PolicyEngine
diff --git a/‎.env-example‎
Lines changed: 15 additions & 1 deletion b/‎.env-example‎
Lines changed: 15 additions & 1 deletion
diff --git a/‎changelog_entry.yaml‎
Lines changed: 14 additions & 0 deletions b/‎changelog_entry.yaml‎
Lines changed: 14 additions & 0 deletions
diff --git a/‎config/README.md‎
Lines changed: 33 additions & 0 deletions b/‎config/README.md‎
Lines changed: 33 additions & 0 deletions
diff --git a/‎config/default.yaml‎
Lines changed: 8 additions & 0 deletions b/‎config/default.yaml‎
Lines changed: 8 additions & 0 deletions
diff --git a/‎policyengine_household_api/api.py‎
Lines changed: 52 additions & 2 deletions b/‎policyengine_household_api/api.py‎
Lines changed: 52 additions & 2 deletions
diff --git a/‎policyengine_household_api/auth/validation.py‎
Lines changed: 98 additions & 2 deletions b/‎policyengine_household_api/auth/validation.py‎
Lines changed: 98 additions & 2 deletions
diff --git a/‎policyengine_household_api/country.py‎
Lines changed: 8 additions & 3 deletions b/‎policyengine_household_api/country.py‎
Lines changed: 8 additions & 3 deletions
diff --git a/‎policyengine_household_api/data/analytics_setup.py‎
Lines changed: 12 additions & 2 deletions b/‎policyengine_household_api/data/analytics_setup.py‎
Lines changed: 12 additions & 2 deletions
@@ -1,2 +1,16 @@
 FLASK_DEBUG=1
-CACHE_REDIS_HOST=redis
+CACHE_REDIS_HOST=redis
+
+# Optional: wipe the local sqlite analytics DB on startup. Only
+# consulted when FLASK_DEBUG=1 and analytics is enabled. Default off
+# so captured debug data is not lost across restarts.
+# RESET_ANALYTICS=1
+
+# Optional: comma-separated list of origins (or regex patterns) allowed
+# by CORS. If unset, the default allowlist is:
+#   - https://policyengine.org
+#   - https://*.policyengine.org  (anchored regex)
+#   - http://localhost[:port]     (any port)
+#   - http://127.0.0.1[:port]
+# Example override:
+# CORS_ALLOWED_ORIGINS=https://foo.example.com,https://bar.example.com
@@ -0,0 +1,14 @@
+- bump: patch
+  changes:
+    fixed:
+    - Flatten every (entity, variable, period) triple in flatten_variables_from_household (#1462).
+    - Tighten /calculate_demo rate limit from 1/second to 1/10 seconds (#1463).
+    - Stop unconditionally wiping the analytics SQLite DB and fix the sqlite:// URI (#1464).
+    - Restrict CORS to PolicyEngine origins by default, anchored so attacker subdomains can't bypass (#1465).
+    - Replace invalid ConnectionError(description=...) with a GCPError class (#1466).
+    - Keep "0"/"1" env-var values as integers instead of collapsing to False/True (#1467).
+    - Verify JWT signatures in the analytics decorator and drop datetime.utcnow (#1468).
+    - Re-raise tracer failures in PolicyEngineCountry.calculate so the endpoint can return a real 500 (#1469).
+    - Validate /calculate payloads and cap axes scans; add per-endpoint rate limit (#1470).
+    - Time-bound and lazy-load the Auth0 JWKS fetch so a startup outage doesn't crash the API, caching only successes so the lazy retry actually retries (#1471).
+    - Replace deprecated dpath.util.search with dpath.search (#1472).
@@ -270,6 +270,39 @@ The following endpoints remain unprotected:
 - When enabled, all protected endpoints validate JWT tokens against Auth0's JWKS
 - The Auth0 domain and audience must match the configured values
 
+## CORS Configuration
+
+Browsers enforce CORS against the API. The default allowlist accepts:
+
+- `https://policyengine.org`
+- Any `https://*.policyengine.org` host (anchored regex)
+- `http://localhost` on any port (dev servers)
+- `http://127.0.0.1` on any port
+
+Override with `CORS_ALLOWED_ORIGINS` (comma-separated strings or
+regexes) or `cors.allowed_origins` in YAML:
+
+```bash
+CORS_ALLOWED_ORIGINS=https://app.example.com,https://admin.example.com
+```
+
+```yaml
+cors:
+  allowed_origins:
+    - https://app.example.com
+    - 'https://.*\.example\.com$'
+```
+
+Always terminate regex patterns with `$` — Flask-CORS matches with
+`re.match`, so an unanchored pattern like `https://.*\.example\.com`
+would accept `https://example.com.attacker.com`.
+
+## Analytics reset (debug only)
+
+`RESET_ANALYTICS=1` (or `analytics.reset: true` in YAML) wipes the
+local SQLite analytics DB on startup. This is **only** consulted when
+`FLASK_DEBUG=1`; production never resets the analytics DB.
+
 ## Usage Examples
 
 ### Production Deployment (Current)
 
@@ -46,3 +46,11 @@ ai:
   enabled: false
   anthropic:
     api_key: ""  # Override with ANTHROPIC_API_KEY
+
+# CORS configuration
+cors:
+  # List of allowed origins (strings or regex patterns). If left null
+  # the API defaults to PolicyEngine production domains. Override with
+  # CORS_ALLOWED_ORIGINS (comma-separated) in environments that serve
+  # additional frontends.
+  allowed_origins: null
@@ -24,6 +24,7 @@
 from policyengine_household_api.decorators.analytics import (
     log_analytics_if_enabled,
 )
+from policyengine_household_api.utils.config_loader import get_config_value
 
 # Endpoints
 from .endpoints import (
@@ -40,7 +41,52 @@
 
 app = application = flask.Flask(__name__)
 
-CORS(app)
+# Reject absurdly large request bodies before any view runs. 10 MiB is
+# well above the largest legitimate household payload we have seen
+# (axes scans push a few hundred KiB) while still capping the memory a
+# single attacker can force us to allocate. Overridable via the
+# ``MAX_CONTENT_LENGTH`` env var (bytes).
+app.config["MAX_CONTENT_LENGTH"] = int(
+    os.getenv("MAX_CONTENT_LENGTH", 10 * 1024 * 1024)
+)
+
+
+def _resolve_cors_origins():
+    """
+    Resolve the CORS allowed origins list.
+
+    Priority:
+      1. CORS_ALLOWED_ORIGINS env var (comma-separated list)
+      2. config value "cors.allowed_origins" (list or comma string)
+      3. Safe default: the PolicyEngine production domains
+
+    Use regex patterns so that wildcard subdomains work with
+    Flask-CORS's `origins` kwarg.
+    """
+    raw = os.getenv("CORS_ALLOWED_ORIGINS") or get_config_value(
+        "cors.allowed_origins", None
+    )
+
+    if raw is None:
+        # Flask-CORS uses re.match, which is a prefix match; anchor with
+        # ``$`` so a hostile host like ``policyengine.org.attacker.com``
+        # cannot satisfy the wildcard pattern. Include ``localhost:*``
+        # so local dev servers can hit the API without extra setup.
+        origins = [
+            "https://policyengine.org",
+            r"https://.*\.policyengine\.org$",
+            r"http://localhost(:[0-9]+)?$",
+            r"http://127\.0\.0\.1(:[0-9]+)?$",
+        ]
+    elif isinstance(raw, str):
+        origins = [o.strip() for o in raw.split(",") if o.strip()]
+    else:
+        origins = list(raw)
+
+    return origins
+
+
+CORS(app, origins=_resolve_cors_origins())
 
 # Use in-memory storage for rate limiting
 # Note that this provides limits per-instance;
@@ -59,6 +105,7 @@
 
 @app.route("/<country_id>/calculate", methods=["POST"])
 @require_auth_if_enabled()
+@limiter.limit("60 per minute")
 @log_analytics_if_enabled
 def calculate(country_id):
     return get_calculate(country_id)
@@ -84,8 +131,11 @@ def readiness_check():
     )
 
 
+# Note: `/calculate_demo` is intentionally public (documented in
+# config/README.md). It is guarded by a conservative rate limit rather
+# than JWT authentication.
 @app.route("/<country_id>/calculate_demo", methods=["POST"])
-@limiter.limit("1 per second")
+@limiter.limit("1 per 10 seconds")
 def calculate_demo(country_id):
     return get_calculate(country_id)
 
 
@@ -1,18 +1,114 @@
 import json
+import logging
+import time
+from threading import Lock
 from urllib.request import urlopen
 
 from authlib.oauth2.rfc7523 import JWTBearerTokenValidator
 from authlib.jose.rfc7517.jwk import JsonWebKey
 
+logger = logging.getLogger(__name__)
+
+JWKS_FETCH_TIMEOUT = 10  # seconds
+# Minimum wait between back-to-back lazy retries after a failure.
+# Keeps us from hammering Auth0 when it is actively degraded.
+JWKS_RETRY_INTERVAL_SECONDS = 30
+
+
+# Module-level cache of successful JWKS fetches, keyed by issuer. Only
+# successes are cached so that a transient failure is retried on the
+# next authenticated request (``lru_cache`` would have memoised the
+# ``None`` return, making the "lazy retry" dead code).
+_jwks_cache: dict = {}
+# Records the monotonic timestamp of the most recent *failed* fetch
+# per-issuer so we can rate-limit retries without caching the failure
+# itself.
+_jwks_last_failure: dict = {}
+_jwks_lock = Lock()
+
+
+def _fetch_jwks_uncached(issuer: str):
+    """Fetch the JWKS for an Auth0 issuer, bypassing the cache.
+
+    Returns an authlib key set on success, ``None`` on failure. Errors
+    are logged rather than raised so that a transient Auth0 outage
+    doesn't crash the process at import time.
+    """
+    jwks_url = f"{issuer}.well-known/jwks.json"
+    try:
+        with urlopen(jwks_url, timeout=JWKS_FETCH_TIMEOUT) as response:
+            return JsonWebKey.import_key_set(json.loads(response.read()))
+    except Exception as e:
+        logger.warning(f"Failed to fetch JWKS from {jwks_url}: {e}")
+        return None
+
+
+def _fetch_jwks(issuer: str):
+    """Fetch JWKS, caching only successful results.
+
+    On failure we record the time but do not memoise the ``None`` — a
+    later call will retry (subject to ``JWKS_RETRY_INTERVAL_SECONDS``
+    backoff) so that the validator self-heals once Auth0 recovers.
+    """
+    with _jwks_lock:
+        cached = _jwks_cache.get(issuer)
+        if cached is not None:
+            return cached
+        last_failure = _jwks_last_failure.get(issuer)
+        if (
+            last_failure is not None
+            and time.monotonic() - last_failure < JWKS_RETRY_INTERVAL_SECONDS
+        ):
+            # Too soon after the last failure — don't hammer Auth0.
+            return None
+
+    # Fetch outside the lock so a slow network call doesn't block
+    # other threads that might be serving requests with a cached key.
+    key_set = _fetch_jwks_uncached(issuer)
+
+    with _jwks_lock:
+        if key_set is not None:
+            _jwks_cache[issuer] = key_set
+            _jwks_last_failure.pop(issuer, None)
+        else:
+            _jwks_last_failure[issuer] = time.monotonic()
+    return key_set
+
+
+def _clear_jwks_cache():
+    """Test helper: wipe the success/failure caches."""
+    with _jwks_lock:
+        _jwks_cache.clear()
+        _jwks_last_failure.clear()
+
 
 class Auth0JWTBearerTokenValidator(JWTBearerTokenValidator):
     def __init__(self, domain, audience):
         issuer = f"https://{domain}/"
-        jsonurl = urlopen(f"{issuer}.well-known/jwks.json")
-        public_key = JsonWebKey.import_key_set(json.loads(jsonurl.read()))
+
+        public_key = _fetch_jwks(issuer)
+        if public_key is None:
+            # Retry on next token validation rather than failing hard
+            # at construction time. A missing key set means token
+            # validation will fail cleanly inside authlib.
+            logger.warning(
+                "JWKS unavailable at construction; will retry on first "
+                "token validation."
+            )
+
         super(Auth0JWTBearerTokenValidator, self).__init__(public_key)
+        self._issuer = issuer
         self.claims_options = {
             "exp": {"essential": True},
             "aud": {"essential": True, "value": audience},
             "iss": {"essential": True, "value": issuer},
         }
+
+    def authenticate_token(self, token_string):
+        # Lazy-refresh the JWKS if the initial fetch failed. Because
+        # ``_fetch_jwks`` only caches successes, this call will retry
+        # the network fetch (subject to a short backoff) until Auth0
+        # responds.
+        if self.public_key is None:
+            self.public_key = _fetch_jwks(self._issuer)
+        return super().authenticate_token(token_string)
@@ -1,4 +1,5 @@
 import importlib
+import logging
 from flask import Response
 import json
 from policyengine_core.taxbenefitsystems import TaxBenefitSystem
@@ -432,8 +433,12 @@ def calculate(
 
             return household, None
 
-        except Exception as e:
-            print(f"Error computing tracer output: {e}")
+        except Exception:
+            # Re-raise so endpoints/household.py (which unpacks
+            # ``(result, computation_tree_uuid)``) can surface a real
+            # 500 instead of a TypeError on ``None`` unpacking.
+            logging.exception("Tracer failed while computing household")
+            raise
 
 
 def create_policy_reform(policy_data: dict) -> dict:
@@ -478,7 +483,7 @@ def apply(self):
 
 
 def get_requested_computations(household: dict):
-    requested_computations = dpath.util.search(
+    requested_computations = dpath.search(
         household,
         "*/*/*/*",
         afilter=lambda t: t is None,
 
@@ -47,11 +47,21 @@ def initialize_analytics_db_if_enabled(app):
         db_url = (
             REPO / "policyengine_household_api" / "data" / "policyengine.db"
         )
-        if Path(db_url).exists():
+        # Only wipe the analytics DB when explicitly requested via
+        # RESET_ANALYTICS=1 (or the ``analytics.reset`` config flag).
+        should_reset = os.getenv("RESET_ANALYTICS", "").lower() in (
+            "1",
+            "true",
+            "yes",
+        ) or get_config_value("analytics.reset", False)
+        if should_reset and Path(db_url).exists():
             Path(db_url).unlink()
         if not Path(db_url).exists():
             Path(db_url).touch()
-        app.config["SQLALCHEMY_DATABASE_URI"] = "sqlite:////" + str(db_url)
+        # sqlite: absolute paths require exactly three slashes plus the
+        # leading "/" from the absolute path (=> "sqlite:////tmp/x.db").
+        # db_url here is already absolute, so use an f-string.
+        app.config["SQLALCHEMY_DATABASE_URI"] = f"sqlite:///{db_url}"
     else:
         app.config["SQLALCHEMY_DATABASE_URI"] = "mysql+pymysql://"
         app.config["SQLALCHEMY_ENGINE_OPTIONS"] = {"creator": getconn}