Listening vs Triggers

"Listening" is the optional, heuristic/judge-driven intervention path. It remains unchanged.

"Triggers" are an explicit, config-driven generation path: - DM and/or mention (configurable via triggers.on_mention) - Word triggers (substring or regex) in any message (triggers.enabled, triggers.words, triggers.use_regex)

If no trigger fires, listen heuristics may still intervene when enabled. These features are independent.

Listening (Experimental)

Enable the bot to passively "listen" in servers and occasionally join conversations when it adds value or a good joke.

Behavior - Heuristics and optional judge decide if the bot should intervene. - Cooldowns and channel allow/deny guards avoid spam. - Send-side rate limiting also applies to all outbound messages to prevent loops. - Replies stream as natural bursts; tone follows the persona.

Persona YAML (listen) - enabled: true|false - allow_channels, deny_channels: lists of IDs or names - cooldown_channel_seconds, cooldown_user_seconds - min_len, trigger_keywords - judge_enabled, judge_model (default gpt-5-nano), judge_threshold, judge_max_context_messages - generation_model_override (else default model), response_max_chars, joke_bias - cost_daily_usd, cost_monthly_usd (optional persona-level hints) - moderation_enabled (default false), moderation_model (default omni-moderation-latest)

Commands (prefix respects persona) - ~listen on / ~listen off: toggle listening for this guild (persisted) - ~listen status: show current state and banned channels - ~listen ban #channel / ~listen unban #channel: update bans (persisted)

Notes - The judge uses a small GPT-5 model by default to keep costs low. - Moderation is optional and off by default.