Skip to content

Personal AI OS 0.1.6

Latest

Choose a tag to compare

@yimengbenxin yimengbenxin released this 02 Nov 14:44

依米 · Personal AI OS 0.1.6

让每个人都有自己的依米:一个私人助理,住在你自己掌控的服务器上,用你选的模型,在手机、手表、眼镜和 Mac 之间跟着你走。

这一版让对话在设备之间接得更顺:通话记录回到了聊天里,手表绑定跟着 iPhone 配对走,模型设置可以直接在手机和 Mac 上改。普通响应多出字段或部分展示值有变化时,客户端可以继续处理自己认识的部分,付款相关的检查依旧严格。任务结果和失败的回报更明确了,归档也新增了记录存储和冷存储分层。

升级前请留意:终端消息合同升到了 v5,需要协调 iPhone、Mac 和手表客户端一起更新;时间线存储会单向升级到 v3,请保留升级前的备份。

有哪些变化

对话和设备

  • 通话记录回到聊天里。 手机和 Mac 的聊天里,转写按通话归在一起,标明来源和时长。通话中插进一条文字消息时,同一通电话可以分成前后相接的两段。服务端的通话记录一到,就替换对应的本地字幕,不会再出现本地和服务端各一份的重复话轮。
  • Mac 先恢复聊天,再连接。 Mac 会先恢复已配对的聊天缓存,连接过程不会再抢先存下一个空白视图、把缓存盖掉。缓存读取出错时,界面会直接提示。
  • 手表绑定跟着 iPhone 配对走。 手表的授权挂在持久的 iPhone 配对上,而不是短期的实时会话上,所以单是换一个实时会话,不会让手表绑定失效。
  • 扫码配对更严格。 二维码里如果夹带了轮询凭据,配对会拒绝它;正常的配对流程不受影响。

模型设置

  • 在手机和 Mac 上改模型。 模型设置里可以添加服务商,调整高级语音参数、向量维度和各服务商专属的请求参数。
  • 选用已安装的唤醒和 OCR 模型。 可以按清单 ID 选择苹果设备和眼镜上已安装的唤醒模型,以及手机上已安装的 OCR 模型。这里只是选择,不会下载模型文件。通话中更换唤醒检测器,要等这通电话结束才生效。
  • 不安全的设置可以被隔离。 已保存的模型设置如果不安全,可以被隔离,设置页会单独显示存储审计提示。
  • 语音和执行读同一个来源。 手机执行器、实时语音和朗读,都从模型设置的权威来源读取当前选择;已安装的资源和服务策略仍然适用。

通话和任务

  • 事件循环卡一下,不再重置语音检测。 事件循环短暂卡住时,通话不再自动重置语音检测,也就不会因此切掉下一句话。
  • 关掉固定开场问候。 云端配置关掉了接通电话时的固定开场问候,思考模式下的那一版也一起关了。
  • 失败有明确的结果。 手机不可用、收不到拍照请求时,网关会立即返回,不再干等到请求超时;数字人的视频通路不可用时,手机会收到明确的失败;微信传输失败会给出明确的结果,不会被记成已发送。
  • 任务回复送达更可靠。 任务已完成、回复已开启且文本有效时,即使工具记录的证明格式变了,回复也能送达;空回复和已取消的回复仍要通过原有检查。旧任务的终止回复有送达时限,不会一直反复尝试。
  • 活动卡和子任务正确收尾。 旧任务没有可用的原生绑定时,活动卡可以按持久保存的终止状态收尾;仍在运行或暂时联系不上的子任务,不会被直接判为完成。父任务的回复完成了,也不会自动终止仍获授权的子任务。
  • 恢复不会被一条坏记录卡住。 恢复时遇到无效或过时的入场记录,会先把它搁到一边,让其他记录继续恢复、新输入继续进来。被搁置的记录不算恢复成功。

兼容新内容,付款保持严格

  • 普通消息多了字段,不再整条拒收。 普通对话或配对响应里多出字段时,苹果客户端不会因此拒收整条响应;必需的身份和格式检查照旧。
  • 认不出的展示内容也有着落。 活动卡把认不出的展示状态显示为“未知”,把认不出的来源显示为“通用”。手表能接受新增的展示字段;认不出的最新卡片类型或跑腿状态会被略过,认识的内容照常显示。通话里认不出的对话模式按“未确认”处理:不能借它切换通话模式,也不能打开数字人,但已经请求的数字人仍可以关闭;认不出的思考状态显示值可以略过,不会让一条本来有效的通话更新被整体拒收。普通结果头可以容忍新的语音通道和输出模式显示值。
  • 哪些地方不放宽。 这并不是对任意协议都兼容。手表的版本、类型、连接状态、范围和凭据校验不变;付款消息保持原有的精确合同;付款私密结果头不享受这种放宽,带着认不出的语音通道值就会被拒收。普通工具请求带了多余字段时,平台会在校验前去掉未声明的字段;而付款、发送和花钱类工具的输入,仍然拒收多余参数。

用量、范围和存储

  • 模型用量按每次尝试记录。 服务商提供用量时,按每次模型尝试保留,供用量和缓存统计使用;服务商没给的数据,不会被编成费用。用量记录失败只写日志,不挡住模型调用;真正的预算暂停照样生效。
  • 对话范围由平台来执行。 产品的对话范围改由平台能力网关和插件中间件执行,不再依赖内核补丁里的规则。工具执行仍要经过正常的判定。
  • 归档:记录存储和冷存储分层。 归档会把符合条件的元数据收进记录存储,把大的原件按目标大小装进分段文件,特别大的单个原件独占一段;登记、同意、设置和意图仍各自保留原来的文件布局。冷存储只在副本核验通过后才迁移数据;普通读取、活跃的输入和新追加的内容,都会保护数据不被提前迁走;空闲但还没关闭的数据流,在旧内容迁走后仍然可以继续追加。
  • 存储预算算得更全。 存储预算会统计字节、inode 和文件名的增长,还可以把计划中的部署峰值需求一并算进去。
  • 删掉不用的模块。 已经不用的独立上下文引擎、公共 RTC 模块和过时的记忆模块被删除了。实时通话和现行的记忆路径都还在。

升级与部署前请注意

  • 协调服务端与 iPhone、Mac 和手表一起更新。 本版终端消息合同升到了 v5。请让服务端与 iPhone、Mac、手表协调更新;旧客户端可能拒收或漏掉新的通话记录信息。
  • 保留升级前的备份:时间线单向升级到 v3。 本版的发布合同把 v3 定为时间线存储的格式和目标,v1/v2 作为迁移的输入。迁移 v1/v2 时,发布流程会先停止写入、备份时间线存储,再用新代码打开它。不兼容的旧版本读不了 v3;只把程序回退到旧版本,并不能恢复这份数据。如果日后要回到不兼容的旧版本,需要用升级前的备份恢复数据。
  • 这类升级出问题时,向前修复。 切换到时间线目标与上一版不兼容的新版本后,如果切换后出现故障,系统会保持现状、等待向前修复,而不是自动退回。存储一旦是 v3,切换、手动回退和看门狗都会拒绝不兼容的旧目标;彼此兼容的版本之间仍然可以回退。
  • 没有时间线存储的部署。 如果时间线原本就没启用或不存在,而且既没有存储文件、也没有 SQLite 附属文件,发布流程可以不做备份直接继续。时间线已启用却找不到存储,或者还残留着附属文件,都会被当成错误。
  • 记录存储的启用顺序。 新版本先以文件布局通过初始健康检查;之后发布流程排空网关、检查快照容量,再切换到记录存储布局,重启并核验网关的记录写入。
  • 模型设置迁移。 发布流程会先预演现有的模型设置迁移,复核运行中的输入,等旧的写入方停下后再正式应用。缺少角色时,可以补入 OCR 角色。全新的单机安装,会把已配置的语音选择和凭据导入模型设置的权威来源;已有部署则需要走这次迁移或补入。模型文件必须事先装好。
  • 手表绑定迁移。 符合条件的有效手表绑定,会连同凭据状态一起迁到持久的 iPhone 配对上。已撤销、不属于同一用户,或与该 iPhone 配对不匹配的绑定,不保证能恢复;凭据也仍会按各自的期限过期。
  • 本版不开启的部分。 会话折叠在本版的发布合同里保持关闭;有了记录存储和冷存储分层,并不代表会话折叠已经开启。可选的备用实时语音模式已经有实现和设置,但本版保持关闭。可选会议助手的 Mac 执行端不在本版之内。
  • 需要另行准备的资源。 模型文件、运行环境和第三方 SDK 需要另行准备。
  • 仍需在你自己的环境里验证。 通话、设备连接、模型服务商和恢复流程,仍需要在你自己的主机和设备上验证。

Yimi · Personal AI OS 0.1.6

Everyone should have their own Yimi: a personal assistant who lives on a server you control, uses the models you choose, and goes with you across phone, watch, glasses and Mac.

This release makes conversations flow more smoothly between devices: call history is back in the chat, Watch binding follows the iPhone pairing, and you can edit model settings directly on the phone and the Mac. When ordinary responses gain fields or selected display values change, clients keep processing the parts they recognize, while payment-related checks stay strict. Task results and failures are reported more clearly, and archives gain record storage and cold tiering.

Before upgrading: the terminal message contract moves to v5, so coordinate the iPhone, Mac and Watch client updates; the timeline store moves one way to v3, so keep a pre-upgrade backup.

What changed

Conversations and devices

  • Call history is back in the chat. Phone and Mac chats group transcripts by call, showing source and duration. If a text message comes in during a call, the call can split into two consecutive sections. Once the server's call records arrive, they replace the matching local captions, so turns no longer show up twice (once local, once from the server).
  • The Mac restores the chat before connecting. The Mac restores its paired chat cache first, so connection activity can no longer save an empty view over it. If the cache can't be read, the app says so.
  • Watch binding follows the iPhone pairing. Watch authorization rests on the durable iPhone pairing rather than the short-lived realtime session, so rotating a realtime session does not on its own invalidate the Watch binding.
  • Stricter QR pairing. Pairing rejects a QR code that carries an embedded poll credential; the normal pairing flow is unaffected.

Model settings

  • Edit models on the phone and Mac. Model settings can add providers and edit advanced voice parameters, embedding dimensions and provider-specific request parameters.
  • Pick installed wake and OCR models. You can select installed wake models for Apple devices and glasses by manifest ID, as well as installed phone OCR models. This only selects them; it does not download model files. Changing the wake detector during a call takes effect after the call ends.
  • Unsafe settings can be quarantined. Saved model settings that are unsafe can be quarantined, and the settings screens show a separate storage-audit notice.
  • Voice and execution read from one source. The phone executor, realtime voice and read-aloud all take their current selections from the model-settings authority; installed resources and service policy still apply.

Calls and tasks

  • A brief event-loop stall no longer resets voice detection. When the event loop stalls briefly, the call no longer automatically resets voice detection, so the next utterance is not cut off because of it.
  • Fixed opening greeting turned off. The cloud configuration turns off the fixed greeting when a call connects, including its thinking-mode version.
  • Failures have explicit outcomes. When the phone is unavailable for a camera request, the gateway returns right away instead of waiting for the request to time out; when the avatar's video route is unavailable, the phone gets an explicit failure; WeChat transport failures produce an explicit outcome instead of being recorded as sent.
  • Task replies arrive more reliably. A completed task with an opened reply and valid text can be delivered even if its tool-transcript proof has changed shape; empty and cancelled replies still go through their existing checks. Terminal replies from old runs have a delivery age limit, so they are not retried forever.
  • Activity cards and subtasks wrap up correctly. When an old run has no usable native binding, its activity card can settle from the durable terminal state; a subtask that is still running or temporarily unreachable is not simply declared finished. A completed parent reply does not automatically stop a still-authorized subtask's tool work.
  • One bad record no longer holds up recovery. During recovery, an invalid or obsolete admission record is set aside so other records can recover and new input keeps flowing. A set-aside record does not count as successfully recovered.

New content tolerated, payments kept strict

  • Extra fields no longer get a whole message rejected. When an ordinary conversation or pairing response carries extra fields, the Apple client no longer rejects the whole response; required identity and schema checks still apply.
  • Unrecognized display content has somewhere to go. Activity cards show unrecognized display states as "unknown" and unrecognized sources as "generic". The Watch accepts additional display fields; an unrecognized latest-card kind or errand state is skipped while known content still shows. An unrecognized call conversation mode is treated as unconfirmed: you cannot switch call modes or open the avatar from it, but an avatar already requested can still be turned off; unrecognized thinking-status display values can be skipped instead of rejecting an otherwise valid call update. Ordinary result headers tolerate new voice-transport and output-mode display values.
  • What is not relaxed. This is not compatibility with arbitrary protocols. Watch version, kind, link, scope and credential checks are unchanged; payment messages keep their exact contract; a payment-private result header gets no such relaxation and is rejected if it carries an unrecognized voice transport. When an ordinary tool request has extra fields, the platform strips undeclared fields before validation, while payment, send and spend tool inputs still reject extra parameters.

Usage, scope and storage

  • Model usage recorded per attempt. When a provider reports usage, it is kept per model attempt for the usage and cache views; data the provider did not supply is never turned into a made-up cost. If recording usage fails, it is logged and the model call goes ahead; a real budget pause still applies.
  • Conversation scope enforced by the platform. Product conversation scope is now enforced by the platform's capability gateway and plugin middleware rather than by rules in kernel patches. Tool execution still needs its normal decision.
  • Archives: record storage and cold tiering. Archives pack eligible metadata into record storage and large originals into segments of a target size, with an oversized single original getting a segment of its own; registry, consent, settings and intent data keep their own file layouts. Cold tiering moves data only after copies are verified; ordinary reads, active input and newly appended content keep data from being moved too early; an idle stream that has not been closed can still be appended to after its older content moves.
  • A more complete storage budget. The storage budget tracks bytes, inodes and filename growth, and can factor in planned deployment peak demand.
  • Unused modules removed. The unused standalone context engine, the public RTC module and obsolete memory modules have been removed. Live calls and the current memory path remain.

Before you upgrade

  • Coordinate the server with iPhone, Mac and Watch updates. This release moves the terminal message contract to v5. Update the server together with the iPhone, Mac and Watch; older clients may reject or miss the new call-history information.
  • Keep a pre-upgrade backup: the timeline moves one way to v3. This release's contract sets v3 as the timeline store's schema and target, with v1/v2 as migration inputs. For a v1/v2 migration, the release flow stops writers and backs up the timeline store before opening it with the new code. An incompatible older release cannot read v3, and rolling the program back to an older release does not bring this data back. To return to an incompatible older release later, you need to restore data from the pre-upgrade backup.
  • If this kind of upgrade fails, fix forward. After switching to a release whose timeline target is incompatible with the previous one, a failure after the switch is held for a forward fix instead of being rolled back automatically. Once the store is v3, switch, manual rollback and watchdog paths all reject an incompatible older target; rollback between compatible releases still works.
  • Deployments without a timeline store. If the timeline was disabled or absent and there is neither a store nor any SQLite sidecar file, the release flow can continue without creating a backup. An enabled timeline whose store is missing, or a leftover sidecar file, is treated as an error.
  • Order for turning on record storage. The new release first passes its initial health check on the files layout; the release flow then drains the gateway, checks snapshot capacity, switches to the record storage layout, restarts, and verifies that the gateway is writing records.
  • Model settings migration. The release flow previews the existing model-settings migration, rechecks the running inputs, and applies it once the old writer has stopped. If roles are missing, the OCR role can be adopted. A fresh single-host install imports the configured voice choices and credentials into the model-settings authority; existing deployments need this migration or adoption. Model files must be installed beforehand.
  • Watch binding migration. A qualifying active Watch binding moves to the durable iPhone pairing along with its credential state. Revoked bindings, bindings that belong to another user and bindings that don't match the iPhone pairing are not guaranteed to come back, and credentials can still expire on their own schedule.
  • Not enabled in this release. Session folding stays off under this release's contract; having record storage and cold tiering does not mean folding is on. The optional alternate realtime-voice mode is implemented and has settings, but stays off in this release. The optional meeting assistant's Mac executor is not part of this release.
  • Resources to prepare separately. Model files, runtimes and third-party SDKs are obtained separately.
  • Still to verify in your own environment. Calls, device connections, model providers and recovery still need to be verified on your own host and devices.