diff --git a/docs/superpowers/specs/2026-04-15-filter-ohlcv-cache-design.md b/docs/superpowers/specs/2026-04-15-filter-ohlcv-cache-design.md
new file mode 100644
index 00000000..4b2c036f
--- /dev/null
+++ b/docs/superpowers/specs/2026-04-15-filter-ohlcv-cache-design.md
@@ -0,0 +1,61 @@
+# Filter Stage OHLCV Cache Design
+
+**Date:** 2026-04-15
+**Status:** Approved
+
+## Problem
+
+The discovery filter stage fires ~120 sequential per-ticker yfinance HTTP calls for each run (~60 candidates × 2 calls each), causing "Too Many Requests" rate limit errors. In the 2026-04-15 run, only 36/61 candidates got prices — the rest were silently dropped, leaving the ranker with just 11 candidates and producing only 2 final picks.
+
+The three offending call sites in `filter.py`:
+1. `get_stock_price(ticker)` — fallback per-ticker price fetch
+2. `check_intraday_movement(ticker)` — `yf.Ticker().history(period="1d", interval="1m")`
+3. `check_if_price_reacted(ticker)` — `yf.Ticker().history(period="1mo")`
+
+The OHLCV cache (1y daily bars, populated nightly) already contains all the data these checks need. Since discovery runs at 7:30am ET (pre-market), yesterday's close is the correct "current price" anyway.
+
+## Solution
+
+Load the OHLCV cache once at the start of the filter stage for the ~60 candidate tickers, then replace all three yfinance call sites with cache lookups.
+
+## Scope
+
+**File changed:** `tradingagents/dataflows/discovery/filter.py` only.
+
+**Unchanged:** `_fetch_batch_volume()`, `_fetch_batch_news()`, `get_fundamentals()`, scanner code, OHLCV cache internals.
+
+## Design
+
+### Cache load
+
+At the top of `CandidateFilter.filter()`, before the per-candidate loop:
+
+```python
+cache_dir = self.config.get("discovery", {}).get("ohlcv_cache_dir", "data/ohlcv_cache")
+candidate_tickers = [c["ticker"] for c in candidates if c.get("ticker")]
+ohlcv_data = download_ohlcv_cached(candidate_tickers, period="1y", cache_dir=cache_dir)
+```
+
+`ohlcv_data` is `Dict[str, DataFrame]` — ticker → daily OHLCV. Loading ~60 tickers from the parquet cache takes under 1 second.
+
+### Replacement map
+
+| Was | Becomes |
+|---|---|
+| `_fetch_batch_prices()` + `get_stock_price()` fallback | `ohlcv_data[ticker]["Close"].iloc[-1]` |
+| `check_intraday_movement()` via `history(1d, 1m)` | `(close[-1] - close[-2]) / close[-2] * 100` (last 2 daily closes) |
+| `check_if_price_reacted()` via `history(1mo)` | `(close[-1] - close[-N]) / close[-N] * 100` where N = `recent_movement_lookback_days` |
+
+### Fallback
+
+If a ticker is missing from `ohlcv_data` (e.g. newly listed, not yet in nightly prefetch), fall through to the existing `get_stock_price()` call. No candidates are lost that weren't already being lost before this change.
+
+### Remove
+
+`_fetch_batch_prices()` method — no longer needed once the cache covers current price.
+
+## Expected outcome
+
+- Zero yfinance rate limit errors during filter enrichment
+- All candidates that pass the scan reach the ranker (was 11/63, should be ~60)
+- No change to filter logic or thresholds — only the data source changes