Zeelin Auto Evolution

Version: 0.1.0

English | 简体中文

Overview

Zeelin Auto Evolution adds a safe, self-improving feedback loop to an agent system. It silently captures corrective user feedback, turns repeated issues into rule proposals, suggests improvements for low-performing skills, and proposes new skills when repeated workflows are not covered. It is intentionally conservative: it can observe automatically, but it must ask for confirmation before changing rules or skills.

What It Does

• Silently captures corrective feedback from user conversations
• Aggregates repeated feedback into candidate rules
• Scores skill executions across accuracy, coverage, efficiency, and satisfaction
• Suggests skill improvements when performance stays low
• Suggests creating new skills when a repeated pattern has no existing coverage

Multilingual Support

This 0.1.0 version supports both Chinese and English feedback signals.

Examples of supported Chinese signals:

• 不是这样
• 你又忘了
• 不对
• 我不是让你这么干

Examples of supported English signals:

• that's not right
• you forgot again
• this is wrong
• that's not what I asked
• don't do it this way

Detection is implemented in scripts/detect_feedback_signal.py.

Trigger Model

The skill is designed around two automatic checkpoints:

1. Per-message detection When a user sends a message, a hook or controller can run the detection script. If the message contains corrective feedback, the controller should finish the current task first and then quietly dispatch a feedback-observer.

2. Session-start evolution scan At session start, or when the user explicitly asks to check evolution proposals, the controller can run scripts/evolution_runner.py to scan accumulated feedback and generate proposals.

Evolution Layers

1. Silent feedback capture
2. Rule graduation after repeated occurrences
3. Skill optimization from low scores
4. New skill proposals for uncovered repeated workflows

Repository Structure

auto_evolution_skill/
├── SKILL.md
├── README.md
├── README.zh-CN.md
├── REFERENCE_ARCHITECTURE.md
├── EXAMPLES.md
├── scripts/
│   ├── detect_feedback_signal.py
│   └── evolution_runner.py
└── templates/
    ├── feedback_index_template.md
    └── feedback_topic_template.md

Quick Start

1. Connect the detection script to your message-submit hook.
2. Dispatch a feedback-observer when corrective feedback is detected.
3. Store structured feedback in .claude/feedback/.
4. Run evolution_runner.py at session start or on demand.
5. Show proposals to the user before applying any rule or skill change.

Related Files

• SKILL.md: core skill behavior and workflow
• REFERENCE_ARCHITECTURE.md: architecture and decision rules
• EXAMPLES.md: example scenarios and outputs
• README.zh-CN.md: Chinese documentation

Notes

• This release does not auto-edit rules.
• The skill is event-driven, not a background daemon.
• High-quality routing still depends on your controller or hook integration.

Changelog

• 0.1.0 (2026-04-20): Added English support, split README into English and Chinese versions, and prepared the package for open-source publishing.