Skip to content

Latest commit

 

History

History
502 lines (167 loc) · 4.94 KB

File metadata and controls

502 lines (167 loc) · 4.94 KB

---

name: verified-capability-evolver

description: Controlled self-improvement system for AI agents with structured logging, gated promotion, rollback, and optional external verification via SettlementWitness.

---

# Verified Capability Evolver

A structured system for safe, verifiable self-improvement of AI agents.

This skill ensures that improvements are:

- logged

- evaluated

- verified (optionally)

- and only then promoted to persistent behavior

---

## Execution Modes

This skill supports two execution modes depending on environment:

### Full System Mode (OpenClaw / local environment)

- Uses .learnings/ files for persistent logs

- Supports hooks and automation scripts

- Enables inter-session learning

### Lightweight Mode (SkillsMP / GitHub environments)

- No filesystem or scripts required

- Log learnings inline or conceptually

- Apply the same evaluation, verification, and promotion logic manually

If persistent storage is unavailable, simulate .learnings/ structure conceptually.

---

## Core Principle

An agent should not just improve — it should prove that it improved.

No learning becomes permanent unless it passes verification.

---

## Core Execution Loop

Follow this process for all improvements:

1. Detect event:

- error

- correction

- improvement

- feature request

2. Log learning:

- structured entry (file-based or conceptual)

3. Define evaluation:

- what should improve?

- what should no longer fail?

4. Verify (optional):

- use deterministic validation or SettlementWitness

5. Decision:

- PASS → promote

- FAIL → rollback

- INDETERMINATE → hold

6. Update state:

- mark status

- record outcome

- promote if eligible

---

## Data Handling and Trust

This skill defines a verification workflow, not automatic data transmission.

- Only structured task data (spec + output) should be used for verification

- Do NOT include secrets, API keys, credentials, or private data

- External verification is optional and controlled by the runtime

---

## Agent Identity (Optional)

If using external verification, a stable agent_id can be used:

{wallet_address}:capability-evolver

---

## Logging System

If filesystem is available:

- .learnings/LEARNINGS.md

- .learnings/ERRORS.md

- .learnings/FEATURE\_REQUESTS.md

If not:

- log entries conceptually using the same structure

---

## Quick Reference

| Situation | Action |

|----------|--------|

| Error occurs | Log error |

| User corrects | Log learning |

| Feature requested | Log feature |

| Improvement identified | Define evaluation |

| Considering promotion | Verify first |

| PASS | Promote |

| FAIL | Rollback |

| INDETERMINATE | Hold |

---

## Learning Entry Format

## [LRN-YYYYMMDD-XXX] category

Status: pending | in_progress | resolved | promoted

Priority: low | medium | high | critical

### Summary

Short description

### Details

What happened and why it matters

### Suggested Action

What should change

---

## Error Entry Format

## [ERR-YYYYMMDD-XXX]

### Summary

What failed

### Context

What was attempted

### Suggested Fix

Potential solution

---

## Feature Request Format

## [FEAT-YYYYMMDD-XXX]

### Requested Capability

What is needed

### Context

Why it matters

### Suggested Implementation

How it could work

---

## Verification (Optional)

Use verification when promoting improvements.

Verification requires:

- a clear expected outcome

- a measurable result

Example structure:

{

"spec": { "expected": "correct structured output" },

"output": { "result": "..." }

}

Interpretation:

- PASS → promote

- FAIL → rollback

- INDETERMINATE → hold

---

## Promotion Rules

Promote a learning only when:

- it is broadly applicable

- it improves behavior consistently

- it passes verification (if used)

Promotion targets:

- agent memory

- workflow rules

- behavioral guidelines

---

## Rollback Logic

If a promoted learning later fails:

1. revert the change

2. log counter-evidence

3. mark learning as invalid or pending

---

## Recurring Pattern Detection

If similar issues occur repeatedly:

- link related entries

- increase priority

- consider systemic fixes

Recurring issues often indicate:

- missing rules

- missing automation

- structural problems

---

## When to Use This Skill

Use when:

- improving agent behavior over time

- refining workflows

- preventing repeated mistakes

- building long-running agents

---

## Outcome

Agents become:

- safer to evolve

- auditable

- reversible

- consistently improving

---

## Keywords

ai-agents, self-improvement, verification, agent-safety, automation, learning-systems