open-gitagent/gitclaw
A universal git-native AI agent framework. Your agent lives inside a git repo — identity, rules, memory, tools, and skills are all version-controlled files.
What's novel
A universal git-native AI agent framework. Your agent lives inside a git repo — identity, rules, memory, tools, and skills are all version-controlled files.
Code Analysis
11 files read · 3 roundsA git-native AI agent framework where the agent's identity, memory, rules, skills, and learning state are all version-controlled files in a git repository, with a reinforcement-learning loop that crystallizes successful task approaches into reusable skills.
Strengths
The learning loop (task tracking → skill evaluation → crystallization → reinforcement with asymmetric confidence adjustment) is a genuinely well-designed, non-trivial implementation that goes well beyond typical agent frameworks. Security considerations are thoughtfully addressed (argv-based git clone to prevent RCE, path traversal guards in hooks, kebab-case validation, quote escaping in commit messages).
Weaknesses
The token estimation (4 chars per token) is a rough heuristic that could lead to inaccurate context window management, and the skill search relies on simple keyword overlap/Jaccard similarity rather than semantic matching, limiting its effectiveness for paraphrased objectives. Test coverage appears thin relative to the breadth of features (8 test files for ~100+ source files), and several critical paths like the MCP manager and plugin loading lack visible test coverage.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
58
Contributors
1
Files
67
Active weeks
2
Repository
Language
TypeScript
Stars
138
Forks
19
License
MIT