openclaw - 💡(How to fix) Fix [Feature]: Extend Skills system with runtime feedback hooks and optional skill update suggestions

Summary

Currently, skills are only loaded at startup and remain static during runtime. This proposal adds an optional runtime feedback mechanism that allows skills to receive execution context and generate improvement suggestions without modifying core agent behavior.

Problem to solve

OpenClaw's current Skills system is initialized at startup and remains static throughout the session.

While this design ensures stability and predictability, it introduces limitations in long-running or multi-step workflows:

• Skills cannot adapt based on real execution outcomes
• There is no standardized mechanism to evaluate whether a skill performed well
• Improvement ideas must be manually written and redeployed
• Advanced users building "self-improving workflows" must implement ad-hoc logging and external systems

As a result, Skills behave as static capabilities rather than evolving operational components.

Proposed solution

Introduce an optional extension layer for Skills without changing the core execution model.

1. Runtime Skill Feedback Hook (optional)

After a skill is executed, emit a structured event:

skillExecutionCompleted:

• skillName
• input context
• output result
• success/failure
• execution duration
• optional error trace

2. Optional Skill Improvement Suggestion Field

Allow plugins or external listeners to attach:

skillImprovementSuggestion:

• suggested description update
• edge cases discovered
• performance notes
• alternative approaches

3. Non-intrusive design

• Core behavior remains unchanged
• No automatic skill modification required
• Suggestions are advisory only
• Backward compatible with existing Skills system

4. Optional storage layer (implementation neutral)

• Could be logged via existing session logs
• Or exposed to plugin layer via hooks

Alternatives considered

A full self-learning or auto-modifying Skills system was considered, but would introduce significant risks in stability, reproducibility, and debugging complexity.

Instead, this proposal focuses on a lightweight feedback layer that enables external systems or plugins to implement learning behavior safely.

Impact

Affected users:

• Advanced automation users
• Multi-agent system builders
• Plugin developers

Severity: Medium

Frequency: High (every skill execution)

Benefits:

• Enables reproducible "learning loop" architectures externally
• Reduces need for custom logging frameworks
• Provides foundation for future self-improving workflows
• Keeps core system stable while enabling ecosystem innovation

Evidence/examples

In multi-agent workflows (e.g., Feishu-based deployments), users often run repeated skill executions with slight variations.

Without runtime feedback:

• failures are not systematically captured
• improvements are not structured
• skill tuning is manual and inconsistent

With this proposal:

• each execution can generate structured metadata
• external systems can aggregate performance patterns
• future improvements can be derived systematically without modifying core OpenClaw behavior

Additional information

This proposal does NOT introduce self-modifying behavior.

It only exposes structured runtime feedback for Skills, enabling external systems and plugins to implement learning, analytics, or optimization layers.

This keeps OpenClaw's core deterministic while unlocking advanced ecosystem-level innovation.