#!/bin/bash
# Analyze each island with letta CLI and post summaries back to the API.
# Usage: ./scripts/analyze-islands.sh
#
# Requires: letta CLI, CLOUDFLARE_API_TOKEN set

set -euo pipefail

API_BASE="https://stigmergic.latha.org"
LETTA_BIN="${LETTA_BIN:-/home/nandi/.local/npm-global/bin/letta}"

# Fetch all islands
islands=$(curl -sf "$API_BASE/api/strata/islands" | python3 -c "
import json, sys
data = json.load(sys.stdin)
for island in data['islands']:
    print(island['id'])
")

total=$(echo "$islands" | wc -l)
current=0

echo "Analyzing $total islands..."

while IFS= read -r island_id; do
    current=$((current + 1))
    echo "[$current/$total] Analyzing island: ${island_id:0:60}..."

    # Fetch this island's full data
    island_json=$(curl -sf "$API_BASE/api/strata/islands" | python3 -c "
import json, sys
data = json.load(sys.stdin)
for island in data['islands']:
    if island['id'] == sys.argv[1]:
        # Build a concise context for the LLM
        vertices = island['vertices']
        edges = island['edges']
        meta = island.get('vertexMeta', {})

        titles = [meta.get(v, {}).get('title', v) for v in vertices[:30]]
        domains = {}
        for v in vertices:
            if v.startswith('http'):
                try:
                    from urllib.parse import urlparse
                    h = urlparse(v).hostname.replace('www.', '')
                    domains[h] = domains.get(h, 0) + 1
                except: pass
        top_domains = sorted(domains.items(), key=lambda x: -x[1])[:5]

        edge_types = {}
        for e in edges:
            t = e.get('connectionType', 'unknown')
            edge_types[t] = edge_types.get(t, 0) + 1

        notes = [e['note'] for e in edges if e.get('note')][:5]

        out = {
            'vertex_count': len(vertices),
            'edge_count': len(edges),
            'top_domains': [d[0] + f' ({d[1]})' for d in top_domains],
            'edge_types': edge_types,
            'sample_titles': titles[:15],
            'notes': notes,
        }
        print(json.dumps(out))
        break
" "$island_id")

    if [ -z "$island_json" ]; then
        echo "  Skipping (no data)"
        continue
    fi

    # Ask letta for a one-sentence summary
    prompt="You are analyzing a knowledge graph island (connected component). Write a single concise sentence (max 200 chars) summarizing what this cluster of links is about. No preamble, just the summary.

Island data:
$island_json"

    summary=$($LETTA_BIN -p "$prompt" --output-format json --system letta --memfs 2>/dev/null | python3 -c "
import json, sys
try:
    data = json.load(sys.stdin)
    if data.get('type') == 'result':
        print(data['result'].strip().strip('\"').strip())
    else:
        print(data.get('result', data.get('text', '')), end='')
except:
    # Fallback: try reading as plain text
    sys.stdout.write(sys.stdin.read().strip())
" 2>/dev/null || echo "")

    if [ -z "$summary" ]; then
        echo "  Skipping (no summary generated)"
        continue
    fi

    # Truncate to 300 chars
    summary=$(echo "$summary" | head -c 300)

    echo "  Summary: $summary"

    # PUT the summary back
    encoded_id=$(python3 -c "import urllib.parse; print(urllib.parse.quote(sys.argv[1], safe=''))" "$island_id")
    curl -sf -X PUT "$API_BASE/api/strata/islands/$encoded_id/summary" \
        -H "Content-Type: application/json" \
        -d "{\"summary\":$(python3 -c "import json; print(json.dumps(sys.argv[1]))" "$summary")}" \
        > /dev/null

    echo "  Saved."

    # Rate limit
    sleep 2
done <<< "$islands"

echo "Done. $current islands analyzed."