[Feature]: Custom system prompt per AI mode

[safranklin]

Feature description

AI modes in waveai.json let you configure model, endpoint, provider, and display settings, but there's no way to set a system prompt. Every chat starts with Wave's built-in context, and you can't change what the AI knows or how it behaves before you start talking to it.

Adding an ai:systemprompt field to the mode config would fix that. A few places where this matters:

• Right now the only thing differentiating modes is the model. The "personality" is always the same. With a system prompt you could have a DevOps mode, a code review mode, and a general assistant mode and switch between them from the dropdown.
• Some people want code-only responses. Others want explanations. A system prompt per mode would let you set that once instead of repeating yourself every chat.
• If you're always working in the same stack (say, Go + sqlc + templ), you could bake the project context into the prompt so the AI doesn't start from zero each time.
• Non-English users could set "Always respond in Spanish" and not have to ask twice.

You could also pair this with model selection in interesting ways. Cheap model + opinionated system prompt for quick Q&A, frontier model + detailed system prompt for architecture review.

Implementation suggestion

Support both inline strings and file paths:

{
  "devops-mode": {
    "display:name": "DevOps Assistant",
    "ai:provider": "openrouter",
    "ai:model": "anthropic/claude-sonnet-4",
    "ai:systemprompt": "You are a senior DevOps engineer. Focus on Kubernetes, Terraform, and CI/CD. Prefer kubectl commands and IaC over manual steps.",
    "ai:capabilities": ["tools"]
  },
  "project-mode": {
    "display:name": "Project Assistant",
    "ai:provider": "openrouter",
    "ai:model": "anthropic/claude-sonnet-4",
    "ai:systempromptfile": "~/.config/waveterm/prompts/project-context.md",
    "ai:capabilities": ["tools"]
  }
}

Looking at the codebase, getSystemPrompt() in pkg/aiusechat/usechat.go already returns a []string that each API backend (OpenAI Chat, OpenAI Responses, Anthropic, Google) formats appropriately. The user-defined prompt could be appended to that array after the built-in system prompt is selected, so it layers on top of Wave's default context rather than replacing it.

The changes would touch:

• pkg/wconfig/settingsconfig.go - add SystemPrompt and SystemPromptFile fields to AIModeConfigType
• pkg/aiusechat/usechat.go - read the field in getSystemPrompt() or just after it, resolve file paths, append to the []string
• schema/waveai.json - add the new fields to the JSON schema

Anything else?

I'd be interested in taking a shot at implementing this. Wanted to check alignment first per the contributing guide.