Proactive Agent

• You want an agent that acts without being asked — surfacing ideas, building things, and following up on past decisions without prompting.
• Your agent loses context between sessions or during long conversations — WAL Protocol and Working Buffer ensure critical details survive compaction and memory flushes.
• You need consistent, safe self-improvement — the ADL/VFM guardrails let your agent evolve its own operating patterns without drifting or breaking stability.
• You're running autonomous background cron jobs — the skill teaches the distinction between prompted vs. truly autonomous execution so scheduled tasks actually fire and complete.

• If your use case is purely stateless or single-turn — the memory architecture and WAL Protocol add overhead that isn't useful when there's no continuity to preserve.
• If you don't want your agent to take initiative — the proactive surprise and reverse-prompting behaviors are core to this skill's identity; a strictly instruction-following agent is a poor fit.

Before composing any response, the agent scans every human message for corrections, decisions, preferences, and specific values, then writes them to SESSION-STATE.md first. This ensures critical details are never lost to context eviction — the trigger is the human's input, not the agent's memory.

Once context reaches 60%, the agent activates a working buffer that logs every exchange — both the human's message and a summary of the agent's response — to a persistent file. After compaction, the buffer is read first so the agent never has to ask "what were we doing?"

Auto-triggers when a session restarts after truncation. The agent reads memory/working-buffer.md, then SESSION-STATE.md, then daily notes in order, extracting and consolidating context before presenting a recovery summary — without asking the human to re-explain.

The Anti-Drift Limits protocol prevents the agent from adding complexity for its own sake, making unverifiable changes, or sacrificing stability for novelty. The Value-First Modification scoring system ensures every behavioral change must prove compounding leverage before being adopted.

Distinguishes between systemEvent crons (which prompt the main session) and isolated agentTurn crons (which spawn a sub-agent to do the work independently). Prevents the failure mode where scheduled checks never actually execute because the main session is occupied.

Covers skill installation vetting (with guidance that ~26% of community skills contain vulnerabilities), protection against external AI agent networks used for context harvesting, and context leakage prevention before posting to shared channels.

For ongoing work spanning days or weeks, Proactive Agent maintains continuity through SESSION-STATE.md and daily memory logs. The WAL Protocol captures every decision and correction as it happens, so the agent picks up exactly where it left off each session.

Using the isolated agentTurn cron pattern, schedule tasks like memory freshening, log reviews, or security audits to run and complete without requiring main session attention — critical for truly hands-off automation.

The Curiosity Loop, Pattern Recognition Loop, and USER.md/SOUL.md architecture let the agent build a persistent model of who you are and what you value. Repeated requests get detected and proposed for automation; past decisions get followed up on weekly.

Instead of waiting to be asked, the agent surfaces ideas via reverse prompting — asking what would be helpful based on what it knows about you. The Heartbeat System schedules regular "proactive surprise" checks so this behavior fires automatically, not just when prompted.