@rulemetric/proxy 0.8.0 → 0.10.0

package/addon/rulemetric_addon.py

@@ -264,6 +264,12 @@ _PENDING_LOCK = threading.Lock()
 # moments of proxy start while Claude Code is in use.
 _CACHED_PLAN: dict[str, str | None] = {"plan": None}
 
+# One-shot guard for the "settings carried no tier" warning (dict so tests can
+# reset). Verified 2026-08-12: for at least some accounts Anthropic sends
+# `internal_tier_rate_limit_tier: null`, so this is a steady state, not a
+# transient — the warning must not fire on every poll.
+_SETTINGS_DRIFT_LOGGED: dict[str, bool] = {"done": False}
+
 _TIER_MAP = (
     ("max_20x", "max_20x"),
     ("max_5x", "max_5x"),
@@ -630,20 +636,30 @@ class RuleMetricAddon:
     # (last exit reason OS_REASON_JETSAM, 200+ restarts/day), which is why
     # the /usage dashboard went stale: the proxy was dead more often than
     # alive and the gateway routed direct. Only flows we actually parse
-    # (detected LLM providers + the oauth/usage poll) may be buffered.
-
-    def _is_oauth_usage_flow(self, flow: http.HTTPFlow) -> bool:
+    # (detected LLM providers + the oauth side-channel polls) may be buffered.
+    #
+    # Both oauth paths must be listed here, not just the usage poll. The guard
+    # streams anything `_flow_is_parsed()` rejects, and a streamed response has
+    # no body — so omitting `account/settings` silently disabled plan detection
+    # entirely: the endpoint reached the proxy 73 times and produced 0
+    # detections, leaving every observation stamped plan='unknown' and
+    # /usage/inferred-caps falling back to Max-20x for every user (found
+    # 2026-08-12). Buffering these two is safe — they are small JSON metadata
+    # responses, not the CDN downloads and telemetry payloads that caused the
+    # jetsam crash loop.
+
+    def _is_oauth_side_channel_flow(self, flow: http.HTTPFlow) -> bool:
         return (
             flow.request.method == "GET"
             and flow.request.pretty_host == self._OAUTH_USAGE_HOST
-            and flow.request.path.split("?", 1)[0] == self._OAUTH_USAGE_PATH
+            and flow.request.path.split("?", 1)[0] in self._OAUTH_PARSED_PATHS
         )
 
     def _flow_is_parsed(self, flow: http.HTTPFlow) -> bool:
         """True when request/response bodies are needed by this addon."""
         if detect_provider(flow.request.pretty_host, flow.request.path) is not None:
             return True
-        return self._is_oauth_usage_flow(flow)
+        return self._is_oauth_side_channel_flow(flow)
 
     def requestheaders(self, flow: http.HTTPFlow) -> None:
         if not self._flow_is_parsed(flow):
@@ -1026,6 +1042,11 @@ class RuleMetricAddon:
 
     _OAUTH_USAGE_HOST = "api.anthropic.com"
     _OAUTH_USAGE_PATH = "/api/oauth/usage"
+    # Side channel carrying the rate-limit tier (the usage body has no plan
+    # field). Must stay in _OAUTH_PARSED_PATHS or its body is streamed away
+    # before _extract_plan_tier can read it — see the streaming guard above.
+    _OAUTH_SETTINGS_PATH = "/api/oauth/account/settings"
+    _OAUTH_PARSED_PATHS = (_OAUTH_USAGE_PATH, _OAUTH_SETTINGS_PATH)
     _OAUTH_USAGE_WINDOWS = (
         ("five_hour", "five_hour_percent"),
         ("seven_day", "seven_day_percent"),
@@ -1042,15 +1063,49 @@ class RuleMetricAddon:
         # Match exact path (strip query string just in case)
         path = request.path.split("?", 1)[0]
 
-        # Side channel: oauth/account/settings carries the rate-limit tier
-        # (e.g. default_claude_max_5x) that the usage body does NOT — cache
-        # it so usage observations get a real plan instead of 'unknown'.
-        if path == "/api/oauth/account/settings" and flow.response is not None \
+        # Side channel: oauth/account/settings is the only response that has
+        # ever carried a rate-limit tier (e.g. default_claude_max_5x); the
+        # usage body has no plan field at all. When present, cache it so usage
+        # observations get a real plan instead of 'unknown'.
+        #
+        # It is frequently ABSENT. Verified 2026-08-12 against a live account:
+        # the field is there but null (`internal_tier_rate_limit_tier: None`,
+        # alongside `internal_tier_seat_tier` and `internal_tier_org_type`,
+        # also null). So buffering this response is necessary for detection but
+        # not sufficient — for such accounts the plan cannot be read from the
+        # wire at all, and /usage/inferred-caps keeps its assumed=true
+        # fallback. The warning below is what makes that visible.
+        if path == self._OAUTH_SETTINGS_PATH and flow.response is not None \
                 and flow.response.status_code == 200:
+            body_text = flow.response.get_text()
             try:
-                tier = _extract_plan_tier(json.loads(flow.response.get_text() or "null"))
+                parsed = json.loads(body_text or "null")
             except (json.JSONDecodeError, ValueError):
-                tier = None
+                parsed = None
+            tier = _extract_plan_tier(parsed)
+            if tier is None and not _SETTINGS_DRIFT_LOGGED["done"]:
+                # Once per process: Claude Code polls this endpoint on every
+                # startup (77 hits in one log), and the tier is null for
+                # accounts Anthropic does not populate it for — warning on
+                # each would bury the log it is meant to make readable.
+                _SETTINGS_DRIFT_LOGGED["done"] = True
+                # Loud on purpose, mirroring the oauth/usage drift warning
+                # below. A silently-absent tier is indistinguishable from a
+                # never-buffered body, and that ambiguity is what hid this for
+                # months — `buffered` disambiguates the two. Restricted to
+                # `*tier*` fields: the rest of this body is account settings
+                # and must never reach the log.
+                tier_fields = (
+                    {k: v for k, v in parsed.items() if "tier" in k.replace("_", "").lower()}
+                    if isinstance(parsed, dict)
+                    else {}
+                )
+                logger.warning(
+                    "oauth/account/settings carried no rate-limit tier — "
+                    "schema drift? buffered=%s tier_fields=%r",
+                    body_text is not None,
+                    tier_fields,
+                )
             if tier and tier != _CACHED_PLAN["plan"]:
                 _CACHED_PLAN["plan"] = tier
                 logger.info("detected plan tier from account/settings: %s", tier)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rulemetric/proxy",
-  "version": "0.8.0",
+  "version": "0.10.0",
   "type": "module",
   "publishConfig": {
     "access": "public"