Coding Partner

This recipe builds a small custom coding assistant.

The point is not to expose an agent loop. The point is to show how a normal Axon agent is authored: src/ defines who the agent is and how tasks are framed, server/ defines how the world reaches the agent, data/ holds durable reference material, modules/ adds packaged capabilities, and policy constrains what the agent may do.

The agent handles interactive chat and headless API review.

Folder structure

coding-partner/
├── axon.config.ts
├── data/
│   └── knowledge/
│       └── repo-guidelines.md
├── modules/
│   └── typescript/
├── server/
│   └── api/
│       ├── chat.post.ts
│       └── repo-review.post.ts
└── src/
    ├── boot.vue
    ├── prompts/
    │   ├── chat-context.vue
    │   ├── command-approval.vue
    │   ├── repo-review.vue
    │   └── components/
    │       ├── identity.vue
    │       ├── personality.vue
    │       └── coding-standards.vue
    └── tools/
        └── github.ts

How it works

The chat route handles interactive conversation. It renders chat-context.vue with the current request and calls axon.stream() on a named thread — the thread name keeps context across messages in the same conversation.

The review route runs a stateless task. It renders repo-review.vue with the target repo and branch, then calls axon.stream() with no thread — each review is independent.

Both routes use the same agent identity and tools. The route context (task scope, policy constraints) changes the agent's behavior for that invocation.

What to read first

Start with boot.vue to see how identity and operating rules are composed, then read chat.post.ts to see how the TUI path prepares a task.