diogenes 0.1.4 → 0.1.5

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: ed59a2db9702739e79ab10146a799260ff41fbbe85c0178d9fd5c0976bd4621f
-  data.tar.gz: 2ca1bf8250b1753b059b48ee226eb8b20c122c2adce7ada559bb1047b5c5a03f
+  metadata.gz: 00c7fad965257b62e461fbd2ae43db54c10bb7cc6c93d454cb2563ae1c6eb6a8
+  data.tar.gz: 69e47cf048ce0a2d31036ad803ebefcb0cf66bb6d6417ca517d2dc6cc52343d5
 SHA512:
-  metadata.gz: 84d21f0c4cf3817bafa44054a479f0c8d91e4a26b19d6835313511c1e5108e0e49f96316391d5b58f5c1f6a569e1432f0481e07b50d59b2fe6401c52d8871f4e
-  data.tar.gz: 2e5a9956eb13f0e19bd86cb71a333822771d4edd302e93c6000981b59ea9c67622a36702b9aee8f4e42b3f848b882b3ed44bcca3500fe2d38137ac4a47322ce6
+  metadata.gz: f3a0ad8ca1f41be1e4edfe1fe38552d954ecdf1d5d2dc6e417c713be956673c2ae67b0781fdb5720b693c9aab999b8a31c9b0eac55ccec5e16a3eb392c0decec
+  data.tar.gz: 42b1551a1bb754c6252d9fa0d04cdc0894251ac76f38656c9c1ae003d5854d59e816f4de3823fca4f2056a97d6a7b8b41026e112be6fa1b3a8ffbb8ecdac2164

data/.release-please-manifest.json

@@ -1,3 +1,3 @@
 {
-    ".": "0.1.4"
+    ".": "0.1.5"
 }

data/CHANGELOG.md

@@ -1,5 +1,12 @@
 ## [Unreleased]
 
+## [0.1.5](https://github.com/meaganewaller/diogenes/compare/diogenes/v0.1.4...diogenes/v0.1.5) (2026-06-27)
+
+
+### Features
+
+* **evaluate:** interactive five-gate evaluation command :fire: ([#9](https://github.com/meaganewaller/diogenes/issues/9)) ([cc43a17](https://github.com/meaganewaller/diogenes/commit/cc43a17b8f94fcd094b4d0fb806d9a3e43567fbe))
+
 ## [0.1.4](https://github.com/meaganewaller/diogenes/compare/diogenes/v0.1.3...diogenes/v0.1.4) (2026-06-27)
 
 

data/lib/diogenes/cli/evaluate.rb

@@ -0,0 +1,20 @@
+# frozen_string_literal: true
+# rbs_inline: enabled
+
+module Diogenes
+  class Cli
+    class Evaluate
+      #: (argv: Array[String], out: IO, err: IO, **untyped) -> Integer
+      def self.run(argv:, out:, err:, **opts)
+        description = argv.join(" ").strip
+
+        if description.empty?
+          err.puts "Usage: diogenes evaluate \"<feature description>\""
+          return 1
+        end
+
+        Evaluation::Session.new(description:, out:, err:, **opts).run
+      end
+    end
+  end
+end

data/lib/diogenes/cli/init.rb

@@ -9,8 +9,10 @@ module Diogenes
       DIOGENES_DIR = ".diogenes" #: String
       TEMPLATES_ROOT = File.expand_path("../templates/init", __dir__) #: String
 
-      #: (cwd: String, out: IO, err: IO) -> Integer
-      def self.run(cwd:, out:, err:)
+      #: (cwd: String, out: IO, err: IO, ?argv: Array[String], **untyped) -> Integer
+      def self.run(cwd:, out:, err:, argv: [], **_opts)
+        raise Diogenes::UserError, "`init` takes no arguments" unless argv.empty?
+
         new(cwd: cwd, out: out, err: err).run
       end
 

data/lib/diogenes/cli.rb

@@ -2,23 +2,26 @@
 # rbs_inline: enabled
 
 require_relative "cli/init"
+require_relative "cli/evaluate"
 
 module Diogenes
   class Cli
     COMMANDS = {
-      "init" => Init
+      "init" => Init,
+      "evaluate" => Evaluate
     }.freeze #: Hash[String, Class]
 
-    #: (?Array[String] argv, ?out: IO, ?err: IO) -> Integer
-    def self.run(argv = ARGV, out: $stdout, err: $stderr)
-      new(argv: argv, out: out, err: err).run
+    #: (?Array[String] argv, ?out: IO, ?err: IO, **untyped) -> Integer
+    def self.run(argv = ARGV, out: $stdout, err: $stderr, **opts)
+      new(argv: argv, out: out, err: err, **opts).run
     end
 
-    #: (argv: Array[String], out: IO, err: IO) -> void
-    def initialize(argv:, out:, err:)
+    #: (argv: Array[String], out: IO, err: IO, **untyped) -> void
+    def initialize(argv:, out:, err:, **opts)
       @argv = argv.dup #: Array[String]
       @out = out #: IO
       @err = err #: IO
+      @in = opts.fetch(:in, $stdin) #: IO
     end
 
     #: () -> Integer
@@ -42,13 +45,10 @@ module Diogenes
 
     private
 
-    # Run a command
-    # --
     #: (Class, Array[String]) -> Integer
     def run_command(command_class, args)
-      raise UserError, "Unexpected arguments: #{args.join(" ")}" if args.any?
-
-      command_class.run(cwd: Dir.pwd, out: @out, err: @err)
+      opts = {in: @in}
+      command_class.run(argv: args, cwd: Dir.pwd, out: @out, err: @err, **opts)
     end
 
     #: () -> Integer

data/lib/diogenes/evaluation/gate.rb

@@ -0,0 +1,25 @@
+# frozen_string_literal: true
+# rbs_inline: enabled
+
+module Diogenes
+  module Evaluation
+    class Gate
+      attr_reader :key       #: Symbol
+      attr_reader :name      #: String
+      attr_reader :principle #: String
+      attr_reader :explanation #: String
+      attr_reader :question  #: String
+      attr_reader :fail_message #: String
+
+      #: (key: Symbol, name: String, principle: String, explanation: String, question: String, fail_message: String) -> void
+      def initialize(key:, name:, principle:, explanation:, question:, fail_message:)
+        @key = key
+        @name = name
+        @principle = principle
+        @explanation = explanation
+        @question = question
+        @fail_message = fail_message
+      end
+    end
+  end
+end

data/lib/diogenes/evaluation/gates.rb

@@ -0,0 +1,77 @@
+# frozen_string_literal: true
+# rbs_inline: enabled
+
+module Diogenes
+  module Evaluation
+    module Gates
+      ALL = [
+        Gate.new(
+          key: :failure_mode,
+          name: "Failure Mode",
+          principle: "Least surprise at scale",
+          explanation: "When AI gets things wrong — and it will — what happens? " \
+            "This gate tests whether failure is detectable and recoverable. " \
+            "Embarrassing failures are manageable. Silent, cumulative, or " \
+            "catastrophic ones are not.",
+          question: "When this feature produces wrong output, can the user " \
+            "detect and recover without your direct intervention? (Y/n)",
+          fail_message: "Failures that go undetected or unrecoverable create compounding " \
+            "harm over time. Add explicit uncertainty signals, limit the feature " \
+            "scope, or move to a human-reviewed workflow."
+        ),
+        Gate.new(
+          key: :user_verifiable,
+          name: "User Verifiable",
+          principle: "Trust requires verification",
+          explanation: "Can a typical user of this feature actually tell when the " \
+            "output is wrong? If they need domain expertise they don't have to " \
+            "evaluate AI output, you're shipping a confidence gap, not a feature.",
+          question: "Can a typical user judge whether the output is correct — " \
+            "without needing domain expertise they don't already have? (Y/n)",
+          fail_message: "Shipping outputs users cannot evaluate creates misplaced trust. " \
+            "Consider adding confidence scores, human review, or limiting the " \
+            "feature to domains where users have the relevant expertise."
+        ),
+        Gate.new(
+          key: :human_in_loop,
+          name: "Human in the Loop",
+          principle: "Human-centered design, genuinely",
+          explanation: "A loop requires a human with time, context, and authority to act. " \
+            "If the human is approving outputs they cannot evaluate, that is a " \
+            "rubber stamp — not a loop.",
+          question: "Is there a specific person with enough time, context, and " \
+            "authority to review and override outputs before they affect users? (Y/n)",
+          fail_message: "A human in the loop who lacks time, context, or authority is " \
+            "safety theater. Either invest in a real review capacity or design " \
+            "the feature so it doesn't require human sign-off."
+        ),
+        Gate.new(
+          key: :observability,
+          name: "Observability",
+          principle: "Craftsmanship — you wouldn't ship blind",
+          explanation: "Silent degradation is worse than no feature. If you won't know " \
+            "when this AI feature starts producing worse outputs in production, " \
+            "you'll find out through a user complaint — after the damage is done.",
+          question: "Do you have (or will you have before shipping) monitoring to " \
+            "detect when this feature's quality degrades in production? (Y/n)",
+          fail_message: "Shipping an AI feature without observability means betting your " \
+            "users will tell you when it breaks. Set up quality monitoring, " \
+            "sampling, or feedback loops before you ship."
+        ),
+        Gate.new(
+          key: :right_tool,
+          name: "Right Tool",
+          principle: "Convention over configuration",
+          explanation: "AI is powerful, but it is not always the right answer. " \
+            "Deterministic software is more reliable, cheaper to operate, and " \
+            "easier to audit. This gate asks you to make the case for AI explicitly.",
+          question: "Have you genuinely considered and ruled out a simpler, " \
+            "deterministic software solution for this specific need? (Y/n)",
+          fail_message: "The best AI feature is sometimes no AI feature. Revisit whether " \
+            "the problem can be solved with rules, search, filtering, or structured " \
+            "data before adding an LLM dependency."
+        )
+      ].freeze #: Array[Gate]
+    end
+  end
+end

data/lib/diogenes/evaluation/session.rb

@@ -0,0 +1,137 @@
+# frozen_string_literal: true
+# rbs_inline: enabled
+
+module Diogenes
+  module Evaluation
+    class Session
+      SEPARATOR = ("─" * 60).freeze #: String
+
+      Result = Struct.new(:gate, :passed, keyword_init: true) do
+        def passed? = passed
+        def failed? = !passed
+      end
+
+      #: (description: String, out: IO, err: IO, **untyped) -> void
+      def initialize(description:, out:, err:, **opts)
+        @description = description
+        @out = out
+        @err = err
+        @in = opts.fetch(:in, $stdin)
+        @results = [] #: Array[untyped]
+      end
+
+      #: () -> Integer
+      def run
+        print_header
+
+        Gates::ALL.each_with_index do |gate, index|
+          run_gate(gate, index + 1)
+
+          if @results.last.failed? && !prompt_continue?
+            print_partial_summary
+            return 1
+          end
+        end
+
+        print_final_summary
+        prompt_decision_record
+
+        @results.all?(&:passed?) ? 0 : 1
+      rescue Interrupt
+        @out.puts "\n\nEvaluation interrupted."
+        print_partial_summary
+        1
+      end
+
+      private
+
+      #: () -> void
+      def print_header
+        @out.puts
+        @out.puts "Evaluating: \"#{@description}\""
+        @out.puts
+      end
+
+      #: (Gate, Integer) -> void
+      def run_gate(gate, number)
+        @out.puts SEPARATOR
+        @out.puts "Gate #{number} of #{Gates::ALL.size} — #{gate.name}"
+        @out.puts "Principle: #{gate.principle}"
+        @out.puts SEPARATOR
+        @out.puts
+        @out.puts wrap(gate.explanation, 70)
+        @out.puts
+
+        passed = ask_yes_no(gate.question)
+        @results << Result.new(gate:, passed:)
+
+        if passed
+          @out.puts "  ✓ Gate #{number} — PASS"
+        else
+          @out.puts "  ✗ Gate #{number} — FAIL"
+          @out.puts
+          @out.puts "  #{gate.fail_message}"
+        end
+        @out.puts
+      end
+
+      #: () -> bool
+      def prompt_continue?
+        ask_yes_no("Continue evaluating remaining gates? (Y/n)")
+      end
+
+      #: () -> void
+      def prompt_decision_record
+        if ask_yes_no("Generate a decision record? (Y/n)")
+          @out.puts "  Decision record generation is coming in the next release."
+        end
+      end
+
+      #: (String) -> bool
+      def ask_yes_no(question)
+        @out.print "  #{question} "
+        response = @in.gets&.strip&.downcase || ""
+        response.empty? || response.start_with?("y")
+      end
+
+      #: () -> void
+      def print_final_summary
+        if @results.all?(&:passed?)
+          @out.puts "All #{@results.size} gates passed! ✓"
+        else
+          @out.puts "Evaluation complete."
+        end
+        @out.puts
+        print_summary_table
+        @out.puts
+      end
+
+      #: () -> void
+      def print_partial_summary
+        @out.puts
+        @out.puts "Gates evaluated:"
+        print_summary_table
+      end
+
+      #: () -> void
+      def print_summary_table
+        Gates::ALL.each_with_index do |gate, index|
+          result = @results[index]
+          status = if result.nil?
+            "(not evaluated)"
+          elsif result.passed?
+            "PASS"
+          else
+            "FAIL"
+          end
+          @out.puts "  Gate #{index + 1} — #{gate.name.ljust(22)} #{status}"
+        end
+      end
+
+      #: (String, Integer) -> String
+      def wrap(text, width)
+        text.gsub(/(.{1,#{width}})(\s+|$)/, "\\1\n").strip
+      end
+    end
+  end
+end

data/lib/diogenes/version.rb

@@ -2,5 +2,5 @@
 # rbs_inline: enabled
 
 module Diogenes
-  VERSION = "0.1.4" #: String
+  VERSION = "0.1.5" #: String
 end

data/sig/generated/diogenes/cli/evaluate.rbs

@@ -0,0 +1,10 @@
+# Generated from lib/diogenes/cli/evaluate.rb with RBS::Inline
+
+module Diogenes
+  class Cli
+    class Evaluate
+      # : (argv: Array[String], out: IO, err: IO, **untyped) -> Integer
+      def self.run: (argv: Array[String], out: IO, err: IO, **untyped) -> Integer
+    end
+  end
+end

data/sig/generated/diogenes/cli/init.rbs

@@ -7,8 +7,8 @@ module Diogenes
 
       TEMPLATES_ROOT: String
 
-      # : (cwd: String, out: IO, err: IO) -> Integer
-      def self.run: (cwd: String, out: IO, err: IO) -> Integer
+      # : (cwd: String, out: IO, err: IO, ?argv: Array[String], **untyped) -> Integer
+      def self.run: (cwd: String, out: IO, err: IO, ?argv: Array[String], **untyped) -> Integer
 
       # : (cwd: String, out: IO, err: IO) -> void
       def initialize: (cwd: String, out: IO, err: IO) -> void

data/sig/generated/diogenes/cli.rbs

@@ -4,19 +4,17 @@ module Diogenes
   class Cli
     COMMANDS: Hash[String, Class]
 
-    # : (?Array[String] argv, ?out: IO, ?err: IO) -> Integer
-    def self.run: (?Array[String] argv, ?out: IO, ?err: IO) -> Integer
+    # : (?Array[String] argv, ?out: IO, ?err: IO, **untyped) -> Integer
+    def self.run: (?Array[String] argv, ?out: IO, ?err: IO, **untyped) -> Integer
 
-    # : (argv: Array[String], out: IO, err: IO) -> void
-    def initialize: (argv: Array[String], out: IO, err: IO) -> void
+    # : (argv: Array[String], out: IO, err: IO, **untyped) -> void
+    def initialize: (argv: Array[String], out: IO, err: IO, **untyped) -> void
 
     # : () -> Integer
     def run: () -> Integer
 
     private
 
-    # Run a command
-    # --
     # : (Class, Array[String]) -> Integer
     def run_command: (Class, Array[String]) -> Integer
 

data/sig/generated/diogenes/evaluation/gate.rbs

@@ -0,0 +1,22 @@
+# Generated from lib/diogenes/evaluation/gate.rb with RBS::Inline
+
+module Diogenes
+  module Evaluation
+    class Gate
+      attr_reader key: Symbol
+
+      attr_reader name: String
+
+      attr_reader principle: String
+
+      attr_reader explanation: String
+
+      attr_reader question: String
+
+      attr_reader fail_message: String
+
+      # : (key: Symbol, name: String, principle: String, explanation: String, question: String, fail_message: String) -> void
+      def initialize: (key: Symbol, name: String, principle: String, explanation: String, question: String, fail_message: String) -> void
+    end
+  end
+end

data/sig/generated/diogenes/evaluation/gates.rbs

@@ -0,0 +1,9 @@
+# Generated from lib/diogenes/evaluation/gates.rb with RBS::Inline
+
+module Diogenes
+  module Evaluation
+    module Gates
+      ALL: Array[Gate]
+    end
+  end
+end

data/sig/generated/diogenes/evaluation/session.rbs

@@ -0,0 +1,53 @@
+# Generated from lib/diogenes/evaluation/session.rb with RBS::Inline
+
+module Diogenes
+  module Evaluation
+    class Session
+      SEPARATOR: String
+
+      class Result < Struct[untyped]
+        attr_accessor gate(): untyped
+
+        attr_accessor passed(): untyped
+
+        def self.new: (?gate: untyped, ?passed: untyped) -> instance
+                    | ({ ?gate: untyped, ?passed: untyped }) -> instance
+      end
+
+      # : (description: String, out: IO, err: IO, **untyped) -> void
+      def initialize: (description: String, out: IO, err: IO, **untyped) -> void
+
+      # : () -> Integer
+      def run: () -> Integer
+
+      private
+
+      # : () -> void
+      def print_header: () -> void
+
+      # : (Gate, Integer) -> void
+      def run_gate: (Gate, Integer) -> void
+
+      # : () -> bool
+      def prompt_continue?: () -> bool
+
+      # : () -> void
+      def prompt_decision_record: () -> void
+
+      # : (String) -> bool
+      def ask_yes_no: (String) -> bool
+
+      # : () -> void
+      def print_final_summary: () -> void
+
+      # : () -> void
+      def print_partial_summary: () -> void
+
+      # : () -> void
+      def print_summary_table: () -> void
+
+      # : (String, Integer) -> String
+      def wrap: (String, Integer) -> String
+    end
+  end
+end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: diogenes
 version: !ruby/object:Gem::Version
-  version: 0.1.4
+  version: 0.1.5
 platform: ruby
 authors:
 - Meagan Waller
@@ -55,11 +55,15 @@ files:
 - hk.pkl
 - lib/diogenes.rb
 - lib/diogenes/cli.rb
+- lib/diogenes/cli/evaluate.rb
 - lib/diogenes/cli/init.rb
 - lib/diogenes/dsl/artifact.rb
 - lib/diogenes/dsl/hook.rb
 - lib/diogenes/dsl/rule.rb
 - lib/diogenes/dsl/skill.rb
+- lib/diogenes/evaluation/gate.rb
+- lib/diogenes/evaluation/gates.rb
+- lib/diogenes/evaluation/session.rb
 - lib/diogenes/templates/init/artifacts/decision_record.md.erb
 - lib/diogenes/templates/init/diogenes.rb
 - lib/diogenes/templates/init/hooks/README.md
@@ -69,11 +73,15 @@ files:
 - sig/diogenes.rbs
 - sig/generated/diogenes.rbs
 - sig/generated/diogenes/cli.rbs
+- sig/generated/diogenes/cli/evaluate.rbs
 - sig/generated/diogenes/cli/init.rbs
 - sig/generated/diogenes/dsl/artifact.rbs
 - sig/generated/diogenes/dsl/hook.rbs
 - sig/generated/diogenes/dsl/rule.rbs
 - sig/generated/diogenes/dsl/skill.rbs
+- sig/generated/diogenes/evaluation/gate.rbs
+- sig/generated/diogenes/evaluation/gates.rbs
+- sig/generated/diogenes/evaluation/session.rbs
 - sig/generated/diogenes/version.rbs
 homepage: https://github.com/meaganewaller/diogenes/tree/main/diogenes
 licenses: