mirror of
https://wget.la/https://github.com/leookun/cursor-byok
synced 2026-08-17 03:27:02 +08:00
kael-odin
50ab63de3d
AnthropicAdapter.Stream built thinking config (buildAnthropicThinkingConfig)
and wrote it into body only inside the `if len(body)==0` normal-construction
block. The RequestBodyOverride branch skipped it entirely — a disabled effort
on the override path left whatever thinking config the override body carried,
violating user intent and diverging from openai.go, where
applyOpenAIThinkingDisable runs unconditionally after both branches.
Also, on the normal path, disabled only wrote thinking:{type:disabled} but
left a stale output_config (set by a prior adaptive turn or by
AnthropicExtraParams) in place — an explicitly disabled request could still
carry output_config.effort=high, a contradictory payload.
Fix:
- Extract applyAnthropicThinkingConfig(body, req), called unconditionally
after the override/normal block, mirroring openai.go:1860.
- disabled: force thinking:{type:"disabled"}, delete output_config, set
thinking_disabled_provider_param=thinking.type knob.
- adaptive (AnthropicThinkingEffort non-empty): write
thinking:{type:adaptive,display:summarized} + output_config.
- empty effort: no-op.
- buildAnthropicThinkingConfig retained for stable-message-count / message
normalization signals inside the normal block.
Tests: anthropic_thinking_disable_test.go covers all three branches, alias
normalization (off→disabled), disabled overriding existing adaptive config,
and the override-path symmetry scenario.
Backend 架构说明
internal/backend 当前支持本地助手模式与直连上游模式。
关于 backend agent「最小事实集合」的第一阶段研究文档，见 ../../docs/backend-agent-minimum-facts-phase1.md。
核心边界：
server- 本地 HTTP/Connect 入口层
- 负责路由、中间件、错误编码和少量本地 mock
forwarder- 本地协议兼容与 LLM 转发内核
- 负责
BidiAppend、RunSSE、history JSON、prompt 编译、provider 流式调用和广播
host- 唯一组装点
- 负责把
server/config.Manager、forwarder.Module和根路由装起来
当前实现不再支持：
- Pro /
cursor-byok - HTTP/protocol trace debug UI
- DB-backed store、会话索引和 searchable conversation memory
目录结构
internal/backend/
README.md
host.go
server/
context.go
errors.go
local.go
middleware.go
policy.go
route.go
url.go
config/
manager.go
store.go
types.go
legacy_runtime.go
resolver.go
upstream/
action.go
mocks.go
types.go
forwarder/
artifacts.go
broker.go
compiler.go
events.go
file_store.go
legacy_stream.go
module.go
projector.go
provider.go
reminders.go
service.go
tool_catalog.go
types.go
agent/
bridge/
exec/
bridge.go
interaction/
bridge.go
core/
types.go
model/
router.go
openai.go
anthropic.go
artifacts.go
provider_limits.go
http_error.go
types.go
prompt/
engine.go
replay.go
protocol/
inbound.go
持久化布局
助手目录固定为：
~/.cursor-local-assistant-v2/config.yaml~/.cursor-local-assistant-v2/data/ca.crt~/.cursor-local-assistant-v2/data/ads/~/.cursor-local-assistant-v2/history/~/.cursor-local-assistant-v2/logs/
约定：
config.yaml是用户配置data/ca.crt是注入给宿主的 CA 证书data/ads/是广告包与资源缓存目录history/是会话事实与全局 usage JSON 目录，不属于日志logs/只保留必要文本运行日志
当前 history/ 目录布局：
history/
usage.json
<conversation_id>/
state.json
context.json
conversation.lock
state.json 只表达当前 loop 状态与持久化内存，不保存可投射给 LLM 的历史内容。当前 loop status 语义为：
idle：没有正在推进的 loop。running：已落入本轮输入或中间上下文，正在等待/发起模型推进。waiting_tool：已落完整 tool call，正在等待工具结果。completed：本轮已正常完成。canceled：本轮被取消，不制造 assistant 输出。provider_error：provider/LLM 调用失败，错误作为 context tag 记录。failed：本地内部失败，例如投影、持久化、usage JSON 写入或桥接收口失败；它不等同于 provider 错误。
请求流
- 请求进入 backend 根路由。
PolicyMiddleware根据routing.mode与X-Server-Upstream-URL选择本地或上游分支。BidiAppend/RunSSE进入forwarder。forwarder先把当前 loop 状态写入state.json，再把已发生语义事件追加到context.json。- 发给 LLM 的 prompt 只由
context.json投射生成；state.json不保存可投射历史。 - provider usage/cache 与聚合统计写入
history/usage.json，不从 conversation 文件现场扫描。 checkpoint只表示同一 backend 进程内的 live state。
模型渠道
- 用户在配置里填写
displayName、baseURL、apiKey、modelID - 运行时渠道唯一 ID 不再由
modelID决定 - 当前唯一 ID 是
url + modelID + key + name的短SHA-256hash（前 16 个十六进制字符） modelID仅表示 provider model