AskaEth
|
0de68bc960
|
fix(model): DeepSeek 推理参数语义反转 + thinking 开关真发(按官方文档核对结果)
CI / Backend (build / vet / test / cross-compile) (push) Has been cancelled
CI / Frontend (tsc + vite build) (push) Has been cancelled
依据 DevDocs/design/04 §3.2「官方核对后的修正」9 条 + 本地 DevDocs/reference/deepseek/facts-for-cyrene.md(官方原句/URL)。
- 🔴 reasoning_effort 白名单改为官方集合 [none, low, high, max] + 别名映射(minimal→low, medium→high, xhigh→high, ultra→max);**删除就近降级**,非法值一律丢弃+WARN。此前 none(关思考) 被降级成 low(开思考) 属语义反转
- 🔴 DeepSeek thinking 开关真发顶层 {"thinking":{"type":"enabled|disabled"}}(官方默认 enabled,不传≠关);此前只打 WARN 导致 mode:off 静默无效
- DeepSeek 的 /anthropic 端点忽略 budget_tokens → 不再对其抬 max_tokens(会误导预算/日志),改为按官方用 output_config.effort
- 注释纠正:penalty 属接口级 deprecated(非仅思考模式);top_p 仅思考模式生效且下限 0.95(不加该字段)
- tool_choice 雷登记:思考模式下 required/指定 tool 会 400(思考默认开)
- 旧模型名清理(deepseek-chat/reasoner 已于 2026-07-24 停用)
- 请求体级测试 + 变异验证(把 none 改回 low / 把 thinking 改回只 WARN → 对应断言 FAIL)
|
2026-10-03 20:22:31 +08:00 |
|
AskaEth
|
38a403afca
|
feat(persona,model,loop): 人格 thinking 配置 + 5 个死字段接线 + 推理旋钮(按 provider 映射)
CI / Backend (build / vet / test / cross-compile) (push) Has been cancelled
CI / Frontend (tsc + vite build) (push) Has been cancelled
按 DevDocs/design/01 §2.1/§2.2 与 design/04 §3.2 实现(规格含联网核过的官方依据:思考内容刻意不角色化、不写分步思考指令、思考控制只走参数级)。
- 人格 schema 新增 thinking:budgetTokens / effort / showToUser(*bool,省略=nil=保持现行为,显式 false=只落库) / mode(default|on|off→Enable 三态);非法 mode/负预算 → ErrInvalidThinking(不静默)
- 5 个死字段接上注入:personality→values→speechStyle→catchphrase→background 拼为独立段落(带【】小标题,空值不拼);**全空时输出与旧实现逐字节一致**(有专门测试)
- model.Request 新增 Thinking{ BudgetTokens, Enable } 与 ReasoningEffort;新增 model/reasoning.go:厂商识别(模型名→baseURL 兜底)、取值白名单(OpenAI 系 none…max / DeepSeek low|high|max)、非法值降级+WARN、**不支持一律 WARN 而非静默**;Anthropic 映射 thinking{budget_tokens} 并**保证 budget<max_tokens**(不足时抬高 max_tokens+WARN,正是历史上思考吃光预算→空回复那个坑)
- loop:新增 Options.ThinkingShowToUser(*bool) 仅门控 thinking_* 帧(正文/工具帧不受影响);场景级 mute 语义不变
- 补一处任务清单外但必需的接线:persona→model.Request 的 thinking 映射(否则 budget/effort/mode 会变成新的死字段);showToUser 刻意不产生 Thinking(纯呈现开关)
- 管理端:personaUpsertRequest 补 thinking DTO + 校验(400);PUT 省略时**保留已存块**(防旧客户端抹配置);写侧 yaml round-trip 已验证
- Docs/api-client v0.40→v0.41(可选请求字段)
- 测试:26 个新用例(loader 解析/拼接逐字节/映射、anthropic 非流式+流式预算约束、openai 透传/降级/丢弃、Qwen 诚实 WARN、loop 门控、DTO round-trip);3 次变异验证(省略与 false 混同 / 去掉预算约束 / 短路门控 各自 FAIL)
- 验收:go build/vet/test 全绿(含 57 包)+ linux/darwin 交叉编译 + 前端 build;唯一 FAIL 是既有 flake TestInterruptCancelsTurnAndClearsPending(3/20 新代码 vs 3/20 纯 HEAD,已归因非本次引入)
|
2026-10-03 19:54:22 +08:00 |
|
AskaEth
|
c7b2cd8ae0
|
fix(model): 解析 stop_reason + max_tokens 截断重试/非空兜底 — 不再静默空回复
CI / Backend (build / vet / test / cross-compile) (push) Has been cancelled
CI / Frontend (tsc + vite build) (push) Has been cancelled
- model.go: Response.StopReason; anthropic wire加stop_reason; openai finish_reason映射(两协议统一,原值透传)
- truncation.go(新,159行): effectiveMaxTokens/retryMaxTokens(×2封顶32768)/isTruncatedEmpty(空正文+无tool_use+stop_reason∈{max_tokens,length})/withEmptyBodyHint/callWithTruncationRetry/callStreamWithTruncationRetry/dropThinking/emitTextSegment
- anthropic.go Chat/ChatStream 走重试+兜底; anthropic_encode.go(新)拆编码(anthropic.go 529→357合规); anthropic_stream message_delta落StopReason
- agent/loop: 把模型层StopReason带到llm.call span的model_stop_reason属性(修掉trace硬编码stop的误导); loop.Result.StopReason语义未动
- 触发仅当截断空正文; tool_use空正文永不触发(native_toolcall_test仍过); 至多1次重试; 兜底回'(思考过长,未能生成回复)'绝不用thinking
- 新增7测试(非流式/流式重试+兜底+tool_use不触发+封顶+openai映射), build/vet/test绿+3平台交叉编译
|
2026-09-08 23:52:47 +08:00 |
|
Aska
|
7e39153a69
|
feat: 修复审计落差 #1/#2/#3/#6 + 管理面板移动端适配
CI / Backend (build / vet / test / cross-compile) (push) Has been cancelled
CI / Frontend (tsc + vite build) (push) Has been cancelled
#1 工具布尔标记生效(agent-106, 安全红线): loop toolSchemas/runToolCalls 加 allow-set 安全门 restrictSchemas/applyToolSecurityGate——群聊/陌生人只暴露允许工具(默认关), 被拒not_allowed软提示+AuditEvent, owner全工具不变; 身份靠 ActorEnv(非提示词)
#2 checkpoint 恢复(agent-109): run.go LoadLatest + resumableStopReason门控(只恢复in-flight工具循环,不出答案不续跑), 幂等不重放副作用, 新会话不误恢复; resume_test 4用例
#3 tool_result(agent-108, 协议层): Anthropic 原生tool_result块(tool_use_id/is_error)非user纯文本 + OpenAI tool消息补tool_call_id; Content加ToolCallID/IsError等; (loop接线层2处留待,见备注)
#6 capacity_awaken推送(agent-107): onCapacity从只Debug改推WS hub(BroadcastCapacityAwaken, sessionId定向+payload对齐前端AwakenPayload)
管理面板移动端(agent-110): AdminApp 汉堡菜单+抽屉侧栏+标题truncate防换行, RBAC守卫不受影响
全仓 build/vet/test绿 + 前端build绿; 前端无涉差异已对齐
备注: #3 loop接线层2处(assistant工具轮append历史 + Message.ToolCallID/IsError)留给后续; #4代审/#5子代理触发/#7拆文件后置(主人说后三不急)
|
2026-09-07 13:43:18 +08:00 |
|
Aska
|
9c2a37c4ac
|
feat: 管理面板补齐(RAG导入/记忆添加/会话接history/诚实表述) + design/16 §4.6 IM多会话隔离+第三方身份+黑白名单
CI / Backend (build / vet / test / cross-compile) (push) Has been cancelled
CI / Frontend (tsc + vite build) (push) Has been cancelled
agent-98 面板批次:
- 会话管理面板改接 history.ListSessions(主web+次级) + 新建次级会话按钮
- 记忆面板手动添加入口 POST /admin/memories(owner)
- 知识库真正实现 IndexDocument(分块+落projectindex) + POST /admin/rag 导入端点
- 三面板空态诚实表述(去除'后端M10+提供')
- RAG嵌入降级到keyword/line(无直属embedder,不伪造语义,符合红线)
agent-99 design/16 §4.6:
- model.Message.SenderName + 迁移0006 sender_name 列
- 第三方身份边界: session.ResolveActor判定(非LLM猜) + [第三方·昵称]前缀(不改源content)
- 群聊/陌生人落独立sess-xxx(persistIsolated),不写主会话,不广播
- 黑/白名单(allowGroups/denyGroups/allowUsers/denyUsers)最前拦截,不进LLM
- 多账号 CHANNELS_JSON 配置 + 每平台多实例
- 管理只读 GET /admin/channels/{platform}/messages(该平台非主人消息流)
实机验证(JWT): 会话管理返回主会话web / 记忆添加成功 / RAG导入documents非空+Cyrene命中; 后端build/vet/test全绿 + 前端build绿
|
2026-09-07 05:13:36 +08:00 |
|
Aska
|
3632e04733
|
feat(model): 双协议 LLM Provider(AnthropicChNativeAdapter /v1/messages + tool_use.input→Args 直接映射 + 流式重组)+ Provider 感知工厂/路由 + LLM_PROVIDER/anthropic 配置(OpenAI 兼容保留)
CI / Backend (build / vet / test / cross-compile) (push) Has been cancelled
CI / Frontend (tsc + vite build) (push) Has been cancelled
|
2026-09-06 09:41:06 +08:00 |
|
Aska
|
d09ccc7570
|
feat(model): LLM 多供应商适配(Adapter/OpenAI 非流式+SSE 流式/per-role 路由/主备降级/多模态检测/调用记录)
|
2026-09-06 04:37:27 +08:00 |
|