10_async_feedback_skill

Overview

This skill enables mid-stream course correction for long-running automation by monitoring a FEEDBACK.md file for user interventions. It lets the agent detect new instructions without restarting the task and optionally acknowledge them. The implementation is in PowerShell and is designed for deterministic, safe pivots during execution.

How this skill works

The agent runs a PowerShell check_feedback script periodically or on demand to inspect the FEEDBACK.md feedback channel for new content. When it detects new instructions, it prints a console message and can append a timestamped acknowledgment to mark the feedback as read. The script returns a success exit code when new feedback is found so caller logic can pivot strategy immediately.

When to use it

• Long-running jobs that may need user guidance mid-execution
• Automated deployments or build pipelines requiring occasional manual input
• Multi-step data processing where upstream decisions can change outcomes
• Agents that must remain non-blocking but responsive to operator edits
• Situations requiring auditable confirmation that feedback was consumed

Best practices

• Keep FEEDBACK.md concise and timestamped so changes are easy to detect
• Run the check periodically at sensible intervals to balance responsiveness and overhead
• Use the acknowledge switch to append a read receipt and avoid duplicate handling
• Treat feedback as authoritative for decision pivots, but validate any critical values before applying
• Combine with safety gates like pre-action guards to prevent unsafe immediate pivots

Example use cases

• A CI pipeline that adjusts test selection mid-run after a stakeholder notes flaky tests
• A data migration that pauses for operator instructions if schema ambiguities are reported
• An infrastructure provisioning job that switches regions when an incident is annotated in FEEDBACK.md
• A long-running analysis where domain experts can inject corrected parameters without restarting