decision_agent 0.1.1 → 0.1.3

data/spec/thread_safety_spec.rb.broken

@@ -0,0 +1,878 @@
+# frozen_string_literal: true
+# encoding: UTF-8
+
+require "spec_helper"
+
+RSpec.describe "Thread-Safety" do
+  describe "Agent with shared evaluators" do
+    let(:rules_json) do
+      {
+        version: "1.0",
+        ruleset: "approval_rules",
+        rules: [
+          {
+            id: "approve_high",
+            if: { field: "amount", op: "gt", value: 1000 },
+            then: { decision: "approve", weight: 0.9, reason: "High value" }
+          },
+          {
+            id: "reject_low",
+            if: { field: "amount", op: "lte", value: 1000 },
+            then: { decision: "reject", weight: 0.8, reason: "Low value" }
+          }
+        ]
+      }
+    end
+
+    let(:evaluator) { DecisionAgent::Evaluators::JsonRuleEvaluator.new(rules_json: rules_json) }
+    let(:agent) { DecisionAgent::Agent.new(evaluators: [evaluator]) }
+
+    it "handles concurrent decisions from multiple threads safely" do
+      threads = []
+      results = Array.new(50)
+
+      # Create 50 threads making concurrent decisions
+      50.times do |i|
+        threads << Thread.new do
+          context = { amount: (i % 2 == 0) ? 1500 : 500 }
+          results[i] = agent.decide(context: context)
+        end
+      end
+
+      threads.each(&:join)
+
+      # Verify all threads completed successfully
+      expect(results.compact.size).to eq(50)
+
+      # Verify results are correct and frozen
+      results.each_with_index do |decision, i|
+        expect(decision).to be_frozen
+        expect(decision.decision).to be_frozen
+        expect(decision.explanations).to be_frozen
+        expect(decision.evaluations).to be_frozen
+        expect(decision.audit_payload).to be_frozen
+
+        # Verify correctness based on input
+        if i % 2 == 0
+          expect(decision.decision).to eq("approve")
+        else
+          expect(decision.decision).to eq("reject")
+        end
+      end
+    end
+
+    it "prevents modification of shared evaluator ruleset" do
+      # Verify the ruleset is frozen
+      expect(evaluator.instance_variable_get(:@ruleset)).to be_frozen
+
+      # Attempt to modify should raise error
+      expect {
+        evaluator.instance_variable_get(:@ruleset)["rules"] << { id: "new_rule" }
+      }.to raise_error(FrozenError)
+    end
+
+    it "prevents modification of evaluators array in Agent" do
+      expect(agent.evaluators).to be_frozen
+
+      expect {
+        agent.evaluators << DecisionAgent::Evaluators::StaticEvaluator.new(decision: true, weight: 1.0)
+      }.to raise_error(FrozenError)
+    end
+  end
+
+  describe "Multiple agents sharing evaluators" do
+    let(:evaluator) do
+      DecisionAgent::Evaluators::JsonRuleEvaluator.new(
+        rules_json: {
+          version: "1.0",
+          ruleset: "shared_rules",
+          rules: [
+            {
+              id: "rule1",
+              if: { field: "value", op: "eq", value: "yes" },
+              then: { decision: "approve", weight: 1.0, reason: "Match" }
+            }
+          ]
+        }
+      )
+    end
+
+    it "allows multiple agents to safely share the same evaluator instance" do
+      agent1 = DecisionAgent::Agent.new(evaluators: [evaluator])
+      agent2 = DecisionAgent::Agent.new(evaluators: [evaluator])
+      agent3 = DecisionAgent::Agent.new(evaluators: [evaluator])
+
+      threads = []
+      results = []
+
+      # Each agent makes decisions in parallel
+      [agent1, agent2, agent3].each do |agent|
+        threads << Thread.new do
+          10.times do
+            decision = agent.decide(context: { value: "yes" })
+            mutex.synchronize { results << decision }
+          end
+        end
+      end
+
+      threads.each(&:join)
+
+      # All 30 decisions should succeed
+      expect(results.size).to eq(30)
+      results.each do |decision|
+        expect(decision.decision).to eq("approve")
+        expect(decision).to be_frozen
+      end
+    end
+  end
+
+  describe "Evaluation immutability" do
+    it "ensures evaluations are deeply frozen" do
+      evaluation = DecisionAgent::Evaluation.new(
+        decision: "approve",
+        weight: 0.8,
+        reason: "Test reason",
+        evaluator_name: "TestEvaluator",
+        metadata: { key: "value" }
+      )
+
+      expect(evaluation).to be_frozen
+      expect(evaluation.decision).to be_frozen
+      expect(evaluation.reason).to be_frozen
+      expect(evaluation.evaluator_name).to be_frozen
+      expect(evaluation.metadata).to be_frozen
+    end
+  end
+
+  describe "Decision immutability" do
+    it "ensures decisions are deeply frozen" do
+      evaluation = DecisionAgent::Evaluation.new(
+        decision: "approve",
+        weight: 1.0,
+        reason: "Test",
+        evaluator_name: "Test"
+      )
+
+      decision = DecisionAgent::Decision.new(
+        decision: "approve",
+        confidence: 0.95,
+        explanations: ["Explanation 1"],
+        evaluations: [evaluation],
+        audit_payload: { timestamp: "2024-01-01" }
+      )
+
+      expect(decision).to be_frozen
+      expect(decision.decision).to be_frozen
+      expect(decision.explanations).to be_frozen
+      expect(decision.evaluations).to be_frozen
+      expect(decision.audit_payload).to be_frozen
+
+      # Nested structures should also be frozen
+      expect(decision.explanations.first).to be_frozen
+      expect(decision.evaluations.first).to be_frozen
+    end
+  end
+
+  describe "Context immutability" do
+    it "freezes context data to prevent mutation" do
+      context_data = { user: { id: 1, name: "Test" }, amount: 100 }
+      context = DecisionAgent::Context.new(context_data)
+
+      expect(context.to_h).to be_frozen
+      expect(context.to_h[:user]).to be_frozen
+
+      # Original data should not be affected
+      expect(context_data).not_to be_frozen
+    end
+  end
+
+  describe "Concurrent file storage operations" do
+    let(:storage_path) { File.join(__dir__, "../tmp/thread_safety_test") }
+    let(:adapter) { DecisionAgent::Versioning::FileStorageAdapter.new(storage_path: storage_path) }
+
+    before do
+      FileUtils.rm_rf(storage_path) if Dir.exist?(storage_path)
+    end
+
+    after do
+      FileUtils.rm_rf(storage_path) if Dir.exist?(storage_path)
+    end
+
+    it "handles concurrent version creation safely" do
+      threads = []
+      results = []
+
+      # Create 10 versions concurrently
+      10.times do |i|
+        threads << Thread.new do
+          version = adapter.create_version(
+            rule_id: "concurrent_rule",
+            content: { rule: "version_#{i}" },
+            metadata: { created_by: "thread_#{i}" }
+          )
+          mutex.synchronize { results << version }
+        end
+      end
+
+      threads.each(&:join)
+
+      # All versions should be created successfully
+      expect(results.size).to eq(10)
+
+      # Version numbers should be unique and sequential
+      version_numbers = results.map { |v| v[:version_number] }.sort
+      expect(version_numbers).to eq((1..10).to_a)
+
+      # Each thread created its version as active
+      # Due to thread scheduling, all might be created as active initially
+      # The last one written should be active in the file system
+      final_active = adapter.get_active_version(rule_id: "concurrent_rule")
+      expect(final_active).not_to be_nil
+      expect(final_active[:status]).to eq("active")
+    end
+
+    it "handles concurrent read and write operations safely" do
+      # Create initial version
+      adapter.create_version(
+        rule_id: "read_write_test",
+        content: { rule: "initial" },
+        metadata: { created_by: "setup" }
+      )
+
+      threads = []
+      read_results = []
+      write_results = []
+
+      # Mix of read and write operations
+      10.times do |i|
+        if i % 2 == 0
+          # Read operations
+          threads << Thread.new do
+            versions = adapter.list_versions(rule_id: "read_write_test")
+            read_mutex.synchronize { results << versions }
+          end
+        else
+          # Write operations
+          threads << Thread.new do
+            version = adapter.create_version(
+              rule_id: "read_write_test",
+              content: { rule: "version_#{i}" },
+              metadata: { created_by: "thread_#{i}" }
+            )
+            write_mutex.synchronize { results << version }
+          end
+        end
+      end
+
+      threads.each(&:join)
+
+      # All operations should complete successfully
+      expect(read_results.size).to eq(5)
+      expect(write_results.size).to eq(5)
+
+      # Reads should never return inconsistent data
+      read_results.each do |versions|
+        expect(versions).to be_an(Array)
+        versions.each do |version|
+          expect(version).to have_key(:id)
+          expect(version).to have_key(:version_number)
+          expect(version).to have_key(:status)
+        end
+      end
+    end
+  end
+
+  describe "EvaluationValidator" do
+    it "validates frozen evaluations" do
+      evaluation = DecisionAgent::Evaluation.new(
+        decision: "approve",
+        weight: 0.8,
+        reason: "Valid",
+        evaluator_name: "TestEvaluator"
+      )
+
+      expect {
+        DecisionAgent::EvaluationValidator.validate!(evaluation)
+      }.not_to raise_error
+    end
+
+    it "raises error for unfrozen evaluations" do
+      # Create an evaluation and unfreeze it (for testing purposes)
+      evaluation = DecisionAgent::Evaluation.allocate
+      evaluation.instance_variable_set(:@decision, "approve")
+      evaluation.instance_variable_set(:@weight, 0.8)
+      evaluation.instance_variable_set(:@reason, "Test")
+      evaluation.instance_variable_set(:@evaluator_name, "Test")
+
+      expect {
+        DecisionAgent::EvaluationValidator.validate!(evaluation)
+      }.to raise_error(DecisionAgent::EvaluationValidator::ValidationError, /must be frozen/)
+    end
+
+    it "validates arrays of evaluations" do
+      evaluations = [
+        DecisionAgent::Evaluation.new(
+          decision: "approve",
+          weight: 0.8,
+          reason: "Valid 1",
+          evaluator_name: "Evaluator1"
+        ),
+        DecisionAgent::Evaluation.new(
+          decision: "reject",
+          weight: 0.6,
+          reason: "Valid 2",
+          evaluator_name: "Evaluator2"
+        )
+      ]
+
+      expect {
+        DecisionAgent::EvaluationValidator.validate_all!(evaluations)
+      }.not_to raise_error
+    end
+  end
+
+  describe "Stress Testing" do
+    let(:rules_json) do
+      {
+        version: "1.0",
+        ruleset: "stress_test",
+        rules: [
+          {
+            id: "rule1",
+            if: { field: "value", op: "gt", value: 50 },
+            then: { decision: "high", weight: 0.9, reason: "High value" }
+          },
+          {
+            id: "rule2",
+            if: { field: "value", op: "lte", value: 50 },
+            then: { decision: "low", weight: 0.8, reason: "Low value" }
+          }
+        ]
+      }
+    end
+
+    let(:evaluator) { DecisionAgent::Evaluators::JsonRuleEvaluator.new(rules_json: rules_json) }
+    let(:agent) { DecisionAgent::Agent.new(evaluators: [evaluator]) }
+
+    it "handles 100 threads making 100 decisions each (10,000 total)" do
+      thread_count = 100
+      decisions_per_thread = 100
+      total_decisions = thread_count * decisions_per_thread
+      results = []
+      mutex = Mutex.new
+
+      threads = thread_count.times.map do |thread_id|
+        Thread.new do
+          decisions_per_thread.times do |i|
+            context = { value: (thread_id * decisions_per_thread + i) % 100 }
+            decision = agent.decide(context: context)
+            mutex.synchronize { mutex.synchronize { results << decision } }
+          end
+        end
+      end
+
+      threads.each(&:join)
+
+      expect(results.size).to eq(total_decisions)
+      expect(results).to all(be_frozen)
+      expect(results.map(&:decision).uniq.sort).to eq(["high", "low"])
+    end
+
+    it "handles rapid-fire decisions from single thread (no race conditions)" do
+      results = []
+
+      1000.times do |i|
+        decision = agent.decide(context: { value: i % 100 })
+        mutex.synchronize { results << decision }
+      end
+
+      expect(results.size).to eq(1000)
+      expect(results).to all(be_frozen)
+
+      # Verify determinism - same input produces same output
+      decision1 = agent.decide(context: { value: 75 })
+      decision2 = agent.decide(context: { value: 75 })
+      expect(decision1.decision).to eq(decision2.decision)
+    end
+
+    it "handles concurrent decisions with complex nested contexts" do
+      complex_contexts = 50.times.map do |i|
+        {
+          value: i,
+          user: {
+            id: i,
+            profile: {
+              age: 20 + i % 50,
+              score: 0.5 + (i % 10) * 0.05
+            }
+          },
+          metadata: {
+            tags: ["tag#{i % 5}", "tag#{i % 3}"],
+            timestamps: [Time.now.to_i - i, Time.now.to_i]
+          }
+        }
+      end
+
+      results = []; mutex = Mutex.new
+      threads = complex_contexts.map do |context|
+        Thread.new do
+          decision = agent.decide(context: context)
+          mutex.synchronize { results << decision }
+        end
+      end
+
+      threads.each(&:join)
+
+      expect(results.size).to eq(50)
+      expect(results).to all(be_frozen)
+      results.each do |decision|
+        expect(decision.audit_payload).to be_frozen
+        expect(decision.audit_payload[:context]).to be_frozen
+      end
+    end
+  end
+
+  describe "Edge Cases" do
+    let(:evaluator) do
+      DecisionAgent::Evaluators::JsonRuleEvaluator.new(
+        rules_json: {
+          version: "1.0",
+          ruleset: "edge_cases",
+          rules: [
+            {
+              id: "rule1",
+              if: { field: "status", op: "eq", value: "active" },
+              then: { decision: "proceed", weight: 1.0, reason: "Active status" }
+            }
+          ]
+        }
+      )
+    end
+
+    let(:agent) { DecisionAgent::Agent.new(evaluators: [evaluator]) }
+
+    it "handles empty context safely across threads" do
+      results = []; mutex = Mutex.new
+
+      10.times do
+        threads = 10.times.map do
+          Thread.new do
+            # Empty context should not match any rules, causing no evaluations
+            begin
+              decision = agent.decide(context: {})
+              mutex.synchronize { results << decision }
+            rescue DecisionAgent::NoEvaluationsError
+              # Expected when no rules match
+            end
+          end
+        end
+        threads.each(&:join)
+      end
+
+      # Should either have no results (NoEvaluationsError) or all frozen
+      if results.any?
+        expect(results).to all(be_frozen)
+      end
+    end
+
+    it "handles nil values in context safely across threads" do
+      contexts = [
+        { status: nil },
+        { status: "active" },
+        { status: "" },
+        { status: false }
+      ]
+
+      results = []; mutex = Mutex.new
+      threads = contexts.flat_map do |context|
+        10.times.map do
+          Thread.new do
+            begin
+              decision = agent.decide(context: context)
+              mutex.synchronize { results << decision }
+            rescue DecisionAgent::NoEvaluationsError
+              # Expected for non-matching contexts
+            end
+          end
+        end
+      end
+
+      threads.each(&:join)
+
+      # Only contexts with status: "active" should produce decisions
+      matching_results = results.select { |d| d.decision == "proceed" }
+      expect(matching_results.size).to be <= 10
+      expect(results).to all(be_frozen)
+    end
+
+    it "handles unicode and special characters in context" do
+      special_contexts = [
+        { status: "active", name: "Test 测试 тест" },
+        { status: "active", emoji: "🚀🎉" },
+        { status: "active", special: "!@#$%^&*()" },
+        { status: "active", json: '{"nested": "value"}' }
+      ]
+
+      results = []; mutex = Mutex.new
+      threads = special_contexts.flat_map do |context|
+        5.times.map do
+          Thread.new do
+            decision = agent.decide(context: context)
+            mutex.synchronize { results << decision }
+          end
+        end
+      end
+
+      threads.each(&:join)
+
+      expect(results.size).to eq(20)
+      expect(results).to all(be_frozen)
+      results.each do |decision|
+        expect(decision.audit_payload[:context]).to be_frozen
+      end
+    end
+  end
+
+  describe "Multiple Evaluators" do
+    let(:evaluator1) do
+      DecisionAgent::Evaluators::JsonRuleEvaluator.new(
+        rules_json: {
+          version: "1.0",
+          ruleset: "evaluator1",
+          rules: [
+            {
+              id: "rule1",
+              if: { field: "score", op: "gt", value: 0.7 },
+              then: { decision: "approve", weight: 0.8, reason: "High score" }
+            }
+          ]
+        }
+      )
+    end
+
+    let(:evaluator2) do
+      DecisionAgent::Evaluators::JsonRuleEvaluator.new(
+        rules_json: {
+          version: "1.0",
+          ruleset: "evaluator2",
+          rules: [
+            {
+              id: "rule2",
+              if: { field: "verified", op: "eq", value: true },
+              then: { decision: "approve", weight: 0.9, reason: "Verified user" }
+            }
+          ]
+        }
+      )
+    end
+
+    let(:evaluator3) do
+      DecisionAgent::Evaluators::StaticEvaluator.new(
+        decision: "approve",
+        weight: 0.5,
+        reason: "Default approval"
+      )
+    end
+
+    let(:agent) do
+      DecisionAgent::Agent.new(
+        evaluators: [evaluator1, evaluator2, evaluator3],
+        scoring_strategy: DecisionAgent::Scoring::WeightedAverage.new
+      )
+    end
+
+    it "handles multiple evaluators safely across threads" do
+      contexts = [
+        { score: 0.8, verified: true },
+        { score: 0.9, verified: false },
+        { score: 0.5, verified: true },
+        { score: 0.6, verified: false }
+      ]
+
+      results = []; mutex = Mutex.new
+      threads = contexts.flat_map do |context|
+        25.times.map do
+          Thread.new do
+            decision = agent.decide(context: context)
+            mutex.synchronize { results << decision }
+          end
+        end
+      end
+
+      threads.each(&:join)
+
+      expect(results.size).to eq(100)
+      expect(results).to all(be_frozen)
+
+      # All should have multiple evaluations from different evaluators
+      results.each do |decision|
+        expect(decision.evaluations.size).to be >= 1
+        expect(decision.evaluations).to all(be_frozen)
+      end
+    end
+
+    it "prevents modification of shared evaluators array" do
+      expect(agent.evaluators).to be_frozen
+      expect(agent.evaluators.size).to eq(3)
+
+      expect {
+        agent.evaluators << DecisionAgent::Evaluators::StaticEvaluator.new(
+          decision: "reject",
+          weight: 1.0,
+          reason: "Test"
+        )
+      }.to raise_error(FrozenError)
+    end
+  end
+
+  describe "Different Scoring Strategies" do
+    let(:evaluator) do
+      DecisionAgent::Evaluators::JsonRuleEvaluator.new(
+        rules_json: {
+          version: "1.0",
+          ruleset: "scoring_test",
+          rules: [
+            {
+              id: "rule1",
+              if: { field: "value", op: "gt", value: 50 },
+              then: { decision: "high", weight: 0.9, reason: "High value" }
+            }
+          ]
+        }
+      )
+    end
+
+    it "handles Consensus strategy thread-safely" do
+      agent = DecisionAgent::Agent.new(
+        evaluators: [evaluator],
+        scoring_strategy: DecisionAgent::Scoring::Consensus.new(minimum_agreement: 0.5)
+      )
+
+      results = []; mutex = Mutex.new
+      threads = 50.times.map do |i|
+        Thread.new do
+          decision = agent.decide(context: { value: i + 25 })
+          mutex.synchronize { results << decision }
+        end
+      end
+
+      threads.each(&:join)
+
+      expect(results.size).to eq(50)
+      expect(results).to all(be_frozen)
+    end
+
+    it "handles MaxWeight strategy thread-safely" do
+      agent = DecisionAgent::Agent.new(
+        evaluators: [evaluator],
+        scoring_strategy: DecisionAgent::Scoring::MaxWeight.new
+      )
+
+      results = []; mutex = Mutex.new
+      threads = 50.times.map do |i|
+        Thread.new do
+          decision = agent.decide(context: { value: i + 25 })
+          mutex.synchronize { results << decision }
+        end
+      end
+
+      threads.each(&:join)
+
+      expect(results.size).to eq(50)
+      expect(results).to all(be_frozen)
+    end
+
+    it "handles Threshold strategy thread-safely" do
+      agent = DecisionAgent::Agent.new(
+        evaluators: [evaluator],
+        scoring_strategy: DecisionAgent::Scoring::Threshold.new(threshold: 0.75)
+      )
+
+      results = []; mutex = Mutex.new
+      threads = 50.times.map do |i|
+        Thread.new do
+          decision = agent.decide(context: { value: i + 25 })
+          mutex.synchronize { results << decision }
+        end
+      end
+
+      threads.each(&:join)
+
+      expect(results.size).to eq(50)
+      expect(results).to all(be_frozen)
+    end
+  end
+
+  describe "Race Condition Prevention" do
+    let(:evaluator) do
+      DecisionAgent::Evaluators::JsonRuleEvaluator.new(
+        rules_json: {
+          version: "1.0",
+          ruleset: "race_test",
+          rules: [
+            {
+              id: "rule1",
+              if: { field: "counter", op: "eq", value: 0 },
+              then: { decision: "zero", weight: 1.0, reason: "Counter is zero" }
+            }
+          ]
+        }
+      )
+    end
+
+    let(:agent) { DecisionAgent::Agent.new(evaluators: [evaluator]) }
+
+    it "prevents race conditions when reading same frozen objects" do
+      results = []; mutex = Mutex.new
+      decision = agent.decide(context: { counter: 0 })
+
+      # Multiple threads reading the same frozen decision
+      threads = 100.times.map do
+        Thread.new do
+          # These reads should be safe because decision is frozen
+          mutex.synchronize { results << { }
+            decision: decision.decision,
+            confidence: decision.confidence,
+            evaluations_count: decision.evaluations.size
+          }
+        end
+      end
+
+      threads.each(&:join)
+
+      expect(results.size).to eq(100)
+      # All threads should see the same values
+      expect(results.map { |r| r[:decision] }.uniq).to eq(["zero"])
+      expect(results.map { |r| r[:evaluations_count] }.uniq).to eq([1])
+    end
+
+    it "ensures deterministic hashes are consistent across threads" do
+      hashes = []; mutex = Mutex.new
+      context = { value: 42, user: { id: 123 } }
+
+      threads = 50.times.map do
+        Thread.new do
+          decision = agent.decide(context: context.dup)
+          hashes << decision.audit_payload[:deterministic_hash]
+        end
+      end
+
+      threads.each(&:join)
+
+      # All decisions with same context should have same hash
+      expect(hashes.uniq.size).to be <= 2  # May differ if rule matches/doesn't match
+    end
+  end
+
+  describe "Memory Safety" do
+    let(:evaluator) do
+      DecisionAgent::Evaluators::JsonRuleEvaluator.new(
+        rules_json: {
+          version: "1.0",
+          ruleset: "memory_test",
+          rules: [
+            {
+              id: "rule1",
+              if: { field: "active", op: "eq", value: true },
+              then: { decision: "proceed", weight: 1.0, reason: "Active" }
+            }
+          ]
+        }
+      )
+    end
+
+    let(:agent) { DecisionAgent::Agent.new(evaluators: [evaluator]) }
+
+    it "prevents memory leaks from unfrozen nested structures" do
+      results = []
+
+      100.times do |i|
+        decision = agent.decide(
+          context: {
+            active: true,
+            metadata: {
+              level1: {
+                level2: {
+                  level3: {
+                    data: "value_#{i}"
+                  }
+                }
+              }
+            }
+          }
+        )
+        mutex.synchronize { results << decision }
+      end
+
+      # Verify all nested structures are frozen
+      results.each do |decision|
+        expect(decision.audit_payload).to be_frozen
+        expect(decision.audit_payload[:context]).to be_frozen
+
+        # Check deep nesting
+        if decision.audit_payload[:context][:metadata]
+          expect(decision.audit_payload[:context][:metadata]).to be_frozen
+
+          if decision.audit_payload[:context][:metadata][:level1]
+            expect(decision.audit_payload[:context][:metadata][:level1]).to be_frozen
+          end
+        end
+      end
+    end
+
+    it "does not mutate original context data" do
+      original_context = { active: true, count: 0 }
+      original_context_copy = original_context.dup
+
+      threads = 10.times.map do
+        Thread.new do
+          agent.decide(context: original_context)
+        end
+      end
+
+      threads.each(&:join)
+
+      # Original context should be unchanged
+      expect(original_context).to eq(original_context_copy)
+      expect(original_context).not_to be_frozen
+    end
+  end
+
+  describe "Error Handling in Concurrent Context" do
+    it "handles evaluator errors gracefully in multi-threaded context" do
+      failing_evaluator = Class.new(DecisionAgent::Evaluators::Base) do
+        def evaluate(context, feedback: {})
+          raise StandardError, "Intentional failure" if context[:fail]
+          DecisionAgent::Evaluation.new(
+            decision: "success",
+            weight: 1.0,
+            reason: "Success",
+            evaluator_name: "FailingEvaluator"
+          )
+        end
+      end.new
+
+      agent = DecisionAgent::Agent.new(evaluators: [failing_evaluator])
+      results = []; mutex = Mutex.new
+      errors = []; mutex = Mutex.new
+
+      threads = 50.times.map do |i|
+        Thread.new do
+          begin
+            decision = agent.decide(context: { fail: i.even? })
+            mutex.synchronize { results << decision }
+          rescue DecisionAgent::NoEvaluationsError => e
+            errors << e
+          end
+        end
+      end
+
+      threads.each(&:join)
+
+      # Half should succeed (odd i), half should raise NoEvaluationsError (even i)
+      expect(results.size).to be > 0
+      expect(errors.size).to be > 0
+      expect(results).to all(be_frozen)
+    end
+  end
+end