메뉴
Use when implementing server-side message and content-block compatibility for a Claude Code Anthropic gateway, including messages arrays, system blocks, tool_result pairing, image/document/thinking blocks, cache_control, and tolerant validation.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Use when implementing server-side tolerance for Claude Code context-management fields in an Anthropic gateway, including context_management edits, cache_edits blocks, cache references, prompt cache controls, and non-streaming compact requests.
Use when implementing server-side HTTP request handling for a Claude Code compatible Anthropic API gateway, especially /v1/messages, /v1/messages/count_tokens, /v1/files, /v1/models, headers, body fields, beta handling, and gateway routing.
Use when implementing server-side SSE streaming compatibility for a Claude Code Anthropic gateway, including message_start/content_block_delta/message_delta events, tool_use streaming, stream:false fallback, timeouts, and malformed stream behavior.
Use when implementing server-side tool schema and tool-call compatibility for a Claude Code Anthropic gateway, including tools arrays, tool_choice, tool_use SSE blocks, tool_result messages, deferred tools, MCP-style names, and tool parameter JSON.
Use when implementing a server-side Anthropic-compatible LLM gateway for Claude Code, including /v1/messages, SSE streaming, count_tokens, files, models, tool_use/tool_result, beta headers, and gateway error compatibility.
SOC 직업 분류 기준
| name | claude-code-api-conversation |
| description | Use when implementing server-side message and content-block compatibility for a Claude Code Anthropic gateway, including messages arrays, system blocks, tool_result pairing, image/document/thinking blocks, cache_control, and tolerant validation. |
根据本 skill，实现 /v1/messages 服务端对 messages、system 和 content blocks 的兼容解析。你要校验并转发 Anthropic 消息结构，不要实现 Claude Code 本地 transcript 存储。
详细字段见 references/message-wire-format.md。
请求中的 message 只有两类：
{ "role": "user", "content": "hello" }
{ "role": "assistant", "content": [{ "type": "text", "text": "hi" }] }
Claude Code 通常发送 content block array。网关应同时接受字符串和数组。
| Block | 方向 | 处理 |
|---|---|---|
text | user/assistant | 普通文本 |
tool_use | assistant history | 模型曾请求工具 |
tool_result | user | 工具执行结果 |
image | user | base64 图片 |
document | user | base64 PDF |
thinking | assistant history | 保留或忽略 |
redacted_thinking | assistant history | 保留或忽略 |
server_tool_use | assistant history | 保留或忽略 |
cache_edits | user | 不支持缓存时忽略 |
system 可能是字符串，也可能是：
[
{ "type": "text", "text": "...", "cache_control": { "type": "ephemeral", "scope": "org" } }
]
不支持 prompt cache 时忽略 cache_control，但保留 text。
如果历史中有 assistant tool_use，下一条或后续 user message 通常包含同 id 的 tool_result。服务端应尽量保留这条链路给上游模型。
最低要求：
tool_result 不要导致 500；可以忽略或转成文本提示。tool_result 的 tool_use 不要阻断整个请求；可以按历史内容保留。tool_result.tool_use_id 与 tool_use.id 匹配时，应作为工具结果传给上游。image.source 是 base64，media_type 常见 image/png、image/jpeg、image/gif、image/webp。document.source.media_type 常见 application/pdf。Claude Code 请求里可能携带 cache、tool search、advisor、caller 等扩展字段。外部网关不需要实现这些内部能力，但应尽量忽略未知字段，只在真正无法转换时返回 400。