Feat/silent processing

[ufobat]

Description

Adds silent_processing to agents.defaults. When enabled, the agent processes every inbound message fully — tool calls execute, session history is written — but if the LLM produces no text output, no message is sent to the channel. The automatic empty-response fallback ("The model returned an empty response...") is suppressed. Explicit LLM text responses are sent as normal.

This enables observer agents that monitor group conversations and perform background work (e.g. write_file calls to maintain memory) without producing chat noise on every silent turn.

Type of Change

• New feature (non-breaking, opt-in via config)

🤖 AI Code Generation

• 🤖 Fully AI-generated — reviewed and tested by contributor

Related Issue

Closes #2126

Technical Context

PicoClaw's processOptions.DefaultResponse was already a string field, always set to the hardcoded constant. This PR makes the value conditional on a new AgentDefaults.SilentProcessing flag — when true, DefaultResponse is "". The existing guard if finalContent == "" && ts.opts.DefaultResponse != "" then leaves finalContent empty, and publishResponseIfNeeded already has an early-return for empty strings.

Two session history fixes: the assistant message write is gated on finalContent != "" to avoid writing empty assistant messages (which cause provider errors on the next turn); the Session.Save call is separated from the assistant write so the session is always flushed to disk — without this, user messages added to the in-memory session during a silent turn would be lost on restart.

Both the empty-response fallback and the tool-iteration-limit message are suppressed in silent mode. This is intentional: a background observer agent must not surface internal diagnostics to the channel.

Test Environment

• Hardware: Raspberry Pi 4 (ARM64, Debian)
• Provider: Ollama (local, Qwen model)
• Channel: Telegram group chat
• Verified: agent writes to memory/MEMORY.md via write_file on each group message; no response appears in Telegram when LLM produces no text output

Evidence

Config used for verification:

{
  "agents": {
    "defaults": {
      "silent_processing": true
    }
  }
}

Checklist

• Code follows the existing style and passes make check
• Tests added: empty response suppressed in silent mode, text response still sent in silent mode, default fallback preserved in normal mode
• Docs updated: docs/configuration.md (new section with group observer pattern), docs/chat-apps.md (silent observer example); both clarify the distinction between silent_processing and group_trigger.mention_only
• No breaking changes — existing deployments unaffected (silent_processing defaults to false)
• Rebased onto current main

[CLAassistant]

All committers have signed the CLA.

[ufobat]

@lxowalle @Zhaoyikaiii — tagging per CONTRIBUTING.md reviewer list for agent changes.

[sipeed-bot]

@ufobat Hi! This PR has been inactive for over a week. If there's no update in the next 7 days, it will be closed automatically. If you're still working on it, just leave a comment to keep it open!