ikUdZddlZddlZddlZddlZddlZddlmZmZm Z m Z m Z ddl m Z ddlmZmZddlZej$ej&ej(eZ ddlmZ ddlmZdd lmZd Z ddl"m#Z#e#de#dGdde Z$GddZ%e%dddddde$jLgddd e%dd d!d"d#de$jLgd$d%d& e%d'd d(d"d#de$jLgdd%d& e%d)ddd*d+d,e$jNd-d.gd/d0 e%d1ddd*d2d,e$jNd-d.gd/d0 e%d3d4d5d6d7d8e$jLd-d.gd/d e%d9d:d;ddd%d e%d?d@dAdBdCd=e$jNgd>d%d e%dDd@dAdBdEd,e$jNgdFd%d e%dGd@dAdBdde$jLgdHdd e%dId@dAdBdJd8e$jLd-d.gd/d e%dKd@dAdBdLdMe$jLgdNd/d e%dOd@dAdBdPdQe$jPgdRd%d e%dSd@dAdBdTdUe$jPgdVd%d dWZ)ee*e%fe+dX<dYdYdYdZdZd[d[dYdYd\d] Z,ee*e*fe+d^<Gd_d`Z-daZ.edbk(re.yy#e$r'ej5dej6d Y;wxYw#e$rd Z ejCd YKwxYw#e$rejCdYQwxYw)cu Genesis Multi-Model Swarm Unified interface for Kimi K2.5, MiniMax M2.5, Gemini 2.0 Flash, and other swarm models. COMMAND CENTRE DOCTRINE: This module enables Claude (orchestrator) to dispatch work to cost-effective agent swarms (Kimi, MiniMax, Gemini) while staying lightweight and responsive. Cost Hierarchy (per MTok): - Gemini 2.0 Flash: ~$0.10 (fastest, best for computer use + agentic vision) - DeepSeek-R1 Distill: ~$0.12 (fast deep reasoning, 32B distill) - DeepSeek-R1: ~$0.55 (o1-level deep reasoning, 671B MoE — USE FOR STRATEGY) - Kimi K2.5: $1.07 (9x cheaper than Claude, #1 agentic benchmarks) - MiniMax M2.5: ~$1.00 (80.2% SWE-Bench, free via NVIDIA NIM/Kilo.ai) - Haiku: $1.00 (good for simple extractions) - Sonnet: $9.00 (management/coordination) - Opus: $15.00 (architecture/strategic only) Deep Reasoning Routing Guide: - Use deepseek-r1 for: multi-step strategy, architecture decisions, complex RLM design, GTM planning, patent analysis, anything requiring extended chain-of-thought - Use deepseek-r1-distill for: code review, math problems, faster reasoning tasks - Use gemini-3-flash-preview for: computer use, agentic vision, screen understanding - Use kimi-k2-moonshot for: agent swarms, long context, tool-use heavy tasks - NEVER use Claude for tasks DeepSeek-R1 can handle — 27x cost difference Author: Genesis System Date: 2026-02-24 N)DictListLiteralOptionalAny)Enum)ThreadPoolExecutor as_completed)level)OpenAIz1OpenAI SDK not installed. Run: pip install openai)typesTFz\google-genai SDK not installed. Gemini native API unavailable. Run: pip install google-genai) load_dotenvz/mnt/e/genesis-system/.envz(/mnt/e/genesis-system/config/secrets.envz@python-dotenv not installed. Using system environment variables.c$eZdZdZdZdZdZdZdZy) ModelTierz'Model cost tiers for routing decisions. ultra_cheapcheapdeep_reasoningstandardpremiumN) __name__ __module__ __qualname____doc__ ULTRA_CHEAPCHEAPDEEP_REASONINGSTANDARDPREMIUM//mnt/e/genesis-system/core/multi_model_swarm.pyrrJs1K E%NHGr!rceZdZdZ ddedededededed ed eed ed e efd Z e de efdZ e de fdZy) SwarmModelz1Represents a model available for swarm execution.Nnameprovider api_key_envbase_urlmodel_id cost_per_mtoktier capabilities max_tokens rpm_limitc ||_||_||_||_||_||_||_||_| |_| |_ y)N r%r&r'r(r)r*r+r,r-r.) selfr%r&r'r(r)r*r+r,r-r.s r"__init__zSwarmModel.__init__VsM   &    * ($"r!returnc@tj|jS)zGet API key from environment.)osgetenvr'r1s r"api_keyzSwarmModel.api_keynsyy))**r!c|jduS)z*Check if model is available (has API key).Nr8r7s r" availablezSwarmModel.availabless||4''r!)N)rrrrstrfloatrrintrr2propertyr8boolr;r r!r"r$r$Ss;#'### #  #  ###3i##C=#0+#++(4((r!r$z Kimi K2.5z NVIDIA NIMNVIDIA_NIM_API_KEYz#https://integrate.api.nvidia.com/v1zmoonshotai/kimi-k2.5gQ?)coderesearch long_contextvision agent_swarmr<r0zKimi K2.5 (Moonshot Direct)z Moonshot AIMOONSHOT_API_KEYzhttps://api.moonshot.cn/v1zkimi-k2)rCrDrErFrG reasoning zKimi K2.5 Official KIMI_API_KEYz MiniMax M2.1zhttps://api.nvidia.com/v1zminimaxai/minimax-m2.1rCrIi(z MiniMax M2zminimaxai/minimax-m2z MiniMax M2.5z MiniMax AIMINIMAX_API_KEYz#https://api.minimax.io/anthropic/v1z minimax-m2.5g?u,Gemini 3 Flash (Direct API — $200 credits)GoogleGEMINI_API_KEY_NEWzgemini-3-flash-previewg?)rCrDrFfast computer_useagentic_visionthink_act_observez$Gemini 3 Flash (OpenRouter fallback) OpenRouterOPENROUTER_API_KEYzhttps://openrouter.ai/api/v1zgoogle/gemini-3-flash-previewz&Gemini 2.0 Flash Exp (OpenRouter free)z google/gemini-2.0-flash-exp:free)rCrDrFrRrSrTzKimi K2.5 (OpenRouter))rCrDrEzMiniMax M2.5 (OpenRouter)zminimax/minimax-m2.5zMiniMax M2.1 (OpenRouter)zminimax/minimax-m2.1g?)rCrIagentz DeepSeek-R1zdeepseek/deepseek-r1g?)rstrategyrCmath architecturerEDeepSeek-R1 Distill 32Bz%deepseek/deepseek-r1-distill-qwen-32bgQ?)rrCrZrR)zkimi-k25kimi-k2-moonshotkimi-k25-official minimax-m21z minimax-m2z minimax-m25 gemini-flashopenrouter-gemini-flashzopenrouter-gemini-flash-expzopenrouter-kimizopenrouter-minimaxzopenrouter-minimax-m21 deepseek-r1deepseek-r1-distill SWARM_MODELSr`r]rbra) rSrTscreenrGtool_userrYbulkrRrC TASK_ROUTINGceZdZdZdZdZdZedee fdZ d(de d e d e d e d e e d ede fdZ d)de d e d e e de ede de f dZ d*de de e de ed e de f dZ d+de de d e de fdZ d,de d e ded e de f dZ d-de d e de de fdZdee efde dee effdZ d.deee efde de deee effdZd eee efd!e de fd"Z d/ded#d$ed%de e fd&Zdee effd'Zy)0MultiModelSwarmao Unified swarm interface for cost-effective parallel execution. Usage: swarm = MultiModelSwarm() # Direct model execution result = swarm.execute("Write a CSV parser", model="kimi-k2-moonshot") # Gemini 2.0 Flash (native, no OpenAI SDK overhead) result = swarm.execute("Explain this code", model="gemini-flash") # Computer use / agentic vision result = swarm.execute_computer_use("What app is open?", image_path="screen.png") # Auto-routing by task type result = swarm.route_by_task_type("Analyse this strategy", task_type="strategy") c`i|_d|_|j|jy)z'Initialize swarm with available models.N)clients_gemini_client_initialize_clients_initialize_geminir7s r"r2zMultiModelSwarm.__init__js)*, "   " !r!ctstjdytjdxs,tjdxstjd}|stjdy t j ||_tjdy#t$r"}tjd |Yd}~yd}~wwxYw) z Initialise native Gemini API client via google-genai SDK. Key priority: GEMINI_API_KEY_NEW > GEMINI_API_KEY > GOOGLE_API_KEY Model: gemini-3-flash-preview (best for computer use + agentic vision). zYgoogle-genai SDK not available; Gemini native API disabled. Run: pip install google-genaiNrQGEMINI_API_KEYGOOGLE_API_KEYzSNo Gemini API key found. Set GEMINI_API_KEY_NEW, GEMINI_API_KEY, or GOOGLE_API_KEY.r:z9Gemini 2.0 Flash native client ready (GEMINI_API_KEY_NEW)z"Gemini native client init failed: ) GENAI_AVAILABLEloggerwarningr5r6genaiClientrminfo Exception)r1 gemini_keyes r"roz"MultiModelSwarm._initialize_geminius NN0   II* + +yy)* +yy)*   NNM   E"',,z"BD  KKS T E NN?sC D D Es90B** C3CCctjD]\}}|dk(r |jr` t|j|j |j |<tjd|jd|jdxtjd |jd |jdy#t$r/}tjd|jd|Yd}~d}~wwxYw) z:Initialize OpenAI-compatible clients for available models.r`)r8r(zOK z available ()zFAIL z failed to initialize: NzSKIP z not available (no API key: )rditemsr;r r8r(rlrtrxr%r&ryrudebugr')r1 model_keymodelr{s r"rnz#MultiModelSwarm._initialize_clientss , 2 2 4  IuN*S.4 % !&/DLL+KK#ejj\enn=MQ OP EJJ<'CEDUDUCVVWX !SNNU5::,6MaS#QRRSsAC C8 %C33C8r3ct|jj}|j|j d|S)zFReturn all usable model keys (OpenAI-compat + native Gemini if ready).r`)listrlkeysrmappend)r1rs r"available_modelsz MultiModelSwarm.available_modelss8DLL%%'(    * KK ' r!Ntaskr system_prompt temperaturer-streamc |dk(r|j|||S||jvrAdj|j}t d|d|dt |j dt |}|j|} tjd|jd |jd i} |jd k(rd d d} | jjj|jd|dd|dg||xs |j || xsd} |r| S| j"dj$j&} t)| dr[| j*} | j,dz |jz}tjd| j,d|dd|j| S#t.$r+}tj1d|jd|d}~wwxYw)a Execute a task using the specified model. Args: task: Task description / prompt model: Model key from SWARM_MODELS system_prompt: System instruction temperature: Sampling temperature (0-1) max_tokens: Max output tokens (None = model default) stream: Stream response chunks (not supported for native Gemini) Returns: Generated text response r`)rrr-z, zModel 'z' not available. Available: z Setup: export z ='your_key'zExecuting task on z ($/MTok)rVzhttps://agileadapt.comzGenesis System)z HTTP-RefererzX-Titlesystem)rolecontentuserN)rmessagesrr-r extra_headersrusagei@BzTokens: z | Cost: $z.4fz | Model: zError executing task on : )_execute_gemini_nativerljoinr ValueErrorrdr'rtrxr%r*r&chat completionscreater)r-choicesmessagerhasattrr total_tokensryerror)r1rrrrr-rr; model_infoclientrresponseresultrcostr{s r"executezMultiModelSwarm.executes0 N "..+%/   $ $"7"78I% )>)>+3t6 H%%a(0088Fx) **Y6*:R:RR u112*T#JG(oo.0 M  LL3JOO3DBqcJ K  s AF'B F'' G0&GG image_data image_mimec|j*tjd|j|d||Std}tj d|j d d}|rtj|}|r|d |n|}|rtjj| tjj|| g} |jjj|j| | } | jS|jjj|j|| } | jS#t $r} tj#d | d} ~ wwxYw)a Execute via the native google-genai SDK using gemini-3-flash-preview. Falls back to OpenRouter gemini-flash if the native client is unavailable. Args: task: Prompt / question system_prompt: Prepended to the user prompt max_tokens: Maximum output tokens image_data: Raw image bytes for vision / computer-use tasks image_mime: MIME type of image (default: image/png) Returns: Generated text NzJNative Gemini client unavailable; falling back to openrouter-gemini-flash.ra)rrrr-r`z,Executing via Gemini 2.0 Flash native API ($r)max_output_tokens )text)data mime_type)rcontentsconfigzGemini native API error: )rmrtrurrdrxr* genai_typesGenerateContentConfigPart from_text from_bytesmodelsgenerate_contentr)rryr) r1rrr-rrrr full_promptrrr{s r"rz&MultiModelSwarm._execute_gemini_natives.    & NN\ <</+%   ".1  ::;S;S:TTZ [  F$::&0;H]O4v6TK $$..K.@$$//Z:/V ..55FF$--%!G==  ..55FF$--(!G ==   LL4QC8 9  s$B%E =E E/E**E/ image_pathc ddl}d}d}|r@|j|\}}|xsd}t|d5} | j}dddn|r|}tj d|j ||||S#1swY3xYw)uB Execute a computer use / agentic vision task via Gemini 2.0 Flash. Gemini 2.0 Flash is the recommended model for computer use — best-in-class vision understanding at ultra-low cost ($0.10/MTok). Args: task: Instruction/question about the screen content image_path: Path to screenshot file (PNG/JPEG) image_data: Raw image bytes (alternative to image_path) system_prompt: Vision-optimised system prompt Returns: Gemini's interpretation / next-action recommendation rN image/pngrbz>Routing computer_use task -> Gemini 2.0 Flash (agentic_vision))rrrr) mimetypes guess_typeopenreadrtrxr) r1rrrrr raw_bytesmime_fs r"execute_computer_usez$MultiModelSwarm.execute_computer_useYs4   **:6GD!&;Dj$' %1FFH  % % "I TU**'  +   % %s A99B task_typec tj|d}tjd|d||dvr|j ||S|j d|||d|S)a Execute a task using automatic model routing based on task_type. Routing table (TASK_ROUTING): computer_use / agentic_vision / screen -> gemini-flash agent_swarm / tool_use -> kimi-k2-moonshot deep_reasoning / strategy -> deepseek-r1 bulk / fast / code -> gemini-flash / openrouter-gemini-flash Args: task: The task prompt task_type: Key in TASK_ROUTING (or any string; unknown defaults to gemini-flash) system_prompt: System instruction **kwargs: Additional args forwarded to execute() Returns: Generated response from the routed model r`zroute_by_task_type: 'z' -> rSrTre)rr)rrrr )rhgetrtrxrr)r1rrrkwargsrs r"route_by_task_typez"MultiModelSwarm.route_by_task_typeso2  N; +I;eE7CD D D,,$m,T Tt||ZU-ZSYZZr!rRc|rdnd}||jvr tdtjd|rdnd|j |||d|S) u0 Dispatch a deep reasoning task to DeepSeek-R1 via OpenRouter. DeepSeek-R1 provides o1-level reasoning at ~$0.55/MTok — 27x cheaper than Opus. Use for: strategy, architecture, multi-step analysis, complex planning. Args: task: The reasoning task or question system_prompt: Chain-of-thought system context fast: If True, use distilled 32B variant ($0.12/MTok) max_tokens: Max output tokens (default 8192) Returns: Reasoning output from DeepSeek-R1 rcrbz"%8 #DEE]]((*F KK.s6{m6B CM(( I!z3MultiModelSwarm.parallel_execute..asGaakY.FGs zParallel execution complete: z succeeded in .1fr) rrtrxtimer submitrr rrrsum)r1rrrresultstotal start_timeexecutorrfuture_to_task completedfuturer status_iconelapsed successess r"parallel_executez MultiModelSwarm.parallel_execute8sx$E  j'H UVWXYY[ K 8 H"' 4 4iOQZZN I&~6 Q v&&,X&6)&Cd   {!E7"[Mzz*c233vg6GI   "))+ *G7GG  +I;awnWUXMYZ [ +  s E %E4A=E E  Er output_dirc tj|dt|ddgd}|D]}|jdd}|jddj dd }|jd d}|j dd j d d }|d |d } tj j || } |jd d} | rt| dd5} | jd|d| jd|jddd| jd|d| jd|jddd| j| ddd|dxxdz cc<nt| dd5} | jd|d| jd|jddd| jd| jd|jdddddd|dxxdz cc<|d j||jdd|| |jdd!!tj j |d"} t| dd5} tj|| d#$dddtjd%|d&|dd'|dd(| S#1swY^xYw#1swYxYw#1swYIxYw))z~Save each agent's output to individual markdown files in output_dir. Returns path to the _summary.json file. T)exist_okr)rrfailuresagentsrrrrrr z.mdrrwzutf-8)encodingz # Agent:  z # Model: z # Status: z# Task: rrNrr z# Status: FAILED z # Error: rr r)rrrfilerz _summary.json)indentzResults saved to rz succeeded, z failed)r5makedirsrrreplacepathrrwriterjsondumprtrx)r1rr summaryrrrrsafe_idfilenamefilepathrr summary_paths r" save_resultszMultiModelSwarm.save_resultsgs J.\ #  Fzz*i8HJJw 2::3DEZZ)4F&&sC088cBG!!E7#.Hww||J9Hjj2.G(C':%aGGiz45GGi 7I(F'GrJKGGj34GGhvzz&)'D&ETJKGGG$ %  $)$(C':LaGGiz45GGi 7I(F'GrJKGG02GGi 7I(F'GrJK L  #q(# H  $ $$GY7  G, & 5 Dww||J@ ,g 6 ,! IIgq + ,   |2{#$L1D0EW N ?%%LL  , ,s& BK A1K:KK K K$)rCrDrFrIrRrSrTrGpriority)rqualityspeedc|dvr|jrdSdS|dk(rd|jvryd|jvrytjDcgc]6\}}||jvr#||jvs|dk(r|j||f8}}}|sy|dk(r|j d n1|d k(r|j d n|d k(r|j d |r|ddSdScc}}w)u Recommend the best available model for a task type. Hard overrides: computer_use / agentic_vision -> gemini-flash (Gemini 2.0 Flash native) agent_swarm -> kimi-k2-moonshot Args: task_type: Descriptor of the task priority: Ranking criterion — cost, quality, or speed Returns: Model key string, or None if nothing matches rr`rarGr]r^Nrc |djSNr r*xs r"z1MultiModelSwarm.recommend_model..s!A$*<*<r!)keyr#c"|dj Sr'r(r)s r"r+z1MultiModelSwarm.recommend_model..s1Q4+=+=*=r!r$c |djSr')r-r)s r"r+z1MultiModelSwarm.recommend_model..s!A$//r!r)rmrlrdr~r,sort)r1rr"r,r candidatess r"recommend_modelzMultiModelSwarm.recommend_models , D D%)%8%8> W>W W  %!T\\1)"dll2*,8+=+=+? 'S%E...t||#>)d.A.A.M%L   v  OO  OO 9O :#-z!}Q747% s;C&c |j}tt}|Dcic]3}|tvr)t|jdt|jd5}}|Dcgc]}|tvr|t|jf!}}t|||rt||z dzddnd|t |dd d |j d utd Scc}wcc}w) zIReturn swarm statistics including routing table and Gemini native status.$z/MTokdr%z0%c |dSr'r r)s r"r+z+MultiModelSwarm.get_stats..s QqTr!)NN)r,defaultrN)r total_modelsavailability_ratercheapest gemini_native routing_table)rrrdr%r*minrmrh)r1available_keysr8kcost_breakdown all_costss r" get_statszMultiModelSwarm.get_statss..<( $ L  O Al1o&C&C%DE"J J  $ L  Q-- .  !$N 3( ',6<PQRS!00<)    s 8C$C)r_#You are a helpful coding assistant.gffffff?NF)rNNr)NNzYou are a computer vision assistant with expertise in understanding screenshots, UI elements, and agentic web/desktop tasks. Describe precisely what you see and suggest the next action.)r)zIYou are a deep reasoning AI. Think step by step. Be thorough and precise.FrJ)rC)r_)r)rrrrr2rornr@rr=rr>rr?rArbytesrrrrrrrrr r!rr1rBr r!r"rjrjVs&"E@2$s)#B $(QQQ Q  Q SM QQ Qt<$(&*% FFFSM F UO F  F FV%)&* K - - SM- UO -  -  - n< [ [ [ [ [T X( ( (  ( (  ( bC ,,, , ,d$S>$$ c3h $R+ -DcN#-- - d38n  -^<d38n%<< <P9? 08 < 08 45 08 #08d 4S> r!rjc ntdtdtdtt}td|jD]U}|tvr=t|}td|jd|j d|j dHtd|Wt|jsMtdtjD]*\}}td |jd |j,y |j}td |d d|ddtd|dtd|dttdtjD]\}}td|dd|t|jd}|tvrt|jn|}td|d |jd|d} td| ttdy #t$r} td | Yd } ~ 2d } ~ wwxYw)!z Test script.zP================================================================================zGENESIS MULTI-MODEL SWARMzAvailable Models:z OK z - $z/MTok (r}z+WARNING: No models available. Set API keys:z export z='your_key' # NzStats: rrr8z models availablez Cheapest: r:zGemini native: r;zRouting table:z 20sz -> rzTesting z...z:Write a one-sentence explanation of async/await in Python.r4)rr-zResult: zError: ) printrjrrdr%r*r&r~r'rBrhrry) swarmrmr,rstatsr test_model model_namerr{s r"mainrOs- (O %& (O G  E ++'  $Y'A E!&&aoo%6gajj\K L E)% & '  G  ! ! ;<&,,. NJC Ie//0 |L M N OO E GE,-.an0E/FFW XY JuZ() *+ OE/23 45 G  , 2 2 43 9 9S/i[123 G''*J)3|)C Z %% HZL $% H  !" G (O  sms"H H4H//H4__main__)/rr5sysrrrtypingrrrrrenumrconcurrent.futuresr r logging basicConfigINFO getLoggerrrtopenair ImportErrorrexit google.genairvrrrsrudotenvrrr$rrrrdr=__annotations__rhrjrOr r!r"r_s<  55?',,'   8 $  1OW",-:; #(#(X (6' __R  # *&- ___ $ !"- __R  (,)  " "k*  (,'  " "k*  %6 __k* " ;()  " "r  * 3(/0  " "r $. 5(/3  " "] $" %(/' __9 % ((/' __k* ) ((/' __3   (/'  % %c & &(/8  % %? oC' d3 ?#CT%$$((##$$/  d38n$a  a  P3l zFG  LLDE CHHQKO NN (W NNUVWs6(I>/J->K >(J*)J*-K K  K&%K&