Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 40 additions & 3 deletions docs/architecture/rfcs/manager-runtime-profile-v0.md
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,17 @@ These defaults do not change ordinary Goal Chat models or grant other Agents too
sandbox. Trusted-owner remains limited to the existing verified owner audience. Governed
Task execution, workspace filesystem grants and native managed steward identity preparation
retain their existing Codex owner; selecting Antigravity does not widen those contracts.
Argon tiers must be offered by the local account and match the requested reasoning effort.
Registered Argon and Carbon tiers must be offered by the local account and match the
`reasoning_effort` (`low`, `medium`, or `high`). Model discovery reads the same
account metadata without opening a cascade; an unavailable catalog is an explicit error,
not a static list of supposedly available models. Antigravity's default remains Argon;
operators who want another offered tier (such as Carbon `gemini-4-carbon-high`) for an
endpoint on a specific machine configure it explicitly via top-level `executor_model` or
optional per-endpoint `endpoint_models` entries on `steward_executor_machine_defaults_v1`.
Ordered `eligible_endpoints` and bidirectional flexible failover between `codex` and
`antigravity` route exhausted Codex stewards onto the configured fallback endpoint (and back
to Codex once restored) with model-isolated reasoning effort and native steward monitor/report
parity on the qualified fallback Session.
A missing original cascade fails resume without silently starting a replacement. Stop waits
for the native cascade to become idle; timeout or partial output is not successful completion.
During execution, native `IDLE` alone is not completion: it can appear between input, tool
Expand Down Expand Up @@ -131,6 +141,32 @@ owner-private bindings reuse Core's fresh audience/resource verification before
reusing a host; no channel-name inference, persisted proof or manager-specific ACL replaces it.
A revoked binding or changed App/owner stops admission.

Saving machine defaults does not move an existing steward to another executor endpoint.
The original private conversation's model menu offers **Apply to this conversation** for
the saved preferred/pinned executor, model and reasoning effort. Its preview opens no
native thread. Apply requires an idle Session with no active or reserved queued work,
registered worker, parked native work or answer-review route. It prepares the selected
host without dispatching a model Turn, after closing the old idle transport and capturing
the complete visible history. When only the Codex model or reasoning effort changes
and the current working checkpoint is available, apply reuses the model-resume path
to verify and continue the same native thread without reinjecting history. If that
checkpoint fails verification, apply stops before closing the original transport.
An executor migration restores the visible history on the new native thread.
Only provider-observed model and effort can be committed to the same Session.
The source binding remains unchanged;
Core verifies its live owner, source and resource grant again, and subsequent inputs use
the explicitly accepted Session allocation with retained original native provenance.

The Session store checks the original identity, selected binding, saved configuration,
queue and captured transcript again under its publication locks. A changed source,
revoked owner, new queued input, returned message or unconfirmed host setting rejects
publication and closes the unpublished transport. Canonical history and queued work
remain available for recovery. After a lost response, read the actual Session first:
an already applied allocation needs no second preparation. The menu displays that actual
executor/model readback; saving defaults, preparing a host, executing a model Turn and
installing the changed service remain distinct observations. Other conversations,
ordinary model settings and flexible quota recovery retain their existing owners.

This slice implements only the private-owner M1 journey in
[capable-manager-semantic-handoff-v0](capable-manager-semantic-handoff-v0.md), targeting A1–A3/A12.
It does not implement the M2 collaboration request, the M3 outbox, or treat manager Session fields
Expand All @@ -156,8 +192,9 @@ delivery. Follow [roadmap](loopx-overall-roadmap-v0.md) R2 for those remaining q
4. A profile change rotates the upstream thread while preserving the LoopX Session and visible
history.
5. An unrelated machine-configuration change does not rotate the thread.
6. A non-Codex endpoint fails closed for `trusted_owner` and recommends selecting Codex or
restoring `restricted`.
6. Codex and Antigravity support the verified private owner's `trusted_owner` profile while
retaining their native permission and approval boundaries. An unsupported endpoint fails
closed and recommends a supported executor or restoring `restricted`.
7. Desktop and mobile Dashboard show the effective profile; invalid configuration shows its
fallback state.
8. An external audience without an existing scoped grant remains `read-only` and exposes the
Expand Down
31 changes: 28 additions & 3 deletions docs/architecture/rfcs/manager-runtime-profile-v0.zh-CN.md
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,15 @@ LoopX 管家最初只有受限的规划会话:它可以读取 LoopX 提供的
planner 模式,不继承上一回合的工具配置。这是 provider 工具限制,不是 Codex 或操作
系统 sandbox。强能力模式仍限于既有核验过的本人受众。受治理 Task 执行、工作区文件
系统授权和原生受管理管家身份准备仍由既有 Codex owner 承接;选择 Antigravity 不扩展
这些契约。Argon 档位必须由本地账号提供,并与请求的思考深度一致。原始 cascade 丢失
这些契约。已注册的 Argon 与 Carbon 档位必须由本地账号提供,并与请求的思考深度
(`low`、`medium` 或 `high`)一致。模型发现读取相同的账号元数据,不创建 cascade;
目录不可用时明确报错,不用静态清单冒充可用模型。Antigravity 的默认模型仍为 Argon;
需要在本机为特定执行器优先指定其他已提供档位(例如 Carbon `gemini-4-carbon-high`)时,
可通过 `executor_model` 或 `steward_executor_machine_defaults_v1` 上可选的按执行器映射
`endpoint_models` 显式配置。配合有序的 `eligible_endpoints` 与 `codex`、`antigravity` 之间
的双向弹性故障转移,可在 Codex 耗尽时按可观测配额切换至配置的回退执行器(并在恢复后
切回),同时隔离模型专属的推理强度配置并在合格回退 Session 上保持原生管家 monitor 与报告
能力对齐。原始 cascade 丢失
时续接失败,不静默创建替代会话。停止等待原生 cascade 回到 idle;超时或部分输出
不代表成功完成。
执行时,原生 `IDLE` 本身不代表完成:输入、工具与 planner 步骤之间也可能出现此状态。
Expand Down Expand Up @@ -106,6 +114,22 @@ profile 或权限状态。原生本人私聊在启动或复用宿主前重新核
的 M1 私有 Owner 旅程,目标验收为 A1–A3/A12。它不实现 M2 collaboration request、M3
outbox,也不把管家 session 字段当成工作、请求或送达权威。

保存机器默认配置不会把既有管家移到另一执行器 endpoint。原私有对话的模型菜单提供
**应用到此对话**,采用已保存的 preferred/pinned 执行器、模型与思考深度。预览不创建
原生线程。应用要求 Session 空闲,没有活动或已预留的排队工作、注册 worker、停放的
原生工作或答案复核路径。它先关闭旧空闲 transport、捕获完整可见历史,再准备所选宿主,
不派发模型 Turn。如果只调整 Codex 内的模型或思考深度,且当前工作检查点可用,应用会
沿用同一模型续接路径,核验后恢复原生线程,省去重新注入历史。检查点核验失败时,应用在
关闭原 transport 之前停止。跨执行器迁移会在新的原生线程恢复完整可见历史。
同一 Session 只保存宿主实际回报的模型与思考深度。来源绑定保持不变;Core 再次校验当前 owner、来源和资源授权,后续输入使用该 Session
明确接受的 allocation,并保留原始原生执行器的沿革。

Session store 在发布锁内再次核验原身份、选定绑定、已保存配置、队列与捕获的完整文本。
来源改变、owner 撤权、新排队输入、返回消息或宿主设置未确认都会拒绝发布,并关闭未发布的
transport。原历史与排队工作仍可恢复。响应丢失后先读取实际 Session;已经应用的 allocation
无需再次准备。菜单显示实际执行器/模型回读;保存默认值、准备宿主、执行模型 Turn 和安装
变更后的服务分别记录。其他会话、普通模型设置及 flexible 配额恢复沿用原有 owner。

### 实现与后继

`43d362532` 引入机器模式。原生本人管家私聊在打开、续接和授权范围刷新时,将 Core
Expand All @@ -123,8 +147,9 @@ outbox,也不把管家 session 字段当成工作、请求或送达权威。
`AGENTS.md` 不再包含只读限制。
4. profile 改变会旋转上游 thread,但保留 LoopX Session 和可见历史。
5. 不相关机器配置变化不会旋转 thread。
6. 非 Codex endpoint 对 `trusted_owner` fail closed,并给出切换 endpoint 或恢复
`restricted` 的动作提示。
6. Codex 与 Antigravity 支持已验证私有本人的 `trusted_owner` 模式,并保留各自原生权限和
审批边界。不支持的 endpoint 必须拒绝,并给出切换到受支持执行器或恢复 `restricted`
的动作提示。
7. 桌面和移动 Dashboard 显示有效 profile;配置损坏时显示回退状态。
8. 外部 audience 在没有既有 scoped grant 时继续 `read-only`,并显示真实降级状态。
9. 已核验的原生本人管家私聊沿用实际机器策略,模式改变或服务重启保留 Session 与历史;
Expand Down
Loading
Loading