@auto-hwp/ai-protocol

The vendor-neutral, isomorphic LLM protocol for auto-hwp chat-editing — the L2' of the SDK layers. It is types + pure transforms + validation only: the doc-context builder, the system prompt (an excerpt of INTENT-SCHEMA), and the request/response validators. The server proxy and the browser client import the SAME module, so the wire contract can never drift between them.

Zero fetch. Zero LLM client. Zero keys. Which model, which vendor, streaming or not — all the host's choice (R6). This package never talks to a network.

// ── server proxy (your key lives here) ──────────────────────────────
import { buildSystemPrompt, buildUserMessage, validateRequest, validateResponse } from "@auto-hwp/ai-protocol";

const check = validateRequest(await req.json());
if (!check.ok) return badRequest(check.error);

const text = await callYourModel({                 // ← YOUR vendor, YOUR key
  system: buildSystemPrompt(),                      // INTENT-SCHEMA excerpt; pass { allowedIntents } to narrow
  user: buildUserMessage(check.value),              // R5 <document-content> fence applied here
});
return json({ intents: validateResponse(text, { onDrop: console.warn }) }); // whitelist + structure

// ── browser client (no key, no model) ───────────────────────────────
import { buildDocContext } from "@auto-hwp/ai-protocol";
const docContext = buildDocContext(meta, anchors); // the string POSTed to your proxy

API

| Export | Role | |--------|------| | EditRequest / EditResponse / Intent / Anchor / INTENT_VERSION | the wire types (schema v0). | | buildDocContext(meta, anchors, { maxLen? }) | the doc-context string (header + anchor lines; anchor text is untrusted). | | buildUserMessage(req) | the LLM user turn with the R5 <document-content> fence. | | buildSystemPrompt({ allowedIntents? }) | the system prompt — an excerpt of INTENT-SCHEMA (source-line comments preserved), optionally a subset. | | validateRequest(body, limits?) | input guard (length/count caps). | | validateResponse(textOrArray, { allowedIntents?, onDrop? }) | parse LLM text → whitelist + structure check. | | DEFAULT_ALLOWED_INTENTS / DEFAULT_LIMITS | the frozen whitelist + caps. |

Bring your own vendor

The only host-specific glue is the model call itself. Everything before it (prompt + user message) and after it (parse + whitelist) is here and shared. apps/hwp-lab/src/app/api/hwp-edit/route.ts is a documented reference proxy (Anthropic) — copy it for any vendor.

Develop

npm run build       # tsc → dist (JS + .d.ts)
npm run typecheck   # tsc --noEmit
npm test            # vitest (node env): context/prompt/validate — isomorphic