get_trade_recommendation MCP tool times out (cumulative latency, not a single broken dependency)

[JohnFunkCode]

Component: quantcore-stock-price MCP wrapper → quantcore-api → RecommendationsService
Environment: Prod (quantcore-prod-20260606) — reproduced 2026-06-18 with symbol=AVGO
Severity: High — the flagship "one-shot recommendation" tool is currently unusable; it returns Upstream request timed out on every call.

Symptom

get_trade_recommendation(AVGO) fails with Upstream request timed out on every attempt (4/4 retries). Connectivity and auth are healthy — get_stock_price(AVGO) and mcp_health_check on the same prod stack return immediately, so the token (QUANTCORE_MCP_TOKEN) is valid and the wrappers are reachable.

Root cause

The failure is not a single hung dependency. Every sub-analysis the tool aggregates was tested individually against AVGO on prod and all succeeded:

get_stock_price, get_rsi, get_macd, get_stochastic, get_volume_analysis, get_candlestick_patterns, get_unusual_calls, get_delta_adjusted_oi, get_dark_pool, get_short_interest, get_bid_ask_spread, get_news_sentiment, get_stop_loss_analysis.

The aggregator runs all of these sequentially within a single request, and several are independently slow due to multiple yfinance round-trips:

• get_unusual_calls and get_delta_adjusted_oi each scan 3 option expirations.
• get_stop_loss_analysis alone runs 7 nested sub-analyses (price/BB, VWAP, MACD, RSI, delta-adjusted OI / gamma wall, historical drawdown, short interest).

Summed serially, total wall-clock exceeds the upstream/Cloud Run request timeout (~60s) and the gateway kills the request before it returns.

Proposed fixes (most durable first)

1. Parallelize the fan-out in RecommendationsService (e.g. asyncio.gather or a thread pool) so sub-analyses run concurrently. Wall-clock drops to ~the slowest single call instead of the sum.
2. Cache/share the OHLCV download. Many sub-tools each re-fetch the same price history from yfinance. Fetch once, pass the DataFrame down — removes most redundant latency.
3. Raise the Cloud Run / gateway timeout (--timeout=300) as a stopgap only; Build Alert System #1+added alert system #2 are the real fix.

Side-finding (separate, lower priority)

get_news_sentiment(AVGO) returns INSUFFICIENT_DATA (0 stored articles), so the recommendation's sentiment leg would be empty even once the timeout is fixed. News likely needs collect_news(AVGO) run first / a backfill job. Worth a separate ticket.

Acceptance criteria

• get_trade_recommendation returns a full recommendation for a liquid symbol (e.g. AVGO) within the gateway timeout.
• No regression in the individual sub-tool outputs.

🤖 Generated with Claude Code