[FEATURE] improve feature-dev plugin to auto breakdown a big feature to more smaller phases to avoid hitting the 200k token window limit.

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Start a brand new session, with auto‑compact turned off, and use feature-dev plugin.
I did not expect the implementation of my feature would consume many tokens. I have selected the best solutions proposed by Claude, which are also the most complicated solutions.
So it went into the implementation plan—working autonomously. Until I checked and saw that the token window was almost hit 200k. Luckily, the mission has completed with 0.7% ≈ 1467 tokens left.

<img width="2338" height="1578" alt="Image" src="https://github.com/user-attachments/assets/b7503f71-181b-4f0c-8a3f-46ad0f96003f" />

Proposed Solution

Break down to smaller steps:

• feature-dev:plan,
• feature-dev:apply,
• feature-dev:review

Alternative Solutions

After the planning step, estimate the total token usage and check if it can hit the limit. If it can hit the limit, propose to write down the plan and continue with a new session.

Priority

High - Significant impact on productivity

Feature Category

Developer tools/SDK

Use Case Example

I am working on a terminal music application called music-cli. By default, it only has one model for music generation. So I start the feature with a simple prompt to add more supported models (from Hugging Face). It asked me questions about model sections, configuration, architecture design, etc. Then it reached a situation where I selected the most complex solutions.

All feature were implemented in version 0.5.0:

• AudioLDM models: audioldm-s-full-v2, audioldm-l-full for sound effects and ambient audio
• Bark models: bark, bark-small for speech synthesis
• MusicGen models: All existing models continue to work
• Add ai models command to list all available AI models
• Add LRU cache for AI models with configurable size (default: 2 models)
• Add download progress bar during model downloads
• Add GPU memory management with automatic cleanup on model eviction
• Default model: musicgen-small

GitHub repo: https://github.com/luongnv89/music-cli

Additional Context

I have autocompact turn off
I am using Claude Code 2.0.76, with Opus 4.5 as the model.