Update app.py

app.py

@@ -7,6 +7,7 @@ then asynchronously forwards the full payload to the appropriate downstream URL.
 
 import os
 import json
+import asyncio
 import logging
 from contextlib import asynccontextmanager
 
@@ -205,25 +206,94 @@ async def block_ip_cloudflare(ip: str) -> None:
 # Downstream forwarding helper
 # ──────────────────────────────────────────────
 async def forward_request(target_url: str, payload: dict, serv_code: str) -> tuple[dict, int]:
-    """POST the full payload to the chosen downstream service."""
+    """
+    POST the full payload to the chosen downstream service.
+    Retries up to 3 times on ConnectError or TimeoutException with exponential backoff.
+    Uses split timeouts to handle Render cold-starts gracefully.
+    """
     fwd_headers = {
         "Content-Type": "application/json",
         "serv_code":    serv_code,
     }
-    try:
-        async with httpx.AsyncClient(timeout=120.0) as client:
-            resp = await client.post(target_url, json=payload, headers=fwd_headers)
-            logger.info("Downstream %s → HTTP %s", target_url, resp.status_code)
-            try:
-                return resp.json(), resp.status_code
-            except Exception:
-                return {"raw_response": resp.text}, resp.status_code
-    except httpx.TimeoutException:
-        logger.error("Timeout forwarding to %s", target_url)
-        return {"error": f"Downstream timeout: {target_url}"}, 504
-    except Exception as exc:
-        logger.error("Error forwarding to %s: %s", target_url, exc)
-        return {"error": str(exc)}, 502
+
+    # Split timeouts: generous connect window for Render cold-starts,
+    # generous read window for slow inference on downstream services.
+    timeout = httpx.Timeout(
+        connect=60.0,   # Render free-tier cold-start can take 30–50 s
+        read=120.0,     # downstream inference may be slow
+        write=30.0,
+        pool=10.0,
+    )
+
+    last_exc: Exception | None = None
+
+    for attempt in range(1, 4):   # attempts 1, 2, 3
+        try:
+            logger.info(
+                "Forward attempt %d/3 → %s", attempt, target_url
+            )
+            async with httpx.AsyncClient(timeout=timeout) as client:
+                resp = await client.post(target_url, json=payload, headers=fwd_headers)
+                logger.info(
+                    "Downstream %s → HTTP %s (attempt %d)", target_url, resp.status_code, attempt
+                )
+                try:
+                    return resp.json(), resp.status_code
+                except Exception as parse_exc:
+                    logger.warning(
+                        "Could not parse JSON from %s (HTTP %s): %s — returning raw text",
+                        target_url, resp.status_code, parse_exc,
+                    )
+                    return {"raw_response": resp.text}, resp.status_code
+
+        except httpx.TimeoutException as exc:
+            last_exc = exc
+            logger.warning(
+                "Attempt %d/3 TIMEOUT for %s | type=%s | detail=%s",
+                attempt, target_url, type(exc).__name__, exc,
+            )
+
+        except httpx.ConnectError as exc:
+            # This is the primary cause of 502s with Render cold-starts:
+            # the service is sleeping and refuses/resets the connection.
+            last_exc = exc
+            logger.warning(
+                "Attempt %d/3 CONNECT ERROR for %s | type=%s | detail=%s",
+                attempt, target_url, type(exc).__name__, exc,
+            )
+
+        except httpx.HTTPStatusError as exc:
+            # Downstream returned a 4xx/5xx — no point retrying.
+            logger.error(
+                "Downstream %s returned HTTP error (attempt %d): status=%s body=%s",
+                target_url, attempt, exc.response.status_code, exc.response.text,
+            )
+            return {"error": f"Downstream HTTP error: {exc.response.status_code}"}, exc.response.status_code
+
+        except Exception as exc:
+            # Unexpected / non-retryable error — log full traceback and bail.
+            logger.exception(
+                "Attempt %d/3 UNEXPECTED ERROR for %s | type=%s | detail=%s",
+                attempt, target_url, type(exc).__name__, exc,
+            )
+            return {"error": f"Unexpected forwarding error: {exc}"}, 502
+
+        # Exponential backoff before next attempt (2 s, 4 s)
+        if attempt < 3:
+            backoff = 2 ** attempt
+            logger.info("Backing off %ds before retry …", backoff)
+            await asyncio.sleep(backoff)
+
+    # All 3 attempts exhausted
+    logger.error(
+        "All 3 forward attempts failed for %s | last_error_type=%s | last_error=%s",
+        target_url, type(last_exc).__name__, last_exc,
+    )
+    return {
+        "error": f"Downstream unreachable after 3 attempts: {target_url}",
+        "last_error_type": type(last_exc).__name__,
+        "last_error": str(last_exc),
+    }, 502
 
 
 # ──────────────────────────────────────────────
@@ -320,4 +390,4 @@ async def chat(request: Request):
         "decision":  decision,
         "forwarded": target_url,
         "response":  response_body,
-    })
+    })