Building a Multi-Agent Debate Simulation

if (!requireNamespace("pacman", quietly = TRUE)) {
  install.packages("pacman")
}

suppressPackageStartupMessages(
  pacman::p_load(
    ellmer,
    glue,
    purrr,
    jsonlite
  )
)

# Base URL for candidate questionnaires on GitHub
QUESTIONNAIRE_BASE_URL <- "https://raw.githubusercontent.com/jaytimm/newmexico-political-data/main/data/abq_mayor_2025_questionnaires"

# Candidate file mapping with GitHub URLs
CANDIDATE_FILES <- list(
  "Daniel Chavez" = paste0(QUESTIONNAIRE_BASE_URL, "/Official.Daniel.Chavez.Response_250609.md"),
  "Darren White" = paste0(QUESTIONNAIRE_BASE_URL, "/Official.White_.Darren.Response._250609.md"),
  "Mayling Armijo" = paste0(QUESTIONNAIRE_BASE_URL, "/Armijo.Mayling.Response_250609.md"),
  "Eddie Varela" = paste0(QUESTIONNAIRE_BASE_URL, "/Eddie-Varela-Mayoral-Candidate-Response.md"),
  "Louie Sanchez" = paste0(QUESTIONNAIRE_BASE_URL, "/Sanchez.Louie_.Response_250609.md")
)

# Load all candidate questionnaires from GitHub
candidates_data <- map(names(CANDIDATE_FILES), function(name) {
  url <- CANDIDATE_FILES[[name]]
  tryCatch({
    questionnaire_text <- readLines(url, warn = FALSE) |> paste(collapse = "\n")
    list(
      name = name,
      file = basename(url),
      questionnaire = questionnaire_text
    )
  }, error = function(e) {
    warning("Failed to load questionnaire for ", name, ": ", e$message)
    return(NULL)
  })
}) |>
  purrr::compact()

run_debate <- function(topics = debate_topics, max_rounds = 15) {
  topic_keys <- unlist(unname(topics))
  topic_names <- names(topics)
  
  transcript <- list()
  round <- 1
  current_topic_idx <- 1
  current_topic_key <- topic_keys[current_topic_idx]
  
  candidates_spoken <- map(topic_keys, ~ setNames(rep(FALSE, length(candidate_names)), candidate_names))
  names(candidates_spoken) <- topic_keys
  
  topic_list_text <- paste(topic_names, collapse = ", ")
  intro_text <- paste(
    "Good evening. Welcome to the Albuquerque Mayoral Debate.",
    "",
    paste("Tonight we have", length(candidate_names), "candidates:", 
          paste(candidate_names, collapse = ", ")),
    "",
    paste("This debate will focus on", length(topics), "critical issues:"),
    topic_list_text,
    "",
    "Let's begin.",
    sep = "\n"
  )
  transcript <- append(transcript, list(list(speaker = "MODERATOR", text = intro_text)))
  
  while (round <= max_rounds) {
    if (current_topic_idx <= length(topic_keys) && all(candidates_spoken[[current_topic_key]])) {
      current_topic_idx <- current_topic_idx + 1
      if (current_topic_idx > length(topic_keys)) break
      current_topic_key <- topic_keys[current_topic_idx]
    }
    
    current_transcript <- tail_transcript(transcript, n = 5)
    remaining <- names(candidates_spoken[[current_topic_key]])[!candidates_spoken[[current_topic_key]]]
    topic_status <- glue(
      "CURRENT TOPIC: {toupper(topic_names[current_topic_idx])}. ",
      "Remaining: {if(length(remaining)>0) paste(remaining, collapse=', ') else 'All have spoken'}"
    )
    
    # Check if anyone has spoken on current topic
    current_topic_entries <- transcript[which(sapply(transcript, function(e) {
      # Find when current topic started (after previous topic ended or debate started)
      topic_start_idx <- if (current_topic_idx == 1) 1 else {
        prev_topic_key <- topic_keys[current_topic_idx - 1]
        # Find last entry where all candidates spoke on previous topic
        last_prev_idx <- 0
        for (i in seq_along(transcript)) {
          if (transcript[[i]]$speaker != "MODERATOR" && 
              any(sapply(candidate_names, function(n) transcript[[i]]$speaker == n))) {
            last_prev_idx <- i
          }
        }
        last_prev_idx + 1
      }
      # Check if this entry is after topic started
      entry_idx <- which(sapply(seq_along(transcript), function(i) identical(transcript[[i]], e)))
      entry_idx >= topic_start_idx
    }))]
    
    has_spoken_on_topic <- any(sapply(current_topic_entries, function(e) 
      e$speaker != "MODERATOR" && e$speaker %in% candidate_names))
    
    if (has_spoken_on_topic) {
      orchestrator_input <- glue(
        "Current debate transcript:\n\n{current_transcript}\n\n",
        "{topic_status}\n\n",
        "Candidates have already spoken on this topic. Reference relevant things they said in the transcript. Build on their statements, ask follow-ups about their proposals, or challenge their ideas. Output your command in the required format."
      )
    } else {
      orchestrator_input <- glue(
        "Current debate transcript:\n\n{current_transcript}\n\n",
        "{topic_status}\n\n",
        "This is the FIRST question on this topic - no candidates have spoken yet. Ask a fresh opening question about {topic_names[current_topic_idx]}. Do NOT reference things candidates said (they haven't spoken on this topic yet). Output your command in the required format."
      )
    }
    
    command <- parse_orchestrator_command(orchestrator, orchestrator_input)
    
    if (command$action == "end_debate") break
    
    target_name <- command$candidate
    target_cand <- candidate_lookup[[target_name]]
    
    if (is.null(target_cand)) {
      remaining <- names(candidates_spoken[[current_topic_key]])[!candidates_spoken[[current_topic_key]]]
      if (length(remaining) == 0) remaining <- candidate_names
      target_name <- sample(remaining, 1)
      target_cand <- candidate_lookup[[target_name]]
    }
    
    context <- format_transcript(transcript)
    
    if (command$action == "request_rebuttal" && nchar(command$responding_to) > 0) {
      responding_to_name <- command$responding_to
      prompt_text <- command$moderator_statement
      full_prompt <- glue(
        "Here is the complete debate transcript so far:\n\n{context}\n\n",
        "{prompt_text}\n\n",
        "You are responding directly to {responding_to_name}. Be conversational and debate-like.",
        "CRITICAL: You can ONLY reference what other candidates have said in the transcript above."
      )
      response <- target_cand$agent$chat(full_prompt)
      transcript <- append(transcript, list(list(speaker = target_name, text = response)))
    } else {
      moderator_text <- command$moderator_statement
      transcript <- append(transcript, list(list(speaker = "MODERATOR", text = moderator_text)))
      full_prompt <- glue(
        "Here is the complete debate transcript so far:\n\n{context}\n\n",
        "MODERATOR: {moderator_text}\n\n",
        "CRITICAL: You can ONLY reference what other candidates have said in the transcript above. ",
        "Engage with what other candidates have said - challenge their ideas, disagree with their proposals, or build on their statements. ",
        "Do not reference ideas or positions that haven't been mentioned in this debate yet. ",
        "Respond naturally based on the transcript and your own positions."
      )
      response <- target_cand$agent$chat(full_prompt)
      transcript <- append(transcript, list(list(speaker = target_name, text = response)))
    }
    
    candidates_spoken[[current_topic_key]][target_name] <- TRUE
    
    if (all(map_lgl(candidates_spoken, ~ all(.x)))) break
    
    round <- round + 1
    Sys.sleep(0.5)
  }
  
  # Moderator closing statement
  closing_text <- "Thank you to all the candidates for participating in tonight's debate. We appreciate your time and perspectives on these critical issues facing Albuquerque."
  transcript <- append(transcript, list(list(speaker = "MODERATOR", text = closing_text)))
  
  return(transcript)
}