coding-proxy 0.5.1a6__tar.gz → 0.5.1a9__tar.gz

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: coding-proxy
-Version: 0.5.1a6
+Version: 0.5.1a9
 Summary: A High-Availability, Transparent, and Smart Multi-Vendor Proxy for Claude Code. Support Claude Plans, GitHub Copilot, Google Antigravity, ZAI/GLM, MiniMax, Qwen, Xiaomi, Kimi, Doubao...
 Project-URL: Source Code, https://github.com/ThreeFish-AI/coding-proxy
 Project-URL: User Guide, https://github.com/ThreeFish-AI/coding-proxy/blob/master/docs/user-guide.md

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/docs/arch/design-patterns.md

@@ -541,6 +541,8 @@ $$
 
 > **参数默认值**：参见 [配置参考 -- RetryConfig](./config-reference.md#elastic-params)
 
+**Vendor 级应用（Zhipu 429/529 内部退避重试）**：[`vendors/zhipu.py`](../../src/coding/proxy/vendors/zhipu.py) 复用 `RetryConfig` / `calculate_delay()`，在 `send_message` / `send_message_stream` 内对 **429 Rate Limit（限流）** 与 **529 Overloaded（并发过载）** 两类服务端瞬态过载信号做就地指数退避重试（共用 `_BACKOFF_RETRY_STATUS = {429, 529}` 作单一事实源，max=5、1s→2s→4s→8s、Full Jitter，优先尊重 server `retry-after`）。耗尽重试后将状态码原样返回，交由上层 `should_trigger_failover`（§3.6 VendorTier）与 [CircuitBreaker](#circuit-breaker) 处理，从而降低 failover 频率。
+
 ---
 
 ## 3.13 Rate Limit Deadline Tracking（速率限制截止追踪）

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "coding-proxy"
-version = "0.5.1a6"
+version = "0.5.1a9"
 description = "A High-Availability, Transparent, and Smart Multi-Vendor Proxy for Claude Code. Support Claude Plans, GitHub Copilot, Google Antigravity, ZAI/GLM, MiniMax, Qwen, Xiaomi, Kimi, Doubao..."
 readme = "README.md"
 requires-python = ">=3.12"

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/src/coding/proxy/config/config.default.yaml

@@ -228,51 +228,51 @@ failover:
 # === 模型映射 ===
 model_mapping:
   # GitHub Copilot 可用模型（通过 /api/copilot/models 诊断获取）
-  - pattern: "claude-sonnet-.*"
-    vendors: ["copilot"]
-    target: "claude-sonnet-4.6"
-    is_regex: true
   - pattern: "claude-opus-.*"
     vendors: ["copilot"]
     target: "claude-opus-4.6"
     is_regex: true
+  - pattern: "claude-sonnet-.*"
+    vendors: ["copilot"]
+    target: "claude-sonnet-4.6"
+    is_regex: true
   - pattern: "claude-haiku-.*"
     vendors: ["copilot"]
     target: "claude-haiku-4.5"
     is_regex: true
   # Google Antigravity 模型映射
-  - pattern: "claude-sonnet-.*"
-    vendors: ["antigravity"]
-    target: "Claude Sonnet 4.6"
-    is_regex: true
   - pattern: "claude-opus-.*"
     vendors: ["antigravity"]
     target: "Claude Opus 4.6 Thinking"
     is_regex: true
+  - pattern: "claude-sonnet-.*"
+    vendors: ["antigravity"]
+    target: "Claude Sonnet 4.6"
+    is_regex: true
   - pattern: "claude-haiku-.*"
     vendors: ["antigravity"]
     target: "Gemini 3 Flash"
     is_regex: true
   # 智谱 GLM 映射（官方原生 Anthropic 兼容端点）
   # glm-5v-turbo 完整承载 Claude Code 满血能力（tool calling / thinking / 多模态）
-  - pattern: "claude-sonnet-.*"
+  - pattern: "claude-opus-.*"
     vendors: ["zhipu"]
-    target: "glm-5v-turbo"
+    target: "glm-5.2"
     is_regex: true
-  - pattern: "claude-opus-.*"
+  - pattern: "claude-sonnet-.*"
     vendors: ["zhipu"]
     target: "glm-5.1"
     is_regex: true
   - pattern: "claude-haiku-.*"
     vendors: ["zhipu"]
-    target: "glm-4.5-air"
+    target: "glm-5.1"
     is_regex: true
   # MiniMax 模型映射（原生 Anthropic 兼容端点）
-  - pattern: "claude-sonnet-.*"
+  - pattern: "claude-opus-.*"
     vendors: ["minimax"]
     target: "minimax-m2.7"
     is_regex: true
-  - pattern: "claude-opus-.*"
+  - pattern: "claude-sonnet-.*"
     vendors: ["minimax"]
     target: "minimax-m2.7"
     is_regex: true
@@ -281,11 +281,11 @@ model_mapping:
     target: "minimax-m2.7"
     is_regex: true
   # 阿里 Qwen 模型映射（原生 Anthropic 兼容端点）
-  - pattern: "claude-sonnet-.*"
+  - pattern: "claude-opus-.*"
     vendors: ["alibaba"]
     target: "qwen3.6-plus"
     is_regex: true
-  - pattern: "claude-opus-.*"
+  - pattern: "claude-sonnet-.*"
     vendors: ["alibaba"]
     target: "qwen3.6-plus"
     is_regex: true
@@ -294,11 +294,11 @@ model_mapping:
     target: "qwen3.6-plus"
     is_regex: true
   # 小米 MiMo 模型映射（原生 Anthropic 兼容端点）
-  - pattern: "claude-sonnet-.*"
+  - pattern: "claude-opus-.*"
     vendors: ["xiaomi"]
     target: "mimo-v2-pro"
     is_regex: true
-  - pattern: "claude-opus-.*"
+  - pattern: "claude-sonnet-.*"
     vendors: ["xiaomi"]
     target: "mimo-v2-pro"
     is_regex: true
@@ -307,11 +307,11 @@ model_mapping:
     target: "mimo-v2-pro"
     is_regex: true
   # Kimi 模型映射（原生 Anthropic 兼容端点）
-  - pattern: "claude-sonnet-.*"
+  - pattern: "claude-opus-.*"
     vendors: ["kimi"]
     target: "kimi-k2.5"
     is_regex: true
-  - pattern: "claude-opus-.*"
+  - pattern: "claude-sonnet-.*"
     vendors: ["kimi"]
     target: "kimi-k2.5"
     is_regex: true
@@ -339,6 +339,12 @@ model_mapping:
 # 未配置定价的模型在统计中 Cost 列显示 "-"
 pricing:
   # ── Anthropic Claude ──
+  - vendor: anthropic
+    model: claude-fable-5
+    input_cost_per_mtok: $10.0
+    output_cost_per_mtok: $50.0
+    cache_write_cost_per_mtok: $12.50
+    cache_read_cost_per_mtok: $1.00
   - vendor: anthropic
     model: claude-opus-4-8
     input_cost_per_mtok: $5.0
@@ -527,20 +533,10 @@ pricing:
     cache_read_cost_per_mtok: $0.05
   # ── 智谱 GLM ──
   - vendor: zhipu
-    model: glm-4.5-air # 待区分长短上下文定价
-    input_cost_per_mtok: ¥0.80
-    output_cost_per_mtok: ¥2.00
-    cache_read_cost_per_mtok: ¥0.16
-  - vendor: zhipu
-    model: glm-4.7 # 待区分长短上下文定价
-    input_cost_per_mtok: ¥2.00
-    output_cost_per_mtok: ¥8.00
-    cache_read_cost_per_mtok: ¥0.40
-  - vendor: zhipu
-    model: glm-5v-turbo # 待区分长短上下文定价
-    input_cost_per_mtok: ¥5.00
-    output_cost_per_mtok: ¥22.00
-    cache_read_cost_per_mtok: ¥1.20
+    model: glm-5.2 # 待区分长短上下文定价
+    input_cost_per_mtok: ¥6.00
+    output_cost_per_mtok: ¥24.00
+    cache_read_cost_per_mtok: ¥1.30
   - vendor: zhipu
     model: glm-5.1 # 待区分长短上下文定价
     input_cost_per_mtok: ¥6.00
@@ -556,6 +552,21 @@ pricing:
     input_cost_per_mtok: ¥4.00
     output_cost_per_mtok: ¥18.00
     cache_read_cost_per_mtok: ¥1.00
+  - vendor: zhipu
+    model: glm-5v-turbo # 待区分长短上下文定价
+    input_cost_per_mtok: ¥5.00
+    output_cost_per_mtok: ¥22.00
+    cache_read_cost_per_mtok: ¥1.20
+  - vendor: zhipu
+    model: glm-4.7 # 待区分长短上下文定价
+    input_cost_per_mtok: ¥2.00
+    output_cost_per_mtok: ¥8.00
+    cache_read_cost_per_mtok: ¥0.40
+  - vendor: zhipu
+    model: glm-4.5-air # 待区分长短上下文定价
+    input_cost_per_mtok: ¥0.80
+    output_cost_per_mtok: ¥2.00
+    cache_read_cost_per_mtok: ¥0.16
   # ── MiniMax ──
   - vendor: minimax
     model: minimax-m2.7
@@ -685,4 +696,11 @@ native_api:
 #       tiers: ["copilot", "anthropic", "zhipu"]
 #
 # 未配置时（默认），所有 Session 使用全局 tiers 顺序。
-session_policies: []
+session_policies:
+  policies: []
+  # 标题前缀 → 供应商自动绑定。
+  # 当 Session 标题以指定前缀开头时，自动将该 Session 绑定到对应供应商。
+  # 匹配规则按列表顺序求值，首次匹配生效。
+  title_vendor_bindings:
+    - prefix: "# 目标"
+      vendor: "zhipu"

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/src/coding/proxy/config/schema.py

@@ -44,7 +44,7 @@ from .routing import (  # noqa: F401
 
 # ── 子模块 re-export ────────────────────────────────────────────
 from .server import DatabaseConfig, LoggingConfig, ServerConfig  # noqa: F401
-from .session_policy import SessionPoliciesConfig  # noqa: F401
+from .session_policy import SessionPoliciesConfig, TitleVendorBinding  # noqa: F401
 from .vendors import (  # noqa: F401
     AlibabaConfig,
     AnthropicConfig,
@@ -350,4 +350,5 @@ __all__ = [
     "NativeApiConfig",
     # session policy
     "SessionPoliciesConfig",
+    "TitleVendorBinding",
 ]

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/src/coding/proxy/config/session_policy.py

@@ -50,6 +50,22 @@ class SessionPolicy(BaseModel):
     )
 
 
+class TitleVendorBinding(BaseModel):
+    """标题前缀 → 供应商自动绑定规则."""
+
+    prefix: str = Field(
+        min_length=1,
+        description=(
+            "标题前缀匹配模式（大小写敏感的 startswith 匹配）。"
+            "禁止空字符串——空前缀会匹配所有标题,导致全量误绑定。"
+        ),
+    )
+    vendor: str = Field(
+        min_length=1,
+        description="匹配后绑定的目标供应商名称",
+    )
+
+
 class SessionPoliciesConfig(BaseModel):
     """顶层 Session 策略配置容器."""
 
@@ -57,3 +73,11 @@ class SessionPoliciesConfig(BaseModel):
         default_factory=list,
         description="Session 路由策略列表，按定义顺序求值，首次匹配生效",
     )
+    title_vendor_bindings: list[TitleVendorBinding] = Field(
+        default_factory=list,
+        description=(
+            "标题前缀 → 供应商自动绑定规则。"
+            "当 Session 标题以指定前缀开头时，自动绑定到对应供应商。"
+            "匹配规则按列表顺序求值，首次匹配生效。"
+        ),
+    )

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/src/coding/proxy/routing/executor.py

@@ -11,10 +11,13 @@ import logging
 import re
 import time
 from collections.abc import AsyncIterator
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import httpx
 
+if TYPE_CHECKING:
+    from ..config.session_policy import TitleVendorBinding
+
 from ..vendors.base import (
     NoCompatibleVendorError,
     RequestCapabilities,
@@ -610,6 +613,7 @@ class _RouteExecutor:
         session_manager: RouteSessionManager,
         reauth_coordinator: Any | None = None,
         session_policy_resolver: SessionPolicyResolver | None = None,
+        title_vendor_bindings: list[TitleVendorBinding] | None = None,
     ) -> None:
         self._router = router
         self._tiers = tiers
@@ -617,6 +621,8 @@ class _RouteExecutor:
         self._session_mgr = session_manager
         self._reauth_coordinator = reauth_coordinator
         self._policy_resolver = session_policy_resolver or SessionPolicyResolver()
+        self._title_vendor_bindings = title_vendor_bindings or []
+        self._validate_title_vendor_bindings()
 
         # Tier 名称 → OAuth provider 名称的映射
         self._tier_provider_map: dict[str, str] = {
@@ -624,6 +630,26 @@ class _RouteExecutor:
             "antigravity": "google",
         }
 
+    def _validate_title_vendor_bindings(self) -> None:
+        """启动期校验标题绑定引用的 vendor 均存在,缺失则告警.
+
+        与手动绑定 API（拒绝未知 vendor）的语义对齐：此处不硬失败，
+        仅记录警告——避免单条误配置阻断整个代理启动；运行时
+        `_resolve_effective_tiers` 会静默跳过未知 vendor 回退默认顺序。
+        """
+        if not self._title_vendor_bindings:
+            return
+        valid = {t.name for t in self._tiers}
+        for binding in self._title_vendor_bindings:
+            if binding.vendor not in valid:
+                logger.warning(
+                    "title_vendor_bindings 引用了未知 vendor %r（前缀 %r）；"
+                    "可用 vendor: %s。该绑定将在运行时被静默跳过。",
+                    binding.vendor,
+                    binding.prefix,
+                    sorted(valid),
+                )
+
     # ── 公开执行入口 ──────────────────────────────────────
 
     def _resolve_effective_tiers(self, session_key: str) -> list[VendorTier]:
@@ -650,6 +676,27 @@ class _RouteExecutor:
                 seen.add(tier.name)
         return ordered
 
+    def _apply_title_based_policy(self, session_key: str, title: str) -> None:
+        """根据 Session 标题前缀自动绑定供应商.
+
+        当标题以预配置的前缀开头时，通过 SessionPolicyResolver.upsert()
+        将该 Session 绑定到指定供应商，后续请求无需再走默认路由。
+
+        仅在新 Session 首次提取标题时调用，避免覆盖手动绑定的策略。
+        """
+        if not title or not self._title_vendor_bindings:
+            return
+        for binding in self._title_vendor_bindings:
+            if title.startswith(binding.prefix):
+                self._policy_resolver.upsert(session_key, [binding.vendor])
+                logger.info(
+                    "Session title prefix %r matched → auto-bind to %s (session=%s)",
+                    binding.prefix,
+                    binding.vendor,
+                    session_key[:12],
+                )
+                return
+
     def _prepare_body_for_tier(
         self,
         body: dict[str, Any],
@@ -748,6 +795,7 @@ class _RouteExecutor:
                 await self._recorder.set_session_title(
                     canonical_request.session_key, title
                 )
+                self._apply_title_based_policy(canonical_request.session_key, title)
         else:
             # 延迟标题补写: 若 session 尚无标题,尝试从当前请求中提取并回写。
             title = _extract_session_title(canonical_request)
@@ -934,6 +982,7 @@ class _RouteExecutor:
                 await self._recorder.set_session_title(
                     canonical_request.session_key, title
                 )
+                self._apply_title_based_policy(canonical_request.session_key, title)
         else:
             # 延迟标题补写: 若 session 尚无标题,尝试从当前请求中提取并回写。
             title = _extract_session_title(canonical_request)

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/src/coding/proxy/routing/router.py

@@ -14,6 +14,7 @@ from collections.abc import AsyncIterator
 from typing import TYPE_CHECKING, Any
 
 if TYPE_CHECKING:
+    from ..config.session_policy import TitleVendorBinding
     from ..pricing import PricingTable
 
 from .executor import _RouteExecutor
@@ -38,6 +39,7 @@ class RequestRouter:
         reauth_coordinator: Any | None = None,
         compat_session_store: CompatSessionStore | None = None,
         session_policy_resolver: SessionPolicyResolver | None = None,
+        title_vendor_bindings: list[TitleVendorBinding] | None = None,
     ) -> None:
         if not tiers:
             raise ValueError("至少需要一个供应商层级")
@@ -56,6 +58,7 @@ class RequestRouter:
             session_manager=self._session_mgr,
             reauth_coordinator=reauth_coordinator,
             session_policy_resolver=session_policy_resolver,
+            title_vendor_bindings=title_vendor_bindings,
         )
 
     def set_pricing_table(self, table: PricingTable) -> None:

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/src/coding/proxy/server/app.py

@@ -161,6 +161,7 @@ def create_app(config: ProxyConfig | None = None) -> FastAPI:
         reauth_coordinator,
         compat_session_store,
         session_policy_resolver=SessionPolicyResolver(config.session_policies.policies),
+        title_vendor_bindings=config.session_policies.title_vendor_bindings,
     )
 
     app = FastAPI(title="coding-proxy", version=__version__, lifespan=lifespan)

{coding_proxy-0.5.1a6 → coding_proxy-0.5.1a9}/src/coding/proxy/vendors/zhipu.py

@@ -1,4 +1,4 @@
-"""智谱 GLM 供应商 — 原生 Anthropic 兼容端点代理（兼容转换 + 429 重试）.
+"""智谱 GLM 供应商 — 原生 Anthropic 兼容端点代理（兼容转换 + 429/529 重试）.
 
 官方端点 (https://open.bigmodel.cn/api/anthropic) 支持大部分
 Anthropic Messages API 协议，本模块做以下适配：
@@ -13,7 +13,8 @@ Anthropic Messages API 协议，本模块做以下适配：
 - reasoning_effort 参数：静默忽略
 - metadata 字段：暂不处理（待进一步诊断确认兼容性）
 
-额外提供 429 Rate Limit 专用重试挽回机制：
+额外提供 429/529 专用重试挽回机制：
+  - 429 Rate Limit（限流）与 529 Overloaded（并发过载）共用同一退避策略
   - max_attempt = 5（1 初始 + 4 重试）
   - 指数退避 + Full Jitter（1s → 2s → 4s → 8s）
   - 优先尊重 server retry-after header
@@ -56,14 +57,20 @@ _RATE_LIMIT_RETRY = RetryConfig(
     jitter=True,
 )
 
+# 触发指数退避重试的瞬态状态码（单一事实源）
+#   429 Rate Limit（限流）；529 Overloaded（并发过载）
+# 二者均为服务端瞬态过载信号，适用同一退避挽回策略。
+_BACKOFF_RETRY_STATUS = frozenset({429, 529})
+
 
 class ZhipuVendor(NativeAnthropicVendor):
-    """智谱 GLM 原生 Anthropic 兼容端点供应商（薄透传 + 429 重试挽回）.
+    """智谱 GLM 原生 Anthropic 兼容端点供应商（薄透传 + 429/529 重试挽回）.
 
     通过官方 /api/anthropic 端点转发请求，
     仅替换模型名和认证头，其余原样透传。
 
-    429 Rate Limit 时自动重试（指数退避），降低 failover 频率。
+    429 Rate Limit（限流）/ 529 Overloaded（并发过载）时自动重试
+    （指数退避），降低 failover 频率。
     并发限流由 BaseVendor._concurrency_controller 统一管控。
     """
 
@@ -124,24 +131,25 @@ class ZhipuVendor(NativeAnthropicVendor):
 
         return body, new_headers
 
-    # ── 非流式：429 重试 ────────────────────────────────────
+    # ── 非流式：429/529 重试 ────────────────────────────────
 
     async def send_message(
         self,
         request_body: dict[str, Any],
         headers: dict[str, str],
     ) -> VendorResponse:
-        """非流式请求，429 时自动重试."""
+        """非流式请求，429/529 时自动重试."""
         max_attempts = self._rl_retry.max_attempts
 
         for attempt in range(max_attempts):
             resp = await super().send_message(request_body, headers)
-            if resp.status_code != 429:
+            if resp.status_code not in _BACKOFF_RETRY_STATUS:
                 return resp
 
             if attempt == max_attempts - 1:
                 logger.warning(
-                    "Zhipu 429 rate limit exhausted after %d attempts",
+                    "Zhipu %d overload exhausted after %d attempts",
+                    resp.status_code,
                     max_attempts,
                 )
                 return resp
@@ -150,7 +158,8 @@ class ZhipuVendor(NativeAnthropicVendor):
                 resp.response_headers, attempt
             )
             logger.info(
-                "Zhipu 429 rate limit, retry %d/%d in %.1fms",
+                "Zhipu %d overload, retry %d/%d in %.1fms",
+                resp.status_code,
                 attempt + 1,
                 max_attempts - 1,
                 delay,
@@ -159,16 +168,16 @@ class ZhipuVendor(NativeAnthropicVendor):
 
         return resp  # pragma: no cover
 
-    # ── 流式：429 重试 ──────────────────────────────────────
+    # ── 流式：429/529 重试 ──────────────────────────────────
 
     async def send_message_stream(
         self,
         request_body: dict[str, Any],
         headers: dict[str, str],
     ) -> AsyncIterator[bytes]:
-        """流式请求，429 时自动重试.
+        """流式请求，429/529 时自动重试.
 
-        安全性：429 在 BaseVendor.send_message_stream 中于
+        安全性：429/529 在 BaseVendor.send_message_stream 中于
         status code 检查阶段即 raise（在任何 chunk yield 之前），
         因此重试不会导致已发出数据不一致。
         """
@@ -176,25 +185,33 @@ class ZhipuVendor(NativeAnthropicVendor):
 
         for attempt in range(max_attempts):
             try:
-                # 429 在 status code 检查阶段即 raise（在任何 chunk 之前），
+                # 429/529 在 status code 检查阶段即 raise（在任何 chunk 之前），
                 # 因此 __anext__ 安全：要么拿到首个 chunk，要么抛异常。
                 ait = super().send_message_stream(request_body, headers)
                 head = await ait.__anext__()
             except StopAsyncIteration:
                 return
             except httpx.HTTPStatusError as exc:
-                if exc.response is None or exc.response.status_code != 429:
+                if (
+                    exc.response is None
+                    or exc.response.status_code not in _BACKOFF_RETRY_STATUS
+                ):
                     raise
+                status_code = exc.response.status_code
                 if attempt == max_attempts - 1:
                     logger.warning(
-                        "Zhipu 429 stream rate limit exhausted after %d attempts",
+                        "Zhipu %d stream overload exhausted after %d attempts",
+                        status_code,
                         max_attempts,
                     )
                     raise
 
-                delay = self._compute_retry_delay_from_response(exc.response, attempt)
+                delay = self._compute_retry_delay_from_headers(
+                    exc.response.headers, attempt
+                )
                 logger.info(
-                    "Zhipu 429 stream rate limit, retry %d/%d in %.1fms",
+                    "Zhipu %d stream overload, retry %d/%d in %.1fms",
+                    status_code,
                     attempt + 1,
                     max_attempts - 1,
                     delay,
@@ -215,24 +232,15 @@ class ZhipuVendor(NativeAnthropicVendor):
         headers: dict[str, str] | None,
         attempt: int,
     ) -> float:
-        """计算重试延迟（毫秒），优先使用 server retry-after."""
-        rl_info = parse_rate_limit_headers(headers, 429, None)
-        server_delay_s = compute_effective_retry_seconds(rl_info)
-        if server_delay_s is not None:
-            return min(server_delay_s * 1000, self._rl_retry.max_delay_ms)
-        return calculate_delay(attempt, self._rl_retry)
+        """计算重试延迟（毫秒），优先使用 server retry-after.
 
-    def _compute_retry_delay_from_response(
-        self,
-        response: httpx.Response,
-        attempt: int,
-    ) -> float:
-        """计算重试延迟（毫秒），从 httpx.Response 提取 header."""
-        rl_info = parse_rate_limit_headers(
-            response.headers,
-            response.status_code,
-            response.text[:500] if response.text else None,
-        )
+        非流式与流式 429/529 重试共用此方法（单一延迟逻辑）。
+        以"限流退避"语义（429）解析 header：``parse_rate_limit_headers``
+        仅对 429/403 解析 retry-after，故此处固定传 429，
+        使 529 也能尊重 server retry-after，与 429 行为一致。
+        无 server 信号时回退到指数退避 + Full Jitter。
+        """
+        rl_info = parse_rate_limit_headers(headers, 429, None)
         server_delay_s = compute_effective_retry_seconds(rl_info)
         if server_delay_s is not None:
             return min(server_delay_s * 1000, self._rl_retry.max_delay_ms)