Skip to content

fix: 修复缓存低命中与 Claude Code 兼容请求#292

Merged
KilimiaoSix merged 1 commit into
mainfrom
fix-cache-low
May 29, 2026
Merged

fix: 修复缓存低命中与 Claude Code 兼容请求#292
KilimiaoSix merged 1 commit into
mainfrom
fix-cache-low

Conversation

@KilimiaoSix
Copy link
Copy Markdown
Collaborator

Summary

  • 允许代理只保留 x-codex-turn-state 且仍带 prompt_cache_key 时继续复用本地 prompt cache 账号绑定,并补充 anchor_mode 诊断日志。
  • /v1/messages/count_tokens 提供本地 token 估算,避免 Claude Code 计数请求被误转发为真实 Responses 生成。
  • 针对 Anthropic-native 兼容路径增加 Cloudflare 缓解:禁用兼容流请求 zstd、challenge 冷却/失败切换、低额度账号尾部兜底、跳过仍在 challenge 冷却的候选。
  • 将 Claude Code 2.x mid-conversation system message 归一为 Responses developer role,避免 ChatGPT Codex 后端触发 Cloudflare challenge。

Validation

  • cargo fmt --check && cargo test -p codexmanager-service --lib
  • cargo test -p codexmanager-service --test gateway_logs anthropic
  • cargo test -p codexmanager-service --test gateway_logs prompt_cache
  • 已构建并部署测试镜像 registry.cn-hangzhou.aliyuncs.com/kilimiao/codex-manager:fix-cache-low-anthropic-system-role-20260529-amd6410.10.121.253codex-manager 容器。
  • 本机真实 Claude Code 连续请求验证通过;第二次请求返回 cache_read_input_tokens=27136 / input_tokens=30227,未再出现 Cloudflare 403。

@KilimiaoSix KilimiaoSix merged commit 5aee059 into main May 29, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant