-
供应商状态
+
供应商状态
@@ -1396,7 +1410,7 @@ def line(x0: float, y0: float, x1: float, y1: float) -> None:
if (tier.weekly_quota_guard) quotaHTML += renderQuotaBar(tier.weekly_quota_guard);
const rlInfo = tier.rate_limit || {};
- const rlHtml = rlInfo.limited ? `
限速中` : '';
+ const rlHtml = rlInfo.is_rate_limited ? `
限速中` : '';
// Pointer Events 重排：仅需 data-vendor 定位；手柄作为拖拽发起点（无原生 draggable）
const dragAttrs = ` data-vendor="${tier.name}"`;
@@ -1556,6 +1570,34 @@ def line(x0: float, y0: float, x1: float, y1: float) -> None:
}).catch(function() {});
}
+// ── 供应商状态：一键复位（熔断 / 限流 / 配额守卫状态；不动额度用量与优先级）────
+var _vendorResetInFlight = false;
+function resetVendorStatus(btn) {
+ if (_vendorResetInFlight) return; // 防并发，同 persistTierOrder 范式
+ _vendorResetInFlight = true;
+ var label = btn.textContent;
+ btn.disabled = true;
+ btn.textContent = '复位中…';
+ // 无 body → 服务端跳过重排序，仅对全部 tier 复位弹性设施状态
+ fetch('/api/reset', { method: 'POST' }).then(function(res) {
+ if (!res.ok) throw new Error(res.status);
+ btn.textContent = '✓ 已复位';
+ // 复位已生效，状态刷新失败不应误报失败、诱导重复点击 → 单独吞掉
+ return fetchJSON('/api/status').then(function(status) {
+ updateVendorStatus(status); // 定向重渲染，不等 10 分钟轮询
+ }).catch(function() {});
+ }).catch(function(e) {
+ console.error('vendor reset failed:', e);
+ btn.textContent = '✗ 失败';
+ }).finally(function() {
+ setTimeout(function() {
+ btn.textContent = label;
+ btn.disabled = false;
+ _vendorResetInFlight = false;
+ }, 1500);
+ });
+}
+
// ── Model Calling 实时状态 ────────────────────────────────
function updateModelCalling(status) {
var wrap = document.getElementById('model-calling-wrap');
diff --git a/src/coding/proxy/vendors/antigravity.py b/src/coding/proxy/vendors/antigravity.py
index b4d7199..c2b261b 100644
--- a/src/coding/proxy/vendors/antigravity.py
+++ b/src/coding/proxy/vendors/antigravity.py
@@ -24,6 +24,7 @@
VendorCapabilities,
VendorResponse,
_sanitize_headers_for_synthetic_response,
+ decode_error_body,
)
# GoogleOAuthTokenManager 已从 antigravity_token_manager.py 合并至本文件末尾
@@ -532,14 +533,16 @@ async def send_message_stream(
if response.status_code >= 400:
self._on_error_status(response.status_code)
error_body = await response.aread()
+ # 检测用完整解码文本(判定性逻辑不应受日志展示上限影响);
+ # 日志展示则用 decode_error_body 的 500 字符截断版本。
self._mark_scope_error_if_needed(
- error_body.decode("utf-8", errors="ignore"),
+ error_body.decode("utf-8", errors="replace"),
)
logger.warning(
"%s stream error: status=%d body=%s",
self.get_name(),
response.status_code,
- error_body[:500],
+ decode_error_body(error_body),
)
raise httpx.HTTPStatusError(
f"{self.get_name()} API error: {response.status_code}",
diff --git a/src/coding/proxy/vendors/base.py b/src/coding/proxy/vendors/base.py
index d1434bc..1c4c9c3 100644
--- a/src/coding/proxy/vendors/base.py
+++ b/src/coding/proxy/vendors/base.py
@@ -25,6 +25,7 @@
UsageInfo,
VendorCapabilities,
VendorResponse,
+ decode_error_body,
decode_json_body,
extract_error_message,
sanitize_headers_for_synthetic_response,
@@ -328,7 +329,7 @@ async def send_message_stream(
"%s stream error: status=%d body=%s",
self.get_name(),
response.status_code,
- error_body[:500],
+ decode_error_body(error_body),
)
raise httpx.HTTPStatusError(
f"{self.get_name()} API error: {response.status_code}",
diff --git a/src/coding/proxy/vendors/copilot.py b/src/coding/proxy/vendors/copilot.py
index b159976..a8e7f0f 100644
--- a/src/coding/proxy/vendors/copilot.py
+++ b/src/coding/proxy/vendors/copilot.py
@@ -25,6 +25,7 @@
VendorResponse,
_decode_json_body,
_extract_error_message,
+ decode_error_body,
)
from .copilot_models import ( # noqa: F401
CopilotMisdirectedRequest,
@@ -423,7 +424,7 @@ async def _stream_from_client(
"%s stream error: status=%d body=%s",
self.get_name(),
response.status_code,
- error_body[:500],
+ decode_error_body(error_body),
)
raise httpx.HTTPStatusError(
f"{self.get_name()} API error: {response.status_code}",
diff --git a/tests/test_antigravity.py b/tests/test_antigravity.py
index cc93127..675d995 100644
--- a/tests/test_antigravity.py
+++ b/tests/test_antigravity.py
@@ -275,6 +275,51 @@ def test_mark_scope_error_if_needed():
assert diagnostics["token_manager"]["error_kind"] == "insufficient_scope"
+@pytest.mark.asyncio
+async def test_stream_scope_detection_scans_full_body_not_log_truncation():
+ """流式 scope 检测必须扫描完整 body,而非日志展示用的 500 字符截断.
+
+ 回归守护:``decode_error_body`` 的 ``[:500]`` 截断仅服务日志展示;若把它
+ 同时喂给 ``_mark_scope_error_if_needed`` 的子串检测,则标记出现在 500 字符
+ 之后的错误体会漏检,token 不会被标记 ``needs_reauth``。本用例构造标记位于
+ 500 字符之后的 403 错误体,断言检测仍然触发(旧的截断实现下会失败)。
+ """
+ marker = "ACCESS_TOKEN_SCOPE_INSUFFICIENT"
+ # message 填充 600 字符,把 details 中的 marker 挤到第 500 字符之后
+ body = (
+ f'{{"error":{{"message":"{"x" * 600}","status":"PERMISSION_DENIED",'
+ f'"details":[{{"reason":"{marker}"}}]}}}}'
+ ).encode()
+ assert body.decode().index(marker) > 500 # 前置条件:marker 确在 500 之后
+
+ vendor = AntigravityVendor(AntigravityConfig(), FailoverConfig(), ModelMapper([]))
+ vendor._token_manager.get_token = AsyncMock(return_value="tok")
+ vendor._discover_project_id = AsyncMock(return_value="") # 避免真实网络发现
+ vendor._client = httpx.AsyncClient(
+ base_url=vendor._base_url,
+ transport=httpx.MockTransport(
+ lambda _req: httpx.Response(
+ 403, content=body, headers={"content-type": "application/json"}
+ )
+ ),
+ )
+
+ with pytest.raises(httpx.HTTPStatusError):
+ async for _ in vendor.send_message_stream(
+ {
+ "model": "claude-sonnet-4-20250514",
+ "messages": [{"role": "user", "content": "Hi"}],
+ },
+ {},
+ ):
+ pass
+
+ await vendor.close()
+
+ diagnostics = vendor.get_diagnostics()
+ assert diagnostics["token_manager"]["error_kind"] == "insufficient_scope"
+
+
def test_antigravity_supports_request_with_tools_thinking_and_metadata():
vendor = AntigravityVendor(AntigravityConfig(), FailoverConfig(), ModelMapper([]))
supported, reasons = vendor.supports_request(
diff --git a/tests/test_app_routes.py b/tests/test_app_routes.py
index 5c0aa53..e059d65 100644
--- a/tests/test_app_routes.py
+++ b/tests/test_app_routes.py
@@ -1067,6 +1067,52 @@ def test_reset_reorder_also_resets_circuit_breaker_and_rate_limit():
assert not anthropic_tier.is_rate_limited
+def test_reset_preserves_quota_usage():
+ """核心：/api/reset 复位守卫状态但**不**清空滑动窗口用量.
+
+ 旧实现在 QuotaGuard.reset() 里连带 `_entries.clear()` + `_total = 0`，
+ 使 Dashboard 的配额百分比归零且无法自愈（基线仅在进程启动时回填）。
+ """
+ app = _make_reorder_app()
+ router = app.state.router
+ tier = router.tiers[0]
+ qg = tier.quota_guard
+ assert qg is not None
+ qg._enabled = True
+ qg._budget = 100_000
+ qg.record_usage(40_000)
+ qg.notify_cap_error() # 制造 cap 卡死态
+ assert qg.get_info()["state"] == "quota_exceeded"
+
+ with TestClient(app) as client:
+ resp = client.post("/api/reset")
+ assert resp.status_code == 200
+
+ after = qg.get_info()
+ assert after["state"] == "within_quota" # 状态已复位
+ assert after["window_usage_tokens"] == 40_000 # 用量原样保留
+ assert after["usage_percent"] == 40.0
+ assert qg.can_use_primary() is True # cap 卡死标志已清除
+
+
+def test_reset_preserves_quota_usage_with_reorder():
+ """带 -v 重排序的 /api/reset 同样不得清空用量（CLI `coding reset -v ...` 路径）."""
+ app = _make_reorder_app()
+ qg = app.state.router.tiers[0].quota_guard
+ assert qg is not None
+ qg._enabled = True
+ qg._budget = 100_000
+ qg.record_usage(45_000)
+
+ with TestClient(app) as client:
+ resp = client.post(
+ "/api/reset", json={"vendors": ["anthropic", "zhipu", "copilot"]}
+ )
+ assert resp.status_code == 200
+
+ assert qg.get_info()["window_usage_tokens"] == 45_000
+
+
def test_reorder_tiers_shared_reference():
"""验证 reorder_tiers 使用切片赋值，Executor 立即可见."""
from coding.proxy.routing.router import RequestRouter
@@ -1235,3 +1281,32 @@ def test_dashboard_serves_pointer_drag_reorder_and_no_cache():
# 已移除脆弱的原生 HTML5 DnD（不再渲染 draggable 属性 / 监听 dragstart）
assert 'draggable="true"' not in html
assert "addEventListener('dragstart'" not in html
+
+
+def test_dashboard_serves_vendor_status_reset_control():
+ """供应商状态卡片标题栏交付「状态复位」按钮，且走无 body 的 POST /api/reset.
+
+ 无 body 是语义关键：服务端据此跳过重排序（routes.py 的 vendor_names 守卫），
+ 因此复位不会改动供应商优先级。
+ """
+ with _make_app() as client:
+ html = client.get("/dashboard").text
+ # 按钮随卡片标题栏交付
+ assert 'id="btn-vendor-reset"' in html
+ assert "btn-card-action" in html
+ assert "resetVendorStatus(this)" in html
+ # 处理器在位，且为无 body 的 POST（不触发重排序）
+ assert "function resetVendorStatus(btn)" in html
+ assert "fetch('/api/reset', { method: 'POST' })" in html
+ assert "_vendorResetInFlight" in html # 防并发守卫
+
+
+def test_dashboard_rate_limit_badge_reads_correct_key():
+ """「限速中」徽章须读后端真实键名 is_rate_limited（tier.get_rate_limit_info）.
+
+ 旧实现读 rlInfo.limited —— 后端从未产出该键，徽章永远不渲染。
+ """
+ with _make_app() as client:
+ html = client.get("/dashboard").text
+ assert "rlInfo.is_rate_limited" in html
+ assert "rlInfo.limited " not in html
diff --git a/tests/test_quota_guard.py b/tests/test_quota_guard.py
index 9d3ed9a..a2b6df0 100644
--- a/tests/test_quota_guard.py
+++ b/tests/test_quota_guard.py
@@ -90,14 +90,80 @@ def test_probe_success_restores_within_quota():
assert qg.get_info()["state"] == "within_quota"
-def test_reset_clears_all_state():
+def test_reset_preserves_window_usage():
+ """复位只清状态机，滑动窗口用量计数须原样保留."""
qg = _make_guard()
qg.record_usage(500)
qg.notify_cap_error()
qg.reset()
info = qg.get_info()
assert info["state"] == "within_quota"
- assert info["window_usage_tokens"] == 0
+ assert info["window_usage_tokens"] == 500
+
+
+def test_reset_clears_cap_error_stall():
+ """cap 错误卡死态：复位清除 _cap_error_active，预算未超时立即恢复放行."""
+ qg = _make_guard(token_budget=1000)
+ qg.record_usage(100)
+ qg.notify_cap_error()
+ assert qg.can_use_primary() is False
+ qg.reset()
+ assert qg.can_use_primary() is True
+ assert qg.get_info()["window_usage_tokens"] == 100
+
+
+def test_reset_does_not_unblock_genuinely_exhausted_quota():
+ """用量确实超阈值时，复位不放行也不伪造用量——保持 EXCEEDED，不做状态回环."""
+ qg = _make_guard(token_budget=1000, threshold_percent=99.0)
+ qg.record_usage(995)
+ assert qg.can_use_primary() is False
+ qg.reset()
+ assert qg.get_info()["state"] == "quota_exceeded"
+ assert qg.can_use_primary() is False
+ assert qg.get_info()["window_usage_tokens"] == 995
+
+
+def test_reset_does_not_delay_probe_when_over_budget():
+ """回归：用量仍超阈值时复位不得推后探测时钟.
+
+ 若 reset() 走 WITHIN_QUOTA → 下一次判定再回落 QUOTA_EXCEEDED 的回环，
+ `_transition_to(EXCEEDED)` 会把 `_last_probe` 刷成当前时刻，令探测恢复
+ 凭空推迟一个 probe_interval；反复点击「状态复位」即可无限饿死探测。
+ """
+ qg = _make_guard(
+ token_budget=1000, threshold_percent=99.0, probe_interval_seconds=300
+ )
+ base = time.monotonic()
+ with patch("coding.proxy.routing.quota_guard.time") as mock_time:
+ mock_time.monotonic.return_value = base
+ qg.record_usage(995)
+ assert qg.can_use_primary() is False # → EXCEEDED，_last_probe = base
+
+ # 距下次探测仅剩 10s 时复位
+ mock_time.monotonic.return_value = base + 290
+ qg.reset()
+
+ # 原定探测点（base+300）之后应照常放行探测，而非被推到 base+590
+ mock_time.monotonic.return_value = base + 301
+ assert qg.can_use_primary() is True
+
+
+def test_reset_still_clears_cap_stall_when_over_budget():
+ """用量超阈值但由 cap 错误拉长了探测间隔时，复位须还原 probe_interval."""
+ qg = _make_guard(
+ token_budget=1000, threshold_percent=99.0, probe_interval_seconds=300
+ )
+ base = time.monotonic()
+ with patch("coding.proxy.routing.quota_guard.time") as mock_time:
+ mock_time.monotonic.return_value = base
+ qg.record_usage(995)
+ qg.notify_cap_error(retry_after_seconds=3600) # 探测间隔被拉到 3960s
+ assert qg.can_use_primary() is False
+
+ qg.reset()
+ mock_time.monotonic.return_value = base + 301
+ # 若 _effective_probe_interval 未还原为 300，此处仍会被拒
+ assert qg.can_use_primary() is True
def test_get_info_returns_correct_data():
diff --git a/tests/test_vendor_helpers.py b/tests/test_vendor_helpers.py
new file mode 100644
index 0000000..322da7a
--- /dev/null
+++ b/tests/test_vendor_helpers.py
@@ -0,0 +1,107 @@
+"""供应商工具函数测试 — 聚焦 decode_error_body 的编码正确性与流式错误日志回归.
+
+背景：上游 4xx/5xx 错误的流式日志曾将 ``bytes`` 响应体经 ``%s`` 直接格式化，
+Python 对 ``bytes`` 走 ``repr()``，导致非 ASCII 的 UTF-8 字节被转义为
+``\\xe6\\x82\\xa8`` 之类的不可读序列（中文乱码）。本测试锚定该根因并守护修复。
+"""
+
+import logging
+
+import httpx
+import pytest
+
+from coding.proxy.config.schema import AnthropicConfig, FailoverConfig
+from coding.proxy.model.vendor import decode_error_body
+from coding.proxy.vendors.anthropic import AnthropicVendor
+
+# ── 单元测试：decode_error_body ──────────────────────────────
+
+_ZH_TEXT = "您的账户已达到速率限制,请您控制请求频率"
+
+
+class TestDecodeErrorBody:
+ def test_chinese_bytes_decoded_readable(self):
+ """中文 UTF-8 bytes 应解码为可读字符,不残留转义序列或 bytes 前缀."""
+ raw = f'{{"message":"[1302][{_ZH_TEXT}]"}}'.encode()
+ out = decode_error_body(raw)
+ assert _ZH_TEXT in out
+ assert "\\x" not in out # 无字节转义
+ assert not out.startswith("b'") # 非 bytes repr
+ assert isinstance(out, str)
+
+ def test_truncates_by_characters(self):
+ """limit 语义为字符数;超长输入按字符截断."""
+ raw = ("你" * 1000).encode()
+ out = decode_error_body(raw, limit=100)
+ assert len(out) == 100
+ assert out == "你" * 100 # 未在多字节边界切断
+
+ def test_invalid_bytes_do_not_raise(self):
+ """非法字节以 errors='replace' 降级为占位符,绝不抛异常."""
+ raw = b"\xff\xfe invalid"
+ out = decode_error_body(raw)
+ assert isinstance(out, str)
+ assert "�" in out # U+FFFD replacement character
+
+ def test_empty_bytes(self):
+ assert decode_error_body(b"") == ""
+
+ def test_ascii_passthrough(self):
+ raw = b'{"error":{"type":"rate_limit_error"}}'
+ assert decode_error_body(raw) == '{"error":{"type":"rate_limit_error"}}'
+
+ def test_regression_repr_vs_decode(self):
+ """回归对照:复现旧 bug(%s 对 bytes 走 repr)并证明新实现修复之."""
+ raw = _ZH_TEXT.encode()
+ # 旧行为:bytes 经 %s → repr → 转义字节序列(刻意保留 % 格式化以精确
+ # 复现旧 logger.warning("...%s...", error_body) 的缺陷路径,故 noqa UP031)
+ assert "\\x" in "%s" % raw # noqa: UP031
+ # 新行为:先解码 → 可读中文,无转义
+ assert "\\x" not in decode_error_body(raw)
+
+
+# ── 集成测试:BaseVendor 流式错误日志路径 ────────────────────
+
+
+@pytest.mark.asyncio
+async def test_stream_error_log_decodes_chinese(caplog):
+ """经 BaseVendor.send_message_stream 的流式错误日志应输出可读中文.
+
+ 以 AnthropicVendor(纯继承 BaseVendor,无重试/包装)具象化,
+ 注入挂载 MockTransport 的 AsyncClient 返回 400 + 含中文的 body,
+ 断言 WARNING 日志文本含中文且不含字节转义序列。
+ """
+ error_payload = f'{{"error":{{"message":"{_ZH_TEXT}"}}}}'.encode()
+
+ def _handler(request: httpx.Request) -> httpx.Response:
+ return httpx.Response(
+ 400,
+ content=error_payload,
+ headers={"content-type": "application/json; charset=utf-8"},
+ )
+
+ vendor = AnthropicVendor(AnthropicConfig(), FailoverConfig())
+ # 注入 MockTransport 客户端(base_url 与真实一致,仅替换 transport)
+ vendor._client = httpx.AsyncClient(
+ base_url=vendor._base_url,
+ transport=httpx.MockTransport(_handler),
+ )
+
+ with caplog.at_level(logging.WARNING):
+ with pytest.raises(httpx.HTTPStatusError):
+ async for _ in vendor.send_message_stream(
+ {"model": "claude-opus-4-6", "messages": []},
+ {"authorization": "Bearer sk-test"},
+ ):
+ pass
+
+ await vendor.close()
+
+ stream_errors = [
+ r.getMessage() for r in caplog.records if "stream error" in r.getMessage()
+ ]
+ assert stream_errors, "未捕获到 stream error 日志"
+ msg = stream_errors[0]
+ assert _ZH_TEXT in msg # 中文可读
+ assert "\\x" not in msg # 无字节转义
+ assert "body=b'" not in msg # 非 bytes repr
diff --git a/uv.lock b/uv.lock
index b9197bf..ca13d2b 100644
--- a/uv.lock
+++ b/uv.lock
@@ -74,7 +74,7 @@ wheels = [
[[package]]
name = "coding-proxy"
-version = "0.5.2a5"
+version = "0.5.2a8"
source = { editable = "." }
dependencies = [
{ name = "aiosqlite" },