[Epic] Issue-Driven AI-Based Codegen & Routing Mechanism

[0xaskr]

Goal

Establish a mechanism similar to @sii-xinglong/Glaucis, leveraging Claude for kernel code generation/fusion codegen, with issue-driven optimization workflows and shape-based routing to the most capable AI implementation.

Motivation

Automate and optimize kernel code generation for the project, facilitate experimentation with AI-driven backends, and transparently route computation to the most capable solution per shape/configuration.

Scope

In Scope

• Claude-based automated code generation and fusion codegen
• Issue-driven optimization workflow (issue → parsing → codegen trigger)
• Shape-based routing layer directing configurations to AI implementations
• Scoring mechanism evaluating backend support, performance, and compatibility

Out of Scope

• Non-AI manual kernel implementations
• Production deployment infrastructure (initial phase is experimentation)

Sub-Tasks

• Design issue-driven workflow: issue creation → parsing → codegen trigger
• Implement shape-based routing layer mapping dimensions/configurations to AI implementations
• Build scoring mechanism to evaluate AI backend capabilities per shape
• Integrate Claude for code generation and fusion codegen
• Document end-to-end example flows from issue creation to routed codegen
• Ensure extensibility to other AI backends

Success Criteria

• Users can submit optimization requests as GitHub Issues and trigger automated codegen
• Routing layer correctly directs shapes to the highest-scoring AI implementation
• Generated code covers necessary fusion opportunities and kernel optimizations
• Workflow is well-documented with clear example flows

Milestones / Timeline

Milestone	Target Date	Status
Issue-driven workflow design
Routing layer + scoring mechanism
Claude codegen integration
Documentation + examples

Dependencies

• None

Additional Context

• Reference mechanism: @sii-xinglong/Glaucis
• Consider extensibility to other AI backends in the future