aiox-core 5.0.4 → 5.0.7

package/pro/squads/squad-creator-pro/workflows/wf-model-tier-qualification.yaml

@@ -1,282 +1,303 @@
-# wf-model-tier-qualification.yaml
-# Workflow automatizado para qualificar tasks em Haiku/Sonnet/Opus
-# Version: 2.0
-# Created: 2026-02-11
-# Updated: 2026-02-11 - v2.0 rewrite with veto conditions, parallel execution,
-#          auto-improve loop, test input registry, and cost-first logic
-
 workflow:
   id: wf-model-tier-qualification
-  name: "Model Tier Qualification"
-  version: "2.0"
-  purpose: "Testar empiricamente se uma task pode rodar em modelo mais barato mantendo qualidade"
+  name: Model Tier Qualification
+  version: "2.2"
+  purpose: Testar empiricamente se uma task pode rodar em modelo mais barato mantendo qualidade
   orchestrator: "@pedro-valerio"
-  mode: "autonomous"
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# PHILOSOPHY (Alan's Decision Patterns - ADR-001)
-# ═══════════════════════════════════════════════════════════════════════════════
-
+  mode: autonomous
+  description: Legacy phase-based workflow for model-tier qualification.
+  type: qualification
+  state_machine:
+    states: [init, pre-flight, parallel-execution, compare-auto-decide, auto-improve-loop, generate-report, complete, failed]
+    transitions:
+      - trigger: start
+        source: init
+        dest: pre-flight
+      - trigger: pre_flight_complete
+        source: pre-flight
+        dest: parallel-execution
+      - trigger: pre_flight_failed
+        source: pre-flight
+        dest: failed
+      - trigger: parallel_execution_complete
+        source: parallel-execution
+        dest: compare-auto-decide
+      - trigger: parallel_execution_failed
+        source: parallel-execution
+        dest: failed
+      - trigger: compare_auto_decide_qualified
+        source: compare-auto-decide
+        dest: generate-report
+      - trigger: compare_auto_decide_needs_compensation
+        source: compare-auto-decide
+        dest: auto-improve-loop
+      - trigger: compare_auto_decide_failed
+        source: compare-auto-decide
+        dest: failed
+      - trigger: auto_improve_loop_complete
+        source: auto-improve-loop
+        dest: generate-report
+      - trigger: auto_improve_loop_escalate
+        source: auto-improve-loop
+        dest: generate-report
+      - trigger: auto_improve_loop_failed
+        source: auto-improve-loop
+        dest: failed
+      - trigger: generate_report_complete
+        source: generate-report
+        dest: complete
+      - trigger: generate_report_failed
+        source: generate-report
+        dest: failed
+      - trigger: fail
+        source: "*"
+        dest: failed
+  sequence:
+    - step: phase-0
+      id: phase-0
+      phase: 1
+      phase_name: PRE-FLIGHT
+      agent: pedro-valerio
+      action: execute_pre-flight
+      task_ref: qualify-task
+      next: phase-1
+    - step: phase-1
+      id: phase-1
+      phase: 2
+      phase_name: PARALLEL EXECUTION
+      agent: pedro-valerio
+      action: execute_parallel-execution
+      task_ref: qualify-task
+      next: phase-2
+    - step: phase-2
+      id: phase-2
+      phase: 3
+      phase_name: COMPARE & AUTO-DECIDE
+      agent: pedro-valerio
+      action: execute_compare-auto-decide
+      task_ref: an-compare-outputs
+      next: phase-3
+    - step: phase-3
+      id: phase-3
+      phase: 4
+      phase_name: AUTO-IMPROVE LOOP
+      agent: pedro-valerio
+      action: execute_auto-improve-loop
+      task_ref: qualify-task
+      next: phase-4
+    - step: phase-4
+      id: phase-4
+      phase: 5
+      phase_name: GENERATE REPORT
+      agent: pedro-valerio
+      action: execute_generate-report
+      task_ref: qualify-task
+    - workflow_end:
+        id: complete
+        action: workflow_complete
 philosophy:
   core: |
     "Não assumir. Provar com dados."
     "Prioridade é $$ economia, não token economia."
     "Se permitir caminho errado, está errado."
-
   decision_patterns:
     - id: DP_01
-      name: "Self-Contained First"
-      rule: "Tudo dentro do squad-creator. Sem dependências externas."
+      name: Self-Contained First
+      rule: Tudo dentro do squad-creator. Sem dependências externas.
     - id: DP_02
-      name: "Prove que Funciona"
-      rule: "Validação empírica obrigatória. Documentação ≠ Funcionamento."
+      name: Prove que Funciona
+      rule: Validação empírica obrigatória. Documentação ≠ Funcionamento.
     - id: DP_03
-      name: "Questione o Caminho Errado"
-      rule: "Se executor CONSEGUE errar → redesenhar."
+      name: Questione o Caminho Errado
+      rule: Se executor CONSEGUE errar → redesenhar.
     - id: DP_04
-      name: "Testar Antes de Escalar"
-      rule: "Nunca batch sem validar 1 primeiro."
+      name: Testar Antes de Escalar
+      rule: Nunca batch sem validar 1 primeiro.
     - id: DP_05
-      name: "Cost-First"
-      rule: "Haiku usando 3x mais tokens = OK se custa 15x menos em $$."
+      name: Cost-First
+      rule: Haiku usando 3x mais tokens = OK se custa 15x menos em $$.
     - id: DP_06
-      name: "Mantenha Tudo Junto"
-      rule: "Outputs, logs, reports → tudo em test-cases/{task_name}/."
+      name: Mantenha Tudo Junto
+      rule: Outputs, logs, reports → tudo em test-cases/{task_name}/.
     - id: DP_07
-      name: "Documente o Processo"
-      rule: "O COMO é mais valioso que o resultado."
+      name: Documente o Processo
+      rule: O COMO é mais valioso que o resultado.
     - id: DP_08
-      name: "Auto-Improve"
-      rule: "Se falhar, aplicar compensações e re-testar AUTOMATICAMENTE."
-
+      name: Auto-Improve
+      rule: Se falhar, aplicar compensações e re-testar AUTOMATICAMENTE.
   cost_reference:
-    haiku:  { input_per_mtok: 0.25,  output_per_mtok: 1.25  }
-    sonnet: { input_per_mtok: 3.00,  output_per_mtok: 15.00 }
-    opus:   { input_per_mtok: 15.00, output_per_mtok: 75.00 }
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# VETO CONDITIONS (Global)
-# ═══════════════════════════════════════════════════════════════════════════════
-
+    haiku:
+      input_per_mtok: 0.25
+      output_per_mtok: 1.25
+    sonnet:
+      input_per_mtok: 3
+      output_per_mtok: 15
+    opus:
+      input_per_mtok: 15
+      output_per_mtok: 75
 veto_conditions:
   - id: MTQ_VC_001
-    trigger: "Task sem test_input definido no registry"
-    action: "BLOCK - Não testar. Adicionar input ao registry primeiro."
-
+    trigger: Task sem test_input definido no registry
+    action: BLOCK - Não testar. Adicionar input ao registry primeiro.
   - id: MTQ_VC_002
-    trigger: "Opus baseline output vazio ou erro"
-    action: "BLOCK - Não prosseguir sem baseline válido."
-
+    trigger: Opus baseline output vazio ou erro
+    action: BLOCK - Não prosseguir sem baseline válido.
   - id: MTQ_VC_003
-    trigger: "Haiku output não parseável (YAML inválido, estrutura quebrada)"
-    action: "BLOCK - Falha de formato. Compensar com COMP_002 ou COMP_003."
-
+    trigger: Haiku output não parseável (YAML inválido, estrutura quebrada)
+    action: BLOCK - Falha de formato. Compensar com COMP_002 ou COMP_003.
   - id: MTQ_VC_004
     trigger: "Haiku decisão oposta ao Opus (ex: PASS vs FAIL)"
-    action: "BLOCK - Haiku NÃO qualificado. Decisão errada é inaceitável."
-
+    action: BLOCK - Haiku NÃO qualificado. Decisão errada é inaceitável.
   - id: MTQ_VC_005
-    trigger: "Test case sem output salvo em arquivo"
-    action: "BLOCK - Tudo deve ser verificável por humano."
-
+    trigger: Test case sem output salvo em arquivo
+    action: BLOCK - Tudo deve ser verificável por humano.
   - id: MTQ_VC_006
-    trigger: "Mais de 2 rounds de compensação sem atingir threshold"
-    action: "ESCALATE - Promover para Sonnet e testar."
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# TEST INPUT REGISTRY
-# ═══════════════════════════════════════════════════════════════════════════════
-# Cada task precisa de um input REAL para teste.
-# Sem input definido → VETO (MTQ_VC_001)
-
+    trigger: Mais de 2 rounds de compensação sem atingir threshold
+    action: ESCALATE - Promover para Sonnet e testar.
 test_input_registry:
-  # HAIKU CANDIDATES (15 tasks)
   pv-axioma-assessment:
-    target: "squads/squad-creator-pro/workflows/wf-clone-mind.yaml"
-    type: "workflow_file"
+    target: squads/squad-creator-pro/workflows/wf-clone-mind.yaml
+    type: workflow_file
     tested: true
-
   pv-modernization-score:
-    target: "squads/squad-creator-pro/workflows/wf-clone-mind.yaml"
-    type: "workflow_file"
+    target: squads/squad-creator-pro/workflows/wf-clone-mind.yaml
+    type: workflow_file
     tested: false
-
   validate-squad:
-    target: "squads/copy"  # Real squad to validate
-    type: "squad_directory"
+    target: squads/copy
+    type: squad_directory
     tested: false
-
   validate-extraction:
-    target: "outputs/minds/pedro_valerio/extraction-report.yaml"
-    type: "extraction_output"
+    target: .aiox/squad-runtime/minds/pedro_valerio/extraction-report.yaml
+    type: extraction_output
     tested: false
-    note: "Needs existing extraction output to validate"
-
+    note: Needs existing extraction output to validate
   qa-after-creation:
-    target: "squads/copy"
-    type: "squad_directory"
+    target: squads/copy
+    type: squad_directory
     tested: false
-
   an-fidelity-score:
-    target: "outputs/minds/pedro_valerio"
-    type: "mind_directory"
+    target: .aiox/squad-runtime/minds/pedro_valerio
+    type: mind_directory
     tested: false
-
   an-clone-review:
-    target: ".claude/agents/mmos-victoria.md"
-    type: "agent_file"
+    target: .claude/agents/mmos-victoria.md
+    type: agent_file
     tested: false
-
   an-diagnose-clone:
-    target: ".claude/agents/mmos-victoria.md"
-    type: "agent_file"
+    target: .claude/agents/mmos-victoria.md
+    type: agent_file
     tested: false
-
   an-validate-clone:
-    target: ".claude/agents/mmos-victoria.md"
-    type: "agent_file"
+    target: .claude/agents/mmos-victoria.md
+    type: agent_file
     tested: false
-
   an-assess-sources:
-    target: "outputs/minds/pedro_valerio/sources/"
-    type: "sources_directory"
+    target: .aiox/squad-runtime/minds/pedro_valerio/sources/
+    type: sources_directory
     tested: false
-    note: "Needs mind with sources to assess"
-
+    note: Needs mind with sources to assess
   refresh-registry:
-    target: "squads/"
-    type: "squads_root"
+    target: squads/
+    type: squads_root
     tested: false
-
   squad-analytics:
-    target: "squads/"
-    type: "squads_root"
+    target: squads/
+    type: squads_root
     tested: false
-
   migrate-workflows-to-yaml:
-    target: "squads/squad-creator-pro/workflows/"
-    type: "workflows_directory"
+    target: squads/squad-creator-pro/workflows/
+    type: workflows_directory
     tested: false
-
-  install-commands:
-    target: "squads/squad-creator-pro/"
-    type: "squad_directory"
+  install-skills:
+    target: squads/squad-creator-pro/
+    type: squad_directory
     tested: false
-
-  sync-ide-command:
-    target: "squads/squad-creator-pro/"
-    type: "squad_directory"
+  sync-ide-skills:
+    target: squads/squad-creator-pro/
+    type: squad_directory
     tested: false
-
-  # SONNET CANDIDATES (test_with_haiku = true, 4 tasks)
   create-documentation:
-    target: "squads/copy"
-    type: "squad_directory"
+    target: squads/copy
+    type: squad_directory
     tested: false
-    note: "Sonnet candidate, also test in Haiku"
-
+    note: Sonnet candidate, also test in Haiku
   create-template:
-    target: "squads/squad-creator-pro/tasks/pv-axioma-assessment.md"
-    type: "task_file"
+    target: squads/squad-creator-pro/tasks/pv-axioma-assessment.md
+    type: task_file
     tested: false
-    note: "Create template from existing task"
-
+    note: Create template from existing task
   collect-sources:
-    target: "Pedro Valério"
-    type: "person_name"
+    target: Pedro Valério
+    type: person_name
     tested: false
-    note: "Web search for sources"
-
+    note: Web search for sources
   auto-acquire-sources:
-    target: "Pedro Valério"
-    type: "person_name"
+    target: Pedro Valério
+    type: person_name
     tested: false
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# INPUTS
-# ═══════════════════════════════════════════════════════════════════════════════
-
 inputs:
   required:
     - name: task_name
       type: string
-      description: "Nome da task (sem .md)"
-      example: "pv-axioma-assessment"
-      validation: "Must exist in test_input_registry"
-
+      description: Nome da task (sem .md)
+      example: pv-axioma-assessment
+      validation: Must exist in test_input_registry
   optional:
     - name: target_tier
       type: enum
-      values: ["haiku", "sonnet"]
-      default: "haiku"
-
+      values:
+        - haiku
+        - sonnet
+      default: haiku
     - name: threshold
       type: number
-      default: 0.90
-      description: "Threshold mínimo de qualidade vs baseline"
-
+      default: 0.9
+      description: Threshold mínimo de qualidade vs baseline
     - name: max_compensation_rounds
       type: number
       default: 2
-      description: "Máximo de rounds de compensação antes de escalar"
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# PHASES
-# ═══════════════════════════════════════════════════════════════════════════════
-
+      description: Máximo de rounds de compensação antes de escalar
 phases:
-  # ─────────────────────────────────────────────────────────────────────────────
-  # PHASE 0: PRE-FLIGHT VALIDATION
-  # ─────────────────────────────────────────────────────────────────────────────
   - id: phase_0
-    name: "PRE-FLIGHT"
-    purpose: "Validar que temos tudo para rodar o teste"
-    duration: "< 30s"
-
+    name: PRE-FLIGHT
+    purpose: Validar que temos tudo para rodar o teste
+    duration: < 30s
     steps:
       - id: step_0_1
-        name: "Check test_input_registry"
-        action: "Lookup task_name in test_input_registry"
-        veto_if: "task_name not in registry → MTQ_VC_001"
+        name: Check test_input_registry
+        action: Lookup task_name in test_input_registry
+        veto_if: task_name not in registry → MTQ_VC_001
         output:
-          test_input: "Input from registry"
-
+          test_input: Input from registry
       - id: step_0_2
-        name: "Check task file exists"
-        action: "Verify squads/squad-creator-pro/tasks/{task_name}.md exists"
-        veto_if: "File not found"
-
+        name: Check task file exists
+        action: Verify squads/squad-creator-pro/tasks/{task_name}.md exists
+        veto_if: File not found
       - id: step_0_3
-        name: "Read task file"
-        action: "Read COMPLETE task file content"
+        name: Read task file
+        action: Read COMPLETE task file content
         output:
-          task_prompt: "Full task content"
-
+          task_prompt: Full task content
       - id: step_0_4
-        name: "Create test directory"
-        action: "mkdir -p squads/squad-creator-pro/test-cases/{task_name}/"
-
+        name: Create test directory
+        action: mkdir -p squads/squad-creator-pro/test-cases/{task_name}/
     checkpoint:
       id: CP_PREFLIGHT
       blocking: true
       criteria:
-        - test_input: "resolved from registry"
-        - task_prompt: "loaded"
-        - test_directory: "created"
-
-  # ─────────────────────────────────────────────────────────────────────────────
-  # PHASE 1: PARALLEL EXECUTION (Opus + Haiku simultaneously)
-  # ─────────────────────────────────────────────────────────────────────────────
+        - test_input: resolved from registry
+        - task_prompt: loaded
+        - test_directory: created
   - id: phase_1
-    name: "PARALLEL EXECUTION"
-    purpose: "Rodar Opus e Haiku AO MESMO TEMPO para economizar tempo"
-    duration: "2-5 min"
-    execution: "parallel"
-
+    name: PARALLEL EXECUTION
+    purpose: Rodar Opus e Haiku AO MESMO TEMPO para economizar tempo
+    duration: 2-5 min
+    execution: parallel
     steps:
       - id: step_1_1
-        name: "Execute in Opus (baseline)"
-        execution: "parallel_a"
+        name: Execute in Opus (baseline)
+        execution: parallel_a
         action: |
           Task(
             subagent_type: "general-purpose",
@@ -287,13 +308,14 @@ phases:
                      Include metadata: tokens used, duration."
           )
         output:
-          opus_result: "Full output saved to file"
-          opus_tokens: { input: 0, output: 0 }
+          opus_result: Full output saved to file
+          opus_tokens:
+            input: 0
+            output: 0
           opus_latency_ms: 0
-
       - id: step_1_2
-        name: "Execute in Haiku (candidate)"
-        execution: "parallel_b"
+        name: Execute in Haiku (candidate)
+        execution: parallel_b
         action: |
           Task(
             subagent_type: "general-purpose",
@@ -304,34 +326,29 @@ phases:
                      Include metadata: tokens used, duration."
           )
         output:
-          haiku_result: "Full output saved to file"
-          haiku_tokens: { input: 0, output: 0 }
+          haiku_result: Full output saved to file
+          haiku_tokens:
+            input: 0
+            output: 0
           haiku_latency_ms: 0
-
     checkpoint:
       id: CP_EXECUTION
       blocking: true
       criteria:
-        - opus_result: "file exists and not empty (MTQ_VC_002)"
-        - haiku_result: "file exists and not empty"
-
-  # ─────────────────────────────────────────────────────────────────────────────
-  # PHASE 2: COMPARE & AUTO-DECIDE
-  # ─────────────────────────────────────────────────────────────────────────────
+        - opus_result: file exists and not empty (MTQ_VC_002)
+        - haiku_result: file exists and not empty
   - id: phase_2
-    name: "COMPARE & AUTO-DECIDE"
-    purpose: "Comparar outputs, calcular custo, decidir próximo passo"
-    duration: "1-3 min"
-
+    name: COMPARE & AUTO-DECIDE
+    purpose: Comparar outputs, calcular custo, decidir próximo passo
+    duration: 1-3 min
     steps:
       - id: step_2_1
-        name: "Read both outputs"
+        name: Read both outputs
         action: |
           Read opus-baseline.yaml and haiku-round-1.yaml
           Parse both as structured data
-
       - id: step_2_2
-        name: "Compare across 5 dimensions"
+        name: Compare across 5 dimensions
         action: |
           Score each dimension (0-10):
 
@@ -358,12 +375,11 @@ phases:
              - Count recommendations
              - More = better (Haiku giving more is a BONUS)
         output:
-          dimension_scores: "Per-dimension breakdown"
-          quality_score: "Weighted average as percentage (0-100%)"
-          gaps_identified: "List of gap types for compensation matching"
-
+          dimension_scores: Per-dimension breakdown
+          quality_score: Weighted average as percentage (0-100%)
+          gaps_identified: List of gap types for compensation matching
       - id: step_2_3
-        name: "Calculate cost"
+        name: Calculate cost
         action: |
           opus_cost = (opus_tokens.input * 15.00 / 1_000_000) + (opus_tokens.output * 75.00 / 1_000_000)
           haiku_cost = (haiku_tokens.input * 0.25 / 1_000_000) + (haiku_tokens.output * 1.25 / 1_000_000)
@@ -372,9 +388,8 @@ phases:
           opus_cost_usd: 0
           haiku_cost_usd: 0
           savings_percentage: 0
-
       - id: step_2_4
-        name: "Decision gate"
+        name: Decision gate
         action: |
           # DECISION TREE (unidirectional, no going back)
 
@@ -395,27 +410,20 @@ phases:
           ELSE:
             decision = "NEEDS_COMPENSATION"
             → GOTO phase_3
-
         output:
-          decision: "QUALIFIED | NEEDS_COMPENSATION | ESCALATE_TO_SONNET"
+          decision: QUALIFIED | NEEDS_COMPENSATION | ESCALATE_TO_SONNET
           round: 1
-
     checkpoint:
       id: CP_DECISION
       blocking: true
       output:
-        saved_to: "test-cases/{task_name}/comparison-round-1.yaml"
-
-  # ─────────────────────────────────────────────────────────────────────────────
-  # PHASE 3: AUTO-IMPROVE LOOP (compensate + re-test)
-  # ─────────────────────────────────────────────────────────────────────────────
+        saved_to: test-cases/{task_name}/comparison-round-1.yaml
   - id: phase_3
-    name: "AUTO-IMPROVE LOOP"
-    purpose: "Aplicar compensações ao task E re-testar automaticamente"
-    duration: "3-8 min per round"
-    condition: "decision == 'NEEDS_COMPENSATION'"
+    name: AUTO-IMPROVE LOOP
+    purpose: Aplicar compensações ao task E re-testar automaticamente
+    duration: 3-8 min per round
+    condition: decision == 'NEEDS_COMPENSATION'
     max_iterations: 2
-
     auto_improve_cycle:
       description: |
         Este é o loop que Alan pediu:
@@ -427,11 +435,12 @@ phases:
         6. Se >= threshold → QUALIFIED_WITH_COMPENSATION
         7. Se < threshold E rounds < max → repetir loop
         8. Se < threshold E rounds >= max → ESCALATE
-
     compensations:
       - id: COMP_001
-        name: "Scoring Calibration"
-        triggers: ["score_variance", "conservative_scoring"]
+        name: Scoring Calibration
+        triggers:
+          - score_variance
+          - conservative_scoring
         priority: 1
         action: |
           Append to task file:
@@ -441,10 +450,12 @@ phases:
           Se o processo TEM a característica → mínimo 7.0.
           Gaps vão para recommendations, NÃO reduzem score.
           Evidence deve ser POSITIVA primeiro.
-
       - id: COMP_002
-        name: "Output Example"
-        triggers: ["format_issues", "missing_sections", "structure_mismatch"]
+        name: Output Example
+        triggers:
+          - format_issues
+          - missing_sections
+          - structure_mismatch
         priority: 2
         action: |
           Append to task file:
@@ -454,10 +465,11 @@ phases:
           ```yaml
           {first_100_lines_of_opus_baseline}
           ```
-
       - id: COMP_003
-        name: "Completion Checklist"
-        triggers: ["incomplete_output", "missing_fields"]
+        name: Completion Checklist
+        triggers:
+          - incomplete_output
+          - missing_fields
         priority: 3
         action: |
           Append to task file:
@@ -468,10 +480,11 @@ phases:
           - [ ] Each has score + evidence
           - [ ] Overall score calculated correctly
           - [ ] Recommendations are specific and actionable
-
       - id: COMP_004
-        name: "Reasoning Depth"
-        triggers: ["shallow_reasoning", "generic_evidence"]
+        name: Reasoning Depth
+        triggers:
+          - shallow_reasoning
+          - generic_evidence
         priority: 4
         action: |
           Append to task file:
@@ -481,25 +494,22 @@ phases:
           1. List what the process DOES WELL (positive first)
           2. Note gaps (for recommendations only)
           3. Cite specific files, line counts, IDs
-
     steps:
       - id: step_3_1
-        name: "Match gaps to compensations"
+        name: Match gaps to compensations
         action: |
           For each gap in gaps_identified:
             Find matching COMP_* by triggers
             Add to compensations_to_apply (ordered by priority)
-
       - id: step_3_2
-        name: "Apply compensations to task file"
+        name: Apply compensations to task file
         action: |
           Read squads/squad-creator-pro/tasks/{task_name}.md
           Append selected compensations
           Save modified task
           Log: test-cases/{task_name}/compensation-round-{N}.yaml
-
       - id: step_3_3
-        name: "Re-execute in Haiku"
+        name: Re-execute in Haiku
         action: |
           Task(
             subagent_type: "general-purpose",
@@ -508,13 +518,11 @@ phases:
                      Execute with input: {test_input.target}
                      Save to squads/squad-creator-pro/test-cases/{task_name}/haiku-round-{N+1}.yaml"
           )
-
       - id: step_3_4
-        name: "Re-compare"
-        action: "Same comparison as phase_2.step_2_2"
-
+        name: Re-compare
+        action: Same comparison as phase_2.step_2_2
       - id: step_3_5
-        name: "Loop decision"
+        name: Loop decision
         action: |
           IF quality_score_new >= threshold * 100:
             decision = "QUALIFIED_WITH_COMPENSATION"
@@ -527,23 +535,17 @@ phases:
           ELSE:
             round += 1
             → GOTO step_3_1 (loop)
-
     veto_conditions:
       - id: MTQ_VC_LOOP
-        trigger: "Same compensation applied twice without improvement"
-        action: "STOP loop. Escalate."
-
-  # ─────────────────────────────────────────────────────────────────────────────
-  # PHASE 4: GENERATE REPORT
-  # ─────────────────────────────────────────────────────────────────────────────
+        trigger: Same compensation applied twice without improvement
+        action: STOP loop. Escalate.
   - id: phase_4
-    name: "GENERATE REPORT"
-    purpose: "Criar relatório final com TODA a evidência para review humano"
-    duration: "1-2 min"
-
+    name: GENERATE REPORT
+    purpose: Criar relatório final com TODA a evidência para review humano
+    duration: 1-2 min
     steps:
       - id: step_4_1
-        name: "Generate qualification report"
+        name: Generate qualification report
         action: |
           Create test-cases/{task_name}/qualification-report.yaml with ALL data:
 
@@ -604,25 +606,22 @@ phases:
               - "comparison-round-1.yaml"
               - "compensation-round-1.yaml"
               - "qualification-report.yaml"
-
       - id: step_4_2
-        name: "Update test-case.yaml"
+        name: Update test-case.yaml
         action: |
           Update test-cases/{task_name}/test-case.yaml with results:
           - results.haiku.score
           - results.haiku.cost_usd
           - results.haiku.qualified
           - results.final_tier
-
       - id: step_4_3
-        name: "Update model-routing.yaml"
-        condition: "decision is QUALIFIED or QUALIFIED_WITH_COMPENSATION"
+        name: Update model-routing.yaml
+        condition: decision is QUALIFIED or QUALIFIED_WITH_COMPENSATION
         action: |
           node squads/squad-creator-pro/scripts/model-tier-validator.cjs \
             update-routing {task_name} {recommended_tier} "{reason}"
-
       - id: step_4_4
-        name: "Display summary"
+        name: Display summary
         action: |
           ╔══════════════════════════════════════════════════════════════════╗
           ║  MODEL TIER QUALIFICATION RESULT                                 ║
@@ -639,28 +638,20 @@ phases:
           ║                                                                  ║
           ║  Files: test-cases/{task_name}/                                  ║
           ╚══════════════════════════════════════════════════════════════════╝
-
     veto_conditions:
       - id: MTQ_VC_REPORT
-        trigger: "Report file not created"
-        action: "BLOCK - Cannot complete without evidence file."
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# BATCH EXECUTION
-# ═══════════════════════════════════════════════════════════════════════════════
-
+        trigger: Report file not created
+        action: BLOCK - Cannot complete without evidence file.
 batch_mode:
   description: |
     Rodar qualificação em TODAS as tasks candidatas.
     Usa test_input_registry para saber O QUE testar.
     Gera consolidated report no final.
-
   execution_order:
-    # Wave 1: High-confidence Haiku candidates (testáveis agora)
     wave_1:
       description: "Tasks com test_input definido e confidence: high"
       tasks:
-        - pv-axioma-assessment     # ALREADY TESTED ✅
+        - pv-axioma-assessment
         - pv-modernization-score
         - validate-squad
         - validate-extraction
@@ -670,34 +661,28 @@ batch_mode:
         - an-diagnose-clone
         - an-validate-clone
         - an-assess-sources
-
-    # Wave 2: Admin tasks (simpler, likely all pass)
     wave_2:
-      description: "Tasks administrativas (script-based)"
+      description: Tasks administrativas (script-based)
       tasks:
         - refresh-registry
         - squad-analytics
         - migrate-workflows-to-yaml
-        - install-commands
-        - sync-ide-command
-
-    # Wave 3: Sonnet candidates tested with Haiku
+        - install-skills
+        - sync-ide-skills
     wave_3:
-      description: "Tasks sonnet com test_with_haiku=true"
+      description: Tasks sonnet com test_with_haiku=true
       tasks:
         - create-documentation
         - create-template
         - collect-sources
         - auto-acquire-sources
-
   parallelization:
     description: |
       Dentro de cada wave, tasks são executadas SEQUENCIALMENTE
       (para evitar confusão de outputs e permitir learning).
       Mas dentro de cada task, Opus+Haiku rodam em PARALELO (Phase 1).
-
   consolidated_report:
-    output: "squads/squad-creator-pro/test-cases/QUALIFICATION-DASHBOARD.yaml"
+    output: squads/squad-creator-pro/test-cases/QUALIFICATION-DASHBOARD.yaml
     format: |
       qualification_dashboard:
         date: "{today}"
@@ -727,74 +712,26 @@ batch_mode:
           - "Create skills for qualified Haiku tasks"
           - "Test Sonnet candidates"
           - "Review Opus-required tasks"
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# FAILURE PATHS
-# ═══════════════════════════════════════════════════════════════════════════════
-
 failure_paths:
-  - scenario: "Task file not found"
-    action: "Skip task, log error, continue batch"
-
-  - scenario: "Opus execution fails"
-    action: "STOP for this task. Cannot qualify without baseline. Log and skip."
-
-  - scenario: "Haiku execution fails (timeout, error)"
-    action: "Retry once. If fails again → OPUS_REQUIRED."
-
-  - scenario: "Haiku gives opposite decision (PASS vs FAIL)"
-    action: "IMMEDIATE FAIL. MTQ_VC_004. No compensation can fix wrong decisions."
-
-  - scenario: "Compensation makes quality WORSE"
-    action: "Revert compensation. Try next compensation type. If none left → ESCALATE."
-
-  - scenario: "test_input doesn't exist (file deleted, path wrong)"
-    action: "STOP for this task. Update registry. Log and skip."
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# OUTPUTS
-# ═══════════════════════════════════════════════════════════════════════════════
-
+  - scenario: Task file not found
+    action: Skip task, log error, continue batch
+  - scenario: Opus execution fails
+    action: STOP for this task. Cannot qualify without baseline. Log and skip.
+  - scenario: Haiku execution fails (timeout, error)
+    action: Retry once. If fails again → OPUS_REQUIRED.
+  - scenario: Haiku gives opposite decision (PASS vs FAIL)
+    action: IMMEDIATE FAIL. MTQ_VC_004. No compensation can fix wrong decisions.
+  - scenario: Compensation makes quality WORSE
+    action: Revert compensation. Try next compensation type. If none left → ESCALATE.
+  - scenario: test_input doesn't exist (file deleted, path wrong)
+    action: STOP for this task. Update registry. Log and skip.
 outputs:
   per_task:
-    - "test-cases/{task_name}/opus-baseline.yaml"
-    - "test-cases/{task_name}/haiku-round-1.yaml"
-    - "test-cases/{task_name}/haiku-round-{N}.yaml (if compensated)"
-    - "test-cases/{task_name}/comparison-round-{N}.yaml"
-    - "test-cases/{task_name}/compensation-round-{N}.yaml (if applied)"
-    - "test-cases/{task_name}/qualification-report.yaml"
-
+    - test-cases/{task_name}/opus-baseline.yaml
+    - test-cases/{task_name}/haiku-round-1.yaml
+    - test-cases/{task_name}/haiku-round-{N}.yaml (if compensated)
+    - test-cases/{task_name}/comparison-round-{N}.yaml
+    - test-cases/{task_name}/compensation-round-{N}.yaml (if applied)
+    - test-cases/{task_name}/qualification-report.yaml
   batch:
-    - "test-cases/QUALIFICATION-DASHBOARD.yaml"
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# METADATA
-# ═══════════════════════════════════════════════════════════════════════════════
-
-metadata:
-  version: "2.0"
-  created: "2026-02-11"
-  author: "@pedro-valerio"
-  changelog:
-    - version: "2.0"
-      date: "2026-02-11"
-      changes:
-        - "REWRITE: Added phase_0 (pre-flight validation)"
-        - "REWRITE: Phase 1 now runs Opus + Haiku in PARALLEL"
-        - "REWRITE: Merged old phases 2+3 into phase_2 (compare + decide)"
-        - "REWRITE: Phase 3 is now AUTO-IMPROVE LOOP (compensate + re-test)"
-        - "REWRITE: Phase 4 is report generation"
-        - "NEW: 6 veto conditions (MTQ_VC_001 to MTQ_VC_006)"
-        - "NEW: test_input_registry with 19 tasks mapped"
-        - "NEW: Failure paths for 6 scenarios"
-        - "NEW: Batch mode with 3 waves and consolidated dashboard"
-        - "NEW: Cost-first decision logic from ADR-001"
-        - "NEW: Alan's 8 decision patterns embedded in philosophy"
-        - "FIX: Comparison is now more deterministic (section counts, parse tests)"
-        - "FIX: Scoring Calibration compensation included by default"
-    - version: "1.0"
-      date: "2026-02-11"
-      changes:
-        - "Initial workflow based on empirical testing session"
-        - "4 compensation strategies defined"
-        - "2-round testing with automatic improvement"
+    - test-cases/QUALIFICATION-DASHBOARD.yaml