat()
  1299. signals = _escalated_payload()
  1300. service.last_triage = signals
  1301. service._last_brief_signals = signals
  1302. token = _seed_brief(service, "Unknown: hello")
  1303. ts = service._now()
  1304. service._run_brief_update(0, signals, ts, token)
  1305. assert [e for e in _triage_events(emitted) if e.get("stage") == "content"] == []
  1306. assert service.last_triage == signals
  1307. assert "brief" not in service.last_triage
  1308. assert service._pending_brief == ["Unknown: hello"]
  1309. assert service._brief_outstanding is False
  1310. def test_brief_error_requeues_and_clears_flag():
  1311. """A raising brief_fn leaves the flag clear and the deltas buffered."""
  1312. def boom(prompt):
  1313. raise RuntimeError("llm down")
  1314. service, emitted, _, _, _ = _triage_service(brief_fn=boom)
  1315. service.toggle_chat()
  1316. token = _seed_brief(service, "Unknown: hello")
  1317. ts = service._now()
  1318. service._run_brief_update(0, _escalated_payload(), ts, token)
  1319. assert _triage_events(emitted) == []
  1320. assert service.last_triage is None
  1321. assert service._pending_brief == ["Unknown: hello"]
  1322. assert service._brief_outstanding is False
  1323. def test_brief_empty_buffer_returns_early():
  1324. """An empty buffer returns early: no call, no event, flag cleared."""
  1325. calls: list[str] = []
  1326. service, emitted, _, _, _ = _triage_service(
  1327. brief_fn=lambda prompt: calls.append(prompt) or "RAW"
  1328. )
  1329. service.toggle_chat()
  1330. service._brief_outstanding = True
  1331. token = service._brief_token
  1332. ts = service._now()
  1333. assert service._pending_brief == []
  1334. service._run_brief_update(0, _escalated_payload(), ts, token)
  1335. assert calls == []
  1336. assert _triage_events(emitted) == []
  1337. assert service._brief_outstanding is False
  1338. def test_brief_suppressed_after_reset():
  1339. service, emitted, _, _, _ = _triage_service(brief_fn=lambda prompt: "RAW")
  1340. service.toggle_chat()
  1341. service._pending_brief = ["Unknown: hello"]
  1342. service._reset_triage()
  1343. emitted.clear()
  1344. service._run_brief_update(0, _escalated_payload(), service._now(), 0)
  1345. assert _triage_events(emitted) == []
  1346. assert service.last_triage is None
  1347. assert service._pending_brief == ["Unknown: hello"]
  1348. def test_brief_works_after_leaving_chat():
  1349. seen = []
  1350. service, emitted, _, _, _ = _triage_service(
  1351. brief_fn=lambda prompt: seen.append(prompt) or "RAW"
  1352. )
  1353. service.toggle_chat()
  1354. assert service.modes.exit_chat() is True
  1355. token = _seed_brief(service, "Unknown: hello")
  1356. signals = _escalated_payload()
  1357. service._last_brief_signals = signals
  1358. service._run_brief_update(0, signals, service._now(), token)
  1359. assert seen != []
  1360. assert [e.get("stage") for e in _triage_events(emitted)] == ["content"]
  1361. assert service.last_triage is not None
  1362. assert service.last_triage["brief"] == "RAW"
  1363. def test_deltas_arriving_during_call_survive_and_flush():
  1364. """Deltas appended mid-call are not consumed; the finally flushes them."""
  1365. service, emitted, clock, _, _ = _triage_service()
  1366. service.toggle_chat()
  1367. signals = _escalated_payload()
  1368. service._last_brief_signals = signals
  1369. token = _seed_brief(service, "Unknown: first")
  1370. appended: list[int] = []
  1371. def _brief(prompt, on_io=None, finalize=None):
  1372. if not appended:
  1373. appended.append(1)
  1374. service.handle_event(_transcript(clock, "during call"))
  1375. return _brief_reply("first brief")
  1376. service._brief_fn = _brief
  1377. service._run_brief_update(0, signals, clock(), token)
  1378. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1379. assert len(contents) == 1
  1380. assert contents[0]["triage"]["brief"] == "first brief"
  1381. flushed = [j for j in service._inference._queue if j.name == "brief-update"]
  1382. assert len(flushed) == 1
  1383. assert service._brief_outstanding is True
  1384. assert service._pending_brief == ["Unknown: during call"]
  1385. flushed[0].run()
  1386. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1387. assert len(contents) == 2
  1388. assert contents[1]["triage"]["brief"] == "first brief"
  1389. assert service._pending_brief == []
  1390. assert service._brief_outstanding is False
  1391. def test_in_flight_brief_after_clear_emits_and_writes_nothing():
  1392. """Stale in-flight job: no event, no file write, no requeue, flag untouched."""
  1393. brief_file = FakeBriefFile()
  1394. service, emitted, clock, _, lane = _triage_service(
  1395. brief_fn=lambda prompt: "RAW", brief_file=brief_file
  1396. )
  1397. service.toggle_chat()
  1398. service.handle_event(_transcript(clock, "hello"))
  1399. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1400. service.tick()
  1401. _, _, _, gen = lane.submits[0]
  1402. service._on_triage_result(gen, _escalated_payload())
  1403. job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  1404. assert service._brief_outstanding is True
  1405. service.clear()
  1406. assert service._brief_outstanding is False
  1407. token_after_clear = service._brief_token
  1408. emitted.clear()
  1409. job.run()
  1410. assert [e for e in _triage_events(emitted)] == []
  1411. assert brief_file.writes == []
  1412. assert service._pending_brief == []
  1413. assert service.last_triage is None
  1414. assert service._brief_token == token_after_clear
  1415. assert service._brief_outstanding is False
  1416. def test_clear_resets_triage():
  1417. resets: list[int] = []
  1418. chats: list[int] = []
  1419. brief_file = FakeBriefFile()
  1420. service, emitted, clock, _, lane = _triage_service(
  1421. triage=FakeTriage(escalate=True),
  1422. brief_fn=lambda prompt: "RAW",
  1423. brief_file=brief_file,
  1424. brief_reset=lambda: resets.append(1),
  1425. chat_reset=lambda: chats.append(1),
  1426. )
  1427. service.toggle_chat()
  1428. service.handle_event(_transcript(clock, "churn"))
  1429. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1430. service.tick()
  1431. _, _, _, gen = lane.submits[0]
  1432. service._on_triage_result(gen, _escalated_payload())
  1433. assert service._inference.depth() == 2
  1434. assert service._pending_brief == ["Unknown: churn"]
  1435. assert service._brief_outstanding is True
  1436. token_before = service._brief_token
  1437. assert service.clear() is True
  1438. assert service._triage_generation == 1
  1439. assert service.last_triage is None
  1440. assert service._last_triage_window == ()
  1441. assert service.triage_history == []
  1442. assert service._pending_brief == []
  1443. assert service._brief_token == token_before + 1
  1444. assert service._brief_outstanding is False
  1445. assert service._last_brief_signals is None
  1446. assert resets == [1]
  1447. assert brief_file.clears == 1
  1448. assert chats == [1]
  1449. assert lane.drops == 1
  1450. assert service._inference.depth() == 0
  1451. assert list(service._inference._queue) == []
  1452. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1453. assert len(cleared) == 1
  1454. assert cleared[0]["triage"] is None
  1455. assert {"type": "cleared"} in emitted
  1456. signals_before = len(
  1457. [e for e in _triage_events(emitted) if e.get("stage") == "signals"]
  1458. )
  1459. service._on_triage_result(gen, _escalated_payload())
  1460. assert service.last_triage is None
  1461. assert (
  1462. len([e for e in _triage_events(emitted) if e.get("stage") == "signals"])
  1463. == signals_before
  1464. )
  1465. submits_before = len(lane.submits)
  1466. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1467. service.tick()
  1468. assert len(lane.submits) == submits_before
  1469. def test_exit_chat_does_not_reset_triage():
  1470. chats: list[int] = []
  1471. bresets: list[int] = []
  1472. brief_file = FakeBriefFile()
  1473. service, emitted, clock, _, lane = _triage_service(
  1474. brief_fn=lambda prompt: "RAW",
  1475. chat_reset=lambda: chats.append(1),
  1476. brief_reset=lambda: bresets.append(1),
  1477. brief_file=brief_file,
  1478. )
  1479. service.toggle_chat()
  1480. service.handle_event(_transcript(clock, "hello"))
  1481. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1482. service.tick()
  1483. assert len(lane.submits) == 1
  1484. _, _, _, gen = lane.submits[0]
  1485. service._on_triage_result(gen, _escalated_payload())
  1486. assert service.last_triage is not None
  1487. assert service._triage_generation == 0
  1488. assert service.toggle_chat() is True
  1489. assert service.modes.mode.value == "idle"
  1490. assert service._triage_generation == 0
  1491. assert service.last_triage is not None
  1492. assert service._last_triage_window == ("hello",)
  1493. assert service.triage_history == ["hello"]
  1494. assert service._pending_brief == ["Unknown: hello"]
  1495. assert service._last_brief_signals is not None
  1496. assert service._brief_outstanding is True
  1497. assert lane.drops == 0
  1498. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1499. assert cleared == []
  1500. assert chats == []
  1501. assert bresets == []
  1502. assert brief_file.clears == 0
  1503. def test_toggle_chat_enter_does_not_touch_triage():
  1504. """Entering chat resets only the chat buffers; the brief engine is independent."""
  1505. bresets: list[int] = []
  1506. brief_file = FakeBriefFile()
  1507. service, emitted, clock, _, lane = _triage_service(
  1508. brief_fn=lambda prompt: "RAW",
  1509. brief_reset=lambda: bresets.append(1),
  1510. brief_file=brief_file,
  1511. )
  1512. service.handle_event(_transcript(clock, "hello"))
  1513. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1514. service.tick()
  1515. _, _, _, gen = lane.submits[0]
  1516. service._on_triage_result(gen, _escalated_payload())
  1517. triage_before = service.last_triage
  1518. pending_before = list(service._pending_brief)
  1519. assert service.toggle_chat() is True
  1520. assert service.modes.mode.value == "chat"
  1521. assert service._triage_generation == 0
  1522. assert service.last_triage is triage_before
  1523. assert service._last_triage_window == ("hello",)
  1524. assert service.triage_history == ["hello"]
  1525. assert service._pending_brief == pending_before
  1526. assert service._last_brief_signals is not None
  1527. assert service._brief_outstanding is True
  1528. assert lane.drops == 0
  1529. assert bresets == []
  1530. assert brief_file.clears == 0
  1531. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1532. assert cleared == []
  1533. def test_clear_without_triage_emits_no_triage_event():
  1534. service, emitted, _ = _service()
  1535. brief_file = FakeBriefFile()
  1536. bresets: list[int] = []
  1537. service._brief_file = brief_file
  1538. service._brief_reset = lambda: bresets.append(1)
  1539. service.clear()
  1540. assert {"type": "cleared"} in emitted
  1541. assert _triage_events(emitted) == []
  1542. assert brief_file.clears == 1
  1543. assert bresets == []
  1544. def test_stale_brief_after_clear_does_not_clear_newer_dispatch_flag():
  1545. """A stale in-flight job must not clear a newer dispatch's outstanding flag."""
  1546. brief_file = FakeBriefFile()
  1547. service, emitted, clock, _, lane = _triage_service(
  1548. brief_fn=lambda prompt: "RAW", 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. _, _, _, gen = lane.submits[0]
  1555. service._on_triage_result(gen, _escalated_payload())
  1556. stale_job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  1557. assert service._brief_outstanding is True
  1558. service.clear()
  1559. # A fresh turn dispatches a new brief AFTER clear bumped the token.
  1560. service.handle_event(_transcript(clock, "fresh"))
  1561. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1562. service.tick()
  1563. _, _, _, gen2 = lane.submits[1]
  1564. service._on_triage_result(gen2, _escalated_payload())
  1565. assert service._brief_outstanding is True
  1566. newer_token = service._brief_token
  1567. pending_before = list(service._pending_brief)
  1568. triage_before = service.last_triage
  1569. emitted.clear()
  1570. stale_job.run()
  1571. assert _triage_events(emitted) == []
  1572. assert brief_file.writes == []
  1573. assert service._pending_brief == pending_before
  1574. assert service._brief_outstanding is True
  1575. assert service._brief_token == newer_token
  1576. assert service.last_triage is triage_before
  1577. def test_clear_resets_chat_latch():
  1578. service, emitted, _ = _service()
  1579. service.toggle_chat()
  1580. service._chat_queue_full = True
  1581. service._pending_chat = ["stuck text"]
  1582. service.clear()
  1583. assert service._chat_queue_full is False
  1584. assert service._pending_chat == []
  1585. assert {"type": "cleared"} in emitted
  1586. def test_triage_end_to_end_signals_content_cleared():
  1587. """Two settled turns collapse into ONE brief job (single-outstanding guard).
  1588. The second `_on_triage_result` is refused as `"outstanding"`, so the queue
  1589. holds one `brief-update` that consumes both deltas and emits one `content`.
  1590. The panel keeps the FIRST turn's signals until the next settled turn (lag).
  1591. """
  1592. prompts: list[str] = []
  1593. service, emitted, clock, triage, lane = _triage_service(
  1594. brief_fn=lambda prompt: prompts.append(prompt) or "the brief"
  1595. )
  1596. service.toggle_chat()
  1597. service.handle_event(_transcript(clock, "first thought"))
  1598. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1599. service.tick()
  1600. assert len(lane.submits) == 1
  1601. _, _, _, gen0 = lane.submits[0]
  1602. service._on_triage_result(
  1603. gen0, {"order": ["urgency"], "items": {}, "escalated": False, "input_tokens": 1}
  1604. )
  1605. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals"]
  1606. assert prompts == []
  1607. assert service.last_triage is not None
  1608. assert service.last_triage["escalated"] is False
  1609. assert [j.name for j in service._inference._queue] == ["chat", "brief-update"]
  1610. triage._escalate = True
  1611. service.handle_event(_transcript(clock, "we might churn"))
  1612. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1613. service.tick()
  1614. assert len(lane.submits) == 2
  1615. _, _, _, gen1 = lane.submits[1]
  1616. service._on_triage_result(gen1, _escalated_payload())
  1617. assert [e.get("stage") for e in _triage_events(emitted)] == ["signals", "signals"]
  1618. assert [j.name for j in service._inference._queue] == [
  1619. "chat",
  1620. "brief-update",
  1621. "chat",
  1622. ]
  1623. for job in [j for j in service._inference._queue if j.name == "brief-update"]:
  1624. job.run()
  1625. contents = [e for e in _triage_events(emitted) if e.get("stage") == "content"]
  1626. assert len(contents) == 1
  1627. assert len(prompts) == 1
  1628. deltas, _ = triage.prompt_calls[0]
  1629. assert deltas == ("Unknown: first thought", "Unknown: we might churn")
  1630. assert contents[0]["triage"]["escalated"] is False
  1631. assert contents[0]["triage"]["brief"] == "the brief"
  1632. assert service.last_triage == contents[0]["triage"]
  1633. assert service.status()["triage"] == contents[0]["triage"]
  1634. service.clear()
  1635. cleared = [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  1636. assert len(cleared) == 1
  1637. assert cleared[0]["triage"] is None
  1638. assert service.last_triage is None
  1639. service._on_triage_result(gen1, _escalated_payload())
  1640. assert [e.get("stage") for e in _triage_events(emitted)] == [
  1641. "signals",
  1642. "signals",
  1643. "content",
  1644. "cleared",
  1645. ]
  1646. assert service.toggle_chat() is True
  1647. service.handle_event(_transcript(clock, "idle thought"))
  1648. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1649. service.tick()
  1650. assert len(lane.submits) == 3
  1651. assert lane.submits[2][0].endswith("- idle thought")
  1652. def test_tick_emits_status_on_triage_state_change():
  1653. service, emitted, _, triage, _ = _triage_service(
  1654. triage=FakeTriage(state_pair=("loading", "warming up"))
  1655. )
  1656. service.tick()
  1657. statuses = [e for e in emitted if e["type"] == "status"]
  1658. assert len(statuses) == 1
  1659. assert statuses[0]["triage_state"] == "loading"
  1660. assert statuses[0]["triage_reason"] == "warming up"
  1661. assert "triage" in statuses[0]
  1662. triage._state_pair = ("ok", "")
  1663. service.tick()
  1664. statuses = [e for e in emitted if e["type"] == "status"]
  1665. assert len(statuses) == 2
  1666. assert statuses[1]["triage_state"] == "ok"
  1667. assert statuses[1]["triage_reason"] == ""
  1668. assert "triage" in statuses[1]
  1669. service.tick()
  1670. assert len([e for e in emitted if e["type"] == "status"]) == 2
  1671. def test_no_status_emitted_when_triage_disabled():
  1672. service, emitted, _ = _service()
  1673. service.tick()
  1674. service.tick()
  1675. assert [e for e in emitted if e["type"] == "status"] == []
  1676. def test_tick_heals_triage_panel_after_lane_ready():
  1677. import time as _time
  1678. from pathlib import Path
  1679. from stt_daemon.assistant.laya_lane import LayaLane
  1680. from stt_daemon.assistant.triage import TriageAnalyzer
  1681. from stt_daemon.tui.state import ViewState
  1682. class _FakeRouter:
  1683. def preload(self, names):
  1684. return self
  1685. def predict(self, state, questions, model=None, lang=None, **kwargs):
  1686. if "_warmup" in questions:
  1687. return {"model": model, "answers": {}, "usage": {}, "routing": {}}
  1688. return {
  1689. "model": model,
  1690. "answers": {
  1691. "frustration": {
  1692. "noul": 0.1,
  1693. "confidence": 0.9,
  1694. "answer_confidence": 0.9,
  1695. },
  1696. "urgency": {
  1697. "choice": "low",
  1698. "confidence": 0.9,
  1699. "answer_confidence": 0.9,
  1700. },
  1701. },
  1702. "usage": {"input_tokens": 5},
  1703. }
  1704. def _factory(**kwargs):
  1705. return _FakeRouter()
  1706. cfg = Config()
  1707. cfg.assistant.turn_gap_s = 0.5
  1708. cfg.triage.questions_file = str(
  1709. Path(__file__).resolve().parent.parent / "questions.example.json"
  1710. )
  1711. clock = Clock()
  1712. emitted: list[dict] = []
  1713. def emit(type_, **fields):
  1714. emitted.append({"type": type_, **fields})
  1715. analyzer = TriageAnalyzer(cfg.triage)
  1716. assert analyzer.questions, "example questions must load"
  1717. holder: dict = {}
  1718. lane = LayaLane(
  1719. cfg.triage,
  1720. agent_factory=_factory,
  1721. on_result=lambda gen, payload: holder["svc"]._on_triage_result(gen, payload),
  1722. on_state=lambda state, reason: analyzer.refresh(state, reason),
  1723. )
  1724. service = AssistantService(
  1725. cfg,
  1726. FakeSubscriber(),
  1727. emit=emit,
  1728. time_fn=clock,
  1729. triage=analyzer,
  1730. laya_lane=lane,
  1731. chat_fn=_ok_reply,
  1732. )
  1733. holder["svc"] = service
  1734. try:
  1735. lane.start()
  1736. deadline = _time.monotonic() + 5.0
  1737. while not lane.ready() and _time.monotonic() < deadline:
  1738. _time.sleep(0.005)
  1739. assert lane.ready(), "lane did not become ready"
  1740. assert analyzer.state() == ("ok", "")
  1741. service.tick()
  1742. statuses = [e for e in emitted if e["type"] == "status"]
  1743. assert statuses, "tick emitted no status after lane ready"
  1744. assert statuses[-1]["triage_state"] == "ok"
  1745. service.handle_event(_transcript(clock, "hello"))
  1746. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1747. service.tick()
  1748. deadline = _time.monotonic() + 5.0
  1749. while service.last_triage is None and _time.monotonic() < deadline:
  1750. _time.sleep(0.005)
  1751. assert service.last_triage is not None, "no triage signals from lane"
  1752. view = ViewState()
  1753. for event in emitted:
  1754. view.ingest(event)
  1755. assert view.triage_state == "ok"
  1756. lines = view.triage_lines()
  1757. assert lines
  1758. assert not any("unavailable" in line for line in lines)
  1759. assert not any("loading" in line for line in lines)
  1760. finally:
  1761. lane.stop()
  1762. def _dump_brief(prompt, on_io=None, finalize=None):
  1763. """Brief fake that mimics LlmClient.chat's on_io phases."""
  1764. if on_io is not None:
  1765. on_io({"phase": "request", "url": "http://llm/v1", "body": {"prompt": prompt}})
  1766. on_io(
  1767. {
  1768. "phase": "response",
  1769. "payload": {"content": "parsed"},
  1770. "error": None,
  1771. "duration_ms": 250.0,
  1772. }
  1773. )
  1774. return _brief_reply("BRIEF")
  1775. def _run_triage_flow(service, emitted, clock, lane, payload):
  1776. service.toggle_chat()
  1777. service.handle_event(_transcript(clock, "hello"))
  1778. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1779. service.tick()
  1780. _, _, _, gen = lane.submits[0]
  1781. service._on_triage_result(gen, payload)
  1782. return gen
  1783. def _dump_files(tmp_path):
  1784. return list(tmp_path.rglob("*.json"))
  1785. def test_dump_writes_full_laya_and_llm_io(tmp_path):
  1786. service, emitted, clock, triage, lane = _triage_service(brief_fn=_dump_brief)
  1787. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1788. service.cfg.triage.dump_io = True
  1789. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1790. jobs = [j for j in service._inference._queue if j.name == "brief-update"]
  1791. assert len(jobs) == 1
  1792. jobs[0].run()
  1793. files = _dump_files(tmp_path)
  1794. assert len(files) == 1
  1795. assert re.fullmatch(r"\d{6}-\d{3}\.json", files[0].name)
  1796. data = json.loads(files[0].read_text(encoding="utf-8"))
  1797. assert data["ts"] == clock()
  1798. assert isinstance(data["iso"], str) and data["iso"]
  1799. assert data["laya"]["request"]["state"] == "- hello"
  1800. assert data["laya"]["request"]["questions"] == triage.questions
  1801. assert data["laya"]["request"]["model"] == "multilingual"
  1802. assert data["laya"]["request"]["max_len"] == 1024
  1803. assert data["laya"]["response"] == _escalated_payload()
  1804. assert data["laya"]["error"] is None
  1805. assert data["llm"]["request"]["url"] == "http://llm/v1"
  1806. assert data["llm"]["request"]["body"] == {
  1807. "prompt": "CURRENT LEDGER:\n(none yet)\n\nCURRENT BRIEF:\n(none yet)\n\n"
  1808. "NEW UTTERANCES:\n- Unknown: hello"
  1809. }
  1810. assert data["llm"]["response"] == {"content": "parsed"}
  1811. assert data["llm"]["error"] is None
  1812. assert data["llm"]["duration_ms"] == 250.0
  1813. assert service._triage_dump._pending == {}
  1814. def test_dump_laya_error_writes_laya_only_and_keeps_deltas(tmp_path):
  1815. """A Laya predict failure finalizes the Laya half with `llm: null`.
  1816. Nothing is dispatched, so no LLM half is ever recorded and the buffered
  1817. deltas survive for the next settled turn.
  1818. """
  1819. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1820. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1821. service.cfg.triage.dump_io = True
  1822. _run_triage_flow(service, emitted, clock, lane, None)
  1823. assert [j.name for j in service._inference._queue] == ["chat"]
  1824. assert service._pending_brief == ["Unknown: hello"]
  1825. files = _dump_files(tmp_path)
  1826. assert len(files) == 1
  1827. data = json.loads(files[0].read_text(encoding="utf-8"))
  1828. assert data["laya"]["response"] is None
  1829. assert data["laya"]["error"] == "predict failed"
  1830. assert data["llm"] is None
  1831. assert service._triage_dump._pending == {}
  1832. def test_dump_without_brief_has_laya_only(tmp_path):
  1833. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1834. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1835. service.cfg.triage.dump_io = True
  1836. service.cfg.triage.extract_enabled = False
  1837. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1838. files = _dump_files(tmp_path)
  1839. assert len(files) == 1
  1840. data = json.loads(files[0].read_text(encoding="utf-8"))
  1841. assert data["laya"]["response"] == _escalated_payload()
  1842. assert data["llm"] is None
  1843. assert service._triage_dump._pending == {}
  1844. def test_dump_queue_full_has_laya_only(tmp_path):
  1845. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1846. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1847. service.cfg.triage.dump_io = True
  1848. service._inference = InferenceQueue(size=2)
  1849. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1850. assert service._inference.depth() == 1
  1851. assert [j.name for j in service._inference._queue] == ["chat"]
  1852. assert service._pending_brief == ["Unknown: hello"]
  1853. files = _dump_files(tmp_path)
  1854. assert len(files) == 1
  1855. data = json.loads(files[0].read_text(encoding="utf-8"))
  1856. assert data["laya"]["response"] == _escalated_payload()
  1857. assert data["llm"] is None
  1858. assert service._triage_dump._pending == {}
  1859. def test_dump_disabled_writes_nothing(tmp_path):
  1860. service, emitted, clock, _, lane = _triage_service(
  1861. brief_fn=lambda prompt: "RAW"
  1862. )
  1863. assert service._triage_dump.enabled is False
  1864. service._triage_dump = TriageDumper(False, directory=str(tmp_path))
  1865. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1866. jobs = [j for j in service._inference._queue if j.name == "brief-update"]
  1867. assert len(jobs) == 1
  1868. jobs[0].run()
  1869. assert _dump_files(tmp_path) == []
  1870. def test_dump_stale_result_discards_without_file(tmp_path):
  1871. service, emitted, clock, triage, lane = _triage_service(brief_fn=_dump_brief)
  1872. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1873. service.cfg.triage.dump_io = True
  1874. service.toggle_chat()
  1875. service.handle_event(_transcript(clock, "hello"))
  1876. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1877. service.tick()
  1878. _, _, _, gen = lane.submits[0]
  1879. real_normalize = triage.normalize
  1880. def stale_normalize(payload):
  1881. # Simulate a clear landing between the first check/begin and the re-check.
  1882. service._triage_generation += 1
  1883. return real_normalize(payload)
  1884. triage.normalize = stale_normalize
  1885. service._on_triage_result(gen, _escalated_payload())
  1886. assert _dump_files(tmp_path) == []
  1887. assert service._triage_dump._pending == {}
  1888. def test_dump_stale_brief_discards_without_file(tmp_path):
  1889. service, _, clock, _, _ = _triage_service(brief_fn=_dump_brief)
  1890. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1891. service.cfg.triage.dump_io = True
  1892. ts = clock()
  1893. service._triage_dump.begin(ts, laya={"request": {}, "response": {}, "error": None})
  1894. service._pending_brief = ["Unknown: hello"]
  1895. service._triage_generation += 1 # stale before the brief even starts
  1896. service._run_brief_update(0, _escalated_payload(), ts, service._brief_token)
  1897. assert _dump_files(tmp_path) == []
  1898. assert service._triage_dump._pending == {}
  1899. assert service._pending_brief == ["Unknown: hello"]
  1900. def test_dump_brief_cleared_mid_flight_writes_nothing(tmp_path):
  1901. service, _, clock, _, _ = _triage_service(brief_fn=_dump_brief)
  1902. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1903. service.cfg.triage.dump_io = True
  1904. def stale_brief(prompt, on_io=None, finalize=None):
  1905. # Simulate a clear landing mid-LLM-call (after the request phase fired).
  1906. service._triage_generation += 1
  1907. if on_io is not None:
  1908. on_io({"phase": "request", "url": "http://llm/v1", "body": {}})
  1909. on_io({"phase": "response", "payload": {"c": 1}, "error": None})
  1910. return _brief_reply("RAW")
  1911. service._brief_fn = stale_brief
  1912. service._pending_brief = ["Unknown: hello"]
  1913. service._brief_outstanding = True
  1914. token = service._brief_token
  1915. ts = clock()
  1916. service._triage_dump.begin(ts, laya={"request": {}, "response": {}, "error": None})
  1917. service._run_brief_update(0, _escalated_payload(), ts, token)
  1918. assert _dump_files(tmp_path) == []
  1919. assert service._triage_dump._pending == {}
  1920. assert service._pending_brief == []
  1921. assert service._brief_outstanding is False
  1922. def test_dump_gated_out_writes_laya_only_and_no_pending(tmp_path):
  1923. """A turn the gate leaves out writes `llm: null` and leaks no pending record."""
  1924. service, emitted, clock, triage, lane = _triage_service(
  1925. triage=FakeTriage(escalate=True, gate=False), brief_fn=_dump_brief
  1926. )
  1927. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1928. service.cfg.triage.dump_io = True
  1929. _run_triage_flow(service, emitted, clock, lane, _escalated_payload())
  1930. assert [j.name for j in service._inference._queue] == ["chat"]
  1931. assert len(triage.gate_calls) == 1
  1932. files = _dump_files(tmp_path)
  1933. assert len(files) == 1
  1934. data = json.loads(files[0].read_text(encoding="utf-8"))
  1935. assert data["laya"]["response"] == _escalated_payload()
  1936. assert data["llm"] is None
  1937. assert service._triage_dump._pending == {}
  1938. def test_dump_coalesced_update_writes_laya_only_then_llm_only_flush(tmp_path):
  1939. """A coalesced turn writes a Laya-only record; the later flush a separate llm-only one."""
  1940. service, emitted, clock, _, lane = _triage_service(brief_fn=_dump_brief)
  1941. service._triage_dump = TriageDumper(True, directory=str(tmp_path))
  1942. service.cfg.triage.dump_io = True
  1943. service.toggle_chat()
  1944. service.handle_event(_transcript(clock, "first"))
  1945. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1946. service.tick()
  1947. _, _, _, gen0 = lane.submits[0]
  1948. service._on_triage_result(gen0, _escalated_payload())
  1949. service.handle_event(_transcript(clock, "second"))
  1950. clock.advance(service.cfg.assistant.turn_gap_s + 0.1)
  1951. service.tick()
  1952. _, _, _, gen1 = lane.submits[1]
  1953. service._on_triage_result(gen1, _escalated_payload())
  1954. assert service._brief_outstanding is True
  1955. appended: list[int] = []
  1956. def _brief(prompt, on_io=None, finalize=None):
  1957. if not appended:
  1958. appended.append(1)
  1959. service.handle_event(_transcript(clock, "during call"))
  1960. if on_io is not None:
  1961. on_io({"phase": "request", "url": "http://llm/v1", "body": {"prompt": prompt}})
  1962. on_io({"phase": "response", "payload": {"content": "parsed"}, "error": None})
  1963. return _brief_reply("BRIEF")
  1964. service._brief_fn = _brief
  1965. clock.advance(1.0)
  1966. job = [j for j in service._inference._queue if j.name == "brief-update"][0]
  1967. service._inference._queue.remove(job)
  1968. job.run()
  1969. flushed = [j for j in service._inference._queue if j.name == "brief-update"]
  1970. assert len(flushed) == 1
  1971. flushed[0].run()
  1972. assert service._triage_dump._pending == {}
  1973. records = [
  1974. json.loads(path.read_text(encoding="utf-8")) for path in _dump_files(tmp_path)
  1975. ]
  1976. assert len(records) == 3
  1977. coalesced = [r for r in records if r["laya"] is not None and r["llm"] is None]
  1978. flushed_records = [r for r in records if r["laya"] is None and r["llm"] is not None]
  1979. both = [r for r in records if r["laya"] is not None and r["llm"] is not None]
  1980. assert len(coalesced) == 1
  1981. assert len(flushed_records) == 1
  1982. assert len(both) == 1
  1983. assert coalesced[0]["ts"] != flushed_records[0]["ts"]
  1984. assert flushed_records[0]["llm"]["request"]["body"] == {
  1985. "prompt": "CURRENT LEDGER:\n(none yet)\n\nCURRENT BRIEF:\nBRIEF\n\n"
  1986. "NEW UTTERANCES:\n- Unknown: during call"
  1987. }
  1988. def test_chat_client_failure_emits_llm_error_and_service_survives():
  1989. """Adversarial (a): a chat client raising on `chat` surfaces as an `llm_error`
  1990. event and the service keeps functioning afterwards."""
  1991. from stt_daemon.assistant.llm import ConversationEngine, LlmError
  1992. class BoomClient:
  1993. def chat(self, messages, max_tokens=None, timeout_s=None, **kwargs):
  1994. raise LlmError("boom")
  1995. engine = ConversationEngine(Config().llm, BoomClient(), system_prompt="SYS")
  1996. service, emitted, clock = _service()
  1997. service.chat_fn = engine.send
  1998. assert service.toggle_chat() is True
  1999. service._run_chat("hello")
  2000. errors = [e for e in emitted if e["type"] == "error" and e.get("code") == "llm_error"]
  2001. assert errors and "boom" in errors[-1]["message"]
  2002. assert service.toggle_chat() is True
  2003. assert service.modes.mode.value == "idle"
  2004. # ---------------------------------------------------------------------------
  2005. # Conversation session adoption (SESSION/TRANSCRIPT session_dir).
  2006. # ---------------------------------------------------------------------------
  2007. def _session_event(clock, session_dir):
  2008. event = _ev(events.SESSION, clock)
  2009. event["session_id"] = session_dir.rsplit("/", 1)[-1]
  2010. event["session_dir"] = session_dir
  2011. return event
  2012. def _ring(service, clock, *items):
  2013. """Feed a daemon-seeded replay window: status -> items -> ring_end."""
  2014. service.handle_event(_ev(events.STATUS, clock))
  2015. for item in items:
  2016. service.handle_event(item)
  2017. service.handle_event(_ev(events.RING_END, clock))
  2018. def _cleared(emitted):
  2019. return [e for e in _triage_events(emitted) if e.get("stage") == "cleared"]
  2020. def test_status_includes_session_dir_without_brief_file():
  2021. service, _, clock = _service()
  2022. assert service.status()["session_dir"] is None
  2023. service.handle_event(_session_event(clock, "transcripts/2026-09-30-12-47-29"))
  2024. assert service.status()["session_dir"] == "transcripts/2026-09-30-12-47-29"
  2025. def test_live_session_event_sets_session_and_records_directory():
  2026. brief_file = FakeBriefFile()
  2027. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2028. session = "transcripts/2026-09-30-12-47-29"
  2029. service.handle_event(_session_event(clock, session))
  2030. assert service.status()["session_dir"] == session
  2031. assert brief_file.directories == [session]
  2032. assert service._triage_generation == 0
  2033. def test_live_transcript_session_sets_session_without_wiping_utterance():
  2034. brief_file = FakeBriefFile()
  2035. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2036. session = "transcripts/s1"
  2037. service.handle_event(_transcript(clock, "hello", session_dir=session))
  2038. assert service.status()["session_dir"] == session
  2039. assert brief_file.directories == [session]
  2040. assert service.triage_history == ["hello"]
  2041. assert service._pending_brief == ["Unknown: hello"]
  2042. assert service._triage_generation == 0
  2043. def test_brief_write_lands_under_session_dir(tmp_path):
  2044. from stt_daemon.assistant.brief_file import BriefFile
  2045. brief_file = BriefFile(True)
  2046. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2047. session = str(tmp_path / "transcripts" / "s1")
  2048. service.handle_event(_transcript(clock, "hello", session_dir=session))
  2049. brief_file.write("the brief")
  2050. path = tmp_path / "transcripts" / "s1" / "brief.md"
  2051. assert path.read_text(encoding="utf-8") == "the brief"
  2052. def test_live_session_change_resets_triage_and_keeps_chat():
  2053. brief_file = FakeBriefFile()
  2054. resets: list[int] = []
  2055. service, emitted, clock, _, lane = _triage_service(
  2056. brief_file=brief_file, brief_reset=lambda: resets.append(1)
  2057. )
  2058. service.toggle_chat()
  2059. service.handle_event(_transcript(clock, "hello"))
  2060. service.handle_event(_session_event(clock, "transcripts/a"))
  2061. service.last_triage = {"signals": 1}
  2062. service._pending_brief = ["stale"]
  2063. emitted.clear()
  2064. service.handle_event(_session_event(clock, "transcripts/b"))
  2065. assert service.status()["session_dir"] == "transcripts/b"
  2066. assert service.last_triage is None
  2067. assert service._pending_brief == []
  2068. assert service._triage_generation == 1
  2069. assert service.chat_history == ["Unknown: hello"]
  2070. assert resets == [1]
  2071. assert lane.drops == 1
  2072. assert brief_file.directories == ["transcripts/a", "transcripts/b"]
  2073. assert len(_cleared(emitted)) == 1
  2074. def test_sessionless_transcript_does_not_reset():
  2075. service, _, clock, _, _ = _triage_service()
  2076. service.handle_event(_session_event(clock, "transcripts/a"))
  2077. service.last_triage = {"signals": 1}
  2078. service.handle_event(_transcript(clock, "hello"))
  2079. assert service._triage_generation == 0
  2080. assert service.last_triage == {"signals": 1}
  2081. assert service.status()["session_dir"] == "transcripts/a"
  2082. def test_same_dir_readoption_does_not_reset():
  2083. brief_file = FakeBriefFile()
  2084. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2085. session = "transcripts/a"
  2086. service.handle_event(_session_event(clock, session))
  2087. service.last_triage = {"signals": 1}
  2088. service.handle_event(_transcript(clock, "again", session_dir=session))
  2089. assert service._triage_generation == 0
  2090. assert service.last_triage == {"signals": 1}
  2091. assert brief_file.directories == [session]
  2092. def test_session_change_without_brief_file_does_not_raise():
  2093. service, _, clock, _, _ = _triage_service(brief_file=None)
  2094. service.handle_event(_session_event(clock, "transcripts/a"))
  2095. service.handle_event(_session_event(clock, "transcripts/b"))
  2096. assert service.status()["session_dir"] == "transcripts/b"
  2097. assert service._triage_generation == 1
  2098. def test_replay_single_session_adopts_without_reset():
  2099. brief_file = FakeBriefFile()
  2100. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2101. session = "transcripts/a"
  2102. _ring(service, clock, _transcript(clock, "hello", session_dir=session))
  2103. assert service.status()["session_dir"] == session
  2104. assert brief_file.directories == [session]
  2105. assert service._triage_generation == 0
  2106. assert _cleared(emitted) == []
  2107. assert service.triage_history == ["hello"]
  2108. def test_replay_two_sessions_commits_newest_and_resets_once():
  2109. brief_file = FakeBriefFile()
  2110. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2111. a, b = "transcripts/a", "transcripts/b"
  2112. _ring(
  2113. service,
  2114. clock,
  2115. _session_event(clock, a),
  2116. _transcript(clock, "old turn", session_dir=a),
  2117. _session_event(clock, b),
  2118. _transcript(clock, "new turn", session_dir=b),
  2119. )
  2120. assert service.status()["session_dir"] == b
  2121. assert brief_file.directories == [b]
  2122. assert service._triage_generation == 1
  2123. assert service.last_triage is None
  2124. assert service._pending_brief == []
  2125. assert len(_cleared(emitted)) == 1
  2126. def test_reconnect_tailing_current_session_is_noop():
  2127. brief_file = FakeBriefFile()
  2128. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2129. session = "transcripts/a"
  2130. service.handle_event(_session_event(clock, session))
  2131. service.last_triage = {"signals": 1}
  2132. emitted.clear()
  2133. _ring(service, clock, _transcript(clock, "replayed", session_dir=session))
  2134. assert service.status()["session_dir"] == session
  2135. assert brief_file.directories == [session]
  2136. assert service._triage_generation == 0
  2137. assert service.last_triage == {"signals": 1}
  2138. assert _cleared(emitted) == []
  2139. def test_reconnect_older_to_current_resets_once_without_rewind():
  2140. brief_file = FakeBriefFile()
  2141. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2142. a, b = "transcripts/a", "transcripts/b"
  2143. service.handle_event(_session_event(clock, a))
  2144. service.last_triage = {"signals": 1}
  2145. emitted.clear()
  2146. _ring(
  2147. service,
  2148. clock,
  2149. _session_event(clock, a),
  2150. _transcript(clock, "old", session_dir=a),
  2151. _session_event(clock, b),
  2152. _transcript(clock, "new", session_dir=b),
  2153. )
  2154. assert service.status()["session_dir"] == b
  2155. assert brief_file.directories == [a, b]
  2156. assert service._triage_generation == 1
  2157. assert service.last_triage is None
  2158. assert len(_cleared(emitted)) == 1
  2159. def test_reconnect_only_newer_session_resets_once():
  2160. brief_file = FakeBriefFile()
  2161. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2162. a, b = "transcripts/a", "transcripts/b"
  2163. service.handle_event(_session_event(clock, a))
  2164. service.last_triage = {"signals": 1}
  2165. _ring(service, clock, _transcript(clock, "new", session_dir=b))
  2166. assert service.status()["session_dir"] == b
  2167. assert service._triage_generation == 1
  2168. assert service.last_triage is None
  2169. assert brief_file.directories == [a, b]
  2170. def test_reconnect_span_older_current_clears_buffered_older_turns():
  2171. """A RING_END whose ring spanned an older session resets even when the
  2172. newest session already matches the adopted directory (no re-set_directory)."""
  2173. brief_file = FakeBriefFile()
  2174. service, emitted, clock, _, _ = _triage_service(brief_file=brief_file)
  2175. a, b = "transcripts/a", "transcripts/b"
  2176. _ring(service, clock, _session_event(clock, a))
  2177. assert service._triage_generation == 0
  2178. service.handle_event(_transcript(clock, "old turn", session_dir=a))
  2179. assert service._pending_brief == ["Unknown: old turn"]
  2180. service.handle_event(_session_event(clock, b))
  2181. assert service._triage_generation == 1
  2182. service.handle_event(_transcript(clock, "new turn", session_dir=b))
  2183. assert service._pending_brief == ["Unknown: new turn"]
  2184. emitted.clear()
  2185. _ring(
  2186. service,
  2187. clock,
  2188. _session_event(clock, a),
  2189. _transcript(clock, "old turn", session_dir=a),
  2190. _session_event(clock, b),
  2191. _transcript(clock, "new turn", session_dir=b),
  2192. )
  2193. assert service.status()["session_dir"] == b
  2194. assert brief_file.directories == [a, b]
  2195. assert service._pending_brief == []
  2196. assert service._triage_generation == 2
  2197. assert len(_cleared(emitted)) == 1
  2198. def test_daemon_restart_new_session_resets_once():
  2199. brief_file = FakeBriefFile()
  2200. service, _, clock, _, _ = _triage_service(brief_file=brief_file)
  2201. service.handle_event(_session_event(clock, "transcripts/a"))
  2202. service.last_triage = {"signals": 1}
  2203. service.handle_event(_session_event(clock, "transcripts/b"))
  2204. assert service.status()["session_dir"] == "transcripts/b"
  2205. assert service._triage_generation == 1
  2206. assert service.last_triage is None
  2207. assert brief_file.directories == ["transcripts/a", "transcripts/b"]