Problem

When a model generates text before a tool call (e.g., "Let me check that..."), that text block gets delivered to the messaging channel as a separate message. In non-streaming mode, these reasoning preamble messages arrive after the actual response, creating visible inner monologue leaks in channels like Discord.

Current behavior

The message_sending plugin hook receives:

type PluginHookMessageSendingEvent = {
    to: string;
    content: string;
    metadata?: Record<string, unknown>;  // currently: { channel, accountId, mediaUrls }
};

There is no way to distinguish between:

• A text block that precedes a tool call (stopReason: "toolUse") — always inner monologue
• A final response (stopReason: "stop" / "end_turn") — the actual reply

Proposed change

Pass stopReason (and optionally hasToolCalls: boolean) through to the message_sending hook metadata:

metadata: {
    channel,
    accountId,
    mediaUrls,
    stopReason,        // "stop" | "end_turn" | "toolUse" | "tool_use" | ...
    hasToolCalls,      // boolean
}

The relevant code is in applyMessageSendingHook() in the delivery pipeline, which constructs the metadata object. The stopReason is available in the agent command layer but is not currently forwarded to delivery.

Why this matters

This would allow plugins to structurally suppress reasoning preambles:

api.on("message_sending", (event, ctx) => {
    if (event.metadata?.stopReason === "toolUse" || event.metadata?.hasToolCalls) {
        return { cancel: true };
    }
});

This is more reliable than content-based pattern matching (regex for "Let me..." / "Now I need to...") which can false-positive on legitimate responses.

Context

Running an OpenClaw Discord bot that monitors guild channels. In non-streaming mode, text blocks from stopReason: "toolUse" responses consistently leak as separate messages. The existing output-sanitizer plugin catches most cases via META_PLANNING_PATTERNS regex, but a structural filter would be definitive.

Workaround

Re-enable the output-sanitizer plugin with content-based filtering.