widescreen-research: Command for Claude Code

Autonomous Discovery Loop

Pattern: Systematic hypothesis exploration with iterative refinement Based on: "Agentic Science" framework (arXiv:2508.14111) Determinism: High - Each phase has programmatic validation gates

---

Usage

/autonomous-discovery-loop "Research question or domain"

Arguments:

• $ARGUMENTS (required): Research question or domain to explore
• max_iterations (optional): Maximum discovery loops (default: 5)
• hypothesis_breadth (optional): Number of hypotheses per iteration (default: 10)
• convergence_threshold (optional): Stop when X% of hypotheses validated (default: 0.80)

---

Variables

Session State (tracked in .discovery-loop/state.json):

{
  "research_question": "$ARGUMENTS",
  "iteration": 0,
  "max_iterations": 5,
  "hypotheses_generated": [],
  "hypotheses_tested": [],
  "hypotheses_validated": [],
  "evidence_collected": [],
  "current_phase": "observe",
  "convergence_score": 0.0
}

---

Workflow Phases

Phase 0: Initialize Discovery Space

Deterministic Hook: .claude/hooks/research/init-discovery-loop.sh

#!/bin/bash
# Creates research state directory and validates inputs

RESEARCH_QUESTION="$1"
MAX_ITER="${2:-5}"
HYPOTHESIS_BREADTH="${3:-10}"

mkdir -p .discovery-loop/{hypotheses,evidence,tests,synthesis}

cat > .discovery-loop/state.json << EOF
{
  "research_question": "$RESEARCH_QUESTION",
  "iteration": 0,
  "max_iterations": $MAX_ITER,
  "hypothesis_breadth": $HYPOTHESIS_BREADTH,
  "current_phase": "observe",
  "gates": {
    "observe": {"passed": false, "evidence_count": 0},
    "hypothesize": {"passed": false, "hypotheses_count": 0},
    "design": {"passed": false, "tests_defined": 0},
    "execute": {"passed": false, "tests_run": 0},
    "analyze": {"passed": false, "results_recorded": 0}
  }
}
EOF

echo "✅ Discovery loop initialized for: $RESEARCH_QUESTION"

Gate: State file created, inputs validated

---

Phase 1: Observe (LLM Intelligence + Agent Execution)