v0.3.3

Tasks now show the model and effort they run on, callers can name their session across ssh and containers, and Grok's --model and --effort take effect.

Added

• A global --caller <agent>:<session ID> names the calling session where the host's session variable does not reach axiopera, such as over ssh, docker exec or WSL. Without it such an agent was taken for user and shared the terminal's inbox. axiopera guide tells an agent taken for user to add the option. See "指定调用方" in docs/usage.md.
• Each task records the model and effort it actually runs on, also when spawned without --model. status, the result line of each turn, the turn lines of log and the task title in top show them.
• axiopera models <agent> names in its first line what a spawn without --model runs on: the model from Axiopera's configuration, or else the one the agent's CLI picks from your own settings.
• Each task's events.jsonl gets a delivered event when inbox, result or a push hands something to a caller: who received it, how, and which events and results went along. Tools that read the data directory can tell from it what a command or a pushed message gave the agent. Pushes and reminders sent into Codex sessions carry a client message ID starting with axiopera:, which tells them from what the user typed.

Fixed

• Grok: --model and --effort were ignored and the task ran on Grok's default model and effort. They now apply, and an invalid value is reported with the model list.
• Codex: an approval request that Codex withdraws before anyone decides (network approvals, from Codex 0.167) no longer stays waiting for you, and no stale refusal reason is sent into the turn afterwards.
• Codex: messages pushed into a host session no longer begin with an [axiopera:<id>] mark.

Other notes

• Grok's read tier refuses models that run on another harness, such as codex: switching to one drops the read tier's tool limits. Use the work or full tier for those models.
• models --json: Grok's current_model is replaced by native_model and native_effort, which all three agents now report.
• A push that an earlier version left unconfirmed may be sent once more after the upgrade.
• Stop the daemon with axiopera daemon stop before replacing the executable.
• Updating the Codex plugin may require trusting it again.

---

任务现在显示实际使用的模型和推理强度，跨 ssh、容器调用时可以指明自己的会话，Grok 的 --model 和 --effort 开始生效。

新增

• 全局选项 --caller <名称>:<会话 ID>：经 ssh、docker exec、WSL 等调用 axiopera 时，宿主的会话变量传不过去，用它指明调用方会话。此前这类 agent 会被当成 user，与终端共用收件箱。axiopera guide 会提示被当成 user 的 agent 加上这个选项。详见 docs/usage.md 的「指定调用方」。
• 每个任务记录实际使用的模型和推理强度，不带 --model 派发的也有。status、每轮结果行、log 的轮次行和 top 的任务标题都会显示。
• axiopera models <agent> 首行写明不带 --model 派发时使用的模型：有 Axiopera 配置时是配置值，否则是原生 CLI 按你自己的设置选出的模型。
• inbox、result 或推送把内容交给调用方后，在涉及任务的 events.jsonl 中各记一条 delivered 事件：交给了谁、以哪种方式、带了哪些事件和结果。读取数据目录的工具可以据此知道一条命令或一条推送消息交给了 agent 什么。推入 Codex 会话的消息和提醒带有以 axiopera: 开头的客户端消息 ID，可据此与用户输入区分。

修复

• Grok：--model 和 --effort 此前不生效，任务仍按 Grok 的默认模型和推理强度运行。现在生效，取值无效时报错并附模型列表。
• Codex：Codex 在裁决前自行撤回的审批请求（0.167 起的网络审批）不再一直等你处理，之后也不会再把过时的拒绝理由送进轮次。
• Codex：推入宿主会话的消息不再以 [axiopera:<id>] 标记开头。

其它说明

• Grok 的 read 档拒绝运行在其他 harness 上的模型（如 codex）：切换过去会丢掉 read 档的工具限制。这些模型请用 work 或 full 档。
• models --json：Grok 的 current_model 改为 native_model 和 native_effort，三家现在都会报告这两项。
• 旧版本留下的未确认推送，升级后可能再发送一次。
• 替换程序前先执行 axiopera daemon stop。
• 更新 Codex 插件后可能需要重新信任插件。

v0.3.2

Agent sandboxes can no longer start the daemon on macOS or stand in for it on Windows, and sub-agents can no longer stop it.

Fixed

• macOS: inside an agent's sandbox, such as Claude Code's command sandbox, axiopera no longer starts the daemon and asks you to re-run the command outside the sandbox. A daemon started there stayed in the sandbox, passed it on to every sub-agent, and gave them the sandbox's network proxy settings.
• Windows: a process in the Codex sandbox can no longer add an instance of the daemon's named pipe and receive other clients' requests, including the environment that spawn sends. Clients also refuse a pipe served by a sandboxed process.
• If the data directory is read-only or not writable, axiopera asks you to re-run the command outside the agent sandbox instead of reporting only the system error.
• Linux: axiopera top no longer misses a task update that the daemon saves within a few milliseconds of the previous one.

Other notes

• Sub-agents can no longer run axiopera daemon stop, skill install or plugin cleanup; run them from the main session or a terminal. Version-mismatch errors shown to sub-agents no longer suggest daemon stop --force.
• On Linux the client does not check whether it runs in a sandbox. With Grok's --sandbox workspace and a data directory the sandbox can write, it can still start the daemon inside the sandbox.
• Stop the daemon with axiopera daemon stop before replacing the executable.
• Updating the Codex plugin may require trusting it again.

---

agent 沙箱在 macOS 上不能再启动 daemon，在 Windows 上不能再冒充 daemon，子 agent 也不能再停止 daemon。

修复

• macOS：在 agent 的沙箱中（例如 Claude Code 的命令沙箱），axiopera 不再启动 daemon，而是提示到沙箱外重新运行。此前在沙箱中启动的 daemon 会一直留在沙箱里，它启动的每个子 agent 也都受沙箱限制，并沿用沙箱的网络代理设置。
• Windows：Codex 沙箱中的进程不能再为 daemon 的命名管道另建实例、接收其他客户端的请求，包括 spawn 携带的环境变量。客户端也会拒绝由沙箱进程提供的管道。
• 数据目录只读或无权写入时，提示到 agent 沙箱外重新运行，不再只报系统错误。
• Linux：daemon 在几毫秒内连续两次更新同一任务时，axiopera top 不再漏掉后一次更新。

其它说明

• 子 agent 不能再执行 axiopera daemon stop、skill install 和 plugin cleanup，请在主会话或终端中运行。子 agent 遇到版本不一致时，报错不再建议 daemon stop --force。
• Linux 客户端不检查自身是否在沙箱中。使用 Grok 的 --sandbox workspace 且沙箱可写数据目录时，仍可能在沙箱中启动 daemon。
• 替换程序前先执行 axiopera daemon stop。
• 更新 Codex 插件后可能需要重新信任插件。

v0.3.1

Live sub-task progress for Codex, resumable cancelled tasks, and clearer agent guidance.

Added

• axiopera watch --status follows your running and queued tasks, checks every 2 seconds, prints status and progress changes, and exits when none remain. The Codex guide runs it in a background terminal so you can follow sub-task progress.

Improved

• The guide and sub-agent instructions now say that tasks share your working directory unless --cwd selects another, and that parallel writers need separate git worktrees. They also explain that answers reach the task's main session and how Grok's read tier limits shell commands.
• Refusal messages say why a request was refused: the read tier, a closed approval channel, or no one-time approval option.
• The persistent version-mismatch error explains that axiopera daemon stop --force interrupts active tasks from all sessions.
• Release archives contain only the binary and LICENSE, without a top-level directory.

Fixed

• Cancelled tasks can be resumed with send or reply. Cancelling keeps the session and undelivered input. Resuming a parent does not restart its cancelled sub-tasks, so resume each one separately.

Other notes

• The client-daemon protocol changed. Before replacing the executable, stop the daemon with axiopera daemon stop.
• Cancelled sub-tasks now count when lowering a parent's tier. Lower their tier first.
• Updating the Codex plugin may require trusting it again.

---

Codex 可以看到子任务的实时进展，已取消的任务可以续接，agent 说明也更清楚了。

新增

• axiopera watch --status 跟踪正在执行和排队的任务，每 2 秒检查一次，输出状态和进度的变化，任务全部结束后退出。Codex 的 guide 会让它在后台终端运行，方便查看子任务进展。

改进

• guide 和子 agent 说明补充了几点：任务默认共用当前工作目录，--cwd 可另行指定，并行写入的任务应各用一个 git worktree；回答送入所属任务的主会话；Grok 的 read 档会限制 shell 命令。
• 拒绝消息会写明原因：read 档不允许、审批通道已关闭，或后端没有提供单次批准选项。
• daemon 版本持续不匹配时，报错会说明 axiopera daemon stop --force 会中断所有会话中正在运行的任务。
• 发布归档只包含二进制和 LICENSE，不再套一层目录。

修复

• 已取消的任务可以用 send 或 reply 续接。取消时保留会话和未交付输入。续接父任务不会重启已取消的子任务，需要逐个续接。

其它说明

• 客户端与 daemon 的通信协议有变化，替换程序前先执行 axiopera daemon stop。
• 降低父任务档位时，已取消的子任务也会参与检查，需要先降低它们的档位。
• 更新 Codex 插件后可能需要重新信任插件。

v0.3.0

An interactive task viewer, resumable timeouts, and faster task and message handling.

Added

• axiopera top: browse tasks by project and host session, open conversations and tool activity, and scroll through results. top <task> opens a task directly.
• Task conversations now include messages sent to sub-agents and Markdown rendering. Pending questions and approvals stay visible while browsing.

Improved

• New tasks appear first. The list follows new tasks until you browse older ones; detail views can pause or follow live output.
• Faster task refreshes, inbox queries, result previews and log tailing, with lower memory use for large task histories.
• Smaller release binaries through build-profile optimization.
• Daemon and protocol logs rotate to bound their disk usage.

Fixed

• Timed-out tasks can resume with send, with a fresh time limit for each execution. Timeout and cancellation also interrupt backend requests that stop responding.
• Queued input and sibling messages survive daemon restarts. Replies save their answer and queued input together.
• Fixed input ordering during process release, duplicate sibling messages, GC retry and cursor handling, and recovery from incomplete event records.
• Failed push read-back preserves unconfirmed results instead of marking them delivered. Host notification routes are checked again after failures.
• Fixed inaccessible session headings, selection jumps, terminal redraw issues and slow parsing of long unfinished Markdown links.

Other notes

• Stop the daemon with axiopera daemon stop before replacing the executable. Restored input waits for the next send to supply the environment needed to resume.
• Tasks cancelled by older versions specifically because of a timeout become resumable. Manually cancelled tasks remain cancelled.
• Updating the Codex plugin may require trusting it again.

---

新增交互式任务查看界面，支持超时续接，并改善任务和消息处理性能。

新增

• axiopera top：按项目和宿主会话浏览任务，查看对话、工具活动和结果；top <task> 直接打开指定任务。
• 任务对话展示发给子 agent 的消息，并渲染 Markdown。浏览时保留待回答问题和审批请求。

改进

• 新任务排在前面；浏览旧任务前自动跟随新任务，详情可暂停阅读或跟随最新输出。
• 加快任务刷新、收件查询、结果预览和日志尾读，降低大量历史任务下的内存占用。
• 调整构建配置，缩小发布二进制。
• daemon 和协议日志按大小轮转，限制磁盘占用。

修复

• 超时任务可用 send 续接，每次执行重新计时；后端请求无响应时也能超时和取消。
• 排队输入和兄弟留言在 daemon 重启后保留，问题答案与对应输入一起保存。
• 修复进程释放期间的输入顺序、留言重复投递、GC 重试与游标处理，以及残缺事件记录恢复。
• 推送读回失败时保留未确认结果，避免误标为已送达；通知失败后重新检查宿主路由。
• 修复会话标题不可见、选中项跳动、终端重绘，以及长篇未闭合 Markdown 链接导致的卡顿。

其它说明

• 替换程序前先执行 axiopera daemon stop。重启后保留的输入需要下一次 send 提供执行环境，才能继续投递。
• 旧版本因超时而取消的任务会转为可续接状态；手动取消的任务保持取消。
• 更新 Codex 插件后可能需要重新信任插件。

v0.2.2

Axiopera 0.2.2 installs the bundled skill directly from the executable.

Added

• axiopera skill install <claude|codex|grok> installs the skill for a supported harness offline; all installs it for all supported harnesses.
• Identical skill files are skipped. Replacing different contents requires --force.
• Installation supports --json. With all, each harness gets a result; failures do not stop the remaining installations, and any failure returns a nonzero exit code.

---

Axiopera 0.2.2 支持直接通过可执行文件安装内置 Skill。

新增

• axiopera skill install <claude|codex|grok> 为指定 Harness 离线安装 Skill；all 为所有支持的 Harness 安装。
• 相同内容自动跳过，覆盖不同内容需要加 --force。
• 支持 --json 输出。使用 all 时逐项返回结果，失败后继续处理其余项；任何一项失败，命令返回非零退出码。

v0.2.1

Axiopera 0.2.1 adds model discovery and permission changes when resuming tasks, and fixes message delivery for native sub-agents.

Added

• axiopera models <agent> lists available models and reasoning effort levels. Invalid model or effort errors include available choices.
• send --tier changes an idle task's permissions while continuing its original session.

Improved

• Model listings show resolved aliases and distinguish account defaults, backend defaults, and current selections.
• plugin cleanup removes Codex's hook trust baseline left after uninstalling the plugin.

Fixed

• Native Codex and Grok sub-agents receive their own task messages, including within Axiopera tasks.
• Claude sub-agent guidance scopes inbox reads by task ID and keeps agent-team teammates waiting in the foreground.

---

Axiopera 0.2.1 新增模型查询和续接时切换权限，修复原生子 agent 的消息交付。

新增

• axiopera models <agent> 查询可用模型和推理强度；模型或强度无效时，报错附带可选项。
• send --tier 在任务空闲时切换权限，继续原会话。

改进

• 模型列表显示别名对应的实际模型，并区分账号默认、后端默认和当前选择。
• plugin cleanup 清理卸载插件后留下的 Codex hook 信任基线。

修复

• Codex、Grok 原生子 agent 独立接收自己的任务消息，包含在 Axiopera 任务内继续派发的情况。
• Claude 子 agent 的操作说明要求按任务 ID 收件，agent team 的 teammate 使用前台等待。

v0.2.0

Axiopera 0.2.0 lets a busy main agent hear from its sub-tasks mid-turn, and separates waiting for results from taking them.

Breaking changes

• wait now only waits: it blocks until your sub-tasks have news and lists it, without handing anything over. Take results, questions and approval requests with inbox, which never blocks.
• The sub-agents' inbox command and MCP tool, which list their own questions and answers, are now answers.
• Exit codes: 3 means nothing new yet or wait timed out, 4 means nothing new and nothing running. 5 and 6 are gone; output that carries news always exits 0.
• spawn --worktree is removed. Create a worktree with git worktree add and pass it with --cwd.
• gc without --days now removes every stopped task at once.
• The concurrency limit is now 50 top-level tasks instead of 4, and a spawn over it is refused instead of queued.

Added

• An Axiopera skill for Claude Code, Codex and Grok, and a Codex plugin whose hook tells a busy Codex session when its sub-tasks have news.
• Codex app and IDE sessions get a reminder through Codex's queue when there is news.
• axiopera guide --host claude|codex|grok gives each host only the part that applies to it.
• Grok can keep axiopera watch running in a monitor to hear about news mid-turn.
• Stopped tasks untouched for 30 days are removed automatically, together with old log lines.
• Codex sub-agents report when Codex's Windows sandbox is not set up or its automatic review fails.

Improved

• Approval requests come before other news and show the full request; an agent can approve only requests it has been shown. Unanswered requests are now denied after 2 minutes instead of 10.
• Long results, questions, notes and approval requests show their ends, with the full text saved to a file, so replies fit within each host's output limit.
• When nothing is new, inbox and wait say where each running sub-task stands.
• Reminders and watch list each piece of news in one line.
• spawn says how results will reach the caller, and why a Codex session has to poll.

Fixed

• Long stderr lines no longer flood failure reports.

---

Axiopera 0.2.0 让主 agent 在一轮进行中也能得知子任务的新消息，并把等待结果和取走结果分开。

不兼容变更

• wait 现在只负责等：阻塞到子任务有新消息，列出后退出，不取走任何内容。结果、提问和审批请求用 inbox 取，它从不阻塞。
• 子 agent 用来列出自己提问和回答的 inbox 命令及 MCP 工具改名为 answers。
• 退出码：3 表示暂无新消息或 wait 超时，4 表示没有新消息也没有运行中的任务。5 和 6 已移除；带消息的输出一律以 0 退出。
• 移除 spawn --worktree。请用 git worktree add 自行创建工作副本，再用 --cwd 指定。
• 不带 --days 的 gc 现在一次删除所有已停止的任务。
• 并发上限从 4 个顶层任务改为 50 个，超过上限的 spawn 直接报错，不再排队。

新增

• 适用于 Claude Code、Codex 和 Grok 的 Axiopera skill，以及一个 Codex 插件：Codex 会话忙碌时，插件的 hook 会告知子任务的新消息。
• 有新消息时，Codex 桌面端和 IDE 扩展中的会话会通过 Codex 的队列收到提醒。
• axiopera guide --host claude|codex|grok 只输出适用于该宿主的部分。
• Grok 可以在 monitor 中持续运行 axiopera watch，在一轮进行中得知新消息。
• 自动删除 30 天未活动且已停止的任务，以及过期的日志行。
• Codex 的 Windows sandbox 未配置或自动审查失败时，Codex 子 agent 会给出提示。

改进

• 审批请求先于其他消息交出，并给出完整请求；agent 只能批准已经交给它的请求。无人处理的请求 2 分钟后自动拒绝（原为 10 分钟）。
• 过长的结果、提问、留言和审批请求只显示开头和结尾，全文保存到文件，回复因此不会超出各宿主的输出上限。
• 没有新消息时，inbox 和 wait 会说明每个运行中子任务的进度。
• 提醒和 watch 每条新消息只占一行。
• spawn 会说明结果如何送达调用方，以及 Codex 会话为何需要轮询。

修复

• 失败报告中过长的 stderr 行不再刷屏。

v0.1.0

Axiopera’s first release brings lightweight, minimally invasive collaboration to Claude Code, Codex, and Grok Build.

Added

• Delegate tasks across coding agents and collect results, including asking Codex to review Claude’s changes.
• Send follow-up instructions while tasks run, and let agents ask questions and exchange messages.
• Choose a model and permissions for each task, with optional Git worktree isolation.
• Keep long tasks running after the main session ends, and resume them in their original agent CLI.

Other notes

• Available through crates.io and standalone binaries for Windows, Linux, and macOS.
• Windows and Linux have been tested. Agent integration on macOS still needs validation.

---

Axiopera 首次发布，为 Claude Code、Codex 和 Grok Build 提供轻量、低侵入的协作方式。

新增

• 跨编程 agent 派发任务并收取结果，例如让 Codex 审查 Claude 的改动。
• 在任务运行中追加要求，让 agent 提问、互相留言。
• 为每个任务选择模型和权限，可使用独立 Git worktree 隔离改动。
• 长任务可在主会话结束后继续运行，并通过原来的 agent CLI 接手续接。

其它说明

• 支持从 crates.io 安装，并提供 Windows、Linux、macOS 独立可执行文件。
• Windows 和 Linux 已有实测，macOS 上的 agent 集成仍待验证。