rubino-agent 0.3.0

data/lib/rubino/llm/scenario_loader.rb

@@ -0,0 +1,68 @@
+# frozen_string_literal: true
+
+require "yaml"
+
+module Rubino
+  module LLM
+    # Loads a fake-provider scenario YAML by name and returns the parsed
+    # event list. The fake provider drives a deterministic chunk/tool-call
+    # stream from these events so the UI and tool plumbing can be exercised
+    # without burning provider tokens.
+    #
+    # Search order:
+    #   1. <Rubino.configuration.fake_scenarios_dir>/<name>.yml
+    #   2. <gem_root>/lib/rubino/llm/scenarios/<name>.yml
+    #
+    # The YAML file must contain a top-level `events:` key with an array
+    # of event hashes — see scenarios/happy-path.yml for the shape.
+    class ScenarioLoader
+      class NotFound < Rubino::Error; end
+
+      DEFAULT_DIR = File.expand_path("scenarios", __dir__)
+
+      def self.load(name, scenarios_dir: nil)
+        new(scenarios_dir: scenarios_dir).load(name)
+      end
+
+      def initialize(scenarios_dir: nil)
+        @scenarios_dir = scenarios_dir || configured_dir
+      end
+
+      # Returns the array of event hashes under the YAML `events:` key.
+      # Raises NotFound when the scenario can't be located under either
+      # path, citing both so the operator can fix the misconfiguration.
+      def load(name)
+        path = resolve_path(name)
+        raise NotFound, build_not_found_message(name) unless path
+
+        data = YAML.safe_load_file(path, permitted_classes: [Symbol], aliases: true) || {}
+        Array(data["events"] || data[:events])
+      end
+
+      private
+
+      def resolve_path(name)
+        filename = "#{name}.yml"
+        [@scenarios_dir, DEFAULT_DIR].compact.uniq.each do |dir|
+          candidate = File.join(dir, filename)
+          return candidate if File.file?(candidate)
+        end
+        nil
+      end
+
+      def configured_dir
+        cfg = Rubino.configuration
+        return nil unless cfg.respond_to?(:dig)
+
+        cfg.dig("fake_provider", "scenarios_dir") || cfg.dig("providers", "fake", "scenarios_dir")
+      rescue StandardError
+        nil
+      end
+
+      def build_not_found_message(name)
+        tried = [@scenarios_dir, DEFAULT_DIR].compact.uniq.map { |d| File.join(d, "#{name}.yml") }
+        "fake scenario '#{name}' not found. Tried:\n  #{tried.join("\n  ")}"
+      end
+    end
+  end
+end

data/lib/rubino/llm/scenario_selector.rb

@@ -0,0 +1,80 @@
+# frozen_string_literal: true
+
+module Rubino
+  module LLM
+    # Maps a user-supplied prompt to a fake scenario name using a keyword
+    # router ported verbatim from the reference fake-provider SCENARIO_ROUTER.
+    # The ordering of the rules is
+    # significant — see the notes block in fake-provider-spec.md before
+    # touching anything here.
+    class ScenarioSelector
+      ROUTER = [
+        { keywords: ["simula quota exceeded", "quota exceeded", "provider quota"],
+          scenario: "provider-quota-completed" },
+
+        { keywords: ["cron fail", "cron failure", "cron error", "broken cron",
+                     "failing cron", "fallisce cron", "cron fallisce"],
+          scenario: "agent-creates-cron-failure" },
+
+        { keywords: ["crea cron", "create cron", "create a cron", "cron job",
+                     "scheduled task", "schedula", "pianifica",
+                     "schedule daily", "schedule every", "schedule a daily",
+                     "schedule a weekly", "daily cron", "weekly cron",
+                     "every minute cron", "cron giornaliero", "report giornaliero",
+                     "with-cron-success"],
+          scenario: "agent-creates-cron" },
+
+        { keywords: %w[approve approval autorizza conferma permit allow],
+          scenario: "with-approvals" },
+
+        { keywords: ["artifact", "generate report", "create file", "crea file",
+                     "genera report", "write file", "save as"],
+          scenario: "with-artifacts" },
+
+        { keywords: ["upload", "allegato", "allega", "attached file",
+                     "read this file", "analizza file"],
+          scenario: "with-uploads" },
+
+        { keywords: %w[fail error errore fallito broken crash crash],
+          scenario: "failure" },
+
+        { keywords: ["complex-analysis", "report", "analysis report", "comprehensive",
+                     "full refactor", "refactoring completo", "multi-tool",
+                     "multi step analysis"],
+          scenario: "complex-analysis" },
+
+        { keywords: ["multi-step", "analizza", "analyze", "refactor", "debug",
+                     "investigate", "ricerca", "search", "trova",
+                     "how would you", "what do you think"],
+          scenario: "with-reasoning" },
+
+        { keywords: ["analisi", "analysis", "spiegami", "explain in detail",
+                     "descrivi l'architettura", "tell me about",
+                     "how does it work", "full breakdown"],
+          scenario: "analysis" },
+
+        { keywords: ["which approach", "which option", "what do you think",
+                     "choose between", "help me decide", "clarify",
+                     "non so quale", "quale scegliere", "aiutami a scegliere"],
+          scenario: "with-clarify" }
+      ].freeze
+
+      # Resolves the input text to a scenario name. First keyword match wins;
+      # `default:` is returned when nothing matches. Matching is case-insensitive
+      # on both the input and the keyword list. Blank/nil input → default.
+      def self.resolve(input_text, default: "happy-path")
+        return default if input_text.nil?
+
+        text = input_text.to_s.downcase
+        return default if text.strip.empty?
+
+        ROUTER.each do |rule|
+          rule[:keywords].each do |kw|
+            return rule[:scenario] if text.include?(kw.downcase)
+          end
+        end
+        default
+      end
+    end
+  end
+end

data/lib/rubino/llm/scenarios/agent-creates-cron-failure.yml

@@ -0,0 +1,29 @@
+# Agent-creates-cron-failure scenario: the LLM registers a cron job that
+# is wired to the with-cron-failure scenario, so "Run now" exercises the
+# failure path in the real pipeline (run.failed, no delivery).
+#
+# Ported from the reference agent-creates-cron-failure scenario. The fake
+# provider only produces what the LLM would say or call; the cron registration,
+# token expansion, tracing and run accounting all happen downstream once the
+# real `cronjob` tool runs.
+events:
+  - type: thinking
+    text: "User wants a cron that will fail when triggered — useful to test the failure UI."
+  - type: delay_seconds
+    value: 0.05
+
+  - type: content
+    text: "Scheduled **Flaky Job** (every 1m) — Run now will exercise the failure path."
+  - type: delay_seconds
+    value: 0.05
+
+  - type: tool_call
+    id: "call_fake_cron_failure_1"
+    name: "cronjob"
+    arguments:
+      action: "create"
+      name: "Flaky Job"
+      schedule: "every 1m"
+      prompt: "Run a flaky job that fails deterministically."
+      deliver: "web"
+      scenario: "with-cron-failure"

data/lib/rubino/llm/scenarios/agent-creates-cron.yml

@@ -0,0 +1,36 @@
+# Agent-creates-cron scenario: the LLM emits a single tool_call to the real
+# `cronjob` tool (action="create"), exactly as the reference does in its
+# cronjob tooling. Everything downstream (approval,
+# scheduler registration, artifacts, trace token) runs on the real pipeline
+# once the tool_call is consumed.
+#
+# Translated from the reference agent-creates-cron scenario — see the
+# translation rules in the fake-provider scenario spec.
+events:
+  - type: thinking
+    text: "User wants a recurring task."
+  - type: delay_seconds
+    value: 0.05
+  - type: thinking
+    text: " I'll create a cron job with a 1-minute interval so it's easy to see in dev."
+  - type: delay_seconds
+    value: 0.05
+
+  - type: content
+    text: "Scheduling **Daily Summary** (every 1m). "
+  - type: delay_seconds
+    value: 0.05
+  - type: content
+    text: "Look for the trace token in the Cron Jobs index entry and in every session title produced by Run now."
+  - type: delay_seconds
+    value: 0.05
+
+  - type: tool_call
+    id: "call_fake_cronjob_1"
+    name: "cronjob"
+    arguments:
+      action: "create"
+      name: "Daily Summary"
+      schedule: "every 1m"
+      prompt: "Summarize yesterday's activity and produce a 3-bullet recap."
+      deliver: "web"

data/lib/rubino/llm/scenarios/analysis.yml

@@ -0,0 +1,501 @@
+# Analysis scenario: model streams long reasoning token-by-token,
+# reads a file via the real `read` tool, then continues reasoning.
+# Mirrors the reference analysis scenario in the FakeProvider vocabulary.
+# The fake provider terminates when events run out, so no completion event
+# is needed.
+events:
+  - type: thinking
+    text: "Let"
+  - type: thinking
+    text: " me"
+  - type: thinking
+    text: " analyze"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " full"
+  - type: thinking
+    text: " architecture"
+  - type: thinking
+    text: " step"
+  - type: thinking
+    text: " by"
+  - type: thinking
+    text: " step."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "First,"
+  - type: thinking
+    text: " let"
+  - type: thinking
+    text: " me"
+  - type: thinking
+    text: " understand"
+  - type: thinking
+    text: " what"
+  - type: thinking
+    text: " we"
+  - type: thinking
+    text: " are"
+  - type: thinking
+    text: " looking"
+  - type: thinking
+    text: " at."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "The"
+  - type: thinking
+    text: " system"
+  - type: thinking
+    text: " has"
+  - type: thinking
+    text: " multiple"
+  - type: thinking
+    text: " layers"
+  - type: thinking
+    text: " that"
+  - type: thinking
+    text: " interact"
+  - type: thinking
+    text: " in"
+  - type: thinking
+    text: " specific"
+  - type: thinking
+    text: " ways."
+
+  - type: delay_seconds
+    value: 0.8
+
+  - type: thinking
+    text: "\n\n"
+  - type: thinking
+    text: "Layer"
+  - type: thinking
+    text: " one:"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " Acme"
+  - type: thinking
+    text: " API"
+  - type: thinking
+    text: " server."
+  - type: thinking
+    text: "\nIt"
+  - type: thinking
+    text: " exposes"
+  - type: thinking
+    text: " endpoints"
+  - type: thinking
+    text: " for"
+  - type: thinking
+    text: " creating"
+  - type: thinking
+    text: " runs,"
+  - type: thinking
+    text: " streaming"
+  - type: thinking
+    text: " events,"
+  - type: thinking
+    text: " and"
+  - type: thinking
+    text: " handling"
+  - type: thinking
+    text: " approvals."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "It"
+  - type: thinking
+    text: " uses"
+  - type: thinking
+    text: " aiohttp"
+  - type: thinking
+    text: " for"
+  - type: thinking
+    text: " async"
+  - type: thinking
+    text: " HTTP"
+  - type: thinking
+    text: " and"
+  - type: thinking
+    text: " SSE"
+  - type: thinking
+    text: " for"
+  - type: thinking
+    text: " real-time"
+  - type: thinking
+    text: " communication."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "The"
+  - type: thinking
+    text: " key"
+  - type: thinking
+    text: " endpoints"
+  - type: thinking
+    text: " are:"
+  - type: thinking
+    text: " POST"
+  - type: thinking
+    text: " /v1/runs,"
+  - type: thinking
+    text: " GET"
+  - type: thinking
+    text: " /v1/runs/{id}"
+  - type: thinking
+    text: "/events,"
+  - type: thinking
+    text: " POST"
+  - type: thinking
+    text: " /v1/runs/{id}"
+  - type: thinking
+    text: "/stop,"
+  - type: thinking
+    text: " and"
+  - type: thinking
+    text: " POST"
+  - type: thinking
+    text: " /v1/runs/{id}"
+  - type: thinking
+    text: "/approvals"
+  - type: thinking
+    text: "/{aid}."
+
+  - type: delay_seconds
+    value: 1.0
+
+  - type: thinking
+    text: "\n\n"
+  - type: thinking
+    text: "Layer"
+  - type: thinking
+    text: " two:"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " Rails"
+  - type: thinking
+    text: " UI."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "It"
+  - type: thinking
+    text: " connects"
+  - type: thinking
+    text: " via"
+  - type: thinking
+    text: " AcmeApiService"
+  - type: thinking
+    text: " which"
+  - type: thinking
+    text: " wraps"
+  - type: thinking
+    text: " all"
+  - type: thinking
+    text: " HTTP"
+  - type: thinking
+    text: " calls"
+  - type: thinking
+    text: " to"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " API"
+  - type: thinking
+    text: " server."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "AcmeRunJob"
+  - type: thinking
+    text: " processes"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " SSE"
+  - type: thinking
+    text: " stream"
+  - type: thinking
+    text: " and"
+  - type: thinking
+    text: " persists"
+  - type: thinking
+    text: " events"
+  - type: thinking
+    text: " as"
+  - type: thinking
+    text: " RunEvent"
+  - type: thinking
+    text: " records."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "The"
+  - type: thinking
+    text: " timeline"
+  - type: thinking
+    text: " rendering"
+  - type: thinking
+    text: " reads"
+  - type: thinking
+    text: " from"
+  - type: thinking
+    text: " RunEvent"
+  - type: thinking
+    text: " via"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " build_items"
+  - type: thinking
+    text: " method."
+
+  - type: delay_seconds
+    value: 1.2
+
+  - type: tool_call
+    name: read
+    arguments:
+      file_path: "/app/app/jobs/acme_run_job.rb"
+
+  - type: delay_seconds
+    value: 1.0
+
+  - type: thinking
+    text: "\n\n"
+
+  - type: thinking
+    text: "Looking"
+  - type: thinking
+    text: " at"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " job"
+  - type: thinking
+    text: " code,"
+  - type: thinking
+    text: " I"
+  - type: thinking
+    text: " can"
+  - type: thinking
+    text: " see"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " refactoring"
+  - type: thinking
+    text: " removed"
+  - type: thinking
+    text: " content_buffer"
+  - type: thinking
+    text: " and"
+  - type: thinking
+    text: " reasoning_buffer."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "These"
+  - type: thinking
+    text: " were"
+  - type: thinking
+    text: " accumulating"
+  - type: thinking
+    text: " text"
+  - type: thinking
+    text: " in"
+  - type: thinking
+    text: " memory"
+  - type: thinking
+    text: " and"
+  - type: thinking
+    text: " flushing"
+  - type: thinking
+    text: " to"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " database."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "Now"
+  - type: thinking
+    text: " RunEvent"
+  - type: thinking
+    text: " is"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " single"
+  - type: thinking
+    text: " source"
+  - type: thinking
+    text: " of"
+  - type: thinking
+    text: " truth"
+  - type: thinking
+    text: " for"
+  - type: thinking
+    text: " both"
+  - type: thinking
+    text: " content"
+  - type: thinking
+    text: " and"
+  - type: thinking
+    text: " reasoning."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "The"
+  - type: thinking
+    text: " flush_now"
+  - type: thinking
+    text: " method"
+  - type: thinking
+    text: " now"
+  - type: thinking
+    text: " only"
+  - type: thinking
+    text: " handles"
+  - type: thinking
+    text: " tool_calls_cache"
+  - type: thinking
+    text: " for"
+  - type: thinking
+    text: " real-time"
+  - type: thinking
+    text: " tool"
+  - type: thinking
+    text: " status"
+  - type: thinking
+    text: " updates."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "Everything"
+  - type: thinking
+    text: " else"
+  - type: thinking
+    text: " is"
+  - type: thinking
+    text: " event-driven."
+
+  - type: delay_seconds
+    value: 0.8
+
+  - type: thinking
+    text: "\n\n"
+  - type: thinking
+    text: "This"
+  - type: thinking
+    text: " changes"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " data"
+  - type: thinking
+    text: " flow"
+  - type: thinking
+    text: " significantly."
+  - type: thinking
+    text: "\n"
+  - type: thinking
+    text: "When"
+  - type: thinking
+    text: " a"
+  - type: thinking
+    text: " user"
+  - type: thinking
+    text: " sends"
+  - type: thinking
+    text: " a"
+  - type: thinking
+    text: " message:"
+  - type: thinking
+    text: "\n1."
+  - type: thinking
+    text: " AcmeRunJob"
+  - type: thinking
+    text: " starts"
+  - type: thinking
+    text: " and"
+  - type: thinking
+    text: " calls"
+  - type: thinking
+    text: " POST"
+  - type: thinking
+    text: " /v1/runs"
+  - type: thinking
+    text: "\n2."
+  - type: thinking
+    text: " Acme"
+  - type: thinking
+    text: " returns"
+  - type: thinking
+    text: " a"
+  - type: thinking
+    text: " run_id"
+  - type: thinking
+    text: "\n3."
+  - type: thinking
+    text: " The"
+  - type: thinking
+    text: " job"
+  - type: thinking
+    text: " connects"
+  - type: thinking
+    text: " to"
+  - type: thinking
+    text: " the"
+  - type: thinking
+    text: " SSE"
+  - type: thinking
+    text: " stream"
+  - type: thinking
+    text: "\n4."
+  - type: thinking
+    text: " Events"
+  - type: thinking
+    text: " arrive"
+  - type: thinking
+    text: " and"
+  - type: thinking
+    text: " are"
+  - type: thinking
+    text: " persisted"
+  - type: thinking
+    text: "\n5."
+  - type: thinking
+    text: " On"
+  - type: thinking
+    text: " run.completed,"
+  - type: thinking
+    text: " final_output"
+  - type: thinking
+    text: " is"
+  - type: thinking
+    text: " set"
+  - type: thinking
+    text: "\n6."
+  - type: thinking
+    text: " Timeline"
+  - type: thinking
+    text: " renders"
+  - type: thinking
+    text: " from"
+  - type: thinking
+    text: " RunEvent"
+  - type: thinking
+    text: " records"
+
+  - type: delay_seconds
+    value: 1.5
+
+  - type: thinking
+    text: "\n\n"