openclaw-plugin-extra-message-policy

Neutral OpenClaw plugin for cross-platform message policy.

It exists to keep message access, ingest, and response behavior deterministic across Discord, Telegram, and other chat providers without changing OpenClaw core.

What it does

The plugin adds an extra policy layer after the normal platform/OpenClaw access checks:

canRead = allowList && platformCanRead

responseCandidate = canRead && nativeRoutingWouldDispatch

shouldIngest =
  canRead &&
  (
    ingestMode === "all" ||
    (ingestMode === "responseCandidates" && responseCandidate)
  )

shouldRespond = responseCandidate && respond

actualRespond = shouldRespond && platformCanSend

Important points:

• allowList always remains active.
• Platform permissions still matter: Discord/Telegram/etc. decide what the bot can physically read/send.
• For ingestMode: "all", native channel routing must let messages reach plugin hooks. In practice this may mean requireMention: false on the allowed chat, then respond: false here to keep the bot silent.
• ingestMode controls whether the plugin captures all readable messages or only messages that reached the response pipeline.
• respond: false suppresses replies even when normal routing, mentions, or allow-list rules would otherwise allow a response.
• On platforms with granular send permissions, e.g. Discord, the platform may block sending before this policy visibly “fires”. The plugin policy is still the same everywhere.

Use cases

• Memory ingest for channels where the bot should learn context but stay silent.
• Context accumulation for operational/project knowledge.
• Passive monitoring/auditing of visible chat traffic.
• Telegram safety, where per-topic/per-channel outbound permissions are less granular than Discord.

What it does not do

• It does not bypass platform permissions.
• It does not bypass OpenClaw allow-lists.
• It does not modify OpenClaw core.
• It does not call an LLM for every message.
• It does not send messages from the ingest path.

Config model

respond

Boolean. Default: true.

When false, inbound messages can still be ingested, but the plugin marks the dispatch as handled without reply.

ingestMode

One of:

• none — do not ingest.
• responseCandidates — ingest only messages that reached before_dispatch, i.e. messages native routing would consider for response.
• all — ingest messages observed by message_received, i.e. all messages OpenClaw exposes to plugin hooks after allow-list/platform read checks.

Example

{
  "plugins": {
    "extra-message-policy": {
      "enabled": true,
      "defaultPolicy": {
        "respond": true,
        "ingestMode": "responseCandidates"
      },
      "policies": [
        {
          "channelId": "YOUR_RESPONSE_CHANNEL_ID",
          "respond": true,
          "ingestMode": "all"
        },
        {
          "guildId": "YOUR_SILENT_GUILD_ID",
          "respond": false,
          "ingestMode": "all"
        },
        {
          "sessionKeyIncludes": "discord:channel:",
          "respond": false,
          "ingestMode": "all"
        },
        {
          "isGroup": false,
          "senderId": "YOUR_OPERATOR_USER_ID",
          "respond": true,
          "ingestMode": "all"
        },
        {
          "sessionKeyIncludes": "telegram:",
          "respond": false,
          "ingestMode": "all"
        }
      ],
      "jsonlSink": {
        "enabled": true,
        "path": "runtime/extra-message-policy/messages.jsonl"
      }
    }
  }
}

Raw recall

The plugin can also explain the raw archive to the agent, inject small relevant excerpts at request time, and expose an on-demand search_raw_context tool.

This keeps ingest zero-token: no embedding, no summarization, no LLM call is made when messages arrive. Recall only happens during an actual agent request.

{
  "rawRecall": {
    "enabled": true,
    "appendGuidance": true,
    "searchOnTriggerOnly": true,
    "maxDays": 30,
    "maxMatches": 12,
    "maxContextChars": 6000,
    "tool": {
      "enabled": true,
      "maxMatches": 20,
      "maxContextChars": 12000,
      "maxFiles": 5000
    }
  }
}

The automatic preflight recall uses maxDays as a lookback window, then injects only a bounded result block. It does not inject the whole archive.

For deeper recall, the agent can call search_raw_context on demand. By default the tool searches all available JSONL shards under rawRecall.rootPath; callers can pass daysBack, conversationId, channelId, or senderId to narrow the search.

Example trigger: “Where are we on project X?”

The plugin searches local JSONL shards for relevant terms and injects a short Relevant passive raw-recall excerpts block.

Sinks

JSONL sink

Writes one JSON object per ingested message. Useful as a durable raw transcript spool.

For long-running passive ingest, use sharding to avoid a single huge file:

{
  "jsonlSink": {
    "enabled": true,
    "path": "memory/discord/raw",
    "shardBy": "dayConversation"
  }
}

This writes files like:

memory/discord/raw/2026/05/08/channel_YOUR_SECOND_RESPONSE_CHANNEL_ID.jsonl

HTTP sink

Optional generic HTTP POST sink for forwarding records to another memory/monitoring service.

{
  "httpSink": {
    "enabled": true,
    "url": "http://127.0.0.1:8000/ingest",
    "accessToken": "..."
  }
}

The plugin posts the normalized ingest record as JSON.

Hook strategy

• message_received handles ingestMode: "all".
• before_dispatch handles ingestMode: "responseCandidates" and suppresses replies when respond: false.
• message_sending is a defensive outbound guard for any reply associated with a suppressed dispatch.