- Add toggle-able debug logging (AI_DEBUG env var) that logs prompts,
request metadata, raw responses, parsed output, and full error chains
- Replace Node.js native fetch() with https module for Docker Alpine
compatibility (fixes "fetch failed" error with large payloads)
- Reduce max_tokens from 16384 to 4096 (qwen3.5 doesn't need thinking
token budget)
- Strip <think> blocks from model responses
- Add AI_DEBUG to docker-compose.yml and .env.example
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
kimi-k2.5 is a thinking model that times out on complex prompts (>3min).
qwen3.5-397b-a17b responds in ~2s with clean JSON output.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Increase max_tokens from 4096 to 16384 to accommodate reasoning tokens
- Increase timeout from 90s to 180s for thinking model latency
- Add logging for response diagnostics (content length, reasoning, finish reason)
- Better error message when model exhausts tokens on reasoning
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add shared.cd_rates table for cross-tenant market data (CD rates from Bankrate)
- Create standalone Puppeteer scraper script (scripts/fetch-cd-rates.ts) for cron-based rate fetching
- Add investment-planning backend module with 3 endpoints: snapshot, cd-rates, recommendations
- AI service gathers tenant financial data (accounts, investments, budgets, projects, cash flow) and calls OpenAI-compatible API (NVIDIA endpoint) for structured investment recommendations
- Create InvestmentPlanningPage with summary cards, current investments table, market CD rates table, and AI recommendation accordion
- Add Investment Planning to sidebar under Planning menu
- Configure AI_API_URL, AI_API_KEY, AI_MODEL environment variables
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
