llm_core

defmodule LlmCore.LLM.Native do
  @moduledoc """
  In-process agentic provider — runs the agent loop inside the BEAM VM.

    1. Reads the agent .md system prompt
    2. Resolves an LLM API provider (Appliance → Zai → Anthropic)
    3. Calls LlmCore.Agent.Loop.run with LlmToolkit.CodeTools
    4. Returns a standard LlmCore.LLM.Response

  Zero CLI cost. Uses whatever API provider is available.

  ## Provider Resolution Priority

    1. **Appliance** (local, free) — Qwen, etc. on DGX Spark / LM Studio
    2. **Z.ai** (plan-covered) — GLM-5.1, included in coding plan
    3. **Anthropic** (direct API) — Claude models, pay-per-token

  ## Usage

  This provider implements `LlmCore.LLM.Provider` and is selected by
  provider routing when `provider: "native"` is specified or when no
  CLI providers are available.

      {:ok, response} = LlmCore.LLM.Native.send(task,
        system_prompt_file: "/path/to/agent.md",
        cwd: "/path/to/project",
        model: "qwen3-vl-32b-thinking"
      )
  """

  @behaviour LlmCore.LLM.Provider

  alias LlmCore.LLM.{Response, Error}
  alias LlmCore.LLM.Native.Router

  alias LlmCore.Agent.Loop
  alias LlmToolkit.CodeTools

  import Kernel, except: [send: 2]

  @max_iterations 15

  # ── Provider Behaviour ─────────────────────────────────────

  @impl true
  def available?, do: true

  @impl true
  def capabilities do
    %{
      streaming: false,
      passthrough: false,
      tool_use: true,
      native_loop: true,
      models: ["qwen3-vl-32b-thinking", "glm-5.1", "claude-sonnet-4-6"]
    }
  end

  @impl true
  def provider_type, do: :api

  @impl true
  @spec send(String.t(), keyword()) ::
          {:ok, Response.t()} | {:error, Error.t()}
  def send(prompt, opts \\ []) do
    cwd = opts[:cwd] || File.cwd!() || "."
    agent_file = opts[:system_prompt_file]
    model = opts[:model]
    timeout = opts[:timeout]
    llm_provider = opts[:llm_provider]

    start = System.monotonic_time(:millisecond)

    result =
      try do
        do_send(prompt, cwd, agent_file, model, timeout, llm_provider)
      rescue
        e ->
          {:error, "Native dispatch crashed: #{Exception.message(e)}"}
      end

    elapsed = System.monotonic_time(:millisecond) - start

    case result do
      {:ok, llm_response, _messages} ->
        text = llm_response.content || ""

        {:ok,
         Response.new(
           content: text,
           provider: :native,
           metadata: %{elapsed_ms: elapsed, model: llm_response.model, usage: llm_response.usage}
         )}

      {:error, :max_iterations_reached} ->
        {:error,
         Error.new(:provider_error,
           message: "Iteration limit reached (#{@max_iterations})",
           provider: :native
         )}

      {:error, reason} ->
        {:error,
         Error.new(:provider_error,
           message: "Native dispatch error: #{inspect(reason)}",
           provider: :native
         )}
    end
  end

  @impl true
  def stream(_prompt, _opts \\ []) do
    {:error,
     Error.new(:provider_error,
       message: "Streaming not supported for native dispatch (agentic loop is synchronous)",
       provider: :native
     )}
  end

  # ── Execution ──────────────────────────────────────────────

  defp do_send(prompt, cwd, agent_file, model, timeout, llm_provider)

  defp do_send(prompt, cwd, agent_file, model, _timeout, llm_provider) do
    system_prompt = load_agent_prompt(agent_file)
    candidates = resolve_candidates(model, llm_provider)

    messages = [
      %{role: :system, content: system_prompt},
      %{role: :user, content: prompt}
    ]

    tools = CodeTools.available_tools()

    run_fn = fn {provider, resolved_model, provider_opts} ->
      llm_send = build_llm_send(provider, resolved_model, provider_opts)

      Loop.run(
        messages,
        llm_send,
        tools: tools,
        resolve_tool: &CodeTools.resolve(&1, cwd),
        max_iterations: @max_iterations
      )
    end

    try_cascade(candidates, run_fn)
  end

  @doc """
  Walk a list of `{module, model, opts}` candidates and invoke `run_fn`
  on each until one succeeds.

  - `{:ok, response, messages}` from `run_fn` → returned immediately.
  - `{:error, :max_iterations_reached}` → returned immediately (reasoning
    failure, not a provider outage — retrying elsewhere won't help).
  - Any other `{:error, reason}` → logs and advances to the next candidate.
  - Empty list → `{:error, :no_provider_succeeded}`.

  Exposed for direct testing; production call sites go through `send/2`.
  """
  @spec try_cascade([{module(), String.t(), keyword()}], (tuple() -> term())) ::
          {:ok, LlmCore.LLM.Response.t(), [map()]} | {:error, term()}
  def try_cascade([], _run_fn), do: {:error, :no_provider_succeeded}

  def try_cascade([candidate | rest], run_fn) do
    case run_fn.(candidate) do
      {:ok, _response, _messages} = ok ->
        ok

      {:error, :max_iterations_reached} = err ->
        err

      {:error, _reason} = err when rest == [] ->
        err

      {:error, reason} ->
        {mod, _, _} = candidate
        require Logger

        Logger.warning(
          "[Native] Provider #{inspect(mod)} failed (#{inspect(reason)}); trying next in cascade"
        )

        try_cascade(rest, run_fn)
    end
  end

  # ── Agent Prompt ───────────────────────────────────────────

  defp load_agent_prompt(nil) do
    "You are a helpful coding assistant. Complete the task accurately."
  end

  defp load_agent_prompt(path) do
    case File.read(path) do
      {:ok, content} ->
        strip_frontmatter(content)

      {:error, reason} ->
        require Logger
        Logger.warning("[Native] Cannot read agent file #{path}: #{inspect(reason)}")
        "You are a helpful coding assistant. Complete the task accurately."
    end
  end

  defp strip_frontmatter(content) do
    case String.split(content, ~r/^---\s*$/m, parts: 3) do
      [_before, _yaml, body] -> String.trim(body)
      _ -> String.trim(content)
    end
  end

  # ── Provider Resolution ────────────────────────────────────
  #
  # Driven by TOML config ([native] section in priv/config/llm_core.toml).
  #
  # Cascade: ordered list of providers to try. First available wins.
  # Model routing: substring patterns → provider name. First match wins.
  # Default models: per-provider fallback when no model specified.
  #
  # All of this is configurable — change the TOML, not the code.

  # Returns an ordered list of `{mod, model, opts}` candidates.
  #
  # Explicit `llm_provider` → single-element list (caller asked for a specific
  # backend; don't silently cascade to a different one).
  # Otherwise → primary + remaining cascade fallbacks from `Router.candidates/3`.
  defp resolve_candidates(model, llm_provider) when is_binary(llm_provider) do
    case Router.resolve_provider(llm_provider) do
      {:ok, {mod, resolved_model, opts}} ->
        [{mod, model || resolved_model, opts}]

      {:error, :no_provider} ->
        raise "unknown LLM provider: #{llm_provider}"
    end
  end

  defp resolve_candidates(model, nil) do
    config = read_native_config()

    appliance_has =
      is_binary(model) and appliance_available?() and model_available_on_appliance?(model)

    case Router.candidates(model, config, appliance_has_model: appliance_has) do
      [] ->
        raise "no LLM API provider available — check [native] cascade in llm_core.toml"

      candidates ->
        candidates
    end
  end

  # Read [native] config from Application env (loaded from TOML at startup).
  defp read_native_config do
    Application.get_env(:llm_core, :native, default_native_config())
  end

  defp default_native_config do
    %{
      cascade: ["appliance", "zai", "anthropic"],
      default_models: %{
        "appliance" => "qwen3.5-27b-claude-4.6-opus-distilled-mlx",
        "zai" => "glm-5.1",
        "anthropic" => "claude-sonnet-4-6"
      },
      model_routing: [
        %{"pattern" => "claude", "provider" => "anthropic"},
        %{"pattern" => "glm", "provider" => "zai"},
        %{"pattern" => "zai", "provider" => "zai"},
        %{"pattern" => "gpt", "provider" => "openai"},
        %{"pattern" => "openai", "provider" => "openai"}
      ]
    }
  end

  defp appliance_available?, do: LlmCore.LLM.Appliance.available?()

  # ── LLM Send Function ──────────────────────────────────────

  # Check if a specific model is loaded on the local Appliance.
  # Queries /v1/models directly. Result cached in process dict for 60s.
  defp model_available_on_appliance?(model) do
    now = System.monotonic_time(:second)
    cache = Process.get(:appliance_models_cache)

    models =
      case cache do
        {cached_at, cached_models} when now - cached_at < 60 ->
          cached_models

        _ ->
          ids =
            case LlmCore.LLM.Appliance.discover() do
              [{_, %URI{} = uri} | _] ->
                url = String.trim_trailing(URI.to_string(uri), "/") <> "/v1/models"

                case Req.get(url, receive_timeout: 3_000, retry: false) do
                  {:ok, %Req.Response{status: 200, body: %{"data" => data}}} ->
                    Enum.map(data, & &1["id"])

                  _ ->
                    []
                end

              _ ->
                []
            end

          Process.put(:appliance_models_cache, {now, ids})
          ids
      end

    model in models
  end

  defp build_llm_send(provider_mod, resolved_model, provider_opts) do
    fn messages, loop_opts ->
      loop_opts =
        loop_opts
        |> Keyword.put_new(:timeout, 180_000)
        |> Keyword.put(:model, resolved_model)
        |> Keyword.merge(provider_opts)

      provider_mod.send(messages, loop_opts)
    end
  end
end