298 lines
12 KiB
TypeScript
298 lines
12 KiB
TypeScript
import {
|
|
ASSISTANT_SYSTEM_PROMPT_BASE,
|
|
buildAssistantSystemPrompt,
|
|
describeAssistantContext,
|
|
} from "../prompts";
|
|
import { assistantBridge } from "./bridge-client";
|
|
import { DirectAssistantBackend } from "./direct-backend";
|
|
import type { LlmMessage, LlmToolDefinition } from "./llm-protocol";
|
|
import type { ResolvedAssistantModel } from "./models";
|
|
import { toolRegistry, type ToolRegistry } from "./tool-registry";
|
|
import type {
|
|
AssistantBridgeConfig,
|
|
AssistantContext,
|
|
AssistantTurnCallbacks,
|
|
AssistantTurnHandle,
|
|
} from "./types";
|
|
|
|
/** One shared conversation, independent of the voice overlay or chat window. */
|
|
export type AssistantTranscriptEntry = { role: "user" | "assistant"; text: string };
|
|
export type AssistantSessionHistory = {
|
|
messages: LlmMessage[];
|
|
transcript: AssistantTranscriptEntry[];
|
|
/** Model identity for safe replay after Auto resolves differently on restart. */
|
|
engine?: string;
|
|
/** The Hermes server-side session for this conversation (Hermes keeps the history). */
|
|
hermesSessionId?: string;
|
|
};
|
|
|
|
/** Which engine answers utterances, plus what it needs to do so. */
|
|
export type AssistantBackendConfig =
|
|
| { kind: "direct"; llm: ResolvedAssistantModel }
|
|
| { kind: "external"; bridge: AssistantBridgeConfig };
|
|
|
|
/** Trim history from the head once it grows past this many messages. */
|
|
const MAX_HISTORY_MESSAGES = 40;
|
|
|
|
/** The longest tool name every provider accepts (OpenAI allows 64; Anthropic, 128). */
|
|
const MAX_API_TOOL_NAME_LENGTH = 64;
|
|
/** Characters of the sanitized name kept at the start when shortening it. */
|
|
const SHORTENED_TOOL_NAME_HEAD = 16;
|
|
|
|
|
|
/** Hermes brings its own persona and tools; it only needs to know where the reply is shown. */
|
|
const HERMES_SYSTEM_PROMPT =
|
|
"The user is talking to you through smart glasses with a small text display. Reply briefly in plain text, without Markdown.";
|
|
|
|
export class AssistantSession {
|
|
private readonly directBackend = new DirectAssistantBackend();
|
|
private readonly messages: LlmMessage[] = [];
|
|
private turnHandle: AssistantTurnHandle | null = null;
|
|
readonly transcript: AssistantTranscriptEntry[] = [];
|
|
status = "";
|
|
private readonly listeners = new Set<() => void>();
|
|
private turnGeneration = 0;
|
|
// API-safe tool name -> canonical registry name. Provider APIs restrict tool
|
|
// names, so dotted registry names are sanitized and mapped back on calls.
|
|
private readonly toolNameMap = new Map<string, string>();
|
|
// Saved with the history: Faceclaw only sends Hermes the latest message, so a
|
|
// restored conversation must keep addressing the same server-side session.
|
|
private readonly hermesSessionId: string;
|
|
|
|
constructor(
|
|
private config: AssistantBackendConfig,
|
|
private readonly registry: ToolRegistry = toolRegistry,
|
|
history?: AssistantSessionHistory,
|
|
) {
|
|
this.hermesSessionId = typeof history?.hermesSessionId === "string" && history.hermesSessionId
|
|
? history.hermesSessionId
|
|
: `faceclaw-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`;
|
|
if (history) {
|
|
this.messages.push(...history.messages);
|
|
this.transcript.push(...history.transcript);
|
|
if (history.engine !== this.engine()) this.rebuildTextHistory();
|
|
}
|
|
}
|
|
|
|
onChanged(listener: () => void): () => void {
|
|
this.listeners.add(listener);
|
|
return () => this.listeners.delete(listener);
|
|
}
|
|
|
|
private changed(): void {
|
|
for (const listener of this.listeners) {
|
|
try { listener(); } catch (error) { console.warn("assistant history listener failed", error); }
|
|
}
|
|
}
|
|
|
|
history(): AssistantSessionHistory {
|
|
return { messages: this.messages, transcript: this.transcript, engine: this.engine(), hermesSessionId: this.hermesSessionId };
|
|
}
|
|
|
|
configure(config: AssistantBackendConfig): void {
|
|
if (this.isTurnActive()) return;
|
|
// Opaque reasoning/tool provider items are only replayable on their
|
|
// original model. Preserve visible conversation when switching engines.
|
|
if (this.config.kind !== config.kind ||
|
|
(this.config.kind === "direct" && config.kind === "direct" &&
|
|
(this.config.llm.model !== config.llm.model || this.config.llm.provider !== config.llm.provider ||
|
|
(this.config.llm.baseUrl ?? "") !== (config.llm.baseUrl ?? "")))) {
|
|
this.rebuildTextHistory();
|
|
}
|
|
this.config = config;
|
|
}
|
|
|
|
private engine(): string {
|
|
if (this.config.kind !== "direct") return "external";
|
|
const llm = this.config.llm;
|
|
// Keep the legacy engine ID for built-in providers so upgrading does not
|
|
// discard their opaque history. Custom endpoints need endpoint isolation.
|
|
const endpoint = llm.provider === "openai-compatible" ? `:${llm.baseUrl ?? ""}` : "";
|
|
return `${llm.provider}:${llm.model}${endpoint}`;
|
|
}
|
|
|
|
private rebuildTextHistory(): void {
|
|
this.messages.splice(0, this.messages.length, ...this.transcript
|
|
.filter((entry) => entry.text.trim())
|
|
.map((entry) => ({ role: entry.role, content: entry.text })));
|
|
}
|
|
|
|
isTurnActive(): boolean {
|
|
return this.turnHandle !== null;
|
|
}
|
|
|
|
/** Begin a turn from a spoken utterance. Only one turn runs at a time. */
|
|
sendUtterance(text: string, ctx: AssistantContext, callbacks: AssistantTurnCallbacks): void {
|
|
if (this.turnHandle) {
|
|
callbacks.onError("The assistant is still working on the previous request");
|
|
return;
|
|
}
|
|
text = text.trim();
|
|
if (!text) return;
|
|
const generation = ++this.turnGeneration;
|
|
const reply: AssistantTranscriptEntry = { role: "assistant", text: "" };
|
|
this.transcript.push({ role: "user", text }, reply);
|
|
this.status = "Thinking...";
|
|
// Install a sentinel before calling providers: some failures are synchronous.
|
|
this.turnHandle = { cancel: () => {} };
|
|
this.changed();
|
|
|
|
const finish = () => {
|
|
++this.turnGeneration;
|
|
this.turnHandle = null;
|
|
this.changed();
|
|
};
|
|
const wrappedCallbacks: AssistantTurnCallbacks = {
|
|
onTextDelta: (delta, full) => {
|
|
if (generation !== this.turnGeneration) return;
|
|
reply.text = full;
|
|
this.status = "Thinking...";
|
|
this.changed();
|
|
callbacks.onTextDelta(delta, full);
|
|
},
|
|
onToolActivity: (label) => {
|
|
if (generation !== this.turnGeneration) return;
|
|
this.status = `→ ${label}`;
|
|
this.changed();
|
|
callbacks.onToolActivity(label);
|
|
},
|
|
onTurnDone: (result) => {
|
|
if (generation !== this.turnGeneration) return;
|
|
this.status = "";
|
|
finish();
|
|
callbacks.onTurnDone(result);
|
|
},
|
|
onError: (message) => {
|
|
if (generation !== this.turnGeneration) return;
|
|
this.status = message;
|
|
finish();
|
|
callbacks.onError(message);
|
|
},
|
|
};
|
|
|
|
if (this.config.kind === "external") {
|
|
// History and the agent loop live on the agent's machine; the phone just
|
|
// streams this turn. The overlay keeps its own display state.
|
|
const handle = assistantBridge.sendUtterance(text, ctx, wrappedCallbacks);
|
|
if (generation === this.turnGeneration && this.turnHandle) this.turnHandle = handle;
|
|
return;
|
|
}
|
|
|
|
const llm = this.config.llm;
|
|
// Local provider: keep the system prompt (and thus the tool declarations
|
|
// that follow it in the rendered prompt) byte-stable so the on-phone
|
|
// model's KV prefix cache survives across turns; the volatile context
|
|
// (time, foreground app, battery) rides along with the utterance instead.
|
|
const isLocal = llm.provider === "local";
|
|
this.messages.push({
|
|
role: "user",
|
|
content: isLocal ? `${text}\n\n(${describeAssistantContext(ctx)})` : text,
|
|
});
|
|
this.trimHistory();
|
|
// Work on a turn-local array: cancelled tools may finish later, and must
|
|
// never append orphan results into a subsequent turn's context.
|
|
const turnMessages = this.messages.slice();
|
|
const handle = this.directBackend.runTurn({
|
|
provider: llm.provider,
|
|
apiKey: llm.apiKey,
|
|
model: llm.model,
|
|
effort: llm.effort,
|
|
baseUrl: llm.baseUrl,
|
|
webSearch: llm.webSearch,
|
|
sessionId: this.hermesSessionId,
|
|
system: llm.provider === "hermes" ? HERMES_SYSTEM_PROMPT
|
|
: isLocal ? ASSISTANT_SYSTEM_PROMPT_BASE : buildAssistantSystemPrompt(ctx),
|
|
messages: turnMessages,
|
|
buildTools: () => this.buildToolDefinitions(),
|
|
registry: this.registry,
|
|
resolveToolName: (apiName) => this.toolNameMap.get(apiName) ?? apiName,
|
|
callbacks: {
|
|
...wrappedCallbacks,
|
|
onTurnDone: (result) => {
|
|
if (generation !== this.turnGeneration) return;
|
|
this.messages.splice(0, this.messages.length, ...turnMessages);
|
|
wrappedCallbacks.onTurnDone(result);
|
|
},
|
|
onError: (message) => {
|
|
if (generation !== this.turnGeneration) return;
|
|
// Retain the visible partial answer, without incomplete tool pairs.
|
|
if (reply.text) this.messages.push({ role: "assistant", content: reply.text });
|
|
wrappedCallbacks.onError(message);
|
|
},
|
|
},
|
|
});
|
|
if (generation === this.turnGeneration && this.turnHandle) this.turnHandle = handle;
|
|
}
|
|
|
|
cancel(): void {
|
|
if (!this.turnHandle) return;
|
|
++this.turnGeneration;
|
|
this.turnHandle.cancel();
|
|
this.turnHandle = null;
|
|
const reply = this.transcript[this.transcript.length - 1];
|
|
if (reply?.role === "assistant" && reply.text) {
|
|
this.messages.push({ role: "assistant", content: reply.text });
|
|
}
|
|
this.status = "Cancelled";
|
|
this.changed();
|
|
}
|
|
|
|
private buildToolDefinitions(): LlmToolDefinition[] {
|
|
this.toolNameMap.clear();
|
|
return this.registry.listTools().map((spec) => {
|
|
const apiName = this.toApiToolName(spec.name);
|
|
this.toolNameMap.set(apiName, spec.name);
|
|
return { name: apiName, description: spec.description, input_schema: spec.inputSchema };
|
|
});
|
|
}
|
|
|
|
/** Convert a canonical registry name to a provider-legal tool name. */
|
|
private toApiToolName(name: string): string {
|
|
let base = name.replace(/[^a-zA-Z0-9_-]/g, "_");
|
|
if (base.length > MAX_API_TOOL_NAME_LENGTH) {
|
|
// Keep the start (the namespace) and the end (the tool's own name). The
|
|
// hash of the whole name keeps shortened names distinct and stable.
|
|
const hash = fnv32(name).toString(16).padStart(8, "0");
|
|
const tail = MAX_API_TOOL_NAME_LENGTH - SHORTENED_TOOL_NAME_HEAD - hash.length - 2;
|
|
base = `${base.slice(0, SHORTENED_TOOL_NAME_HEAD)}_${hash}_${base.slice(-tail)}`;
|
|
}
|
|
// Disambiguate the rare case where two canonical names sanitize alike.
|
|
let apiName = base;
|
|
for (let i = 2; this.toolNameMap.has(apiName); i++) {
|
|
const suffix = `_${i}`;
|
|
apiName = base.slice(0, MAX_API_TOOL_NAME_LENGTH - suffix.length) + suffix;
|
|
}
|
|
return apiName;
|
|
}
|
|
|
|
private trimHistory(): void {
|
|
if (this.messages.length <= MAX_HISTORY_MESSAGES) return;
|
|
// Drop whole messages from the head. A leading orphan tool_result (whose
|
|
// tool_use we trimmed) would be rejected by the API, so skip past any
|
|
// tool_result-only user message left at the front.
|
|
const overflow = this.messages.length - MAX_HISTORY_MESSAGES;
|
|
this.messages.splice(0, overflow);
|
|
while (this.messages.length && startsWithToolResult(this.messages[0]!)) {
|
|
this.messages.shift();
|
|
}
|
|
}
|
|
}
|
|
|
|
function startsWithToolResult(message: LlmMessage): boolean {
|
|
return (
|
|
Array.isArray(message.content) &&
|
|
message.content.length > 0 &&
|
|
message.content[0]!.type === "tool_result"
|
|
);
|
|
}
|
|
|
|
/** FNV-1a over UTF-16 code units: a small stable hash for shortened tool names. */
|
|
function fnv32(text: string): number {
|
|
let hash = 2166136261;
|
|
for (let i = 0; i < text.length; i++) {
|
|
hash ^= text.charCodeAt(i);
|
|
hash = Math.imul(hash, 16777619);
|
|
}
|
|
return hash >>> 0;
|
|
}
|