feat(agent): LLM proxy integration shape (Pi-AI compatible)

src/lib/llmClient.ts: browser POSTs to /internal/canvas-llm/chat with {messages, max_tokens}. Mirrors the normalized request shape used by earendil-works/pi's @earendil-works/pi-ai. Returns: - ok:true with content+provider when configured and successful - ok:false, configured:false when proxy 404/503/unreachable - ok:false, configured:true when provider returned an error routeAgentInput now falls back to the LLM with a system prompt that includes the user identity, tool catalogue, and the top-3 vault recall hits when no deterministic tool matches. When the proxy is not configured the deterministic vault-hint reply is used (no regression in the demo tier).
2026-06-14 14:55:34 +04:00
parent 9884ddf4ac
commit 15ab8ededa
3 changed files with 146 additions and 14 deletions
@@ -0,0 +1,56 @@
+// Verify the agent memory vault round-trips: signed-in user tells the
+// Assistant a fact, then asks the Assistant to recall it. Second persona
+// in same tenant must NOT see the first persona's notes (isolation check
+// is tenant-scoped not user-scoped since the vault key is per-user).
+import { chromium } from "playwright";
+
+const URL = "https://canvas.flow-master.ai/";
+const args = ["--host-resolver-rules=MAP canvas.flow-master.ai 65.21.71.186", "--ignore-certificate-errors"];
+const b = await chromium.launch({ headless: true, args });
+
+async function asPersona(personaText, work) {
+  const ctx = await b.newContext({ viewport: { width: 1440, height: 900 } });
+  const p = await ctx.newPage();
+  await p.goto(URL, { waitUntil: "networkidle" });
+  await p.waitForTimeout(800);
+  await p.locator(".persona-chip", { hasText: personaText }).click();
+  await p.locator(".dev-login-btn").click();
+  await p.waitForTimeout(2500);
+  const enter = p.locator("button", { hasText: /Enter Mission Control/i }).first();
+  if (await enter.count()) { await enter.click(); await p.waitForTimeout(1200); }
+  await p.locator(".tab", { hasText: /^\s*Assistant\s*$/ }).click();
+  await p.waitForTimeout(1500);
+  const result = await work(p);
+  await ctx.close();
+  return result;
+}
+
+async function ask(p, text) {
+  const before = await p.locator(".agent-turn-agent .agent-turn-body").count();
+  await p.locator(".agent-composer textarea").fill(text);
+  await p.locator(".agent-composer button", { hasText: /Send/i }).click();
+  await p.waitForFunction((n) => document.querySelectorAll(".agent-turn-agent .agent-turn-body").length > n, before, { timeout: 8000 }).catch(() => {});
+  await p.waitForTimeout(600);
+  return (await p.locator(".agent-turn-agent .agent-turn-body").last().textContent()) || "";
+}
+
+const r1 = await asPersona(/Mariana/, async (p) => {
+  const a = await ask(p, "remember that the laptop budget cap is 1500 pounds for store managers");
+  console.log("[CEO remember]:", a.slice(0, 120));
+  await p.waitForTimeout(2000);
+  const b = await ask(p, "recall about laptop budget");
+  console.log("[CEO recall]:", b.slice(0, 200));
+  return { remember: a, recall: b };
+});
+
+const r2 = await asPersona(/Aisha/, async (p) => {
+  const a = await ask(p, "recall about laptop budget");
+  console.log("[HR recall]:", a.slice(0, 200));
+  return { recall: a };
+});
+
+console.log();
+console.log("CEO remembered:", r1.remember.includes("I'll remember") ? "YES" : "NO");
+console.log("CEO recalled own note:", r1.recall.includes("1500") || r1.recall.includes("laptop") ? "YES" : "NO");
+console.log("HR vault isolated from CEO:", !r2.recall.includes("1500") ? "YES" : "NO (leak)");
+await b.close();
@@ -3,6 +3,7 @@ import { wizardApi } from "./wizardApi";
 import { chatApi } from "./chatApi";
 import { curatedPublishedFlows, fetchStartableFlows } from "./flowCuration";
 import { agentMemory, type Memory } from "./agentMemory";
+import { llmClient, type LlmMessage } from "./llmClient";

 export interface ToolResult {
  ok: boolean;
@@ -181,11 +182,24 @@ const TOOLS: ToolDef[] = [
  },
 ];

+async function llmFallback(trimmed: string, ctx: ToolContext, hints: Memory[]): Promise<ToolResult | null> {
+  const memBlock = hints.length ? `\nYour vault has related notes:\n${hints.map((h) => `- ${h.content}`).join("\n")}` : "";
+  const system: LlmMessage = {
+    role: "system",
+    content:
+      `You are the FlowMaster Command Assistant inside canvas.flow-master.ai for ${ctx.userEmail}. ` +
+      `Speak briefly (3 sentences max). You can refer to these tools by name when suggesting actions: ${TOOLS.map((t) => t.name).join(", ")}. ` +
+      `If the user asks something operational, recommend a specific tool. ${memBlock}`,
+  };
+  const user: LlmMessage = { role: "user", content: trimmed };
+  const reply = await llmClient.chat({ messages: [system, user], max_tokens: 320 });
+  if (reply.ok) return { ok: true, display: reply.content };
+  return null;
+}
+
 export async function routeAgentInput(text: string, ctx: ToolContext): Promise<ToolResult> {
  const trimmed = text.trim();
  if (!trimmed) return { ok: false, display: "Tell me what to do." };
-  // Best-effort: every turn becomes an episodic memory in the vault. We never
-  // block the response on it so a vault outage doesn't break the assistant.
  agentMemory.remember(ctx.userEmail, trimmed, ["turn"]).catch(() => {});
  for (const tool of TOOLS) {
    if (tool.matcher.test(trimmed)) {
@@ -196,18 +210,17 @@ export async function routeAgentInput(text: string, ctx: ToolContext): Promise<T
      }
    }
  }
-  // If nothing matched, try to recall related context from the vault before
-  // falling back to the help message.
-  try {
-    const hints: Memory[] = await agentMemory.recall(ctx.userEmail, trimmed, 3);
-    if (hints.length > 0) {
-      const lines = hints.map((h) => `• ${h.content}`).join("\n");
-      return {
-        ok: true,
-        display: `I don't have a command that matches that yet, but your vault has related notes:\n${lines}\n\nTry: 'open mission', 'list processes', 'remember that ...', 'recall about ...', 'create a new process'.`,
-      };
-    }
-  } catch { /* ignore */ }
+  let hints: Memory[] = [];
+  try { hints = await agentMemory.recall(ctx.userEmail, trimmed, 3); } catch { /* ignore */ }
+  const llmReply = await llmFallback(trimmed, ctx, hints);
+  if (llmReply) return llmReply;
+  if (hints.length > 0) {
+    const lines = hints.map((h) => `• ${h.content}`).join("\n");
+    return {
+      ok: true,
+      display: `I don't have a command that matches that yet, but your vault has related notes:\n${lines}\n\nTry: 'open mission', 'list processes', 'remember that ...', 'recall about ...', 'create a new process'.`,
+    };
+  }
  return {
    ok: false,
    display:
@@ -0,0 +1,63 @@
+/**
+ * Browser-side LLM client. Speaks to a backend proxy that brokers to the
+ * configured provider (Anthropic, OpenAI, etc.) using the same request
+ * shape as pi-ai's normalized completions. When the proxy is unreachable
+ * or unconfigured the client returns { configured: false } and the caller
+ * is expected to fall back to a deterministic path.
+ */
+
+import { api } from "./api";
+
+export interface LlmMessage {
+  role: "system" | "user" | "assistant";
+  content: string;
+}
+
+export interface LlmRequest {
+  messages: LlmMessage[];
+  max_tokens?: number;
+}
+
+export interface LlmReply {
+  ok: true;
+  content: string;
+  provider?: string;
+}
+
+export interface LlmUnconfigured {
+  ok: false;
+  configured: false;
+  reason: string;
+}
+
+export interface LlmError {
+  ok: false;
+  configured: true;
+  error: string;
+}
+
+export type LlmResponse = LlmReply | LlmUnconfigured | LlmError;
+
+export const llmClient = {
+  async chat(req: LlmRequest, signal?: AbortSignal): Promise<LlmResponse> {
+    try {
+      const res = await fetch(`${api.config.baseUrl}/internal/canvas-llm/chat`, {
+        method: "POST",
+        headers: {
+          "Content-Type": "application/json",
+          Authorization: `Bearer ${sessionStorage.getItem("fm.mc.token.v1")}`,
+        },
+        body: JSON.stringify({ messages: req.messages, max_tokens: req.max_tokens ?? 512 }),
+        signal,
+      });
+      if (res.status === 404) return { ok: false, configured: false, reason: "LLM proxy not wired" };
+      if (res.status === 503) return { ok: false, configured: false, reason: "LLM provider not configured" };
+      if (!res.ok) return { ok: false, configured: true, error: `Provider error ${res.status}` };
+      const body = await res.json();
+      if (typeof body?.content !== "string") return { ok: false, configured: true, error: "Provider returned unparseable body" };
+      return { ok: true, content: body.content, provider: body.provider };
+    } catch (e) {
+      return { ok: false, configured: false, reason: `LLM proxy unreachable: ${(e as Error).message}` };
+    }
+  },
+};