A public lab for independent development in the AI era
Mafan
Independent developer building AI products.
Exploring how individuals build software, work with AI agents, and design durable systems for work and life.
Turning ideas into things that run.
CodeLeveler
AI Coding Agent
A local-first coding agent runtime that inspects, edits, runs, and verifies real repositories.
MuxLayer
AI Infrastructure
A local model control layer that routes, converts, fails over, and traces requests for coding agents.
ReviewGate
AI Code Review
A pre-merge quality gate for AI-assisted code that prioritizes high-risk findings and suppresses low-value review noise.
02 / Latest writing
Judgments, process, and questions that remain open.
Traditional Tests Pass, but How to Prove the Agent Is Actually Better
Traditional tests only prove the code isn't broken; they can't evaluate whether the Agent has truly improved. Eval is needed: run a fixed task set with unified scoring rules repeatedly, comparing success rate, cost, time, and regressions. The two systems complement each other; before release, pass traditional tests first, then run Eval, to avoid optimizing by feel.
I Built a Coding Agent and Realized Calling the Model Is Just the Beginning
Building a coding agent taught me that calling the model is just the beginning. In long tasks, state management, tool logging, observability, and failure recovery are key. Chat is just the entry point; the system must act like a runtime to keep progressing and validating results. Reliability often matters more than a one-off impressive answer.
ChatGPT and Codex Merge, Protocol Changes Again
After ChatGPT and Codex merged, the protocol changed again: tool definitions moved into additional_tools, and third-party models that don't receive tools may fake calls. The author's local gateway AgentGate has adapted, restoring tools and forwarding upstream; the merged desktop client can still point to localhost to call DeepSeek and other models.
Sharing a Small Tool I Use Myself (Self-Made)
ReviewGate is an open-source MIT-licensed pre-merge filtering tool that screens for high-risk issues from perspectives like security, performance, logic, and AI Smell, categorizing them as block or warn. It prefers to fail rather than give a false green light. It supports --fix, CI integration, and false-positive sharing, and is read-only by default. It's recommended to observe first, then gradually enable.