Skip to main content

Hotfix — prod migrations + chat assistant persistence + chat truncation

1. POST /api/bots/:id/strategy 500 (Change Strategy)​

Root cause: The actual prod Neon database (ep-patient-glade-akbt9n8k, configured as the DATABASE_URL Cloudflare secret on the quantago-api worker) was missing the bot_strategies table and bots.feed_provider column. Earlier migration runs targeted a stale Neon endpoint (ep-gentle-bar-a4xaaekv) that the root .env pointed at, but the worker itself reads a different secret value.

Fix:

  • Applied 0013_bot_strategies.sql and 0014_bot_feed_provider.sql to the real prod DB.
  • Backfilled a bot_strategies row for every existing bot so the resolver finds an active version on first read.
  • Updated root .env to point at the actual prod endpoint so future migrate-neon.sh runs land on the right DB.

2. AI assistant responses not persisted​

Root cause: INSERT INTO bot_chat_messages (... tool_calls ...) passed JSON.stringify(toolTimeline) as a text parameter into a jsonb column without an explicit ::jsonb cast. Postgres has no implicit text→jsonb cast, so every assistant insert silently failed (the try/catch swallowed the error). User turns persisted (no jsonb columns), so reload showed only user messages.

Fix: Add the ::jsonb cast back and store null (not the string "null") when there are no tool calls.

const toolCallsJson = toolTimeline.length > 0 ? JSON.stringify(toolTimeline) : null;
await sql`
  INSERT INTO bot_chat_messages (...)
  VALUES (..., ${toolCallsJson}::jsonb, ${model}, 'done')
`;

3 & 4. Streaming + truncation​

Root cause: max_completion_tokens: 2048 clipped longer answers mid-sentence; combined with the OpenAI auto-fallback warning when gpt-5-mini was rejected this also masked the "no streaming" complaint because chunks arrived all at once after the API errored out.

Fix: Bump cap to 8192, keep max_completion_tokens only (the canonical param across all current OpenAI chat models). The Hono streamSSE writer already flushes per writeSSE, so once the upstream stream is healthy the deltas reach the client incrementally.

5. Match SaaS Stack chat bubble + input styles​

Status: Deferred to follow-up. The current StreamingChatPanel (packages/chat-ui) is functional but visually distinct from the saas-stack ChatInput / ChatMessage components (~1000 LOC, with framer-motion, multi-block hydration, inline tool cards, choice cards, etc.). Porting that UX cleanly is a dedicated task — tracked as a follow-up; this hotfix focuses on the data-loss / functionality bugs above.