OpenPack/AgentWheel package for the same extra message policy concept across agent runtimes.
- OpenClaw: installs the existing
extra-message-policyOpenClaw plugin. - Hermes: installs a Hermes Python plugin that hooks
pre_gateway_dispatchfor listen-only / allow / silent-ingest policy andpre_llm_callfor bounded raw recall context.
Built-in OpenClaw adapter supports plugins directly:
agentwheel install . --adapter openclaw --dry-runAgentWheel 0.9.0's built-in Hermes adapter does not expose plugins; use the included adapter config until upstream adds it:
agentwheel install . --adapter-config adapters/hermes-with-plugins.jsonc --dry-runOpenPack runtime selection is declared in openpack.json through runtimes.
OpenClaw installs settings/openclaw-extra-message-policy.json to grant the
non-bundled plugin conversation access required by its before_prompt_build
hook. Hermes defaults are installed as
settings/hermes-extra-message-policy.json into .hermes/settings.json by
AgentWheel. The Hermes plugin also reads ~/.hermes/settings.json directly
under extra_message_policy.
Default policy is conservative: enabled, respond allowed, no passive ingest sink unless configured.
Both runtime implementations accept mentionRecall beside requireMention in
the default policy and scoped rules. It defaults to true for compatibility.
Set it to false when only mention evidence on the current message may satisfy
requireMention; this does not change message ingest. Hermes currently has no
internal cross-event mention cache, but honors recalled provenance supplied by
an adapter (mention_source: recalled or equivalent evidence metadata).
This package governs chat ingest and chat replies. It does not replace OpenClaw tool allow-lists or review controls for plugins that call outside services.
For example, an OpenClaw workspace can install TweetClaw for public X/Twitter automation while this package keeps Discord, Telegram, or other chat channels silent or recall-only:
openclaw plugins install @xquik/tweetclaw
openclaw config set tools.alsoAllow '["explore", "tweetclaw"]'Keep the responsibilities separate:
- Use Extra Message Policy for channel ingest, raw recall, and reply suppression.
- Use OpenClaw tool allow-lists and TweetClaw review controls for live
tweetclawcalls. - Store only concise decisions, source URLs, tweet IDs, or follow-up notes in raw recall. Do not write API keys, cookies, raw direct-message bodies, or raw follower exports to JSONL or HTTP sinks.
explore lets the agent inspect TweetClaw's endpoint catalog without a live API call. tweetclaw performs Xquik-backed actions such as search tweets, search tweet replies, follower export, user lookup, media upload, media download, direct messages, monitor tweets, webhooks, giveaway draws, post tweets, and post tweet replies.
References: npm package and ClawHub listing.