aaronsb/claude-code-config
Event-driven policy, process, and governance for Claude Code. Ways encode how we do things — prescriptive rules triggered by context, not requested by intent.
What's novel
Event-driven policy, process, and governance for Claude Code. Ways encode how we do things — prescriptive rules triggered by context, not requested by intent.
Code Analysis
25 files read · 6 roundsA context-aware guidance injection system for AI coding agents that uses BM25 and embedding-based semantic matching to fire prescriptive 'ways' (markdown policy documents) at the right moment during a session, with session state tracking, re-disclosure based on token drift, subagent propagation, and
Strengths
Genuinely novel approach to the context-decay problem in LLM agents — the re-disclosure mechanism (token-position-aware re-injection when a way has drifted >25% of context window) is a real engineering solution, not a wrapper. The architecture is clean: thin bash dispatchers → Rust CLI with proper BM25 → C++/llama.cpp embedding engine, with file-based session state that avoids a database dependency.
Weaknesses
Test coverage is thin relative to the codebase size — 9 test files for 51 source files, and the tests appear to be shell-based integration tests rather than unit tests for the Rust/C++ core logic. The hand-rolled JSON parsing in way-embed.cpp, while pragmatic, is a maintenance risk and the C++/Rust boundary (separate processes communicating via files and stdout) adds complexity without a formal protocol.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
308
Contributors
1
Files
200
Active weeks
21
Repository
Language
Shell
Stars
10
Forks
6
License
MIT