maestro-flow 0.4.20 → 0.4.21

package/.agy/skills/team-adversarial-swarm/workflows/wf-swarm-explore.js

@@ -0,0 +1,194 @@
+export const meta = {
+  name: 'wf-swarm-explore',
+  description: 'Parallel ant exploration — N ants explore task space concurrently guided by pheromone hints',
+  whenToUse: 'Single ACO iteration: spawn N ants in parallel, each builds a path through the task space',
+  phases: [
+    { title: 'Explore', detail: 'N ants explore task space in parallel' },
+    { title: 'Validate', detail: 'Cross-validate ant paths for node validity and evidence' },
+  ],
+}
+
+const ANT_RESULT_SCHEMA = {
+  type: 'object',
+  properties: {
+    ant_id: { type: 'string' },
+    iteration: { type: 'number' },
+    path: { type: 'array', items: { type: 'string' } },
+    path_decisions: {
+      type: 'array',
+      items: {
+        type: 'object',
+        properties: {
+          from: { type: 'string' },
+          to: { type: 'string' },
+          rationale: { type: 'string' },
+          guided_by: { type: 'string', enum: ['pheromone', 'heuristic', 'evidence'] },
+          pheromone_weight: { type: 'number' },
+          deviation_from_hint: { type: 'boolean' },
+        },
+        required: ['from', 'to', 'rationale', 'guided_by'],
+      },
+    },
+    self_score: { type: 'number', minimum: 0, maximum: 1 },
+    self_confidence: { type: 'number', minimum: 0, maximum: 1 },
+    candidate_solution: {
+      type: 'object',
+      properties: {
+        type: { type: 'string', enum: ['string', 'object', 'file_ref'] },
+        summary: { type: 'string' },
+        content: { type: 'string' },
+      },
+      required: ['summary'],
+    },
+    evidence: {
+      type: 'array',
+      items: {
+        type: 'object',
+        properties: {
+          source: { type: 'string' },
+          finding: { type: 'string' },
+          strength: { type: 'string', enum: ['strong', 'moderate', 'weak'] },
+        },
+        required: ['source', 'finding'],
+      },
+    },
+    notes: { type: 'string' },
+  },
+  required: ['ant_id', 'iteration', 'path', 'path_decisions', 'self_score', 'self_confidence', 'candidate_solution', 'evidence'],
+}
+
+const VALIDATION_SCHEMA = {
+  type: 'object',
+  properties: {
+    ant_id: { type: 'string' },
+    valid: { type: 'boolean' },
+    issues: { type: 'array', items: { type: 'string' } },
+    evidence_verified: { type: 'number' },
+    evidence_total: { type: 'number' },
+  },
+  required: ['ant_id', 'valid', 'issues'],
+}
+
+const iteration = args?.iteration || 1
+const assignments = args?.assignments || []
+const objective = args?.objective || ''
+const session = args?.session || ''
+const config = args?.config || {}
+const taskSpace = args?.task_space || []
+const wisdom = args?.wisdom || ''
+
+// Phase 1: Parallel ant exploration
+phase('Explore')
+log(`Iteration ${iteration}: launching ${assignments.length} ants in parallel...`)
+
+const antResults = await parallel(
+  assignments.map((assignment, idx) => () =>
+    agent(
+      `You are ANT-${iteration}-${idx + 1} in an ant colony optimization swarm.
+
+## Objective
+${objective}
+
+## Your Assignment
+Start node: ${assignment.start_node}
+Edge preferences (pheromone-derived weights):
+${JSON.stringify(assignment.edge_preferences || {}, null, 2)}
+Max path length: ${assignment.max_path_length || 5}
+
+## Task Space
+Valid nodes: ${JSON.stringify(taskSpace.slice(0, 50))}${taskSpace.length > 50 ? '... (' + taskSpace.length + ' total)' : ''}
+
+## Session
+Session path: ${session}
+${wisdom ? 'Prior iteration learnings:\n' + wisdom : ''}
+
+## Instructions
+1. Read the task space to understand what each node represents
+2. Start from your assigned start_node
+3. At each step, evaluate candidate next nodes:
+   - Use edge_preferences as pheromone guidance (higher = more explored/promising)
+   - BUT use your OWN judgment — deviate when evidence supports a different path
+   - Record whether each decision was guided by pheromone, heuristic, or evidence
+4. Build a path of 1..${assignment.max_path_length || 5} nodes (no revisiting)
+5. Gather EVIDENCE along your path (file:line references, code snippets, test results)
+6. Self-evaluate: score (0-1) how well your path achieves the objective
+7. Extract a candidate_solution from your exploration
+
+Be thorough in evidence gathering. Read actual source files, run greps, verify claims.
+${config.evidence_requirements ? 'Evidence requirements: ' + config.evidence_requirements : ''}`,
+      {
+        label: `ant:${iteration}-${idx + 1}`,
+        phase: 'Explore',
+        schema: ANT_RESULT_SCHEMA,
+        agentType: 'cli-explore-agent',
+      }
+    )
+  )
+)
+
+const validAnts = antResults.filter(Boolean)
+log(`${validAnts.length}/${assignments.length} ants completed exploration`)
+
+// Phase 2: Cross-validate ant paths
+phase('Validate')
+
+if (validAnts.length > 0) {
+  log(`Validating ${validAnts.length} ant paths...`)
+
+  const validations = await parallel(
+    validAnts.map(ant => () =>
+      agent(
+        `Validate this ant's exploration results.
+
+Ant: ${ant.ant_id}
+Path: ${ant.path.join(' → ')}
+Self-score: ${ant.self_score} (confidence: ${ant.self_confidence})
+Evidence count: ${ant.evidence.length}
+Solution summary: ${ant.candidate_solution.summary}
+
+Task space nodes: ${JSON.stringify(taskSpace.slice(0, 30))}
+Session: ${session}
+
+Validate:
+1. Every node in path exists in the task space
+2. Path has no cycles (no repeated nodes)
+3. path_decisions length == path length - 1
+4. At least 1 evidence item exists
+5. If evidence references files — verify they exist (Read/Glob)
+6. self_score > 0.9 requires ≥3 evidence items
+
+Report issues found. Set valid=true only if no blocking issues.`,
+        { label: `validate:${ant.ant_id}`, phase: 'Validate', schema: VALIDATION_SCHEMA }
+      )
+    )
+  )
+
+  const validResults = validations.filter(Boolean)
+  const validCount = validResults.filter(v => v.valid).length
+  log(`Validation: ${validCount}/${validResults.length} ants passed`)
+
+  return {
+    iteration: iteration,
+    ant_results: validAnts.map((ant, i) => ({
+      ...ant,
+      validation: validResults[i] || null,
+    })),
+    metadata: {
+      total_ants: assignments.length,
+      completed_ants: validAnts.length,
+      valid_ants: validCount,
+      avg_self_score: validAnts.length > 0
+        ? Math.round(validAnts.reduce((s, a) => s + a.self_score, 0) / validAnts.length * 100) / 100
+        : 0,
+      avg_path_length: validAnts.length > 0
+        ? Math.round(validAnts.reduce((s, a) => s + a.path.length, 0) / validAnts.length * 10) / 10
+        : 0,
+    },
+  }
+}
+
+return {
+  iteration: iteration,
+  ant_results: [],
+  metadata: { total_ants: assignments.length, completed_ants: 0, valid_ants: 0 },
+}

package/.agy/skills/team-adversarial-swarm/workflows/wf-swarm-score.js

@@ -0,0 +1,188 @@
+export const meta = {
+  name: 'wf-swarm-score',
+  description: 'Adversarial 3-vote scoring — prosecutor/defender/judge per ant for verified scores',
+  whenToUse: 'Score ant results from one iteration using adversarial 3-vote pattern instead of single scorer',
+  phases: [
+    { title: 'Score', detail: '3-vote adversarial scoring per ant (prosecutor/defender/judge)' },
+    { title: 'Calibrate', detail: 'Cross-ant calibration and hallucination detection' },
+  ],
+}
+
+const VOTE_SCHEMA = {
+  type: 'object',
+  properties: {
+    ant_id: { type: 'string' },
+    role: { type: 'string', enum: ['prosecutor', 'defender', 'judge'] },
+    score: { type: 'number', minimum: 0, maximum: 1 },
+    reasoning: { type: 'string' },
+    evidence_verified: { type: 'number' },
+    evidence_total: { type: 'number' },
+    confidence: { type: 'number', minimum: 0, maximum: 100 },
+  },
+  required: ['ant_id', 'role', 'score', 'reasoning', 'confidence'],
+}
+
+const CALIBRATION_SCHEMA = {
+  type: 'object',
+  properties: {
+    scores: {
+      type: 'object',
+      additionalProperties: {
+        type: 'object',
+        properties: {
+          verified_score: { type: 'number' },
+          rationale: { type: 'string' },
+          votes: { type: 'object' },
+          hallucination_flag: { type: 'boolean' },
+          self_vs_verified_delta: { type: 'number' },
+        },
+        required: ['verified_score', 'rationale'],
+      },
+    },
+    calibration: {
+      type: 'object',
+      properties: {
+        mean: { type: 'number' },
+        std: { type: 'number' },
+        min: { type: 'number' },
+        max: { type: 'number' },
+        hallucination_rate: { type: 'number' },
+      },
+      required: ['mean', 'min', 'max'],
+    },
+    ranking: { type: 'array', items: { type: 'string' } },
+    warnings: { type: 'array', items: { type: 'string' } },
+  },
+  required: ['scores', 'calibration', 'ranking'],
+}
+
+const iteration = args?.iteration || 1
+const antResults = args?.ant_results || []
+const objective = args?.objective || ''
+const rubric = args?.rubric || ''
+
+if (antResults.length === 0) {
+  log('No ant results to score')
+  return { scores: {}, calibration: { mean: 0, min: 0, max: 0 }, ranking: [] }
+}
+
+// Phase 1: 3-vote adversarial scoring per ant
+phase('Score')
+log(`Adversarial 3-vote scoring of ${antResults.length} ants...`)
+
+const allVotes = await pipeline(
+  antResults,
+  (ant) => parallel([
+    () => agent(
+      `PROSECUTOR: Score this ant's result HARSHLY. Find flaws.
+
+Objective: ${objective}
+${rubric ? 'Rubric: ' + rubric : 'Default rubric: path_relevance(0.35) + evidence_strength(0.30) + solution_quality(0.25) + path_coherence(0.10)'}
+
+Ant: ${ant.ant_id}
+Path: ${ant.path.join(' → ')} (${ant.path.length} nodes)
+Decisions: ${ant.path_decisions.map(d => d.from + '→' + d.to + ' [' + d.guided_by + '] ' + d.rationale).join('; ')}
+Self-score: ${ant.self_score} (DO NOT anchor on this — score blind first)
+Evidence: ${ant.evidence.map(e => e.source + ': ' + e.finding + ' [' + (e.strength || 'unknown') + ']').join('\n')}
+Solution: ${ant.candidate_solution.summary}
+
+Your job: MINIMIZE the score. Find every weakness.
+- Does the path actually address the objective?
+- Is the evidence real and strong, or vague/unverifiable?
+- Is the solution actionable or hand-wavy?
+- Are there logical gaps in the path decisions?
+
+Score 0.0-1.0. Verify evidence count if possible. Be harsh but fair.`,
+      { label: `prosecutor:${ant.ant_id}`, phase: 'Score', schema: VOTE_SCHEMA }
+    ),
+    () => agent(
+      `DEFENDER: Score this ant's result GENEROUSLY. Find strengths.
+
+Objective: ${objective}
+${rubric ? 'Rubric: ' + rubric : 'Default rubric: path_relevance(0.35) + evidence_strength(0.30) + solution_quality(0.25) + path_coherence(0.10)'}
+
+Ant: ${ant.ant_id}
+Path: ${ant.path.join(' → ')} (${ant.path.length} nodes)
+Decisions: ${ant.path_decisions.map(d => d.from + '→' + d.to + ' [' + d.guided_by + '] ' + d.rationale).join('; ')}
+Self-score: ${ant.self_score} (DO NOT anchor on this — score blind first)
+Evidence: ${ant.evidence.map(e => e.source + ': ' + e.finding + ' [' + (e.strength || 'unknown') + ']').join('\n')}
+Solution: ${ant.candidate_solution.summary}
+
+Your job: MAXIMIZE the score. Find every strength.
+- Does the path show creative or insightful exploration?
+- Is the evidence concrete even if limited?
+- Does the solution provide actionable value?
+- Are path deviations from pheromone justified?
+
+Score 0.0-1.0. Be generous but honest. Don't inflate without basis.`,
+      { label: `defender:${ant.ant_id}`, phase: 'Score', schema: VOTE_SCHEMA }
+    ),
+    () => agent(
+      `JUDGE: Score this ant's result OBJECTIVELY. No bias.
+
+Objective: ${objective}
+${rubric ? 'Rubric: ' + rubric : 'Default rubric: path_relevance(0.35) + evidence_strength(0.30) + solution_quality(0.25) + path_coherence(0.10)'}
+
+Ant: ${ant.ant_id}
+Path: ${ant.path.join(' → ')} (${ant.path.length} nodes)
+Decisions: ${ant.path_decisions.map(d => d.from + '→' + d.to + ' [' + d.guided_by + '] ' + d.rationale).join('; ')}
+Self-score: ${ant.self_score} (DO NOT anchor on this — score blind first)
+Evidence: ${ant.evidence.map(e => e.source + ': ' + e.finding + ' [' + (e.strength || 'unknown') + ']').join('\n')}
+Solution: ${ant.candidate_solution.summary}
+
+Your job: Score PURELY on evidence. No default bias.
+- Apply rubric dimensions systematically
+- Weight each dimension, compute total
+- Verify evidence references if possible (Read files cited)
+- Compare path coherence objectively
+
+Score 0.0-1.0. Confidence reflects evidence coverage.`,
+      { label: `judge:${ant.ant_id}`, phase: 'Score', schema: VOTE_SCHEMA }
+    ),
+  ])
+)
+
+log(`${allVotes.filter(Boolean).length}/${antResults.length} ants scored by 3-vote panel`)
+
+// Phase 2: Calibrate across all ants
+phase('Calibrate')
+log('Cross-ant calibration and hallucination detection...')
+
+const voteDigest = antResults.map((ant, i) => {
+  const votes = allVotes[i]
+  if (!votes) return `${ant.ant_id}: no votes`
+  const validVotes = votes.filter(Boolean)
+  const scores = validVotes.map(v => v.score)
+  const avgScore = scores.length > 0 ? scores.reduce((a, b) => a + b, 0) / scores.length : 0
+  return `${ant.ant_id}: self=${ant.self_score} | prosecutor=${validVotes.find(v => v.role === 'prosecutor')?.score || '?'} defender=${validVotes.find(v => v.role === 'defender')?.score || '?'} judge=${validVotes.find(v => v.role === 'judge')?.score || '?'} | avg=${Math.round(avgScore * 100) / 100} | delta=${Math.round(Math.abs(ant.self_score - avgScore) * 100) / 100}`
+}).join('\n')
+
+const calibration = await agent(
+  `Calibrate adversarial scores across ${antResults.length} ants.
+
+Per-ant votes:
+${voteDigest}
+
+Tasks:
+1. For each ant: compute verified_score as weighted average (prosecutor 0.25, defender 0.25, judge 0.50)
+2. Compare self_score vs verified_score — flag hallucination if delta > 0.3
+3. If all scores within ±0.05 (compressed range) — force differentiation by re-ranking
+4. Compute calibration stats (mean, std, min, max, hallucination_rate)
+5. Produce ranking (best to worst by verified_score)
+6. Warnings: flag if >50% ants are hallucinating, if range is too compressed, etc.
+
+Return the complete calibrated scoring result.`,
+  { label: 'calibrate', phase: 'Calibrate', schema: CALIBRATION_SCHEMA }
+)
+
+return {
+  iteration: iteration,
+  votes: allVotes,
+  calibration: calibration,
+  metadata: {
+    ants_scored: antResults.length,
+    hallucination_rate: calibration ? calibration.calibration.hallucination_rate : null,
+    best_ant: calibration ? calibration.ranking[0] : null,
+    score_range: calibration ? [calibration.calibration.min, calibration.calibration.max] : null,
+  },
+}

package/.agy/skills/team-adversarial-swarm/workflows/wf-swarm-synthesize.js

@@ -0,0 +1,248 @@
+export const meta = {
+  name: 'wf-swarm-synthesize',
+  description: 'Adversarial 3-perspective synthesis of swarm results with arbitrator',
+  whenToUse: 'After swarm converges: synthesize best solution via 3 perspectives + arbitrated final report',
+  phases: [
+    { title: 'Analyze', detail: '3 parallel analysts: why-it-won, stability, caveats' },
+    { title: 'Arbitrate', detail: 'Arbitrator synthesizes perspectives into best-solution report' },
+  ],
+}
+
+const PERSPECTIVE_SCHEMA = {
+  type: 'object',
+  properties: {
+    perspective: { type: 'string' },
+    assessment: { type: 'string' },
+    key_findings: {
+      type: 'array',
+      items: {
+        type: 'object',
+        properties: {
+          finding: { type: 'string' },
+          evidence: { type: 'string' },
+          significance: { type: 'string', enum: ['critical', 'important', 'minor'] },
+        },
+        required: ['finding', 'evidence', 'significance'],
+      },
+    },
+    confidence: { type: 'number', minimum: 0, maximum: 100 },
+    verdict: { type: 'string' },
+  },
+  required: ['perspective', 'assessment', 'key_findings', 'confidence', 'verdict'],
+}
+
+const SYNTHESIS_SCHEMA = {
+  type: 'object',
+  properties: {
+    title: { type: 'string' },
+    best_solution: {
+      type: 'object',
+      properties: {
+        path: { type: 'array', items: { type: 'string' } },
+        score: { type: 'number' },
+        iteration: { type: 'number' },
+        ant_id: { type: 'string' },
+        summary: { type: 'string' },
+        evidence_chain: { type: 'array', items: { type: 'object', properties: { source: { type: 'string' }, finding: { type: 'string' } }, required: ['source', 'finding'] } },
+      },
+      required: ['summary'],
+    },
+    why_it_won: { type: 'string' },
+    pivotal_decisions: {
+      type: 'array',
+      items: {
+        type: 'object',
+        properties: {
+          decision: { type: 'string' },
+          pheromone_guided: { type: 'boolean' },
+          impact: { type: 'string' },
+        },
+        required: ['decision', 'impact'],
+      },
+    },
+    runner_ups: {
+      type: 'array',
+      items: {
+        type: 'object',
+        properties: {
+          ant_id: { type: 'string' },
+          path: { type: 'array', items: { type: 'string' } },
+          score: { type: 'number' },
+          diff_from_best: { type: 'string' },
+        },
+        required: ['ant_id', 'score', 'diff_from_best'],
+      },
+    },
+    convergence_story: { type: 'string' },
+    caveats: { type: 'array', items: { type: 'string' } },
+    adversarial_assessment: {
+      type: 'object',
+      properties: {
+        stability_verdict: { type: 'string' },
+        caveat_severity: { type: 'string', enum: ['none', 'minor', 'significant', 'critical'] },
+        confidence_in_result: { type: 'number' },
+        decisive_perspective: { type: 'string' },
+      },
+      required: ['stability_verdict', 'caveat_severity', 'confidence_in_result'],
+    },
+    executive_summary: { type: 'string' },
+  },
+  required: ['title', 'best_solution', 'why_it_won', 'runner_ups', 'convergence_story', 'caveats', 'adversarial_assessment', 'executive_summary'],
+}
+
+const best = args?.best || {}
+const topK = args?.top_k || []
+const convergenceStory = args?.convergence_story || ''
+const objective = args?.objective || ''
+const totalIterations = args?.total_iterations || 0
+const totalAnts = args?.total_ants || 0
+
+const bestDigest = `Best solution:
+  Ant: ${best.ant_id || 'unknown'}
+  Path: ${(best.path || []).join(' → ')}
+  Score: ${best.score || 'unknown'}
+  Iteration: ${best.iteration || 'unknown'}
+  Summary: ${best.summary || 'none'}
+  Evidence: ${(best.evidence || []).map(e => e.source + ': ' + e.finding).join('; ') || 'none'}`
+
+const topKDigest = topK.map((t, i) =>
+  `#${i + 1}: ${t.ant_id} score=${t.score} path=${(t.path || []).join('→')}`
+).join('\n')
+
+// Phase 1: 3 parallel perspective analysts
+phase('Analyze')
+log('Launching 3-perspective adversarial analysis...')
+
+const perspectives = await parallel([
+  () => agent(
+    `You are the WHY-IT-WON analyst. Explain why the best solution won.
+
+Objective: ${objective}
+${bestDigest}
+
+Runner-ups:
+${topKDigest || 'None available'}
+
+Convergence: ${convergenceStory}
+Total iterations: ${totalIterations}, Total ants: ${totalAnts}
+
+Focus:
+1. Which path decisions were PIVOTAL — where did best diverge from runner-ups?
+2. Which decisions followed pheromone hints vs deviated? Were deviations the key?
+3. Is the evidence chain compelling or circumstantial?
+4. Compare best vs #2: what SPECIFIC factor gave best the edge?
+
+Verdict: one sentence on the quality of the winning strategy.`,
+    { label: 'analyst:why-won', phase: 'Analyze', schema: PERSPECTIVE_SCHEMA }
+  ),
+  () => agent(
+    `You are the STABILITY analyst. Assess whether this result is robust or lucky.
+
+Objective: ${objective}
+${bestDigest}
+
+Runner-ups:
+${topKDigest || 'None available'}
+
+Convergence: ${convergenceStory}
+Total iterations: ${totalIterations}, Total ants: ${totalAnts}
+
+Focus:
+1. Did MULTIPLE ants find similar solutions? (convergence = robust)
+2. Is the best a lone outlier? (divergence from pack = possibly lucky)
+3. Score gap between #1 and #2: large gap = clear winner, small gap = could flip
+4. If the same swarm ran again, would it find the same answer?
+5. Was convergence triggered by genuine consensus or just timeout?
+
+Verdict: "robust" / "fragile" / "uncertain" — with evidence.`,
+    { label: 'analyst:stability', phase: 'Analyze', schema: PERSPECTIVE_SCHEMA }
+  ),
+  () => agent(
+    `You are the CAVEATS analyst. Find every limitation and risk in this result.
+
+Objective: ${objective}
+${bestDigest}
+
+Runner-ups:
+${topKDigest || 'None available'}
+
+Convergence: ${convergenceStory}
+Total iterations: ${totalIterations}, Total ants: ${totalAnts}
+
+Focus:
+1. Search space coverage: was the task space well-explored or did ants cluster?
+2. Evidence quality: single-source claims vs multi-source verification?
+3. Hallucination risk: how many ants were flagged for score inflation?
+4. Solution actionability: can the result be directly applied, or needs more work?
+5. What the swarm DIDN'T explore: are there obvious nodes/paths it missed?
+6. Scaling: would a larger swarm / more iterations have found something better?
+
+Be THOROUGH — every result has caveats. Honest caveats are more valuable than false confidence.
+Verdict: overall risk level of relying on this result.`,
+    { label: 'analyst:caveats', phase: 'Analyze', schema: PERSPECTIVE_SCHEMA }
+  ),
+])
+
+const validPerspectives = perspectives.filter(Boolean)
+const perspectiveDigest = validPerspectives.map(p =>
+  `### ${p.perspective} (confidence: ${p.confidence}%)\n${p.assessment}\nKey findings:\n${p.key_findings.map(f => '- [' + f.significance + '] ' + f.finding).join('\n')}\nVerdict: ${p.verdict}`
+).join('\n\n---\n\n')
+
+log(`${validPerspectives.length} perspective analyses completed`)
+
+// Phase 2: Arbitrator synthesizes
+phase('Arbitrate')
+log('Arbitrator synthesizing final report...')
+
+const synthesis = await agent(
+  `You are the ARBITRATOR. Synthesize 3 analyst perspectives into a definitive swarm result report.
+
+=== OBJECTIVE ===
+${objective}
+
+=== BEST SOLUTION ===
+${bestDigest}
+
+=== RUNNER-UPS ===
+${topKDigest || 'None'}
+
+=== 3 ANALYST PERSPECTIVES ===
+${perspectiveDigest}
+
+=== CONVERGENCE ===
+${convergenceStory}
+Iterations: ${totalIterations}, Ants: ${totalAnts}
+
+SYNTHESIZE:
+1. Build the best_solution record with full evidence chain
+2. Write why_it_won from the first analyst's pivotal decision analysis
+3. Extract pivotal_decisions with pheromone guidance flags
+4. Format runner_ups with diff_from_best
+5. Write convergence_story narrative
+6. Compile ALL caveats from the caveats analyst — don't soften them
+7. adversarial_assessment:
+   - stability_verdict from stability analyst
+   - caveat_severity: none/minor/significant/critical based on caveats count and severity
+   - confidence_in_result: weighted from all 3 perspectives
+   - decisive_perspective: which analyst's findings had the most impact
+8. Write executive_summary (3-4 sentences): what was found, how confident, what to watch out for
+9. Title: concise result title
+
+Max 150 lines in the generated content. Be sharp, not verbose.`,
+  { label: 'arbitrate', phase: 'Arbitrate', schema: SYNTHESIS_SCHEMA }
+)
+
+return {
+  perspectives: validPerspectives,
+  synthesis: synthesis,
+  metadata: {
+    objective: objective,
+    best_score: best.score,
+    best_ant: best.ant_id,
+    total_iterations: totalIterations,
+    total_ants: totalAnts,
+    stability_verdict: synthesis ? synthesis.adversarial_assessment.stability_verdict : null,
+    caveat_severity: synthesis ? synthesis.adversarial_assessment.caveat_severity : null,
+    confidence: synthesis ? synthesis.adversarial_assessment.confidence_in_result : null,
+  },
+}

package/.claude/commands/maestro-ralph-execute.md

@@ -52,6 +52,7 @@ HARD RULES:
 6. **Self-invocation chain** — 持续直到全部 `completion_confirmed` 或 paused
 7. **status.json 每步骤后由 CLI 原子写盘** — resume-safe
 8. **STATUS 枚举受限** — 仅 `DONE | DONE_WITH_CONCERNS | NEEDS_RETRY | BLOCKED`；`NEEDS_CONTEXT` 已废除
+9. **CLI 输出禁止截断** — `maestro ralph next` 的 stdout 包含完整 skill prompt，必须全量捕获。**严禁** `| head`、`| tail`、`2>&1 | head -N` 等任何截断管道。Bash timeout 可加长但不可截断输出
 </invariants>
 
 <state_machine>
@@ -155,7 +156,7 @@ Write enriched args back to status.json.
 
 ### A_EXEC_STEP
 
-1. **Load** — `Bash("maestro ralph next")`
+1. **Load** — `Bash("maestro ralph next --session <session_id>")` — **必须全量捕获 stdout，严禁 `| head`/`| tail` 等截断管道**（stdout 含完整 skill prompt，截断会导致执行内容不完整）
    - 退出码 0 → 按 stdout 内联执行
    - 退出码 2 → 交给 S_LOCATE
    - 退出码 3 → active_step_index 已被占用