← 返回 Skills 市场
yangmeng6666

Agent Mailbox Light

作者 Meng Yang · GitHub ↗ · v1.0.1 · MIT-0
cross-platform ✓ 安全检测通过
149
总下载
0
收藏
0
当前安装
1
版本数
在 OpenClaw 中安装
/install agent-mailbox-light
功能描述
Lightweight cross-agent mailbox using per-workspace file inboxes with best-effort fanout and local read/archive flow. Use when you want a middle layer betwee...
使用说明 (SKILL.md)

Agent Mailbox

Use this skill for a very lightweight internal mail system between OpenClaw agent workspaces.

Design goals

  • No central registry
  • No guaranteed delivery
  • No push notifications
  • No group-message storms
  • Best-effort fanout only
  • Per-workspace local inboxes
  • Token-efficient reads: each agent reads only its own inbox
  • Advisory context only, not an execution bus
  • Local filesystem only, not cloud sync or webhook delivery

This is a middle layer between long-term memory and chat messaging.

Safety boundary

This skill is intentionally narrow.

It is for local mailbox-style context sharing only.

It does not implement:

  • automatic task execution
  • autonomous task dispatch
  • webhook callbacks
  • remote sync
  • cloud delivery
  • receipt tracking
  • retry queues

Mailbox items should be treated as advisory hints for local agent judgment, not commands that must be executed.

Mailbox location

Each workspace owns its own local mailbox under:

\x3Cworkspace>/.agent-mailbox/

Subdirectories:

  • inbox/ — unread or pending mail files
  • archive/ — consumed mail files

Examples:

  • /path/to/workspace/.agent-mailbox/inbox/
  • /path/to/another-workspace/.agent-mailbox/inbox/

Fanout model

Sender scripts scan candidate workspaces using a configurable glob pattern.

Default pattern:

$HOME/.openclaw/workspace*

Override with MAILBOX_GLOB when your workspaces live elsewhere.

If a workspace contains .agent-mailbox/inbox/, write a mail file there.

If the path does not exist or write fails, skip it.

Do not maintain a registry. Do not retry failed deliveries. Do not block on partial failure.

File model

One mail = one file.

Filename format:

\x3Cstamp>--\x3Csender>--\x3Cpriority>--\x3Cslug>.md

Example:

20260322100000--main--warn--cron-list-broken.md

Fields:

  • stampYYYYMMDDHHMMSS
  • sender — short sender id
  • priorityinfo, warn, or critical
  • slug — short readable identifier

Mail content

Use a short header followed by a blank line and body.

Example:

Title: cron list 异常
From: main
Created-At: 2026-03-22T10:00:00+08:00
Priority: warn
Tags: cron,ops

`openclaw cron list` 当前报错,先不要依赖其输出。
建议改用:
- openclaw cron status
- openclaw cron runs

Baseline integration

For best results, integrate mailbox checking into the agent's baseline startup or task-entry workflow.

Recommended baseline rule:

  • If .agent-mailbox/inbox/ exists in the current workspace, check only the newest 1-3 mail files before normal task work
  • Read only enough to determine relevance
  • Compress relevant mail into a very short working summary
  • Archive processed mail immediately so inbox stays small
  • Do not rebroadcast, auto-reply, or auto-write all mail into long-term memory

This makes the mailbox reliable in practice while keeping token cost low.

Read flow

When loading this skill in a workspace:

  1. Ensure the local mailbox exists
  2. List files in local inbox/
  3. Read only the newest few files you need
  4. Summarize them briefly into context
  5. Decide whether to ignore, use for this session, or adopt into memory
  6. Archive or delete consumed mail so it is not repeatedly re-read

Prefer reading only 1–5 newest mails.

Receiver handling policy

Treat mailbox items as single-consumption lightweight context hints.

Default handling:

  1. Read the newest relevant mail
  2. Compress it into a very short summary for the current task
  3. Decide whether it matters only for this turn or should affect longer-lived memory
  4. Archive it after processing so it leaves the inbox

Recommended decision model:

  • critical — read first; usually affect current behavior immediately
  • warn — read and adopt when relevant to current work
  • info — skim quickly; keep only if useful

Possible outcomes after reading:

  • Ignore — not relevant; archive it
  • Session-use only — use it for the current turn/session; archive it
  • Adopt to memory — if it represents a durable fact, explicitly write the distilled version into the proper memory file; then archive it
  • Act now — if it changes current execution, apply it and archive it

Do not automatically:

  • rebroadcast mail
  • reply to the sender through chat
  • write all mail into long-term memory
  • keep already-processed mail in inbox/

Inbox should contain only unprocessed mail. Archive should contain processed mail.

Archive policy

Default: archive after reading, do not keep re-reading inbox mail.

Archive path:

\x3Cworkspace>/.agent-mailbox/archive/

This keeps the inbox small and token cost low.

Good uses

  • low-noise cross-agent operational note
  • temporary warning for other workspaces
  • lightweight fanout without chat delivery
  • “FYI only” advisory
  • coordination hint that does not belong in long-term memory yet
  • local best-effort sharing of small, non-urgent context

Avoid

  • secrets unless truly necessary
  • large documents
  • durable memory that belongs in MEMORY.md or memory/YYYY-MM-DD.md
  • real-time urgent alerts that should use actual messaging
  • any design that assumes guaranteed delivery
  • any design that expects mailbox files to trigger automatic execution
  • any claim of cloud sync, webhook transport, or remote callback support

Recommended AGENTS.md snippet

You can paste this into a workspace's AGENTS.md:

### 📬 Lightweight Mailbox Check

If this workspace has `.agent-mailbox/inbox/`, do a **very light mailbox check** before starting normal task work:

1. Look at only the newest **1-3** mail files
2. Read only what is needed to identify relevance
3. Compress relevant mail into a very short working summary
4. Archive processed mail so inbox stays small
5. Do **not** rebroadcast, auto-reply, or dump all mail into long-term memory

Think of mailbox items as lightweight internal coordination hints, not chat messages and not permanent memory.

Minimal receive command pattern

Use this minimal pattern when an agent wants to consume mailbox items manually from shell.

This pattern is intentionally local and conservative: inspect, summarize, archive. It should not be extended into automatic execution of mailbox contents.

MAIL=$(./skills/agent-mailbox/scripts/list-mailbox.sh . 1 | head -n 1)
if [ -n "$MAIL" ]; then
  sed -n '1,40p' "$MAIL"
  ./skills/agent-mailbox/scripts/archive-mailbox.sh "$MAIL" >/dev/null
fi

Use this behavior convention:

  • check only the newest mail first
  • read only enough to judge relevance
  • archive immediately after processing
  • keep inbox small

If you need a slightly broader pass, read up to 3 mails:

./skills/agent-mailbox/scripts/list-mailbox.sh . 3

Scripts

Use bundled scripts when possible:

  • scripts/init-mailbox.sh — initialize a mailbox in a workspace
  • scripts/send-mailbox.sh — fan out mail to mailbox-enabled workspaces
  • scripts/list-mailbox.sh — list local inbox mail
  • scripts/archive-mailbox.sh — archive a consumed mail file
  • scripts/cleanup-mailbox.sh — prune old archive files

Read the script directly if you need to inspect or change behavior.

安全使用建议
This skill appears to do exactly what it claims: provide a local, file-based mailbox between agent workspaces. Before installing, consider these practical points: (1) ensure MAILBOX_GLOB is pointed only at directories you control (avoid wildcards that include home or system directories); (2) do not send secrets or large documents through this mailbox; (3) prefer sanitized SENDER/SLUG values — the provided SLUG sanitation is good, but SENDER is used as-is in filenames, so treat SENDER as trusted or sanitize it yourself; (4) keep agents from automatically executing mailbox contents — treat items as advisory only; (5) restrict filesystem permissions on workspace directories so a compromised agent cannot broadly modify other users' files. If you need networked delivery, retries, or guaranteed delivery, this skill intentionally does not provide those; choose a different tool for those requirements.
功能分析
Type: OpenClaw Skill Name: agent-mailbox-light Version: 1.0.1 The agent-mailbox-light skill implements a local, file-based messaging system for coordination between agent workspaces. The included bash scripts (e.g., send-mailbox.sh, archive-mailbox.sh) perform standard file operations like creating directories, moving files, and cleanup via find/delete within a specific .agent-mailbox directory structure. The SKILL.md instructions explicitly define safety boundaries, advising the agent to treat mailbox content as advisory context rather than executable commands, and there is no evidence of data exfiltration, remote execution, or malicious intent.
能力评估
Purpose & Capability
Name, description, SKILL.md, and included scripts (init, list, send, archive, cleanup) all consistently implement a local file-based mailbox. There are no unrelated credentials, binaries, or install steps requested.
Instruction Scope
SKILL.md narrowly instructs agents to look only in a workspace-local .agent-mailbox/inbox, to read only a few newest mails, to archive after processing, and to avoid automatic execution or rebroadcast. The included scripts only list, create, move, and delete files in those local directories and do not perform network calls or read unrelated system files.
Install Mechanism
There is no install spec and the skill ships small shell scripts. Nothing is downloaded from external URLs and no archives are extracted — this is low-risk from an install perspective.
Credentials
No credentials or sensitive env vars are requested. Optional env vars (MAILBOX_GLOB, SENDER, PRIORITY, SLUG, TITLE, TAGS, SELF_WORKSPACE) are proportional. Minor note: SLUG is sanitized but SENDER is not sanitized before being embedded into filenames; avoid using untrusted input for SENDER to prevent unexpected file paths/structure within inboxes.
Persistence & Privilege
The skill does not request 'always' persistence, does not alter other skills' configs, and only performs local mailbox file operations. Autonomous invocation being allowed is the platform default and not a unique risk here.
如何使用
  1. 确保已安装 OpenClaw(本地或 Docker 部署)
  2. 在对话框中输入安装命令:/install agent-mailbox-light
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /agent-mailbox-light 触发
  4. 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.0.1
- Major simplification: Replaced previous full CLI/API TypeScript codebase with lightweight shell scripts for core mailbox operations. - Redesigned documentation to clarify new minimal, local, best-effort mailbox model and narrow the feature scope. - Removed cloud sync, webhook, guaranteed delivery, and execution workflow claims; mailbox is now strictly local and advisory. - Added shell scripts for mailbox init, send, list, archive, and cleanup. - Removed all previous Node.js source, package files, and TS docs/examples. - Target use case: very lightweight, low-noise, per-workspace mail fanout for agent coordination without chat or memory bloat.
元数据
Slug agent-mailbox-light
版本 1.0.1
许可证 MIT-0
累计安装 0
当前安装数 0
历史版本数 1
常见问题

Agent Mailbox Light 是什么?

Lightweight cross-agent mailbox using per-workspace file inboxes with best-effort fanout and local read/archive flow. Use when you want a middle layer betwee... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 149 次。

如何安装 Agent Mailbox Light?

在 OpenClaw 或 Claude Code 对话框中运行命令「/install agent-mailbox-light」即可一键安装,无需额外配置。

Agent Mailbox Light 是免费的吗?

是的,Agent Mailbox Light 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。

Agent Mailbox Light 支持哪些平台?

Agent Mailbox Light 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。

谁开发了 Agent Mailbox Light?

由 Meng Yang(@yangmeng6666)开发并维护,当前版本 v1.0.1。

💬 留言讨论