feat: update chat to use patchwork-worker + llm-host arch

chat/package.json

@@ -1,9 +1,9 @@
 {
   "name": "@patchwork/chat",
-  "version": "0.4.0",
+  "version": "0.5.0",
   "description": "chitterchatter: chat with emoji reactions, presence/typing indicators, image paste, voice notes, and message replies.",
   "type": "module",
-  "main": "./main.js",
+  "main": "./dist/index.js",
   "exports": {
     ".": "./dist/index.js"
   },
@@ -25,8 +25,8 @@
   "dependencies": {
     "@automerge/automerge": "^3.2.1",
     "@automerge/automerge-repo": "^2.5.0",
-    "@chee/patchwork-llm": "^0.2.1",
     "@chee/patchwork-transcript": "^0.1.0",
+    "@grjte/patchwork-worker": "^0.0.2",
     "@inkandswitch/patchwork-elements": "^6.0.2",
     "@inkandswitch/patchwork-filesystem": "^0.2.9",
     "cute.txt": "^0.0.6",

libraries/llm/.gitignore

@@ -1 +1,2 @@
 types/
+.pushwork

libraries/llm/README.md

@@ -102,6 +102,32 @@ Pass a stable `sessionKey` (e.g. a doc URL) to `generate`/`stream`; after a
 reload, `resume(sessionKey, { onToken, onDone })` re-attaches to the still-running
 stream in the worker.
 
+### Request preparation (`request.js`)
+
+`generate()` is two halves: preparing the request and posting it to the worker.
+The first half is exported so a host-side worker spec (one that serves this
+library over a stream to sandboxed tools) prepares requests exactly the way
+`generate()` does:
+
+```js
+import { prepareGenerate, buildGeneratePayload } from "@patchwork/llm"
+
+const { cfg, config, extraSystem, builtin, input } = await prepareGenerate(cfg0, {
+  messages,          // chat messages or a string
+  system,            // extra system prompt, already provider-resolved
+  tools,             // tool descriptors, already filtered
+  continuation,      // treat a string as a raw continuation
+  overrides,         // passed to callConfig verbatim
+})
+if (builtin) { /* run builtinGenerate(input, { system: effectiveSystem(cfg, extraSystem), … }) */ }
+else worker.postMessage(buildGeneratePayload(config, input, { id, sessionKey }))
+```
+
+Resolving a provider-conditional `system` map, filtering tools by the user's
+`toolToggles`, and choosing which `overrides` to forward are the caller's job:
+`callConfig` honours provider / apiKey / model / url overrides, so a host serving
+untrusted tools passes only sampling knobs.
+
 ## Events reference
 
 | event        | fields                                                            | local | openrouter | ollama |

libraries/llm/client.js

@@ -13,7 +13,8 @@
 
 import {readConfig, ensureConfig, callConfig, applyPrompts, effectiveSystem} from "./config.js"
 import {builtinGenerate} from "./builtin.js"
-import {resolveTools, toToolSchemas, buildToolsSystem, parseToolCalls, runTool, resolveCfgPrompts, sanitizeToolName} from "./tools.js"
+import {resolveTools, buildToolsSystem, parseToolCalls, runTool, resolveCfgPrompts, sanitizeToolName} from "./tools.js"
+import {prepareGenerate, buildGeneratePayload, NATIVE_TOOL_PROVIDERS, TEMPLATE_TOOL_PROVIDERS} from "./request.js"
 
 /**
  * Per-call options. A superset of every option any exported function accepts;
@@ -66,16 +67,7 @@ import {resolveTools, toToolSchemas, buildToolsSystem, parseToolCalls, runTool,
 
 /** @typedef {{post: (m:any)=>void}} Connection */
 
-/** CallConfig plus the extra mutable fields the client tacks on per-call.
- * @typedef {import("./config.js").CallConfig & {tools?:any, toolSystem?:string, continuation?:boolean}} CallConfigExt */
-
-/** @typedef {{type:string, id:string, sessionKey:string, provider:import("./config.js").ProviderId, config:import("./config.js").CallConfig, text?:string, messages?:any}} GeneratePayload */
-
-// Providers with real function-calling APIs (the worker passes tool schemas and
-// parses structured tool_calls). Everything else (local transformers, Chrome
-// built-in) uses the <tool_call> XML prompt convention, parsed from the text.
-const NATIVE_TOOL_PROVIDERS = new Set(["openrouter", "ollama", "webllm"])
-const TEMPLATE_TOOL_PROVIDERS = new Set(["local"])
+/** @typedef {import("./request.js").CallConfigExt} CallConfigExt */
 
 /** @type {Connection|null} */
 let connection = null
@@ -197,34 +189,21 @@ function getConnection() {
  */
 export async function generate(messages, opts = {}) {
 	const cfg0 = opts.config ?? (await ensureConfig(opts.scope))
-	// Resolve the selected system/pre prompt docs → their text. repo.find is
-	// cached, so this is cheap after first load.
-	const cfg = await resolveCfgPrompts(cfg0)
-	/** @type {CallConfigExt} */
-	const config = callConfig(/** @type {any} */ (cfg), /** @type {any} */ (opts))
-
-	// Tools: native providers get JSON schemas on `config.tools`; the rest get the
-	// <tool_call> XML convention prepended to the system prompt (parsed from text).
-	const hasTools = Array.isArray(opts.tools) && opts.tools.length > 0
-	const native = hasTools && NATIVE_TOOL_PROVIDERS.has(config.provider)
-	const templated = hasTools && TEMPLATE_TOOL_PROVIDERS.has(config.provider)
-	if (native || templated) config.tools = toToolSchemas(opts.tools)
-	if (templated) config.toolSystem = buildToolsSystem(opts.tools)
-	const extraSystem =
-		hasTools && !native && !templated
-			? [buildToolsSystem(opts.tools), opts.system].filter(Boolean).join("\n\n")
-			: opts.system
+	// Prompt resolution, CallConfig, the tools decision, the builtin detour and
+	// the chat-vs-continuation input shape all live in request.js, shared with
+	// host-side worker specs. The whole opts is passed as overrides: a same-realm
+	// caller legitimately owns provider/model/apiKey overrides.
+	const {cfg, config, extraSystem, builtin, input} = await prepareGenerate(cfg0, {
+		messages,
+		system: opts.system,
+		tools: opts.tools,
+		continuation: opts.continuation,
+		overrides: opts,
+	})
 
 	// Built-in (Chrome Prompt API) runs on the main thread, not the worker.
-	if (config.provider === "builtin") {
-		const pre = cfg.resolved?.pre || ""
-		const text =
-			typeof messages === "string"
-				? pre
-					? pre + "\n\n" + messages
-					: messages
-				: messages
-		return builtinGenerate(text, {
+	if (builtin) {
+		return builtinGenerate(input, {
 			temperature: config.temperature,
 			topK: config.topK,
 			system: effectiveSystem(/** @type {any} */ (cfg), extraSystem),
@@ -234,19 +213,6 @@ export async function generate(messages, opts = {}) {
 		}).then((t) => ({text: t, toolCalls: null, stats: null}))
 	}
 
-	// A string input is CHAT by default — wrapped as a user turn, so instruct/chat
-	// models respond normally and the system prompt applies. It's a raw
-	// CONTINUATION only when opts.continuation is set: raw-fed for
-	// local/webllm/ollama, and CONTINUE_SYS-framed for chat-only OpenRouter (see
-	// the worker). Loom passes continuation:true; other callers get plain chat.
-	const asContinuation = !!opts.continuation && typeof messages === "string"
-	const prepared = asContinuation
-		? messages
-		: typeof messages === "string"
-			? [{role: "user", content: messages}]
-			: messages
-	// Prepend the configured system + pre-prompt (and any tool-supplied system).
-	const input = applyPrompts(prepared, /** @type {any} */ (cfg), extraSystem)
 	const conn = getConnection()
 	const id = nextId()
 	const sessionKey = opts.sessionKey || id
@@ -297,11 +263,7 @@ export async function generate(messages, opts = {}) {
 			opts.signal.addEventListener("abort", onAbort)
 		}
 		// A string is a raw continuation prompt; an array is chat messages.
-		/** @type {GeneratePayload} */
-		const payload = {type: "generate", id, sessionKey, provider: config.provider, config}
-		if (typeof input === "string") payload.text = input
-		else payload.messages = input
-		conn.post(payload)
+		conn.post(buildGeneratePayload(config, input, {id, sessionKey}))
 	})
 }
 

libraries/llm/index.js

@@ -21,92 +21,117 @@
  */
 
 export {
-	// config (account doc + patchwork:llm-config provider)
-	readConfig,
-	writeConfig,
-	callConfig,
-	normalizeConfig,
-	subscribeConfig,
-	resolveConfig,
-	ensureSettingsDoc,
-	ensureConfig,
-	// per-tool / per-doc whole-config overrides
-	scopedRaw,
-	hasScopeOverride,
-	readScopedConfig,
-	writeScopeOverride,
-	clearScopeOverride,
-	settingsDocHandle,
-	applyPrompts,
-	effectiveSystem,
-	DEFAULTS,
-	PARAM_KEYS,
-	PROVIDER_CAPS,
-	TOOL_STORAGE_ID,
-	CONFIG_SELECTOR,
-	// catalogues / labels
-	LOCAL_MODELS,
-	WEBLLM_MODELS,
-	fetchOpenRouterModels,
-	fetchOllamaModels,
-	describeConfig,
-} from "./config.js"
+  // config (account doc + patchwork:llm-config provider)
+  readConfig,
+  writeConfig,
+  callConfig,
+  normalizeConfig,
+  subscribeConfig,
+  resolveConfig,
+  ensureSettingsDoc,
+  ensureConfig,
+  // per-tool / per-doc whole-config overrides
+  scopedRaw,
+  hasScopeOverride,
+  readScopedConfig,
+  writeScopeOverride,
+  clearScopeOverride,
+  settingsDocHandle,
+  applyPrompts,
+  effectiveSystem,
+  DEFAULTS,
+  PARAM_KEYS,
+  PROVIDER_CAPS,
+  TOOL_STORAGE_ID,
+  CONFIG_SELECTOR,
+  // catalogues / labels
+  LOCAL_MODELS,
+  WEBLLM_MODELS,
+  fetchOpenRouterModels,
+  fetchOllamaModels,
+  describeConfig,
+} from "./config.js";
 
 export {
-	generate,
-	generateWithTools,
-	stream,
-	predict,
-	scoreTokens,
-	preload,
-	abort,
-	resume,
-	onStatus,
-	registerLocalModel,
-	computeImportance,
-	computeAttentionWeights,
-	extractFeatures,
-	extractCutFeatures,
-	decodeTokens,
-	probeAttention,
-} from "./client.js"
+  generate,
+  generateWithTools,
+  stream,
+  predict,
+  scoreTokens,
+  preload,
+  abort,
+  resume,
+  onStatus,
+  registerLocalModel,
+  computeImportance,
+  computeAttentionWeights,
+  extractFeatures,
+  extractCutFeatures,
+  decodeTokens,
+  probeAttention,
+} from "./client.js";
 
-export {dom, popup} from "./picker.js"
+export { dom, popup } from "./picker.js";
 
-export {builtinSupported, builtinAvailability} from "./builtin.js"
+export {
+  builtinSupported,
+  builtinAvailability,
+  builtinGenerate,
+} from "./builtin.js";
+
+// Construct this library's compute worker. This has to live in the library:
+// `new URL("./worker.js", import.meta.url)` resolves against this module's own
+// URL, so it finds the sibling worker.js wherever the library is served from,
+// and bundlers only emit the worker chunk when they see this exact literal.
+//
+//   import { createWorker } from "@chee/patchwork-llm";
+//   const worker = createWorker();
+//   worker.onmessage = (e) => console.log(e.data.models);
+//   worker.postMessage({ type: "list-local-models" });
+export const createWorker = () =>
+  new Worker(new URL("./worker.js", import.meta.url), { type: "module" });
+
+// Transport-neutral request preparation, shared by generate() and by
+// host-side worker specs that serve this library over a stream.
+export {
+  prepareGenerate,
+  buildGeneratePayload,
+  NATIVE_TOOL_PROVIDERS,
+  TEMPLATE_TOOL_PROVIDERS,
+} from "./request.js";
 
 // LLM tools (user-defined tools the model can be given)
 export {
-	createLLMTool,
-	createToolFile,
-	LLMToolDatatype,
-	sanitizeToolName,
-	resolveTools,
-	toToolSchemas,
-	buildToolsSystem,
-	parseToolCalls,
-	loadHandler,
-	runTool,
-	runHandlerSandboxed,
-	// saved prompts (system + pre), same doc shape as tools
-	createPromptDoc,
-	resolvePromptDocs,
-	resolvePromptText,
-	resolveCfgPrompts,
-	LLMSystemPromptDatatype,
-	LLMPrePromptDatatype,
-	// folders + one-time migration
-	ensureFolderUrl,
-	addToFolder,
-	removeFromFolder,
-	migrateConfig,
-} from "./tools.js"
+  createLLMTool,
+  createToolFile,
+  LLMToolDatatype,
+  sanitizeToolName,
+  resolveTools,
+  toToolSchemas,
+  buildToolsSystem,
+  parseToolCalls,
+  loadHandler,
+  runTool,
+  runHandlerSandboxed,
+  // saved prompts (system + pre), same doc shape as tools
+  createPromptDoc,
+  resolvePromptDocs,
+  resolvePromptText,
+  resolveCfgPrompts,
+  LLMSystemPromptDatatype,
+  LLMPrePromptDatatype,
+  // folders + one-time migration
+  ensureFolderUrl,
+  addToFolder,
+  removeFromFolder,
+  migrateConfig,
+} from "./tools.js";
 
 // Registers <patchwork-llm-config-provider> on import.
 export {
-	PatchworkLLMConfigProvider,
-	definePatchworkLLMConfigProvider,
-} from "./provider.js"
+  PatchworkLLMConfigProvider,
+  definePatchworkLLMConfigProvider,
+} from "./provider.js";
 
 // Built-in prompt templates
-export {PROMPT_TEMPLATES} from "./templates.js"
+export { PROMPT_TEMPLATES } from "./templates.js";

libraries/llm/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chee/patchwork-llm",
-  "version": "0.2.1",
+  "version": "0.3.0",
   "description": "LLM toolkit for Patchwork tools: a <dialog> model picker, a refresh-surviving SharedWorker that runs local (transformers.js) / OpenRouter / Ollama generation, and a streaming API that carries rich telemetry — next-token predictions, temperature, tokens/sec — alongside the text so you can build UIs that show how the model thinks.",
   "type": "module",
   "main": "index.js",
@@ -17,13 +17,15 @@
   },
   "scripts": {
     "build": "pnpm build:types",
-    "build:types": "tsc"
+    "build:types": "tsc",
+    "push": "pnpm build && pushwork sync"
   },
   "files": [
     "types",
     "index.js",
     "config.js",
     "client.js",
+    "request.js",
     "worker.js",
     "picker.js",
     "tools.js",