Use Case

In a Discord server (or Telegram group), different channels often serve different purposes:

• #daily — low-cost summarization channel using a free/cheap model (e.g. openrouter/healer-alpha)
• #dev — coding assistant channel using a capable model (e.g. anthropic/claude-opus-4.6)
• #casual — general chat with a different persona/system prompt

Currently, the gateway uses a single global model + system prompt for all channels. Users who need per-channel differentiation must run separate bot instances or use an external routing layer.

Proposed Design

Add a channel_overrides section under each platform config in config.yaml:

platforms:
  discord:
    enabled: true
    channel_overrides:
      "1234567890":           # channel ID
        model: openrouter/healer-alpha
        provider: openrouter
        system_prompt: "You are a daily news summarizer."
      "9876543210":
        model: anthropic/claude-opus-4.6
        provider: anthropic
        system_prompt: "You are a coding assistant."

Resolution priority (high → low):

1. /model command (ephemeral, per-session)
2. channel_overrides[channel_id] (persistent, per-channel)
3. model.default / agent.system_prompt (global fallback)

Implementation scope:

1. gateway/config.py — extend PlatformConfig with channel_overrides: Dict[str, ChannelOverride]
2. gateway/run.py — _resolve_gateway_model() and _load_ephemeral_system_prompt() accept chat_id parameter, check overrides before global config
3. Agent construction — pass resolved model/provider/system_prompt per dispatch

What it does NOT change:

• No changes to CLI mode
• No changes to the agent loop or prompt caching
• Channels without overrides behave exactly as today

Alternatives Considered

• Multiple bot instances: works but wastes resources, hard to manage
• External routing proxy: adds latency and complexity
• Personality system (/personality): exists but is per-session, not per-channel, and does not support model overrides

Prior Art

Other agent frameworks (e.g. OpenClaw) support modelByChannel config and it is a commonly requested feature for multi-channel deployments.