lora/docs/specialists/prism.md

# Prism — TypeScript Specialist LoRA

Adapter codename: `prism`
Agent: `build-ts`
Base model: Qwen/Qwen3.5-27B

## Objective

Teach the model TypeScript-idiomatic code generation aligned with the build-ts
agent's system prompt. The adapter should internalize:

- Strict TypeScript — no `any` without justification
- `const` over `let`, never `var`
- Named exports, `type` over `interface` (unless extending)
- Package manager detection (bun/pnpm/npm)
- Framework-aware patterns (Next.js, Vite, Astro, SvelteKit, Remix, Bun)
- tsc → lint → test verification cycle

## Data Sources

### Session extraction (~200–350 examples)

Classify from opencode `build` sessions by:
- File paths: `.ts`, `.tsx`, `.js`, `.jsx`, `package.json`, `tsconfig.json`
- Bash commands: `npm`, `pnpm`, `bun`, `npx`, `tsc`, `vitest`, `jest`, `biome`
- Error patterns: `TS2`, `TS7`, `error TS`, type errors
- Framework config files in context: `next.config.*`, `vite.config.*`

### Git repo mining (~150–250 examples)

Target repos:
- `~/.config/opencode/plugins/` — opencode plugins (TypeScript)
- `madcat-os/sere-kit` — UI component library
- `madcat-os/visor` — HUD/display system
- Any `package.json`-rooted repos in madcat-os org

Extract patterns:
- React component implementations (hooks, state management)
- Plugin/tool development (opencode tools, MCP servers)
- Type-safe API integrations
- Build configuration (tsconfig, vite, next)

### Synthetic (~50–100 examples)

- Generic type patterns and constraints
- Discriminated unions and type narrowing
- Zod schema → TypeScript type inference
- React Server Components vs Client Components
- Edge runtime constraints (Cloudflare Workers, Vercel Edge)

## Example Format

```json
{
  "messages": [
    {
      "role": "system",
      "content": "<build-ts agent system prompt>"
    },
    {
      "role": "user",
      "content": "Add a new visor widget component for displaying memory graph connections"
    },
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [
        {
          "type": "function",
          "function": {
            "name": "glob",
            "arguments": {"pattern": "src/components/**/*.tsx"}
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "<file list>"
    },
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [
        {
          "type": "function",
          "function": {
            "name": "read",
            "arguments": {"filePath": "src/components/widgets/LogWidget.tsx"}
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "<existing widget pattern>"
    },
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [
        {
          "type": "function",
          "function": {
            "name": "write",
            "arguments": {
              "filePath": "src/components/widgets/GraphWidget.tsx",
              "content": "<new component following existing patterns>"
            }
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "File written successfully"
    },
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [
        {
          "type": "function",
          "function": {
            "name": "bash",
            "arguments": {"command": "npx tsc --noEmit"}
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "<type check passes>"
    },
    {
      "role": "assistant",
      "content": "GraphWidget created. Follows LogWidget's pattern — named export, typed props, const component. Type check clean."
    }
  ]
}
```

## Evaluation Criteria

1. Generated TypeScript passes `tsc --noEmit` with strict mode
2. No `any` types without explicit justification
3. Correct import/export patterns (named exports)
4. Framework-appropriate patterns (detects Next.js vs Vite etc.)
5. Package manager awareness (correct lockfile, correct run command)
6. Tool call sequence: explore → read → implement → type-check
7. React patterns: proper hook usage, no stale closures

## Training Config Overrides

```python
MAX_SEQ = 8192
LR      = 5e-5
```

## Estimated Size

- 400–600 examples total (largest dataset — most session history)
- ~3M tokens
- Training time: ~3 hrs on H100
- Adapter size: ~305 MB