菜单
Manage World2Agent sensors for OpenClaw. Use when the user asks to install, list, remove, or inspect W2A sensors, or wants an outside-world source such as Hacker News, GitHub, RSS, calendars, or market feeds.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/machinepulse-ai/world2agent-plugins --skill world2agent-manage命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | world2agent-manage |
| description | Manage World2Agent sensors for OpenClaw. Use when the user asks to install, list, remove, or inspect W2A sensors, or wants an outside-world source such as Hacker News, GitHub, RSS, calendars, or market feeds. |
| user-invocable | false |
You manage the user's World2Agent sensors on this OpenClaw machine. Sensors
are long-running probes that watch the outside world (news, repos, markets,
pages, calendars) and dispatch structured signals back into the agent as
# System Event turns.
OpenClaw must have agents.defaults.contextInjection set to
"continuation-skip" in ~/.openclaw/openclaw.json (it's JSON, not YAML).
The plugin refuses to start otherwise. Verify before doing anything else:
jq '.agents.defaults.contextInjection' ~/.openclaw/openclaw.json
If the value is anything other than "continuation-skip", ask the user for
permission to fix it, then run:
jq '.agents.defaults.contextInjection = "continuation-skip"' \
~/.openclaw/openclaw.json > /tmp/oc.tmp && \
mv /tmp/oc.tmp ~/.openclaw/openclaw.json
When the user expresses interest in an outside-world source ("subscribe me to Hacker News", "watch this GitHub repo", "ping me on calendar events"), drive the install end-to-end through dialogue. Do NOT just shell out to the CLI with default config — the auto-generated handler is generic and the agent will not reply meaningfully to signals without the user's preferences baked in.
Map the user's phrase to an npm package name. Common ones:
@world2agent/sensor-hackernews@world2agent/sensor-githubIf unsure, look up what's available:
npm search @world2agent/sensor- --json | jq '.[] | {name, description}'~/.world2agent/sensors.json.If the user describes something that doesn't match any sensor on the hub (e.g. "subscribe to my Notion tasks" with no Notion sensor), say so plainly and offer two paths: pick the closest existing sensor, or write a new sensor following the W2A SDK template (linked from the hub).
Confirm the package name with the user before continuing.
Every sensor package ships a SETUP.md that defines:
To locate SETUP.md you first need the plugin's install directory. Use
openclaw plugins list --json (the canonical way — works in both link mode
and copy mode):
PLUGIN_DIR=$(openclaw plugins list --json | \
jq -r '.plugins[] | select(.id == "world2agent") | .rootDir')
echo "$PLUGIN_DIR"
# → e.g. /Users/<you>/Documents/.../openclaw-plugin-wip
Then check whether the sensor package is already installed there:
SETUP="$PLUGIN_DIR/node_modules/<pkg>/SETUP.md"
ls "$SETUP" 2>/dev/null || echo "not installed yet"
If it isn't installed, install it in-place (no manifest mutation, just
populates node_modules/):
( cd "$PLUGIN_DIR" && npm install --no-save <pkg> )
Then read SETUP.md with the Read tool, passing the absolute path
$PLUGIN_DIR/node_modules/<pkg>/SETUP.md.
(Reading SETUP.md upfront — before sensor add — is preferred so the user
can answer questions before any state mutation. The actual sensor
registration happens in step 6.)
SETUP.md lists 1-3 questions under "Questions to Ask". Ask them ONE AT A TIME, in the user's language, waiting for each answer before continuing. Do NOT batch-ask. Do NOT invent your own questions. Do NOT skip questions even if the user seems impatient — every placeholder in the SKILL.md template corresponds to one of these answers.
SETUP.md provides a template in its "Output" section, with placeholders like
[USER_TOPICS], [USER_NORMAL_STYLE], [USER_DEEP_DIVE_THRESHOLD]. Replace
each placeholder with the user's answer (or the default the user accepted).
Show the filled SKILL.md to the user for confirmation before writing.
Write to ~/.openclaw/skills/<skill_id>/SKILL.md (NOT to Claude Code's
~/.claude/skills/... — that's the channel-side path, irrelevant here).
Compute <skill_id> from the package name: strip leading @, replace /
with -. Example: @world2agent/sensor-hackernews →
world2agent-sensor-hackernews.
mkdir -p ~/.openclaw/skills/<skill_id>
# Then write SKILL.md via the Write tool, with the filled template.
Build the sensor's config JSON object (just the config block from SETUP.md's
"Configuration Parameters" table, with the user's answers). Then call:
openclaw world2agent sensor add <pkg> \
--config-json '<inline json>' \
--skip-generate-skill
The --skip-generate-skill flag is critical: it tells the CLI to keep the
personalized SKILL.md you just wrote in step 5. Without it, the CLI's
fallback would overwrite your work with a generic template.
Optional flags:
--sensor-id <id> for a non-default instance id (only if the user wants
multiple instances of the same sensor)--isolated to run the sensor out-of-process (for unstable third-party
sensors)--deliver-channel <id> --deliver-to <id> to push the agent's reply to a
chat platform (see step 6b).Check whether the user has any inbound chat-platform plugin enabled:
openclaw plugins list --json | jq -r \
'.plugins[] | select(.enabled == true) | select(.channelIds | length > 0) | .id'
If the output is empty, skip this step — there's no IM to push to. The
sensor will still run; replies stay in the OpenClaw session lane (visible via
openclaw sessions --agent main and the dashboard).
If one or more channels are listed (e.g. feishu, lark, whatsapp,
telegram, discord, slack, signal, imessage, line, msteams),
ask the user once — in their language — whether they want this sensor's
replies pushed to one of those chats. If yes, also ask for the recipient
target id (chat id, user id, or platform-specific target — the user knows
this from when they paired the channel; never invent it).
Append to the install command:
--deliver-channel <channel_id> \
--deliver-to <chat_id>
Optional: --deliver-account <id> for multi-account channels,
--deliver-thread <id> to post into a specific thread/topic.
The user can also set this once globally under
plugins.world2agent.deliver in ~/.openclaw/openclaw.json; per-sensor
flags override that default. If the user wants the same target for
everything they're about to install, suggest they set the global default
instead of repeating the flags every time.
If the CLI's reload field returns ok: true, the sensor is already
polling — done.
If reload failed (timeout is the common case), DO NOT run
openclaw gateway restart yourself. Restarting the gateway from inside
this chat would kill the gateway process — including this very chat
session — and the user would see your reply truncated mid-sentence. Always
hand the restart back to the user. Tell them — in their language, not
necessarily English — something equivalent to:
The sensor is registered, but it needs a gateway restart before it starts polling. Please run
openclaw gateway restartin your terminal (I can't run it myself — that command would kill this chat session mid-reply). The first signal will arrive within ~60 seconds after the restart.
Then summarize for the user:
hackernews)~/.openclaw/skills/<skill_id>/SKILL.md)
so they know what to edit later if their preferences changeagent:main:w2a-<sensor_id> (sessionId w2a-<sensor_id>). Tell the
user to switch to that lane in dashboard
(http://127.0.0.1:18789/) — or run
openclaw sessions --agent main --active 60 from CLI — to see
the agent's responses to incoming signals. The SKILL.md they just
configured drives those replies.openclaw world2agent sensor list
Returns the manifest plus the current contextInjection value.
openclaw world2agent sensor remove <sensor_id>
Add --purge only if the user wants the generated handler skill directory
deleted too (this is destructive — confirm first).
If the user hand-edited ~/.world2agent/sensors.json or a personalized
SKILL.md, run:
openclaw world2agent reload
If reload fails, fall back to openclaw gateway restart.
[USER_TOPICS] /
[USER_NORMAL_STYLE] (or whatever the SETUP.md template defines) filled in,
the agent has no anchor for "what's relevant" and will skip most signals
silently — burning tokens on NO_REPLY turns.~/.claude/skills/.... That's the channel-side
(Claude Code) path. OpenClaw reads from ~/.openclaw/skills/....sensor add before writing SKILL.md if you intend to
personalize. The CLI's fallback will skip when SKILL.md exists, but the
cleaner ordering is: write SKILL.md first, then sensor add --skip-generate-skill.After each action, summarize concisely: