Merge PR #2762: fix: make app-layer retry classification structured

fix: make app-layer retry classification structured (408/409/timeout/connection)

nanobot/providers/anthropic_provider.py

@@ -52,6 +52,62 @@ class AnthropicProvider(LLMProvider):
         client_kw["max_retries"] = 0
         self._client = AsyncAnthropic(**client_kw)
 
+    @classmethod
+    def _handle_error(cls, e: Exception) -> LLMResponse:
+        response = getattr(e, "response", None)
+        headers = getattr(response, "headers", None)
+        payload = (
+            getattr(e, "body", None)
+            or getattr(e, "doc", None)
+            or getattr(response, "text", None)
+        )
+        if payload is None and response is not None:
+            response_json = getattr(response, "json", None)
+            if callable(response_json):
+                try:
+                    payload = response_json()
+                except Exception:
+                    payload = None
+        payload_text = payload if isinstance(payload, str) else str(payload) if payload is not None else ""
+        msg = f"Error: {payload_text.strip()[:500]}" if payload_text.strip() else f"Error calling LLM: {e}"
+        retry_after = cls._extract_retry_after_from_headers(headers)
+        if retry_after is None:
+            retry_after = LLMProvider._extract_retry_after(msg)
+
+        status_code = getattr(e, "status_code", None)
+        if status_code is None and response is not None:
+            status_code = getattr(response, "status_code", None)
+
+        should_retry: bool | None = None
+        if headers is not None:
+            raw = headers.get("x-should-retry")
+            if isinstance(raw, str):
+                lowered = raw.strip().lower()
+                if lowered == "true":
+                    should_retry = True
+                elif lowered == "false":
+                    should_retry = False
+
+        error_kind: str | None = None
+        error_name = e.__class__.__name__.lower()
+        if "timeout" in error_name:
+            error_kind = "timeout"
+        elif "connection" in error_name:
+            error_kind = "connection"
+        error_type, error_code = LLMProvider._extract_error_type_code(payload)
+
+        return LLMResponse(
+            content=msg,
+            finish_reason="error",
+            retry_after=retry_after,
+            error_status_code=int(status_code) if status_code is not None else None,
+            error_kind=error_kind,
+            error_type=error_type,
+            error_code=error_code,
+            error_retry_after_s=retry_after,
+            error_should_retry=should_retry,
+        )
+
     @staticmethod
     def _strip_prefix(model: str) -> str:
         if model.startswith("anthropic/"):
@@ -404,15 +460,6 @@ class AnthropicProvider(LLMProvider):
     # Public API
     # ------------------------------------------------------------------
 
-    @staticmethod
-    def _handle_error(e: Exception) -> LLMResponse:
-        msg = f"Error calling LLM: {e}"
-        response = getattr(e, "response", None)
-        retry_after = LLMProvider._extract_retry_after_from_headers(getattr(response, "headers", None))
-        if retry_after is None:
-            retry_after = LLMProvider._extract_retry_after(msg)
-        return LLMResponse(content=msg, finish_reason="error", retry_after=retry_after)
-
     async def chat(
         self,
         messages: list[dict[str, Any]],
@@ -474,6 +521,7 @@ class AnthropicProvider(LLMProvider):
                     f"{idle_timeout_s} seconds"
                 ),
                 finish_reason="error",
+                error_kind="timeout",
             )
         except Exception as e:
             return self._handle_error(e)

nanobot/providers/base.py

@@ -54,6 +54,13 @@ class LLMResponse:
     retry_after: float | None = None  # Provider supplied retry wait in seconds.
     reasoning_content: str | None = None  # Kimi, DeepSeek-R1, MiMo etc.
     thinking_blocks: list[dict] | None = None  # Anthropic extended thinking
+    # Structured error metadata used by retry policy when finish_reason == "error".
+    error_status_code: int | None = None
+    error_kind: str | None = None  # e.g. "timeout", "connection"
+    error_type: str | None = None  # Provider/type semantic, e.g. insufficient_quota.
+    error_code: str | None = None  # Provider/code semantic, e.g. rate_limit_exceeded.
+    error_retry_after_s: float | None = None
+    error_should_retry: bool | None = None
 
     @property
     def has_tool_calls(self) -> bool:
@@ -91,6 +98,52 @@ class LLMProvider(ABC):
         "server error",
         "temporarily unavailable",
     )
+    _RETRYABLE_STATUS_CODES = frozenset({408, 409, 429})
+    _TRANSIENT_ERROR_KINDS = frozenset({"timeout", "connection"})
+    _NON_RETRYABLE_429_ERROR_TOKENS = frozenset({
+        "insufficient_quota",
+        "quota_exceeded",
+        "quota_exhausted",
+        "billing_hard_limit_reached",
+        "insufficient_balance",
+        "credit_balance_too_low",
+        "billing_not_active",
+        "payment_required",
+    })
+    _RETRYABLE_429_ERROR_TOKENS = frozenset({
+        "rate_limit_exceeded",
+        "rate_limit_error",
+        "too_many_requests",
+        "request_limit_exceeded",
+        "requests_limit_exceeded",
+        "overloaded_error",
+    })
+    _NON_RETRYABLE_429_TEXT_MARKERS = (
+        "insufficient_quota",
+        "insufficient quota",
+        "quota exceeded",
+        "quota exhausted",
+        "billing hard limit",
+        "billing_hard_limit_reached",
+        "billing not active",
+        "insufficient balance",
+        "insufficient_balance",
+        "credit balance too low",
+        "payment required",
+        "out of credits",
+        "out of quota",
+        "exceeded your current quota",
+    )
+    _RETRYABLE_429_TEXT_MARKERS = (
+        "rate limit",
+        "rate_limit",
+        "too many requests",
+        "retry after",
+        "try again in",
+        "temporarily unavailable",
+        "overloaded",
+        "concurrency limit",
+    )
 
     _SENTINEL = object()
 
@@ -226,6 +279,80 @@ class LLMProvider(ABC):
         err = (content or "").lower()
         return any(marker in err for marker in cls._TRANSIENT_ERROR_MARKERS)
 
+    @classmethod
+    def _is_transient_response(cls, response: LLMResponse) -> bool:
+        """Prefer structured error metadata, fallback to text markers for legacy providers."""
+        if response.error_should_retry is not None:
+            return bool(response.error_should_retry)
+
+        if response.error_status_code is not None:
+            status = int(response.error_status_code)
+            if status == 429:
+                return cls._is_retryable_429_response(response)
+            if status in cls._RETRYABLE_STATUS_CODES or status >= 500:
+                return True
+
+        kind = (response.error_kind or "").strip().lower()
+        if kind in cls._TRANSIENT_ERROR_KINDS:
+            return True
+
+        return cls._is_transient_error(response.content)
+
+    @staticmethod
+    def _normalize_error_token(value: Any) -> str | None:
+        if value is None:
+            return None
+        token = str(value).strip().lower()
+        return token or None
+
+    @classmethod
+    def _extract_error_type_code(cls, payload: Any) -> tuple[str | None, str | None]:
+        data: dict[str, Any] | None = None
+        if isinstance(payload, dict):
+            data = payload
+        elif isinstance(payload, str):
+            text = payload.strip()
+            if text:
+                try:
+                    parsed = json.loads(text)
+                except Exception:
+                    parsed = None
+                if isinstance(parsed, dict):
+                    data = parsed
+        if not isinstance(data, dict):
+            return None, None
+
+        error_obj = data.get("error")
+        type_value = data.get("type")
+        code_value = data.get("code")
+        if isinstance(error_obj, dict):
+            type_value = error_obj.get("type") or type_value
+            code_value = error_obj.get("code") or code_value
+
+        return cls._normalize_error_token(type_value), cls._normalize_error_token(code_value)
+
+    @classmethod
+    def _is_retryable_429_response(cls, response: LLMResponse) -> bool:
+        type_token = cls._normalize_error_token(response.error_type)
+        code_token = cls._normalize_error_token(response.error_code)
+        semantic_tokens = {
+            token for token in (type_token, code_token)
+            if token is not None
+        }
+        if any(token in cls._NON_RETRYABLE_429_ERROR_TOKENS for token in semantic_tokens):
+            return False
+
+        content = (response.content or "").lower()
+        if any(marker in content for marker in cls._NON_RETRYABLE_429_TEXT_MARKERS):
+            return False
+
+        if any(token in cls._RETRYABLE_429_ERROR_TOKENS for token in semantic_tokens):
+            return True
+        if any(marker in content for marker in cls._RETRYABLE_429_TEXT_MARKERS):
+            return True
+        # Unknown 429 defaults to WAIT+retry.
+        return True
+
     @staticmethod
     def _strip_image_content(messages: list[dict[str, Any]]) -> list[dict[str, Any]] | None:
         """Replace image_url blocks with text placeholder. Returns None if no images found."""
@@ -397,14 +524,28 @@ class LLMProvider(ABC):
     def _extract_retry_after_from_headers(cls, headers: Any) -> float | None:
         if not headers:
             return None
-        retry_after: Any = None
-        if hasattr(headers, "get"):
-            retry_after = headers.get("retry-after") or headers.get("Retry-After")
-        if retry_after is None and isinstance(headers, dict):
-            for key, value in headers.items():
-                if isinstance(key, str) and key.lower() == "retry-after":
-                    retry_after = value
-                    break
+
+        def _header_value(name: str) -> Any:
+            if hasattr(headers, "get"):
+                value = headers.get(name) or headers.get(name.title())
+                if value is not None:
+                    return value
+            if isinstance(headers, dict):
+                for key, value in headers.items():
+                    if isinstance(key, str) and key.lower() == name.lower():
+                        return value
+            return None
+
+        try:
+            retry_ms = _header_value("retry-after-ms")
+            if retry_ms is not None:
+                value = float(retry_ms) / 1000.0
+                if value > 0:
+                    return value
+        except (TypeError, ValueError):
+            pass
+
+        retry_after = _header_value("retry-after")
         if retry_after is None:
             return None
         retry_after_text = str(retry_after).strip()
@@ -421,6 +562,14 @@ class LLMProvider(ABC):
         remaining = (retry_at - datetime.now(retry_at.tzinfo)).total_seconds()
         return max(0.1, remaining)
 
+    @classmethod
+    def _extract_retry_after_from_response(cls, response: LLMResponse) -> float | None:
+        if response.error_retry_after_s is not None and response.error_retry_after_s > 0:
+            return response.error_retry_after_s
+        if response.retry_after is not None and response.retry_after > 0:
+            return response.retry_after
+        return cls._extract_retry_after(response.content)
+
     async def _sleep_with_heartbeat(
         self,
         delay: float,
@@ -469,7 +618,7 @@ class LLMProvider(ABC):
                 last_error_key = error_key
                 identical_error_count = 1 if error_key else 0
 
-            if not self._is_transient_error(response.content):
+            if not self._is_transient_response(response):
                 stripped = self._strip_image_content(original_messages)
                 if stripped is not None and stripped != kw["messages"]:
                     logger.warning(
@@ -492,7 +641,7 @@ class LLMProvider(ABC):
                 break
 
             base_delay = delays[min(attempt - 1, len(delays) - 1)]
-            delay = response.retry_after or self._extract_retry_after(response.content) or base_delay
+            delay = self._extract_retry_after_from_response(response) or base_delay
             if persistent:
                 delay = min(delay, self._PERSISTENT_MAX_DELAY)
 

nanobot/providers/openai_compat_provider.py

@@ -634,16 +634,73 @@ class OpenAICompatProvider(LLMProvider):
             reasoning_content="".join(reasoning_parts) or None,
         )
 
+    @classmethod
+    def _extract_error_metadata(cls, e: Exception) -> dict[str, Any]:
+        response = getattr(e, "response", None)
+        headers = getattr(response, "headers", None)
+        payload = (
+            getattr(e, "body", None)
+            or getattr(e, "doc", None)
+            or getattr(response, "text", None)
+        )
+        if payload is None and response is not None:
+            response_json = getattr(response, "json", None)
+            if callable(response_json):
+                try:
+                    payload = response_json()
+                except Exception:
+                    payload = None
+        error_type, error_code = LLMProvider._extract_error_type_code(payload)
+
+        status_code = getattr(e, "status_code", None)
+        if status_code is None and response is not None:
+            status_code = getattr(response, "status_code", None)
+
+        should_retry: bool | None = None
+        if headers is not None:
+            raw = headers.get("x-should-retry")
+            if isinstance(raw, str):
+                lowered = raw.strip().lower()
+                if lowered == "true":
+                    should_retry = True
+                elif lowered == "false":
+                    should_retry = False
+
+        error_kind: str | None = None
+        error_name = e.__class__.__name__.lower()
+        if "timeout" in error_name:
+            error_kind = "timeout"
+        elif "connection" in error_name:
+            error_kind = "connection"
+
+        return {
+            "error_status_code": int(status_code) if status_code is not None else None,
+            "error_kind": error_kind,
+            "error_type": error_type,
+            "error_code": error_code,
+            "error_retry_after_s": cls._extract_retry_after_from_headers(headers),
+            "error_should_retry": should_retry,
+        }
+
     @staticmethod
     def _handle_error(e: Exception) -> LLMResponse:
+        body = (
+            getattr(e, "doc", None)
+            or getattr(e, "body", None)
+            or getattr(getattr(e, "response", None), "text", None)
+        )
+        body_text = body if isinstance(body, str) else str(body) if body is not None else ""
+        msg = f"Error: {body_text.strip()[:500]}" if body_text.strip() else f"Error calling LLM: {e}"
         response = getattr(e, "response", None)
-        body = getattr(e, "doc", None) or getattr(response, "text", None)
-        body_text = str(body).strip() if body is not None else ""
-        msg = f"Error: {body_text[:500]}" if body_text else f"Error calling LLM: {e}"
         retry_after = LLMProvider._extract_retry_after_from_headers(getattr(response, "headers", None))
         if retry_after is None:
             retry_after = LLMProvider._extract_retry_after(msg)
-        return LLMResponse(content=msg, finish_reason="error", retry_after=retry_after)
+        return LLMResponse(
+            content=msg,
+            finish_reason="error",
+            retry_after=retry_after,
+            **OpenAICompatProvider._extract_error_metadata(e),
+        )
 
     # ------------------------------------------------------------------
     # Public API
@@ -711,6 +768,7 @@ class OpenAICompatProvider(LLMProvider):
                     f"{idle_timeout_s} seconds"
                 ),
                 finish_reason="error",
+                error_kind="timeout",
             )
         except Exception as e:
             return self._handle_error(e)

tests/providers/test_provider_error_metadata.py

@@ -0,0 +1,81 @@
+from types import SimpleNamespace
+
+from nanobot.providers.anthropic_provider import AnthropicProvider
+from nanobot.providers.openai_compat_provider import OpenAICompatProvider
+
+
+def _fake_response(
+    *,
+    status_code: int,
+    headers: dict[str, str] | None = None,
+    text: str = "",
+) -> SimpleNamespace:
+    return SimpleNamespace(
+        status_code=status_code,
+        headers=headers or {},
+        text=text,
+    )
+
+
+def test_openai_handle_error_extracts_structured_metadata() -> None:
+    class FakeStatusError(Exception):
+        pass
+
+    err = FakeStatusError("boom")
+    err.status_code = 409
+    err.response = _fake_response(
+        status_code=409,
+        headers={"retry-after-ms": "250", "x-should-retry": "false"},
+        text='{"error":{"type":"rate_limit_exceeded","code":"rate_limit_exceeded"}}',
+    )
+    err.body = {"error": {"type": "rate_limit_exceeded", "code": "rate_limit_exceeded"}}
+
+    response = OpenAICompatProvider._handle_error(err)
+
+    assert response.finish_reason == "error"
+    assert response.error_status_code == 409
+    assert response.error_type == "rate_limit_exceeded"
+    assert response.error_code == "rate_limit_exceeded"
+    assert response.error_retry_after_s == 0.25
+    assert response.error_should_retry is False
+
+
+def test_openai_handle_error_marks_timeout_kind() -> None:
+    class FakeTimeoutError(Exception):
+        pass
+
+    response = OpenAICompatProvider._handle_error(FakeTimeoutError("timeout"))
+
+    assert response.finish_reason == "error"
+    assert response.error_kind == "timeout"
+
+
+def test_anthropic_handle_error_extracts_structured_metadata() -> None:
+    class FakeStatusError(Exception):
+        pass
+
+    err = FakeStatusError("boom")
+    err.status_code = 408
+    err.response = _fake_response(
+        status_code=408,
+        headers={"retry-after": "1.5", "x-should-retry": "true"},
+    )
+    err.body = {"type": "error", "error": {"type": "rate_limit_error"}}
+
+    response = AnthropicProvider._handle_error(err)
+
+    assert response.finish_reason == "error"
+    assert response.error_status_code == 408
+    assert response.error_type == "rate_limit_error"
+    assert response.error_retry_after_s == 1.5
+    assert response.error_should_retry is True
+
+
+def test_anthropic_handle_error_marks_connection_kind() -> None:
+    class FakeConnectionError(Exception):
+        pass
+
+    response = AnthropicProvider._handle_error(FakeConnectionError("connection"))
+
+    assert response.finish_reason == "error"
+    assert response.error_kind == "connection"

tests/providers/test_provider_retry.py

@@ -254,6 +254,14 @@ def test_extract_retry_after_from_headers_supports_numeric_and_http_date() -> No
     ) == 0.1
 
 
+def test_extract_retry_after_from_headers_supports_retry_after_ms() -> None:
+    assert LLMProvider._extract_retry_after_from_headers({"retry-after-ms": "250"}) == 0.25
+    assert LLMProvider._extract_retry_after_from_headers({"Retry-After-Ms": "1000"}) == 1.0
+    assert LLMProvider._extract_retry_after_from_headers(
+        {"retry-after-ms": "500", "retry-after": "10"},
+    ) == 0.5
+
+
 @pytest.mark.asyncio
 async def test_chat_with_retry_prefers_structured_retry_after_when_present(monkeypatch) -> None:
     provider = ScriptedProvider([
@@ -273,6 +281,153 @@ async def test_chat_with_retry_prefers_structured_retry_after_when_present(monke
     assert delays == [9.0]
 
 
+@pytest.mark.asyncio
+async def test_chat_with_retry_retries_structured_status_code_without_keyword(monkeypatch) -> None:
+    provider = ScriptedProvider([
+        LLMResponse(
+            content="request failed",
+            finish_reason="error",
+            error_status_code=409,
+        ),
+        LLMResponse(content="ok"),
+    ])
+    delays: list[float] = []
+
+    async def _fake_sleep(delay: float) -> None:
+        delays.append(delay)
+
+    monkeypatch.setattr("nanobot.providers.base.asyncio.sleep", _fake_sleep)
+
+    response = await provider.chat_with_retry(messages=[{"role": "user", "content": "hello"}])
+
+    assert response.content == "ok"
+    assert provider.calls == 2
+    assert delays == [1]
+
+
+@pytest.mark.asyncio
+async def test_chat_with_retry_stops_on_429_quota_exhausted(monkeypatch) -> None:
+    provider = ScriptedProvider([
+        LLMResponse(
+            content='{"error":{"type":"insufficient_quota","code":"insufficient_quota"}}',
+            finish_reason="error",
+            error_status_code=429,
+            error_type="insufficient_quota",
+            error_code="insufficient_quota",
+        ),
+        LLMResponse(content="ok"),
+    ])
+    delays: list[float] = []
+
+    async def _fake_sleep(delay: float) -> None:
+        delays.append(delay)
+
+    monkeypatch.setattr("nanobot.providers.base.asyncio.sleep", _fake_sleep)
+
+    response = await provider.chat_with_retry(messages=[{"role": "user", "content": "hello"}])
+
+    assert response.finish_reason == "error"
+    assert provider.calls == 1
+    assert delays == []
+
+
+@pytest.mark.asyncio
+async def test_chat_with_retry_retries_429_transient_rate_limit(monkeypatch) -> None:
+    provider = ScriptedProvider([
+        LLMResponse(
+            content='{"error":{"type":"rate_limit_exceeded","code":"rate_limit_exceeded"}}',
+            finish_reason="error",
+            error_status_code=429,
+            error_type="rate_limit_exceeded",
+            error_code="rate_limit_exceeded",
+            error_retry_after_s=0.2,
+        ),
+        LLMResponse(content="ok"),
+    ])
+    delays: list[float] = []
+
+    async def _fake_sleep(delay: float) -> None:
+        delays.append(delay)
+
+    monkeypatch.setattr("nanobot.providers.base.asyncio.sleep", _fake_sleep)
+
+    response = await provider.chat_with_retry(messages=[{"role": "user", "content": "hello"}])
+
+    assert response.content == "ok"
+    assert provider.calls == 2
+    assert delays == [0.2]
+
+
+@pytest.mark.asyncio
+async def test_chat_with_retry_retries_structured_timeout_kind(monkeypatch) -> None:
+    provider = ScriptedProvider([
+        LLMResponse(
+            content="request failed",
+            finish_reason="error",
+            error_kind="timeout",
+        ),
+        LLMResponse(content="ok"),
+    ])
+    delays: list[float] = []
+
+    async def _fake_sleep(delay: float) -> None:
+        delays.append(delay)
+
+    monkeypatch.setattr("nanobot.providers.base.asyncio.sleep", _fake_sleep)
+
+    response = await provider.chat_with_retry(messages=[{"role": "user", "content": "hello"}])
+
+    assert response.content == "ok"
+    assert provider.calls == 2
+    assert delays == [1]
+
+
+@pytest.mark.asyncio
+async def test_chat_with_retry_structured_should_retry_false_disables_retry(monkeypatch) -> None:
+    provider = ScriptedProvider([
+        LLMResponse(
+            content="429 rate limit",
+            finish_reason="error",
+            error_should_retry=False,
+        ),
+    ])
+    delays: list[float] = []
+
+    async def _fake_sleep(delay: float) -> None:
+        delays.append(delay)
+
+    monkeypatch.setattr("nanobot.providers.base.asyncio.sleep", _fake_sleep)
+
+    response = await provider.chat_with_retry(messages=[{"role": "user", "content": "hello"}])
+
+    assert response.finish_reason == "error"
+    assert provider.calls == 1
+    assert delays == []
+
+
+@pytest.mark.asyncio
+async def test_chat_with_retry_prefers_structured_retry_after(monkeypatch) -> None:
+    provider = ScriptedProvider([
+        LLMResponse(
+            content="429 rate limit, retry after 99s",
+            finish_reason="error",
+            error_retry_after_s=0.2,
+        ),
+        LLMResponse(content="ok"),
+    ])
+    delays: list[float] = []
+
+    async def _fake_sleep(delay: float) -> None:
+        delays.append(delay)
+
+    monkeypatch.setattr("nanobot.providers.base.asyncio.sleep", _fake_sleep)
+
+    response = await provider.chat_with_retry(messages=[{"role": "user", "content": "hello"}])
+
+    assert response.content == "ok"
+    assert delays == [0.2]
+
+
 @pytest.mark.asyncio
 async def test_persistent_retry_aborts_after_ten_identical_transient_errors(monkeypatch) -> None:
     provider = ScriptedProvider([
@@ -295,4 +450,3 @@ async def test_persistent_retry_aborts_after_ten_identical_transient_errors(monk
     assert response.content == "429 rate limit"
     assert provider.calls == 10
     assert delays == [1, 2, 4, 4, 4, 4, 4, 4, 4]
-