面向对象与背景
The project is designed as a readable and runnable engineering reference for understanding how a local coding agent works internally. Beyond coding tasks, it can connect market data tools to analyse stocks and produce investment-oriented suggestions for users.
需要解决的问题
Many agent demos stop at a chat interface and hide orchestration inside one large file. This project needed to make configuration, model providers, the agent loop, tools, permissions, context budgets, and session persistence explicit, while remaining easy to run and extend.
项目目标
- Build a local agent that can start, converse, call tools, and manage persistent sessions.
- Make each stage of the agent runtime understandable and independently modifiable.
- Support multiple model providers and safely extend capabilities through pluggable tools.
- Connect market data sources for stock analysis and clearly framed investment suggestions.
设计约束
- Tool execution needs permission checks and change previews before modifying local files.
- Long-running sessions require context budgeting, compression, recovery, and persistence.
- Provider and data-source credentials must remain outside source code.
- Market outputs are analytical references rather than guaranteed financial advice.
搭建的逻辑框架
- Load environment variables, firstcoder.toml, and global settings through the config module.
- Select an OpenAI-compatible or Anthropic provider and construct the app, session, context, and tool registry.
- Run an agent loop in which the model decides when to call tools, while the runtime checks permissions and previews diffs before execution.
- Write tool results back into context and session state, then compress or restore context when required.
- Keep file, search, Git, web, and market-data capabilities isolated as tools so new providers and data sources can be added without rewriting the core loop.
使用价值
- Delivered a local agent that is runnable, inspectable, and modifiable rather than a conceptual chat shell.
- Separated configuration, providers, orchestration, tools, context, sessions, and application assembly into clear modules.
- Enabled stock-market analysis through extensible data tools and agent-driven research workflows.
- Created a project suitable for learning, technical interviews, and further agent experimentation.
项目总结
- A capable agent depends on explicit runtime mechanics—tool boundaries, permissions, state, context, and recovery—not only model quality.
- Modularity makes it possible to add new tools, providers, and market data sources without destabilising the agent loop.
- Financial analysis needs transparent data sources, assumptions, and risk language to avoid presenting model output as certainty.