I’ve been development it solo since February as a facet venture, and I used it as a possibility to experiment with context potency and agent-driven construction.
3 issues make Willing Code other from different an identical merchandise:
1. Constructed via brokers
Willing Code was once constructed from scratch the usage of state of the art coding brokers. My position was once to behave because the human orchestrator: writing activates and necessities, then reviewing the designs and code produced via brokers.
To stay this clear, the repo comprises an ai-interactions folder with activates and output medical doctors. Extra: https://mochow13.github.io/keen-…
2. Flip reminiscence
To steer clear of filling the context window all the way through multi-turn loops, Willing discards uncooked instrument inputs and outputs after each and every flip. It helps to keep a distilled “flip reminiscence” as a substitute: a easy deterministic Pass struct handed into the following flip.
Extra right here: https://mochow13.github.io/keen-…
3. Talents-driven MCP servers
As a substitute of loading huge MCP server schemas into context in advance, Willing abstracts MCP equipment into native markdown Talents. It best retrieves the precise JSON schema when the LLM requests a particular instrument at runtime.
Main points: https://mochow13.github.io/keen-…
I’ve been the usage of Willing to increase Willing itself, in addition to in my different tasks. I’m taking a look ahead to questions, comments, tips, and evaluations. I’m dedicated to making improvements to the venture over the longer term.
Thank you prematurely!
