Hey everyone,
I’ve been building Codemate — an open source coding agent with two things I haven’t seen in other tools:
1. It learns from its own mistakes
After every task, Codemate extracts structured lessons from what went wrong and how it was fixed. These are classified automatically as either project-specific or globally reusable across projects.
The next time it runs a similar task, those lessons are injected before execution starts. It doesn’t repeat the same mistakes twice — and gets better the more you use it.
2. Parallel multi-agent execution via TaskGraph
Instead of one agent doing everything sequentially, work is split into a dependency graph and executed in parallel: planner → builds the TaskGraph coder + tester → run simultaneously reviewer → waits for both, then checks writer → persists lessons + changelog at the end. Each agent has strict permissions. Tester can only write test files. Writer is the only one that can touch lessons and changelog. Orchestrator can’t directly edit files — everything goes through the TaskGraph.
Other things worth knowing:
- Intent-drift detection: every 5 subtasks, it checks whether execution is still aligned with your original goal. If it drifts, it pauses and asks you.
- Research cache: search results are stored with fingerprint matching and reused across sessions — no redundant searches.
- Self-check loop: if tests fail, it automatically retries up to 5 times before escalating to you.
Built with Bun + TypeScript, MIT licensed, open source.
GitHub: GitHub - Wholiver/codemate: Building Codemate — a coding agent that learns from its own mistakes. TaskGraph scheduling, cross-project memory, and intent-drift detection. Open source, MIT licensed. · GitHub
npm: npm install -g @codemate-ai/cli
Happy to answer any questions!