Increase rate limits

README.md

@@ -89,7 +89,7 @@ Go to **Space Settings → Repository Secrets** and add:
 ```
 
 ```json
-{"error": "Rate limit exceeded. Max 60 requests per 300s window.", "retry_after": 300}
+{"error": "Rate limit exceeded. Max 240 requests per 300s window.", "retry_after": 300}
 ```
 
 ```json
@@ -105,7 +105,7 @@ Go to **Space Settings → Repository Secrets** and add:
 | `STRIX_API_TOKEN` | *(none)* | Required in production. If unset, runs in open/dev mode. |
 
 ### Rate Limiting
-- **60 requests** per **5-minute** sliding window per caller
+- **240 requests** per **5-minute** sliding window per caller (override: `STRIX_RATE_LIMIT_MAX`)
 - In-memory, resets on Space restart
 
 ### Prediction Length (Adaptive)

app.py

@@ -46,8 +46,29 @@ logger = logging.getLogger("strix-chronos")
 # ---------------------------------------------------------------------------
 API_TOKEN = os.environ.get("STRIX_API_TOKEN")  # Set in HF Space secrets
 
-RATE_LIMIT_MAX = 60          # Max requests per window per caller
-RATE_LIMIT_WINDOW = 300      # Window duration in seconds (5 minutes)
+# Rate limit — sized from MEASURED capacity, not intuition.
+#
+# Measured against this Space (Chronos-2 120M, CPU, float32), sequential
+# forecasts from the Strix server: ~8.7s on the first call (cold start / model
+# warm-up), then a steady state of ~1.75-1.8s round-trip. That round-trip
+# INCLUDES the caller's network hop and Gradio overhead, so it is a conservative
+# proxy: serial capacity is therefore at least ~0.55 req/s, i.e. ~165 requests
+# per 300s window at full duty.
+#
+# The old ceiling of 60/300s was 0.2 req/s — roughly a THIRD of what the Space
+# can actually serve, so it was rejecting traffic the hardware was idle for. It
+# bit in normal single-operator use because one MAX Market Intelligence scan
+# issues ~20 forecasts, so two scans in five minutes tripped it.
+#
+# 240/300s = 0.8 req/s leaves the limiter as an abuse ceiling rather than a
+# throughput cap, while staying within what a free CPU Space can sustain. The
+# client also caches, de-duplicates and serialises its calls, so real demand is
+# well under this.
+#
+# Override without a redeploy by setting STRIX_RATE_LIMIT_MAX / _WINDOW in the
+# Space secrets.
+RATE_LIMIT_MAX = int(os.environ.get("STRIX_RATE_LIMIT_MAX", "240"))
+RATE_LIMIT_WINDOW = int(os.environ.get("STRIX_RATE_LIMIT_WINDOW", "300"))
 
 VALID_TIMEFRAMES = {"15m", "1h", "1d", "1wk", "1mo"}
 TICKER_PATTERN = re.compile(r"^[A-Z0-9\.\-\^=]{1,15}$")