aireview 2.0.0 → 2.2.0

data/lib/aireview/review_pipeline.rb

@@ -1,6 +1,7 @@
 # frozen_string_literal: true
 require 'json'
 require_relative 'errors'
+require_relative 'utils'
 require_relative 'stages'
 require_relative 'context_builder'
 require_relative 'candidate_checker'
@@ -8,12 +9,15 @@ require_relative 'result_parser'
 require_relative 'review_renderer'
 require_relative 'review_schemas'
 require_relative 'reviewer'
+require_relative 'jev_shadow'
+require_relative 'jev_stage'
+require_relative 'dry_run_prompts'
 
 module Aireview
   # A review run: context → Generate → anchoring check against the diff →
-  # Critique → report. Invalid JSON is repaired once by the same model; when
-  # the repair is invalid too, the stage restarts on another model with the
-  # original request.
+  # Critique (an LLM, or Jev with llm.critique.engine: jev) → report. Invalid
+  # JSON is repaired once by the same model; when the repair is invalid too,
+  # the stage restarts on another model with the original request.
   class ReviewPipeline
     SchemaError = ResultParser::SchemaError
 
@@ -23,15 +27,22 @@ module Aireview
       Do not use markdown, code fences, comments, or text outside JSON.
       Do not add new review findings.
     PROMPT
-    DRY_RUN_CANDIDATES_JSON = '[{"id":"C1","file":"path/from/diff.rb","line":1,' \
-                              '"quoted_code":"...","problem":"...","why":"...","suggestion":"...",' \
-                              '"category":"bug","severity":"major"}]'
-
-    def initialize(config:, reviewer: nil, context_builder: nil, logger: Logger.new($stderr))
+    # Finish reasons after which an unparsable answer gets no repair.
+    CUT_OFF_REASONS = {
+      max_tokens: 'cut off at the output limit (max_tokens)',
+      content_filter: 'blocked by the provider (content_filter)'
+    }.freeze
+
+    # jev — jev_shadow: and jev_stage: for tests; by default they are built
+    # from the config, the stage only when Jev is the engine.
+    def initialize(config:, reviewer: nil, context_builder: nil, logger: Logger.new($stderr), **jev)
       @config = config
       @parser = ResultParser.new
       @reviewer = reviewer || Reviewer.new(config: config, logger: logger)
       @context_builder = context_builder || ContextBuilder.new(config: config, logger: logger)
+      @jev_shadow = jev[:jev_shadow] ||
+                    JevShadow.new(config: config, scrub: @context_builder.method(:scrub_text), logger: logger)
+      @jev_stage = jev[:jev_stage]
       @logger = logger
     end
 
@@ -58,7 +69,11 @@ module Aireview
                    "(model=#{@reviewer.answered_model('generate')})")
       candidates = check_candidates(context: context, changes: changes, candidates: candidates)
 
-      accepted = critique ? maybe_critique(context: context, candidates: candidates) : skip_critique(candidates)
+      accepted, jev_note = if critique
+                             maybe_critique(context: context, candidates: candidates)
+                           else
+                             skip_critique(candidates)
+                           end
 
       @logger.info("Pipeline finished with #{accepted.size} accepted finding(s)")
 
@@ -67,58 +82,19 @@ module Aireview
         summary: summary,
         coverage: context.coverage,
         fallback_models: @reviewer.fallback_models,
-        critique_weaker: critique && @reviewer.critique_weaker?
+        critique_weaker: critique && @reviewer.critique_weaker?,
+        jev_note: jev_note
       )
     end
 
+    # The prompts of the run and everything --dry-run shows; see DryRunPrompts.
     def dry_run_prompts(merge_request:, changes:, jira_issue: nil, critique: true)
-      @config.require_models!
-
-      context = @context_builder.prepare(
-        merge_request: merge_request,
-        changes: changes,
-        jira_issue: jira_issue,
-        critique: critique
-      )
-      generate_prompt = @context_builder.build_generate_prompt(context)
-      critique_prompt = if critique
-                          @context_builder.build_critique_prompt(context, candidates_json: DRY_RUN_CANDIDATES_JSON)
-                        end
-
-      {
-        generate_prompt: generate_prompt,
-        critique_prompt: critique_prompt,
-        generate_model: @config.generate_model,
-        generate_temperature: @config.generate_temperature,
-        critique_model: @config.critique_model,
-        critique_temperature: @config.critique_temperature,
-        generate_fallbacks: @config.fallback_names('generate'),
-        critique_fallbacks: critique ? @config.fallback_names('critique') : [],
-        sources: setting_sources(critique),
-        config_paths: @config.layer_paths,
-        warnings: @config.warnings,
-        critique_rule: critique ? @config.routing.rule : nil,
-        api_keys: @config.api_key_counts(critique ? STAGES : ['generate']),
-        time_budget: @config.llm_time_budget,
-        overloaded_quarantine: @config.overloaded_quarantine,
-        coverage: context.coverage,
-        sizes: context.sizes
-      }
+      DryRunPrompts.new(config: @config, context_builder: @context_builder, logger: @logger)
+        .build(merge_request: merge_request, changes: changes, jira_issue: jira_issue, critique: critique)
     end
 
     private
 
-    # Where the model, provider and reserves of a stage came from, for --dry-run.
-    def setting_sources(critique)
-      (critique ? %w[generate critique] : %w[generate]).to_h do |stage|
-        [stage.to_sym, {
-          model: @config.stage_model_source(stage),
-          provider: @config.stage_provider_source(stage),
-          fallbacks: @config.stage_fallbacks_source(stage)
-        }]
-      end
-    end
-
     # A stage is a request, parsing and one repair by the same model. An
     # invalid result after the repair, like a repair with no requests left,
     # excludes the model for the stage, and the stage starts over on the next
@@ -148,18 +124,32 @@ module Aireview
       ).check(candidates)
     end
 
+    # Returns the candidates and no report note, like every critique path.
     def skip_critique(candidates, reason = nil)
       @logger.info(['Pipeline critique pass skipped', reason].compact.join(': '))
-      candidates
+      [candidates, nil]
     end
 
     # Critique has nothing to filter without candidates: the LLM request
-    # would waste quota and time.
+    # would waste quota and time. Returns [accepted, note for the report].
     def maybe_critique(context:, candidates:)
       return skip_critique(candidates, 'no candidates') if candidates.empty?
+      return jev_critique(context: context, candidates: candidates) if @config.jev_critique?
 
       @logger.info("Pipeline critique pass started (model=#{@config.critique_model})")
-      critique_candidates(context: context, candidates: candidates)
+      accepted = critique_candidates(context: context, candidates: candidates)
+      @jev_shadow.run(context: context, candidates: candidates, accepted: accepted) if @config.jev_shadow?
+      [accepted, nil]
+    end
+
+    def jev_critique(context:, candidates:)
+      @jev_stage ||= JevStage.new(
+        config: @config, logger: @logger,
+        critic: JevCritic.build(config: @config, scrub: @context_builder.method(:scrub_text), logger: @logger)
+      )
+      llm_critique = ->(subset) { critique_candidates(context: context, candidates: subset) }
+      outcome = @jev_stage.run(context: context, candidates: candidates, llm_critique: llm_critique)
+      [outcome.accepted, outcome.note]
     end
 
     def critique_candidates(context:, candidates:)
@@ -214,6 +204,7 @@ module Aireview
     def parse_string_with_repair(raw:, kind:, expected:, repair_stage:, critique_candidate_ids: nil)
       parse_expected_result(raw, expected, critique_candidate_ids: critique_candidate_ids)
     rescue JSON::ParserError, SchemaError => e
+      raise_if_cut_off(stage: repair_stage, kind: kind, error: e)
       @logger.warn("Invalid #{kind} JSON, requesting one repair: #{e.message}")
       repaired = repair_json(
         raw: raw,
@@ -225,10 +216,19 @@ module Aireview
       begin
         parse_expected_result(repaired, expected, critique_candidate_ids: critique_candidate_ids)
       rescue JSON::ParserError, SchemaError => second_error
+        raise_if_cut_off(stage: repair_stage, kind: "#{kind} repair", error: second_error)
         raise ParseError, "LLM returned invalid #{kind} JSON after repair: #{second_error.message}"
       end
     end
 
+    # An answer the provider cut off or blocked is not a JSON mistake: the
+    # same model would cut the repair off too, so the stage goes to the next
+    # model at once. A valid answer is taken whatever the reason.
+    def raise_if_cut_off(stage:, kind:, error:)
+      reason = CUT_OFF_REASONS[@reviewer.finish_reason(stage)]
+      raise ParseError, "LLM #{kind} was #{reason}: #{error.message}" if reason
+    end
+
     def parse_expected_result(raw, expected, critique_candidate_ids: nil)
       @parser.parse(raw, expected: expected, critique_candidate_ids: critique_candidate_ids)
     end

data/lib/aireview/review_renderer.rb

@@ -47,6 +47,10 @@ module Aireview
         section_list: 'Truncated sections',
         fallback_used: 'Fallback model used',
         critique_weaker: 'Critique ran on a model weaker than Generate: the findings were checked less strictly.',
+        jev: 'The findings were checked by Jev, a fast classifier, without refining their wording.',
+        jev_partial: 'The findings were checked by Jev, a fast classifier, without refining their wording; ' \
+                     'those Jev could not judge were checked by the LLM critique.',
+        jev_failed: 'Jev was unavailable: the findings were checked by the LLM critique.',
         quote_missing: 'quote not found in the diff'
       },
       'ru' => {
@@ -74,6 +78,10 @@ module Aireview
         section_list: 'Усечённые секции',
         fallback_used: 'Использована резервная модель',
         critique_weaker: 'Критика выполнена моделью слабее generate: замечания проверены менее строго.',
+        jev: 'Замечания проверены быстрым классификатором Jev, без уточнения формулировок.',
+        jev_partial: 'Замечания проверены быстрым классификатором Jev, без уточнения формулировок; ' \
+                     'те, что Jev не смог оценить, проверила LLM-критика.',
+        jev_failed: 'Jev был недоступен: замечания проверила LLM-критика.',
         quote_missing: 'цитата не найдена в диффе'
       }
     }.freeze
@@ -86,7 +94,9 @@ module Aireview
     # result is still about the findings; incomplete coverage is written
     # next to it so that the result line does not read as "everything was
     # checked".
-    def render(accepted, summary:, coverage: nil, fallback_models: {}, critique_weaker: false)
+    # jev_note — how Jev took part in the critique (see JevStage::Outcome).
+    # The facts about the run are named one by one on purpose.
+    def render(accepted, summary:, coverage: nil, fallback_models: {}, critique_weaker: false, jev_note: nil) # rubocop:disable Metrics/ParameterLists
       mismatches, important = select_findings(Array(accepted))
       result = mismatches.empty? && important.empty? ? 'ok' : 'needs attention'
 
@@ -106,7 +116,7 @@ module Aireview
         ## #{label(:result)}
 
         #{result}#{partial_note(coverage)}
-        #{coverage_block(coverage)}#{fallback_note(fallback_models)}#{weaker_note(critique_weaker)}
+        #{coverage_block(coverage)}#{fallback_note(fallback_models)}#{weaker_note(critique_weaker)}#{jev_note(jev_note)}
         #{label(:disclaimer)}
       MARKDOWN
     end
@@ -191,6 +201,12 @@ module Aireview
       "\n#{label(:critique_weaker)}\n"
     end
 
+    # Jev decides keep/reject but cannot refine: the reader should know the
+    # wording is the first pass's own.
+    def jev_note(note)
+      note ? "\n#{label(note)}\n" : ''
+    end
+
     def fallback_note(fallback_models)
       return '' if fallback_models.nil? || fallback_models.empty?
 

data/lib/aireview/reviewer.rb

@@ -16,6 +16,7 @@ module Aireview
       @logger = logger
       @router = router || LlmRouter.new(config: config, logger: logger)
       @client = client || LlmClient.new(config: config, logger: logger)
+      @finish_reasons = {}
     end
 
     # pinned — a request only to the model that answered last in the stage
@@ -47,6 +48,12 @@ module Aireview
       @router.critique_weaker?
     end
 
+    # Why the last answer of the stage stopped (:stop, :max_tokens,
+    # :content_filter…), nil when the provider did not say or no answer came.
+    def finish_reason(stage)
+      @finish_reasons[stage.to_s]
+    end
+
     # An invalid result: the model is excluded for the stage, the next
     # request of the stage goes to another. Returns the excluded model or nil.
     def exclude_answered_model(stage:, reason:)
@@ -58,11 +65,13 @@ module Aireview
     # A pinned route giving up is not an API error for the pipeline but
     # "repair impossible": the same fate as an invalid result.
     def call_llm(prompt, pinned:)
+      @finish_reasons.delete(prompt.stage)
       response = @router.call(stage: prompt.stage, request_chars: prompt.chars, pinned: pinned) do |route, timeout|
         @client.request(prompt, candidate: route.candidate, key: route.key, key_index: route.key_index,
                                 timeout: timeout)
       end
-      response.content
+      @finish_reasons[prompt.stage] = response.finish_reason
+      LlmClient.content(response)
     rescue RouteExhaustedError => e
       raise RepairImpossibleError, e.message
     end

data/lib/aireview/stage_chains.rb

@@ -26,7 +26,8 @@ module Aireview
       primary = ModelCandidate.new(
         provider: settings.fetch(:provider).to_s,
         model: settings[:model],
-        max_prompt_chars: settings.fetch(:max_prompt_chars)
+        max_prompt_chars: settings.fetch(:max_prompt_chars),
+        api_base: ModelCandidate.api_base(settings[:api_base], "llm.#{stage}.api_base")
       )
       return [primary] if only_primary
 
@@ -34,7 +35,8 @@ module Aireview
     end
 
     # A reserve without a provider inherits the stage provider, without a
-    # limit its limit.
+    # limit its limit. The address is not inherited: a reserve on another
+    # server names its own.
     def self.fallbacks(stage, items, primary)
       items.each_with_index.map do |item, index|
         item = ModelCandidate.parse_item(item)
@@ -46,7 +48,8 @@ module Aireview
         ModelCandidate.new(
           provider: (item['provider'] || primary.provider).to_s,
           model: item['model'].to_s,
-          max_prompt_chars: limit.nil? ? primary.max_prompt_chars : positive_limit(limit, name)
+          max_prompt_chars: limit.nil? ? primary.max_prompt_chars : positive_limit(limit, name),
+          api_base: ModelCandidate.api_base(item['api_base'], "#{name}.api_base")
         )
       end
     end

data/lib/aireview/version.rb

@@ -1,4 +1,4 @@
 # frozen_string_literal: true
 module Aireview
-  VERSION = '2.0.0'
+  VERSION = '2.2.0'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: aireview
 version: !ruby/object:Gem::Version
-  version: 2.0.0
+  version: 2.2.0
 platform: ruby
 authors:
 - Denis Levenko
 autorequire:
 bindir: bin
 cert_chain: []
-date: 2026-09-19 00:00:00.000000000 Z
+date: 2026-09-27 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: dotenv
@@ -50,18 +50,19 @@ dependencies:
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 1.16.0
+        version: 2.0.0
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 1.16.0
+        version: 2.0.0
 description: 'Reviews self-hosted GitLab merge requests with a two-pass LLM pipeline:
   the first pass finds candidate findings, the second one critiques them and drops
-  the weak ones. Supports Gemini and local Ollama, optional Jira context and posting
-  a single updatable review note back to the merge request.'
+  the weak ones. Supports Gemini, Ollama, OpenAI, Anthropic, OpenRouter and OpenAI-compatible
+  servers, Jev as the critic, optional Jira context and posting a single updatable
+  review note back to the merge request.'
 email:
 executables:
 - aireview
@@ -80,15 +81,21 @@ files:
 - lib/aireview/cli.rb
 - lib/aireview/config.rb
 - lib/aireview/config_fallbacks.rb
+- lib/aireview/config_jev.rb
 - lib/aireview/config_layers.rb
 - lib/aireview/config_limits.rb
 - lib/aireview/config_loader.rb
 - lib/aireview/context_budget.rb
 - lib/aireview/context_builder.rb
 - lib/aireview/diff_fetcher.rb
+- lib/aireview/dry_run_prompts.rb
 - lib/aireview/dry_run_report.rb
 - lib/aireview/errors.rb
 - lib/aireview/gitlab_client.rb
+- lib/aireview/jev_client.rb
+- lib/aireview/jev_critic.rb
+- lib/aireview/jev_shadow.rb
+- lib/aireview/jev_stage.rb
 - lib/aireview/jira_client.rb
 - lib/aireview/llm_client.rb
 - lib/aireview/llm_failure.rb
@@ -101,6 +108,7 @@ files:
 - lib/aireview/output_schemas.rb
 - lib/aireview/prompts/critique.txt
 - lib/aireview/prompts/generate.txt
+- lib/aireview/prompts/jev_questions.yml
 - lib/aireview/publisher.rb
 - lib/aireview/result_parser.rb
 - lib/aireview/review_marker.rb