+ Wj7Rt^RIHt^RIt^RIt^RIt^RIt^RIt]PPRR4t Rt .ROt ]PPR4;'g ] ^,s]P!R4t^t]P!R]P$4tR R ltR t]P!R 4tRR RlltRtRRlt!RR4tR#)uLocalBrain — the ~0.1s conversational front-line (local LLM via Ollama). Measured reality: a reply through the Claude Code CLI has a ~2s floor per turn (CLI overhead, any model). A small local model answers in ~0.11s to first token. So conversation — greetings, questions, banter, quick recall — goes to a local model here; anything needing an ACTION, tools, code, or live system state is declined with <> and the engine routes that turn to the full Claude brain. Instant chat, full power on demand. Backend: Ollama (already used for the screen watcher). Default model qwen2.5:3b (~2GB) — fits beside a game, and Ollama auto-falls to CPU when the GPU is full (still faster than the CLI). Stateless HTTP per call (full history sent each time), so barge-in just stops reading — no stale-stream to drain. Env: LOCAL_MODEL (default qwen2.5:3b), OLLAMA_URL, FAST_CHAT=0 to disable. ) annotationsN OLLAMA_URLzhttp://127.0.0.1:11434z<> LOCAL_MODELu[.!?…]['\")\]]?\s|\nu[*_`#~|]|^\s*[-•]\s+c V^8dQhRRRR/#)textstrreturn)formats"voice/local_brain.py __annotate__r *s++++cJ\PRV4P4#)) _MD_STRIPsubstrip)rs&r _cleanr*s ==T " ( ( **rueYou are Jarvis, a British butler voice assistant. Ahmed just gave you a task and you're starting it NOW. Reply with ONE very short spoken acknowledgment that NAMES the action, present-continuous, at most 8 words, addressing him as "sir". Plain spoken prose only — no markdown, no quotes, no emoji, no tone tags, exactly one line. Examples: "check my email" -> Checking your emails now, sir. "open whatsapp" -> Pulling up WhatsApp now, sir. "what's on my asana board" -> On the Asana board now, sir. "search the web for flights" -> Looking that up now, sir. Output ONLY the acknowledgment line, nothing else.z^\s*\[[^\]]{1,16}\]\s*c(V^8dQhRRRRRRRR/#)rrrtimeoutfloatmodelz str | Noner r )r s"r r r Es(##3##Z#:#rc ZT;'g\pRTRRRR\/RRRT;'gRP4R,/.R R R R R RRR^ //pVPR4'd RV9dR VR&\P !V4P 4p\PP\ R2VRR/R7p\PPWaR7;_uu_4p\P!VP44pRRR4XPR4;'g/PR4;'gRP4p T 'gR#\ P#RT P%4^,4p \'T 4P4PR4PR4P4p T 'd*\(T 9g\+T P-44^ 8dR#T # +'giEL;i \dR#i;i)a)A fast, task-specific spoken ack from the local model (e.g. "Pulling up WhatsApp now, sir."). Non-streaming with a tiny generation budget (~0.3s); returns None on any error / slow Ollama / missing model / odd output so the caller can fall back to a generic canned line. Safe to call from a thread.rmessagesrolesystemcontentuserr:NiNstreamF keep_alive20moptions temperatureg333333? num_predictqwen3instructthink /api/chat Content-Typeapplication/jsondataheadersrNmessage"')MODEL _ACK_PROMPTr startswithjsondumpsencodeurllibrequestRequestOLLAMAurlopenloadsread Exceptionget_ACK_TAGr splitlinesrHANDOFFlensplit) rrrmdlpayloadbodyreqrolines &&& r ack_linerMEs ..5C Xy+ 6 VY(:(:(> Same for a TERSE FOLLOW-UP — "where", "why", "and?", "it didn't work", "it's not there" — right after something was done or shown: he means THAT thing. Never ask "where what, sir?" — if the recent turns don't tell you confidently what he's pointing at, reply <> so the layer that did the thing answers. ALSO CRUCIAL — you have NO long-term memory of past conversations. Anything that asks you to RECALL — "do you remember", "do I have", "when is my", "what did I say about", "did I mention", "what's my", any question about his meetings, appointments, plans, people, notes, or anything he told you before — you CANNOT answer; emit <> so your full self can look it up. Never say "no" or "I don't remember" — hand off. CRUCIAL: questions about YOURSELF also need <> — "what model are you", "are you local or the cloud", "which mic/model are you using", "is the camera on", "how are you built", "what can you do", "what are your limits". You do NOT know these; your full self reads them from a live status file. Do not guess or say you're "an entity running on nothing" — just emit <>. No apology, no explanation — just <> and it is routed to your full self. When unsure whether you can truly answer, prefer <>.cV^8dQhRR/#)rr boolr )r s"r r r s4rca\PPRR4R8XdR#\PP \ R2^R7;_uu_4p\P!VP44pRRR4XPR.4Uu.uFq"PR R 4NK pp\PPR 4'd5\;QJdR V4F 'gK R # R#!R V44#\F&o\V3RlV4R4pV'gK#Vs R # R# +'giL;iuupi \dR#i;i)zTrue if Ollama is up and a usable model is present. When LOCAL_MODEL isn't forced, picks the best installed model from _PREFERRED (module global MODEL is updated so LocalBrain uses the pick). FAST_CHAT10Fz /api/tagsr.Nmodelsnamerrc3"TFApV\8H;'g,VP\PR4^,4xKC R#5i):N)r2r4rE).0ns& r available..s=' %1EzFFQ\\%++c2B12E%FF %s A 4A Tc3f<"TF&qS8XgVPS4'gK"VxK( R#5iN)r4)rXrYwants& r rZr[s(/5aI<<-5s1 1)osenvironr@r8r9r<r;r5r=r>any _PREFERREDnextr2r?)rJtagsmnameshitr^s @r availablerhs.  zz~~k3'3. ^^ # #vhi$8! # D D::affh'DE,0HHXr,BC,Bqvr",BC ::>>- ( (3' %'33 '3 '3' %'' 'D/5/046Cs E DC s_0E%E?EE51E'E;E>E E$E9E>E E E E)(E)cn]tRt^tRRltRRltRRltRRRlltR R ltR R lt R Rlt Rt R#) LocalBraincV^8dQhRR/#rr Noner )r s"r r LocalBrain.__annotate__s$$$$rc ^\Vn.Vn\P!4VnR#r])r2r _messagesasyncioLock_lockselfs&r __init__LocalBrain.__init__s %'\\^ rcV^8dQhRR/#rlr )r s"r r rnsTrc "VPR4RjxL pK LDTPP4R#5i)u$(warm up — reply with just: ready)N)replyrpclear)ru_s& r startLocalBrain.starts9zz"HI  !  I s?" "?"?cV^8dQhRR/#rlr )r s"r r rns  D rc "R#5ir]r rts&r stopLocalBrain.stops sc V^8dQhRRRR/#)rrrr rmr )r s"r r rns!!u!t!rc "VP'd>VPR,R,R8XdVPP4R#R#R#5i)r assistantN)rppop)rurs&&r interruptLocalBrain.interrupts= >>>dnnR08KG NN   H>sAAcV^8dQhRR/#)rrz list[dict]r )r s"r r rns 7 7Z 7rc  RVPRVRRRRRRR//pVPPR 4'dR VP9dR VR &\P!V4P 4p\ P P\ R 2VRR/R7p\ P PV^?D NN ! !669i"H IB#))+D--dD4E4EtLLD% !%!5!5dDMM!JJD! JJt,UU9-3388BGEu  "d?s4yCL1D s#(*0-$7$)!&uuV}}JJL99;Dt+Tl"d"tzz||%%v{It&LMg::MK %! !:!YX!JJL Y:::sUNLNBM="L#M=( ML MML.MM"AM,L.< NMNM,C M=M?MM='9M=!M= N M; NM=M L+&M*L++M. L=9M=<L==M=N MM=MM=M8M'&M8' M5 2M84M5 5M88M=;N=N N N N  NcV^8dQhRR/#rr )r s"r r rns $ $# $rc "RpVPV4RjxL pV'd,Rp\V9gVP4'g R5xR#RV35xKFLADBR#5i)zfStream ('say', sentence); if the first content is the handoff token, yield ('handoff', None) and stop.TNFsay)handoffN)rzrCr)rurfirstsentences&& r classify_and_replyLocalBrain.classify_and_replys\"jj3 $ $(h&hnn.>.>++(# #  $3s+A!AAA,A! A!AA!)rsrprN)g) __name__ __module__ __qualname____firstlineno__rvr}rrrrzr__static_attributes__r rr rjrjs,$   ! 75Nn $ $rrj)zqwen3:4b-instruct-2507zqwen3:4b-instructzqwen3:4bz qwen2.5:3b)g?N)__doc__ __future__rrqr5r_reurllib.requestr8r`r@r;rCrbr2compilerr MULTILINErrr3rArMrrhrjr rr rs"#   &> ?    }%66A 45  JJ0",, ? + 9  ::/ 0#L3 =l4e$e$r