ragex

defmodule Ragex.Agent.Core do
  @moduledoc """
  Main entry point for Ragex Agent operations.

  Orchestrates the full project analysis pipeline:
  1. Analyze project (build knowledge graph, embeddings)
  2. Discover issues (dead code, duplicates, security, smells, complexity)
  3. Generate AI-polished report (AI may use Ragex MCP RAG tools for evidence)
  4. Enable conversation session for follow-up

  ## Report generation and RAG

  During step 3 the AI assistant is given access to a restricted set of
  read-only Ragex MCP query tools (`ToolSchema.rag_query_tools/1`).  This lets
  the AI look up concrete code details — reading a flagged file, checking
  coupling metrics, or finding callers of a complex function — to produce
  evidence-based findings rather than relying solely on pre-computed statistics.
  Heavy re-analysis tools are excluded so the pipeline is not re-triggered.

  ## Usage

      # Full project analysis with report
      {:ok, result} = Agent.Core.analyze_project("/path/to/project")

      # Skip report generation (e.g. before streaming it separately)
      {:ok, result} = Agent.Core.analyze_project("/path/to/project", skip_report: true)

      # Continue conversation (agent uses full tool set)
      {:ok, response} = Agent.Core.chat(result.session_id, "Tell me more about the security issues")

      # Get just the report
      {:ok, report} = Agent.Core.get_report(result.session_id)
  """

  require Logger

  alias Ragex.Agent.{Executor, Memory, Report}
  alias Ragex.AI.Config, as: AIConfig
  alias Ragex.Analyzers.Directory

  alias Ragex.Analysis.Cache, as: AnalysisCache

  alias Ragex.Analysis.{
    DeadCode,
    DependencyGraph,
    Duplication,
    Quality,
    Security,
    Smells,
    Suggestions
  }

  alias Ragex.Embeddings.Persistence, as: EmbeddingsPersistence
  alias Ragex.Graph.Persistence, as: GraphPersistence
  alias Ragex.Graph.Store

  @type analysis_result :: %{
          session_id: String.t(),
          report: String.t(),
          issues: map(),
          summary: map()
        }

  @features Application.compile_env(:ragex, :features, [])
  @include_suggestions Keyword.get(@features, :suggestions, true)
  @include_dead_code Keyword.get(@features, :dead_code, false)

  @doc """
  Analyze a project and generate an AI-polished report.

  The AI report is generated using a restricted set of read-only Ragex MCP
  query tools so the AI can retrieve concrete code evidence.  Pass
  `skip_report: true` to skip report generation (e.g. when you intend to
  stream it later via `stream_generate_report/3`).

  ## Parameters

  - `path` - Project root path
  - `opts` - Options:
    - `:provider` - AI provider (:deepseek_r1, :openai, :anthropic, :ollama)
    - `:model` - Model name override
    - `:include_suggestions` - Include refactoring suggestions (default: true)
    - `:max_files` - Maximum files to analyze (default: 500)
    - `:skip_embeddings` - Skip embedding generation (default: false)
    - `:skip_report` - Skip AI report generation (default: false)
    - `:include_dead_code` - Enable dead code analysis (default: false)
    - `:exclude_patterns` - Patterns to exclude (default: standard ignores)

  ## Returns

  - `{:ok, result}` - Analysis completed with session ID and report
  - `{:error, reason}` - Analysis failed
  """
  @spec analyze_project(String.t(), keyword()) :: {:ok, analysis_result()} | {:error, term()}
  def analyze_project(path, opts \\ []) do
    Logger.info("Starting project analysis: #{path}")

    with {:ok, abs_path} <- validate_path(path) do
      # Switch store to the target project: clears stale data and loads
      # the correct per-project cache (graph + embeddings + file tracker).
      Store.load_project(abs_path)

      graph_stats = Store.stats()

      case {graph_stats.nodes > 0, AnalysisCache.load(abs_path)} do
        {true, {:ok, cached_issues}} ->
          # Graph loaded from cache + issues fresh -> skip everything
          Logger.info("Using cached analysis (#{graph_stats.nodes} nodes, all files unchanged)")

          finalize_analysis(abs_path, cached_issues, opts)

        {true, {:stale, _cached_issues, changed_files}} ->
          # Graph loaded but some files changed -> incremental re-analysis
          Logger.info("Incremental analysis: #{length(changed_files)} files changed")

          with {:ok, _} <- analyze_codebase(abs_path, opts),
               {:ok, issues} <- discover_issues(abs_path, opts) do
            persist_all_state(issues, abs_path)
            finalize_analysis(abs_path, issues, opts)
          end

        _ ->
          # No cache or empty graph -> full analysis
          with {:ok, _} <- analyze_codebase(abs_path, opts),
               {:ok, issues} <- discover_issues(abs_path, opts) do
            persist_all_state(issues, abs_path)
            finalize_analysis(abs_path, issues, opts)
          end
      end
    end
  end

  @doc """
  Continue a conversation with the agent in an existing session.

  ## Parameters

  - `session_id` - Active session ID
  - `message` - User message
  - `opts` - Options (same as analyze_project)

  ## Returns

  - `{:ok, response}` - Agent response
  - `{:error, reason}` - Chat failed
  """
  @spec chat(String.t(), String.t(), keyword()) :: {:ok, map()} | {:error, term()}
  def chat(session_id, message, opts \\ []) do
    Logger.debug("Agent chat: session=#{session_id}")

    with {:ok, _session} <- Memory.get_session(session_id),
         :ok <- Memory.add_message(session_id, :user, message),
         {:ok, result} <- Executor.run(session_id, opts) do
      {:ok,
       %{
         content: result.content,
         tool_calls_made: result.tool_calls_made,
         usage: result.usage
       }}
    end
  end

  @doc """
  Continue a conversation with streaming support.

  Same as `chat/3` but streams the final AI response in real-time via callbacks.
  Intermediate tool-call steps use blocking calls, but the final text response
  is streamed chunk-by-chunk.

  ## Additional Options

  - `:on_chunk` - `(chunk -> :ok)` callback for real-time content/thinking delivery
  - `:on_phase` - `(:thinking | :answering | :done -> :ok)` phase transition callback
  - `:on_tool_progress` - `(map() -> :ok)` callback when tools are being called

  ## Returns

  Same as `chat/3`.
  """
  @spec stream_chat(String.t(), String.t(), keyword()) :: {:ok, map()} | {:error, term()}
  def stream_chat(session_id, message, opts \\ []) do
    Logger.debug("Agent stream_chat: session=#{session_id}")

    with {:ok, _session} <- Memory.get_session(session_id),
         :ok <- Memory.add_message(session_id, :user, message),
         {:ok, result} <- Executor.stream_run(session_id, opts) do
      {:ok,
       %{
         content: result.content,
         tool_calls_made: result.tool_calls_made,
         usage: result.usage
       }}
    end
  end

  @doc """
  Get the generated report from a session.

  If not yet generated, generates it on-demand.
  """
  @spec get_report(String.t(), keyword()) :: {:ok, String.t()} | {:error, term()}
  def get_report(session_id, opts \\ []) do
    with {:ok, session} <- Memory.get_session(session_id) do
      case session.metadata[:report] do
        nil ->
          # Generate report from issues
          issues = session.metadata[:issues] || %{}

          {:ok, report, _ai_status} = generate_report(session_id, issues, opts)
          {:ok, report}

        report ->
          {:ok, report}
      end
    end
  end

  @doc """
  Quick analysis - runs all detectors without AI polishing.

  Useful for programmatic access to raw issue data.
  """
  @spec quick_analyze(String.t(), keyword()) :: {:ok, map()} | {:error, term()}
  def quick_analyze(path, opts \\ []) do
    with {:ok, abs_path} <- validate_path(path),
         {:ok, _} <- analyze_codebase(abs_path, opts),
         {:ok, issues} <- discover_issues(abs_path, opts) do
      {:ok, %{issues: issues, summary: build_summary(issues)}}
    end
  end

  @doc """
  List all active agent sessions.
  """
  @spec list_sessions(keyword()) :: [map()]
  def list_sessions(opts \\ []) do
    Memory.list_sessions(opts)
    |> Enum.map(fn session ->
      %{
        id: session.id,
        project_path: session.metadata[:project_path],
        created_at: session.created_at,
        message_count: length(session.messages)
      }
    end)
  end

  @doc """
  Get session details.
  """
  @spec get_session(String.t()) :: {:ok, map()} | {:error, :not_found}
  def get_session(session_id) do
    with {:ok, session} <- Memory.get_session(session_id) do
      {:ok,
       %{
         id: session.id,
         project_path: session.metadata[:project_path],
         created_at: session.created_at,
         updated_at: session.updated_at,
         message_count: length(session.messages),
         has_report: not is_nil(session.metadata[:report]),
         issues_summary: build_summary(session.metadata[:issues] || %{})
       }}
    end
  end

  @doc """
  Clear/end a session.
  """
  @spec clear_session(String.t()) :: :ok
  def clear_session(session_id) do
    Memory.clear_session(session_id)
  end

  # Private functions

  defp finalize_analysis(path, issues, opts) do
    if Keyword.get(opts, :skip_report, false) do
      summary = build_summary(issues)
      Logger.info("Analysis complete (report skipped): #{summary.total_issues} issues found")

      {:ok,
       %{
         session_id: nil,
         report: nil,
         ai_status: %{status: "skipped"},
         issues: issues,
         summary: summary
       }}
    else
      with {:ok, session} <- create_analysis_session(path, issues, opts),
           {:ok, report, ai_status} <- generate_report(session.id, issues, opts) do
        summary = build_summary(issues)
        Logger.info("Project analysis complete: #{summary.total_issues} issues found")

        {:ok,
         %{
           session_id: session.id,
           report: report,
           ai_status: ai_status,
           issues: issues,
           summary: summary
         }}
      end
    end
  end

  defp persist_all_state(issues, path) do
    # Eagerly save state to disk since Mix tasks don't trigger GenServer terminate/2.
    # Use the analyzed path as the cache key so different projects get separate caches.
    AnalysisCache.save(issues, path)
    EmbeddingsPersistence.save(nil, path)
    GraphPersistence.save(path)
  end

  defp validate_path(path) do
    abs_path = Path.expand(path)

    cond do
      not File.exists?(abs_path) ->
        {:error, {:path_not_found, path}}

      not File.dir?(abs_path) ->
        {:error, {:not_a_directory, path}}

      true ->
        {:ok, abs_path}
    end
  end

  defp analyze_codebase(path, opts) do
    Logger.info("Analyzing codebase structure...")

    exclude_patterns =
      Keyword.get(opts, :exclude_patterns, [
        "_build",
        "deps",
        "node_modules",
        ".git",
        ".elixir_ls",
        "cover",
        "priv/static"
      ])

    max_depth = Keyword.get(opts, :max_depth, 20)
    generate_embeddings = not Keyword.get(opts, :skip_embeddings, false)

    case Directory.analyze_directory(path,
           exclude_patterns: exclude_patterns,
           max_depth: max_depth,
           generate_embeddings: generate_embeddings
         ) do
      {:ok, result} ->
        Logger.info("Analyzed #{result.analyzed} files")
        {:ok, result}

      {:error, reason} ->
        Logger.error("Codebase analysis failed: #{inspect(reason)}")
        {:error, {:analysis_failed, reason}}
    end
  end

  defp discover_issues(path, opts) do
    Logger.info("Discovering issues...")

    include_suggestions = Keyword.get(opts, :include_suggestions, @include_suggestions)
    include_dead_code = Keyword.get(opts, :include_dead_code, @include_dead_code)

    # Run MetastaticBridge quality analysis first to populate QualityStore
    # so that find_complex queries below return actual data.
    Logger.info("Running quality analysis (MetastaticBridge)...")
    safe_analyze(&Quality.analyze_directory/2, [path, [store: true]])

    # Collect cyclomatic and cognitive complexity hotspots
    cyclomatic_complex =
      safe_analyze(&Quality.find_complex/1, [[metric: :cyclomatic, threshold: 10]])

    cognitive_complex =
      safe_analyze(&Quality.find_complex/1, [[metric: :cognitive, threshold: 15]])

    # Merge and deduplicate by path
    all_complex =
      (cyclomatic_complex ++ cognitive_complex)
      |> Enum.uniq_by(fn
        %{path: path} -> path
        item -> item
      end)

    issues = %{
      dead_code:
        if(include_dead_code,
          do: safe_analyze(&DeadCode.find_dead_code/0, []),
          else: []
        ),
      duplicates: safe_analyze(&Duplication.detect_in_directory/2, [path, [threshold: 0.8]]),
      security: safe_analyze(&Security.analyze_directory/2, [path, []]),
      smells: safe_analyze(&Smells.detect_smells/2, [path, []]),
      complexity: all_complex,
      circular_deps: safe_analyze(&DependencyGraph.find_cycles/1, [[]]),
      quality_metrics: safe_analyze(&Quality.statistics/0, [])
    }

    issues =
      if include_suggestions do
        Map.put(issues, :suggestions, safe_analyze(&Suggestions.analyze_target/2, [path, []]))
      else
        issues
      end

    {:ok, issues}
  end

  defp safe_analyze(func, args) do
    case apply(func, args) do
      {:ok, result} -> result
      {:error, _} -> []
      result when is_list(result) -> result
      result when is_map(result) -> result
      _ -> []
    end
  rescue
    e ->
      Logger.warning("Analysis function failed: #{Exception.message(e)}")
      []
  catch
    :exit, reason ->
      Logger.warning("Analysis function exited: #{inspect(reason)}")
      []
  end

  defp create_analysis_session(path, issues, _opts) do
    metadata = %{
      project_path: path,
      issues: issues,
      analyzed_at: DateTime.utc_now()
    }

    Memory.new_session(metadata)
  end

  defp generate_report(session_id, issues, opts) do
    Logger.info("Generating AI-polished report...")

    # Get project path from session metadata for path-aware system prompt
    project_path =
      case Memory.get_session(session_id) do
        {:ok, session} -> session.metadata[:project_path]
        _ -> nil
      end

    # Resolve provider info for AI status tracking
    provider_name = Keyword.get(opts, :provider) || AIConfig.provider_name()
    config = AIConfig.api_config(provider_name)

    if is_nil(config.api_key) or config.api_key == "" do
      Logger.info("No API key for #{provider_name}, skipping AI report")

      ai_status = %{
        status: "no_keys",
        provider: to_string(provider_name),
        model: config.model,
        error: "No API key configured for #{provider_name}"
      }

      {:ok, Report.generate_basic_report(issues), ai_status}
    else
      generate_ai_report(session_id, issues, opts, project_path, provider_name, config)
    end
  end

  defp generate_ai_report(session_id, issues, opts, project_path, provider_name, config) do
    setup_report_prompts(session_id, issues, project_path)

    # Run the report executor with NO tools so the LLM writes the report in a
    # single blocking call.  When tools are enabled DeepSeek R1 (and similar
    # reasoning models) places the full report in its thinking block and then
    # emits tool calls instead of the actual response, causing the executor to
    # loop until the dedup guard fires and force_text_response produces a broken
    # "data not provided" output.
    # Large context window keeps the full issues summary intact.
    report_opts =
      opts
      |> Keyword.put(:tools, [])
      |> Keyword.put(:context_max_chars, 128_000)

    # Run the agent to generate report
    case Executor.run(session_id, report_opts) do
      {:ok, result} ->
        # Save report to session metadata
        Memory.update_metadata(session_id, %{report: result.content})

        ai_status = %{
          status: "success",
          provider: to_string(provider_name),
          model: config.model,
          chars: String.length(result.content),
          tokens: result.usage
        }

        {:ok, result.content, ai_status}

      {:error, reason} ->
        Logger.error("Report generation failed: #{inspect(reason)}")

        ai_status = %{
          status: "failed",
          provider: to_string(provider_name),
          model: config.model,
          error: inspect(reason)
        }

        error_note = "[AI report generation failed: #{inspect(reason)}]\n\n"
        {:ok, error_note <> Report.generate_basic_report(issues), ai_status}
    end
  end

  defp setup_report_prompts(session_id, issues, project_path) do
    system_prompt = Report.system_prompt(project_path)
    Memory.add_message(session_id, :system, system_prompt)

    graph_stats = Store.stats()
    modules = Store.list_nodes(:module, :infinity)
    functions = Store.list_nodes(:function, :infinity)
    issues_summary = Report.format_issues_for_llm(issues)

    # Derive file counts from the analysis results themselves because the
    # knowledge graph may be sparse when only a subdirectory was analysed.
    quality_total = get_in(issues, [:quality_metrics, :total_files]) || 0
    sec_count = length(issues[:security] || [])

    module_count = max(length(modules), quality_total)
    func_count = length(functions)

    user_prompt = """
    COMPLETE analysis data for the audit report is below.  Write the full
    12-section Markdown report now using exclusively this data.
    Do NOT ask for more data.  Sections with zero findings should state
    "No issues detected" — that is a positive outcome.

    ## Analysis Scope

    - Project path: #{project_path || "unknown"}
    - Files analyzed (quality): #{quality_total}
    - Files scanned (security): #{sec_count}
    - Modules in knowledge graph: #{module_count}
    - Functions in knowledge graph: #{func_count}
    - Graph edges: #{graph_stats.edges}
    - Embeddings: #{graph_stats.embeddings}
    - Audit date: #{DateTime.utc_now() |> DateTime.to_date() |> Date.to_string()}

    ## Analysis Results

    #{issues_summary}

    ## Analysis Thresholds Applied

    - Cyclomatic complexity threshold: 10 (flagged if >10)
    - Cognitive complexity threshold: 15 (flagged if >15)
    - Duplication similarity threshold: 80%
    - Dead code minimum confidence: 70%
    """

    Memory.add_message(session_id, :user, user_prompt)
  end

  @doc """
  Generate an AI audit report, optionally notifying a callback when ready.

  Requires the knowledge graph and embeddings to be populated first
  (call `analyze_project/2` with `skip_report: true`).

  The executor runs in blocking mode so that RAG tool calls (read_file,
  semantic_search, hybrid_search, etc.) are executed correctly before the
  final report is written.  Streaming parsers drop `tool_call` deltas, so
  a streaming executor would mis-identify preamble text as the final report
  and never execute the tool calls.

  The `:on_chunk` callback is fired once after the blocking run completes,
  with the full report content, so callers can use it as a completion signal
  (e.g. to stop a spinner).

  ## Options

  - `:on_chunk` - `(chunk -> :ok)` completion callback, fired once with
    `%{content: report_string}` when the report is ready
  - `:provider` - AI provider override
  - `:model` - Model override
  """
  @spec stream_generate_report(String.t(), map(), keyword()) ::
          {:ok, String.t(), map()} | {:error, term()}
  def stream_generate_report(path, issues, opts \\ []) do
    abs_path = Path.expand(path)

    with {:ok, session} <- create_analysis_session(abs_path, issues, opts) do
      provider_name = Keyword.get(opts, :provider) || AIConfig.provider_name()
      config = AIConfig.api_config(provider_name)

      if is_nil(config.api_key) or config.api_key == "" do
        basic = Report.generate_basic_report(issues)
        {:ok, basic, %{status: "no_keys", provider: to_string(provider_name)}}
      else
        setup_report_prompts(session.id, issues, abs_path)

        # No tools: same reasoning as generate_ai_report — reasoning models put
        # the report in their thinking block then emit tool calls, causing the
        # executor to loop and eventually emit a broken response.
        report_opts =
          opts
          |> Keyword.put(:tools, [])
          |> Keyword.delete(:on_chunk)
          |> Keyword.put(:context_max_chars, 128_000)

        case Executor.run(session.id, report_opts) do
          {:ok, result} ->
            Memory.update_metadata(session.id, %{report: result.content})

            # Notify caller that the report is ready (compatibility with on_chunk interface)
            if on_chunk = Keyword.get(opts, :on_chunk) do
              on_chunk.(%{content: result.content})
            end

            ai_status = %{
              status: "success",
              provider: to_string(provider_name),
              model: config.model,
              chars: String.length(result.content),
              tokens: result.usage
            }

            {:ok, result.content, ai_status}

          {:error, reason} ->
            Logger.error("Report generation failed: #{inspect(reason)}")
            error_note = "[AI report generation failed: #{inspect(reason)}]\n\n"
            basic = error_note <> Report.generate_basic_report(issues)

            if on_chunk = Keyword.get(opts, :on_chunk) do
              on_chunk.(%{content: basic})
            end

            {:ok, basic,
             %{status: "failed", provider: to_string(provider_name), error: inspect(reason)}}
        end
      end
    end
  end

  defp build_summary(issues) when is_map(issues) do
    quality = issues[:quality_metrics] || %{}

    %{
      dead_code_count: count_issues(issues[:dead_code]),
      duplicate_count: count_issues(issues[:duplicates]),
      security_count: count_issues(issues[:security]),
      smell_count: count_issues(issues[:smells]),
      complexity_count: count_issues(issues[:complexity]),
      circular_dep_count: count_issues(issues[:circular_deps]),
      suggestion_count: count_issues(issues[:suggestions]),
      quality_files_analyzed: Map.get(quality, :total_files, 0),
      total_issues:
        count_issues(issues[:dead_code]) +
          count_issues(issues[:duplicates]) +
          count_issues(issues[:security]) +
          count_issues(issues[:smells]) +
          count_issues(issues[:complexity]) +
          count_issues(issues[:circular_deps])
    }
  end

  defp build_summary(_), do: %{total_issues: 0}

  defp count_issues(nil), do: 0
  defp count_issues(issues) when is_list(issues), do: length(issues)
  defp count_issues(%{items: items}) when is_list(items), do: length(items)
  defp count_issues(%{count: count}) when is_integer(count), do: count
  defp count_issues(_), do: 0
end