Skip to content
Golden Child Brain

Barkley Configuration

Architecture

Section titled “Architecture”
  • Runtime: Cloudflare Workers (TypeScript)
  • Sandboxes: Cloudflare Sandbox SDK (Durable Objects + containers)
  • AI: Claude Agent SDK for multi-turn sessions
  • State: DO ctx.storage for schedules, session maps, backup handles
  • Backups: R2 for sandbox filesystem snapshots
  • Guidelines: Cloudflare Vectorize with @cf/baai/bge-base-en-v1.5 embeddings

Teaching Barkley

Section titled “Teaching Barkley”

Team members can teach Barkley via Slack:

@Barkley learn: [correction or guideline]

This stores the correction in the Vectorize guidelines database with category tags.

Knowledge Base Sync

Section titled “Knowledge Base Sync”

The golden-child-brain repo is synced to sandboxes:

  • On wake: git pull to get latest knowledge
  • On learning: Create new .md file and git push
  • Semantic search: Query AutoRAG REST API for relevant context

Reddit Monitoring

Section titled “Reddit Monitoring”
  • Schedule ID: sched-3
  • Frequency: Every 4 hours
  • Thread grouping: Daily rolling threads in Slack
  • Deduplication: SQLite database at /workspace/barkley.db (reddit_seen table)
  • API access: Brave Search + PullPush API (Reddit OAuth API recommended but not yet configured)

Known Limitations

Section titled “Known Limitations”
  • Direct Reddit API blocked from Cloudflare sandbox; relies on Brave Search (free tier, rate-limited 1 req/sec)
  • Reddit OAuth API credentials not yet configured
  • Only 4 of Megan’s 8 original corrections are in the codebase; remaining 4 are in Vectorize only