halthelobster/proactive-agent

Overview

This skill transforms task-following agents into proactive, self-improving partners that anticipate needs, persist through context loss, and improve safely over time. v3.0.0 introduces the WAL Protocol, a Working Buffer for the danger zone, Compaction Recovery, unified search, and hardened security patterns. It focuses on delivering continuous value while minimizing context leakage and drift.

How this skill works

The agent inspects every incoming message for corrections, decisions, names, preferences, and specific values and writes critical details to a session-level working file before responding (WAL Protocol). When session context nears limits, it switches to a persistent Working Buffer that captures every exchange surviving compaction. If context is truncated, Compaction Recovery reads the buffer and other sources to restore state. Unified search checks all memory layers before admitting ignorance.

When to use it

• When you want an agent that anticipates work and offers helpful suggestions without being prompted
• When sessions risk losing important context due to token limits or compaction
• When you need robust, auditable memory capture of decisions and corrections
• When deploying agents in environments that require strict security and installation vetting
• When you want an agent that tries multiple approaches before asking for help

Best practices

• Always write critical details to SESSION-STATE.md (WAL) before composing a reply
• Enable the Working Buffer once session context exceeds the danger threshold (~60%)
• Use the Compaction Recovery flow to restore tasks instead of asking the user to repeat context
• Run the security installation checklist before adding external skills or integrations
• Score self-modifications with the VFM protocol and block changes under the ADL thresholds

Example use cases

• A product manager hands off multi-step requirements; the agent logs decisions and proactively drafts follow-ups
• A long debugging session hits token limits; the agent recovers progress from the working buffer and resumes
• An operations agent vets and refuses risky external skill installs based on source checks and patterns
• A personal assistant notices overdue tasks and proactively surfaces an actionable summary with next steps
• An engineering agent tries multiple build/fix strategies before escalating to a human