' j4UdZddlZddlZddlZddlmZmZddlZddlm Z ddl m Z m Z m Z mZddlmZddlmZddlmZdd lmZdd lmZd Zejd d ZejddZejddZejddZ eje!e"d<ddZ#e#GddeZ$defde%de%de$fdZ&GddeZ'dS) aShared base for all Enterprise Brain agents. By default agents use the gateway model string directly (no extra provider). Pass ``api_base`` + optional ``api_key`` to route through a LiteLLM proxy instead. ``EnterpriseAgent`` subclasses ``PydanticAIResponder`` (pylogue), inheriting streaming, tool-rendering, system-prompt management and conversation history out of the box. Callers can use it directly as a pylogue responder, or call ``agent.run(...)`` / ``agent.tool_plain(...)`` as before. N)AnyCallable) data_snapshot)ClauseContains ClauseEquals ModelInfoProvider)Agent)OpenAIChatModel)LiteLLMProvider)OpenAIProvider)PydanticAIResponderzgateway/openai:gpt-5-miniLITELLM_PROVIDER_MODEL_NAMEzgemini/gemini-2.5-flash MLX_MODEL MLX_BASE_URLzhttp://localhost:8080/v1_pylogue_activeF)defaultreturnc t}|dd}tj}d}|jD]"}||}| |j}n#|$ddl}| d|d|dd dSt| }tj d d }td d|rtj|ndt#d|t%|||g} tjtj| g|j|jdS)uRegister a LiteLLM wrapper provider so genai-prices can cost custom deployments. LiteLLM deployments are typically named like ``"divami-gemini/gemini-2.5-flash"``, wrapping a well-known base model (``"gemini-2.5-flash"`` — the part after the last ``/``). This function: 1. Strips the base model name from the deployment string. 2. Looks it up in the bundled genai-prices database. 3. Registers a custom provider (matched by ``provider_id="litellm"``) that claims any model name containing the base name and reuses its pricing. No prices are hardcoded — they are always pulled from the bundled data. /Nrz8genai-prices: could not resolve pricing for base model 'z' (derived from 'z'). Cost will not be tracked.) stacklevel)containsLITELLM_API_BASErzlitellm-wrapperzLiteLLM Wrapperz(?!)litellm)equals)idmatchprices)rname api_patternprovider_match model_matchmodels) providersfrom_auto_update)DEFAULT_LITELLM_MODEL_IDsplitlowerr get_snapshotr' find_modelr!warningswarnrosenvirongetr reescaperrset_custom_snapshot DataSnapshotr() deployment base_modelsnapresolved_pricesprovidermodelr.r%api_baselitellm_providers :/var/www/html/ai-enterprise-brain/backend/chanakya/base.py_register_litellm_model_pricesr@*s*J!!#&&r*0022J  % ' 'DON##J//  #lO E   Hz H H( H H H     !*555Kz~~0"55H  +3@BIh'''#9555 !&     %"'9$.9!2   c"eZdZdZfdZxZS) MLXChatModela ``OpenAIChatModel`` variant for ``mlx_lm.server``. ``mlx_lm.server`` requires the system message to be strictly the *first* message in the list and rejects requests that have system messages scattered through conversation history (which pydantic-ai produces when message_history is used across turns). This subclass post-processes the mapped messages to: 1. Collect all ``role="system"`` messages. 2. Keep only the last one (the most-current instructions). 3. Hoist it to position 0. cKt||d{V}d|D}d|D}|r |dg|S|S)NcDg|]}|ddk|Srolesystemr2.0ms r? z.MLXChatModel._map_messages..{s,OOOQQUU6]]h5N5Nq5N5N5NrAcDg|]}|ddk|SrFrIrJs r?rMz.MLXChatModel._map_messages..|s,SSSv(9R9R19R9R9RrAr)super _map_messages)selfmessagesmodel_request_parametersopenai_messages system_msgsnon_system_msgs __class__s r?rPzMLXChatModel._map_messagesysy % 5 5h@X Y YYYYYYYOO/OOO SSoSSS  7O6o6 6rA)__name__ __module__ __qualname____doc__rP __classcell__rWs@r?rCrCmsB  rArCz!mlx-community/Qwen3.5-9B-MLX-4bit model_namebase_urlc@t|t|S)aCreate an MLX-backed model via mlx_lm.server's OpenAI-compatible endpoint. Prerequisite: an instance of ``mlx_lm.server`` running and serving the desired model at the specified URL (default: ``http://localhost:8080/v1``). You can run it locally using `mlx_lm.server --model mlx-community/Qwen3.5-9B-MLX-4bit --host 0.0.0.0` once you `pip install mlx_lm`. The Qwen model will be automatically downloaded on first run if not present locally. )r_)r^r;)rCr )r^r_s r?make_mlx_modelras# :PX8Y8Y8Y Z Z ZZrAceZdZdZd!defdZd!deffd Zedddeddddddddd d ed ed edzd edzdedzdededzdedzde de de de dzde dzde dzddffdZdedefdZdedefdZdedefdZde de fdZde fd ZxZS)"EnterpriseAgentuPydantic-AI agent with LiteLLM routing and full pylogue responder support. Inherits from ``PydanticAIResponder``, so every instance is directly usable as a pylogue streaming responder (``await agent("prompt", context=ctx)``), with streaming, tool-rendering, system-prompt management, and conversation history all included. ``pylogue_instructions`` (mermaid/HTML hints) are injected **only** when the agent is invoked via the pylogue ``/chat`` WebSocket — i.e. when ``set_context()`` has been called by pylogue's session machinery. REST API calls from ``/query`` and ``/query/stream`` bypass ``set_context`` entirely, so they receive a clean system prompt with no mermaid or HTML instructions. Args: instructions: The agent's persona / system instructions. model_id: Model identifier when NOT routing through LiteLLM or MLX (default: ``"gateway/openai:gpt-5-mini"``). api_base: When provided, requests are routed through the LiteLLM proxy at this URL. api_key: API key for the proxy; falls back to ``LITELLM_API_KEY`` env var. mlx_model: When provided (or ``MLX_MODEL`` env var is set), requests are served by a local ``mlx_lm.server`` instance. Takes precedence over ``model_id`` but not ``api_base``. mlx_base_url: mlx_lm.server OpenAI-compatible endpoint; falls back to ``MLX_BASE_URL`` env var or ``http://localhost:8080/v1``. logfire_env: When provided, logfire is configured with this environment name and pydantic-ai instrumentation is enabled automatically. agent_deps: Optional deps object forwarded to the underlying pydantic_ai.Agent runs (passed to PydanticAIResponder). show_tool_details: Whether to render full tool call/result details in the stream (passed to PydanticAIResponder). retries: Number of retries for the underlying agent. Nrc ddlm}m}||jddt |jdg|t r|ndS)ajInject pylogue instructions only during pylogue WebSocket sessions. ``_pylogue_active`` is a ContextVar set to ``True`` for the exact duration of ``__call__`` (the pylogue WS message handler). REST API calls via ``agent.run()`` never enter ``__call__``, so the var stays ``False`` and they receive a clean system prompt. r)PYLOGUE_INSTRUCTIONScompose_system_prompt base_promptr additional)rgadditional_instructionsuserpylogue_instructions)pylogue.integrations.commonrerf _prompt_stater2listr)rQrjre_csps r?_compose_system_promptz&EnterpriseAgent._compose_system_prompts        t*..}bAA$();)?)? b)Q)Q$R$R9H9L9L9N9N!V!5!5TV     rAtextcKtd} t||23d{V}|WV 6 t|dS#t|wxYw)aPylogue WebSocket message handler. Sets ``_pylogue_active`` for the exact duration of this call so that ``_compose_system_prompt`` injects pylogue instructions for WS sessions but not for REST API calls via ``agent.run()``. TN)rsetrO__call__reset)rQrqcontexttokenchunkrWs r?rtzEnterpriseAgent.__call__s ##D)) )$ww//g>>       e  ?>  ! !% ( ( ( ( (O ! !% ( ( ( (s"A,AA,,BT) model_idr=api_key mlx_model mlx_base_url logfire_env service_name agent_depsshow_tool_detailsretries deps_type output_type capabilities instructionsrzr=r{r|r}r~rrrrrrrc |r*tj||dtj|ptx}rt ||}n@|r<|pt jd}ttt||}n|}t|f|| d| t| ni| t| ni|t| ni}| | | } t || | dS) NF) environmentr scrubbingLITELLM_API_KEY)r=r{)r;)rr)r)r)r)rr)logfire configureinstrument_pydantic_aiDEFAULT_MLX_MODELrar0getenvr r)r r dictrO__init__)rQrrzr=r{r|r}r~rrrrrrr effective_mlxr< resolved_keypydantic_agentrWs r?rzEnterpriseAgent.__init__sf$  -  +L\a b b b b  * , , ,&;*; <= "=,??EE  "Bbi0A&B&BL#(((LQQQEE E  %  -6,Ati((((r  1<0Gt ,,,,R  3?2Jt....PR     Z%7"J JRcdddddrAfuncc6|j|S)z>Register a plain (no RunContext) tool on the underlying agent.)agent tool_plainrQrs r?rzEnterpriseAgent.tool_plainsz$$T***rAc6|j|S)z8Register a tool with RunContext on the underlying agent.)rtoolrs r?rzEnterpriseAgent.toolszt$$$rAc6|j|S)zBRegister a dynamic system-prompt function on the underlying agent.)r system_promptrs r?rzEnterpriseAgent.system_promptsz''---rAargskwargscK|d|j|jj|i|d{V}||_|S)a3Run the agent with automatic conversation-history threading. Passes ``self.message_history`` into the underlying ``Agent.run`` call (unless the caller supplies their own ``message_history`` kwarg) and updates it from the result, so successive ``run()`` calls share context. message_historyN) setdefaultrrrun all_messages)rQrrresults r?rzEnterpriseAgent.runse +T-ABBB%tz~t6v66666666%2244 rAc &|jjdi|S)zDExpose the agent as an A2A ASGI app (delegates to ``Agent.to_a2a``).)rto_a2a)rQrs r?rzEnterpriseAgent.to_a2a(s tz **6***rA)N)rXrYrZr[strrprtDEFAULT_MODEL_IDDEFAULT_MLX_BASE_URLrboolinttypernrrrrrrrr\r]s@r?rcrcs4""H  3    & ) )3 ) ) ) ) ) )&)#" $0"&#'"&!%#'$(!1e1e1e1e 1e * 1e t 1e:1e1e4Z1eDj1e1e 1e1e$;1eD[1e Tk!1e" #1e1e1e1e1e1ep+x+H++++%%h%%%%.(.x.... s c    +s++++++++rArc)rN)(r[ contextvarsr0r3typingrrr genai_pricesrgenai_prices.typesrrrr pydantic_air pydantic_ai.models.openair pydantic_ai.providers.litellmr pydantic_ai.providers.openair pylogue.integrations.pydantic_airrr1r2r)rr ContextVarrr__annotations__r@rCrrarcrrAr?rs<    &&&&&&PPPPPPPPPPPP555555999999777777@@@@@@.:>>*GIbccJNN;33z~~n6PQQ1G 0Fu111'- ====@ ?.:( [ [ [ [ [ [ [ [Y+Y+Y+Y+Y+)Y+Y+Y+Y+Y+rA