A platform for community discussion. Free, open, simple.
# frozen_string_literal: true
module DiscourseAi
module Agents
class SubagentRunner
MAX_SUBAGENT_DEPTH = 2
MAX_PROMPT_LENGTH = 20_000
MAX_RESULT_TOKENS = 30_000
def self.resolve_model(agent_record)
resolve_models([agent_record])[agent_record.id]
end
def self.resolve_models(agent_records)
site_model_id = SiteSetting.ai_default_llm_model.presence&.to_i
model_ids = agent_records.filter_map(&:default_llm_id)
model_ids << site_model_id if site_model_id
models_by_id = LlmModel.where(id: model_ids.uniq).index_by(&:id)
agent_records.each_with_object({}) do |agent_record, result|
model = models_by_id[agent_record.default_llm_id]
model ||= models_by_id[site_model_id] if !agent_record.force_default_llm?
result[agent_record.id] = model if model
end
end
def initialize(parent_agent:, child_id:, prompt:, context:, parent_llm:)
@parent_agent = parent_agent
@child_id =
if child_id.is_a?(Integer)
child_id
elsif child_id.is_a?(String) && child_id.match?(/\A-?\d+\z/)
child_id.to_i
end
@prompt = prompt
@context = context
@parent_llm = parent_llm
end
def run
child_record = nil
validation = validate_request
return validation if validation
child_record = AiAgent.find_by(id: @child_id)
child_class = child_record&.class_instance
return error("unavailable_agent") if !usable_child?(child_record, child_class)
model = self.class.resolve_model(child_record)
return error("missing_model", child_record) if !model
state = @context.subagent_execution_state
return error("cancelled", child_record) if @context.cancel_manager&.cancelled?
return error("token_limit", child_record) if state.remaining_tokens <= 0
yield(@prompt) if block_given?
child_agent = child_class.new
child_context = build_child_context(child_record)
response = collect_response(child_agent, child_context, model)
return error("completion_limit", child_record) if child_context.completion_limit_reached
truncate_response(response, child_record)
rescue => exception
return error("cancelled", child_record) if @context.cancel_manager&.cancelled?
if exception.is_a?(DiscourseAi::Completions::ContextPreparation::Error) &&
exception.reason == "turn_budget_exhausted"
result = error("token_limit", child_record)
if exception.raw_context.present?
result.merge!(
truncate_response(exception.raw_context.to_json, child_record).slice(
:response,
:truncated,
),
)
end
return result
end
Rails.logger.error(
"DiscourseAi: subagent failed parent=#{@parent_agent.id} child=#{@child_id}: #{exception.class}: #{exception.message}",
)
error("execution_failed")
end
private
def validate_request
return error("invalid_agent") if @child_id.nil?
return error("invalid_prompt") if !@prompt.is_a?(String) || @prompt.blank?
return error("prompt_too_long") if @prompt.length > MAX_PROMPT_LENGTH
return error("missing_user") if @context.user.nil?
return error("depth_limit") if @context.subagent_depth.to_i >= MAX_SUBAGENT_DEPTH
parent_record = AiAgent.find_by(id: @parent_agent.id)
return error("unavailable_parent") if !parent_record
return error("agent_not_allowed") if !parent_record.subagent_ids.include?(@child_id)
if !@context.subagent_execution_state&.completion_available?
return error("completion_limit")
end
error("token_limit") if @context.subagent_execution_state.remaining_tokens <= 0
end
def usable_child?(child_record, child_class)
return false if !child_record&.enabled? || !child_class
return false if !@context.user.in_any_groups?(child_record.allowed_group_ids)
return true if !child_record.system?
required_tools = child_class.new.required_tools
required_tools.empty? || (required_tools - Agent.all_available_tools).empty?
end
def build_child_context(child_record)
child_depth = @context.subagent_depth.to_i + 1
feature_context =
@context.feature_context.to_h.merge(
"subagent_parent_agent_id" => @parent_agent.id,
"subagent_agent_id" => child_record.id,
"subagent_depth" => child_depth,
)
child_budget = Bot.effective_max_turn_tokens(child_record.max_turn_tokens)
effective_budget = [child_budget, @context.subagent_execution_state.remaining_tokens].min
child_execution = @context.execution_context.dup
child_execution.work_budget =
@context.execution_context.work_budget.child(limit: effective_budget)
child_context =
BotContext.new(
messages: [{ type: :user, content: @prompt }],
participants: @context.participants,
user: @context.user,
skip_show_thinking: true,
feature_context: feature_context,
message_id: @context.message_id,
channel_id: @context.channel_id,
context_post_ids: @context.context_post_ids,
feature_name: @context.feature_name,
resource_url: @context.resource_url,
site_url: @context.site_url,
site_title: @context.site_title,
site_description: @context.site_description,
time: @context.time,
cancel_manager: @context.cancel_manager,
inferred_concepts: @context.inferred_concepts,
format_dates: @context.format_dates,
bypass_response_format: false,
guardian: @context.guardian,
server_owned_tools: @context.server_owned_tools,
execution_context: child_execution,
subagent_execution_state: @context.subagent_execution_state,
subagent_depth: child_depth,
parent_agent_id: @parent_agent.id,
current_agent_id: child_record.id,
turn_token_budget: effective_budget,
)
child_context.topic_id = @context.topic_id
child_context.post_id = @context.post_id
child_context.private_message = @context.private_message?
child_context.temporal_context = @context.temporal_context
child_context.user_language = @context.user_language
child_context
end
def collect_response(child_agent, child_context, model)
streamed = +""
structured = nil
bot = Bot.as(@context.user, agent: child_agent, model: model)
raw_context =
bot.reply(
child_context,
execution_context: child_context.execution_context,
) do |partial, _, type|
if type == :structured_output
structured = partial
elsif type.blank? && partial.is_a?(String)
streamed << partial
end
end
return structured.to_s if structured
final_response =
raw_context.reverse.find do |entry|
entry.is_a?(Array) && entry[2].nil? && entry.first.is_a?(String)
end
final_response&.first || streamed
end
def truncate_response(response, child_record)
response = response.to_s
tokenizer = @parent_llm.tokenizer
truncated = tokenizer.size(response) > MAX_RESULT_TOKENS
if truncated
response =
tokenizer.truncate(
response,
MAX_RESULT_TOKENS,
strict: SiteSetting.ai_strict_token_counting,
)
end
{
status: "ok",
agent_id: child_record.id,
agent_name: child_record.class_instance.name,
response: response,
truncated: truncated,
}
end
def error(key, child_record = nil)
result = { status: "error", error: I18n.t("discourse_ai.ai_bot.subagent_errors.#{key}") }
if child_record
result[:agent_id] = child_record.id
result[:agent_name] = child_record.class_instance.name
elsif @child_id
result[:agent_id] = @child_id
end
result
end
end
end
end