Claude Code 2.1.172+ 自定义模型名时在 messages 中注入 role:"system" 消息，严格上游(Vertex/Bedrock类)拒绝导致重试熔断 503

[00010110]

问题描述

Claude Code 2.1.172 配合自定义模型名（通过 ANTHROPIC_DEFAULT_OPUS_MODEL 等指向 cchub 渠道模型，如 aws-claude-opus）使用时，请求经 cchub 转发到严格校验的上游（Vertex / Bedrock 类 Anthropic 兼容端点）会被拒绝。cchub 重试耗尽全部供应商后触发熔断，客户端最终收到：

{"statusCode": 503, "errorMessage": "所有供应商暂时不可用，请稍后重试"}

首次请求的流式响应为：

event: error
data: {"type": "error", "error": {"type": "api_error", "message": "Upstream returned an error event (request_id: ...)"}}

由于错误被包装成 api_error / 503，且重试导致熔断殃及后续请求，排查非常困难。

根因（已通过本地抓包确认）

Claude Code 2.1.172（2.1.153 之后引入的行为变化）会根据模型名是否为内置官方 Claude 模型 ID 切换 skills 列表的注入格式：

模型名	skills 注入方式
官方 ID（claude-opus-4-7 等）	user 消息内 <system-reminder> 文本块（旧行为）
任何非官方名（aws-claude-opus、甚至 claude-test-fake-9）	messages 数组中独立的 {"role": "system", "content": "..."} 消息

抓包验证（同一 CC 版本、本地 HTTP 捕获服务器）：

--model claude-opus-4-7   → messages roles: ['user']            ✅
--model aws-claude-opus   → messages roles: ['user', 'system']  ❌
--model claude-test-fake-9 → messages roles: ['user', 'system'] ❌

Anthropic 官方 API 接受 messages 中的 role: "system"，但 Vertex / Bedrock 等转换层严格校验 messages 角色只能为 user / assistant，直接返回 400 类错误（"Bad request - check request format and parameters"）。cchub 目前原样透传该消息，导致：

CC 2.1.172（自定义模型名）发出 role:"system" 消息
    → cchub 原样转发
    → 严格上游 400 拒绝
    → cchub 重试其余供应商全部失败
    → 熔断 → 503 "所有供应商暂时不可用"

由于同一模型可能挂多个渠道（宽容的渠道能成功），故障表现为间歇性，进一步增加排查难度。

影响范围

所有满足以下条件的用户：

• Claude Code ≥ 2.1.172（2.1.153 确认无此行为，具体引入版本未逐版定位）
• 渠道模型名为自定义命名（如 aws-claude-opus、qwen-xx、glm-xx、MiniMax-xx 等，这是 cchub 用户的常见用法）
• 上游为严格校验 messages 角色的 Anthropic 兼容端点

复现步骤

# 对任意 Vertex/Bedrock 类严格上游渠道发送：
curl -s "https://<cchub>/v1/messages" \
  -H "x-api-key: $KEY" -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "<自定义模型名>", "max_tokens": 100,
    "messages": [
      {"role": "user", "content": "hi"},
      {"role": "system", "content": "The following skills are available: ..."}
    ]
  }'
# → 400/503；去掉 role:"system" 那条消息后同样请求成功

或直接用 Claude Code 2.1.172 + ANTHROPIC_MODEL=<自定义模型名> 发送任意消息（只要本地启用了任意 skill/plugin 就会注入该消息）。

建议修复

参照现有 src/app/v1/_lib/proxy/billing-header-rectifier.ts 的模式（同样是处理 Claude Code 新版本注入的、上游不兼容的内容），新增一个 proactive rectifier：

• 在转发前扫描 messages，将 role: "system" 的消息取出
• 其内容合并到顶层 system 数组末尾（string content 转为 {type: "text", text} block）
• 语义上与 Anthropic 官方对该写法的处理一致，对宽容上游无副作用，对严格上游消除 400

如方案可接受，我可以提交 PR。

环境

• claude-code-hub: v0.8.5（self-hosted）
• Claude Code: 2.1.172（user-agent: claude-cli/2.1.172 (external, cli)）
• 对照版本: Claude Code 2.1.153 无此问题

[coderabbitai]

🔗 Related PRs

#517 - feat: 新增 /v1/models 端点，聚合返回用户可用模型列表 [merged]
#693 - fix: 修复 buildProxyUrl 重复拼接版本前缀的问题（Gemini） [merged]
#759 - fix(proxy): 修复请求卡死（AgentPool 驱逐阻塞） [merged]
#961 - fix: 模型列表端点返回所有可用供应商的模型而非仅第一个 [merged]
#1223 - feat(anthropic): detect actual model from thinking signature in stream [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[github-actions]

Thank you for this detailed and well-researched bug report.

Analysis

Your root cause analysis is accurate. I've verified through the codebase that:

1. No existing logic handles role: "system" messages in the messages array when forwarding to Claude-compatible upstreams (Vertex/Bedrock). The only conversion exists in the Gemini adapter (src/app/v1/_lib/gemini/adapter.ts:111), which extracts role: "system" messages into systemInstruction for Gemini providers — but this does not apply to Claude-type providers.

2. The request is forwarded as-is to Claude-type providers. The forwarder (src/app/v1/_lib/proxy/forwarder.ts) applies proactive rectifiers (billing header at line ~2451, cache TTL at line ~2517) but nothing transforms the messages array itself for Claude providers.

3. Format detection (src/app/v1/_lib/proxy/format-mapper.ts:131) already acknowledges that Claude messages may contain role: "system" entries, but no rectification occurs post-detection.

Potentially affected code

• src/app/v1/_lib/proxy/forwarder.ts (~lines 2449-2511) — where proactive rectifiers are applied for Claude providers. This is the integration point for the suggested fix.
• src/app/v1/_lib/proxy/billing-header-rectifier.ts — the existing rectifier pattern (proactive, in-place mutation, result object with applied/removedCount/extractedValues) that the suggested fix should follow.
• src/app/v1/_lib/gemini/adapter.ts:111 — reference implementation for extracting role: "system" messages from the array and merging them into a top-level system prompt.

Assessment of suggested fix

The proposed approach is architecturally sound and consistent with the existing rectifier pattern:

• A new proactive rectifier that extracts role: "system" entries from the messages array and merges their content into the top-level system field (appending as {type: "text", text} blocks)
• Integrated in forwarder.ts alongside rectifyBillingHeader, enabled by default with a system settings toggle
• Semantically equivalent to how Anthropic's official API interprets role: "system" messages — no behavior change for permissive upstreams, eliminates 400 errors for strict upstreams

Your PR targeting the dev branch would be welcome. The existing billing-header-rectifier provides a clear template for implementation and testing.

---

Automated response from Claude AI