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).

src/lib/agentTools.ts

@@ -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: