feat(server): inference state-graph + supervisor, memory tools, MCP client, schema, routes

- Add state-graph.ts: typed state machine for inference lifecycle
- Add supervisor.ts: agent supervisor pattern for multi-agent coordination
- Add export-formatter.ts: structured export formatting
- Add manage_memory.ts: memory CRUD tool for agent persistence
- Add get_wiki_article.ts: codecontext wiki article retrieval
- Extend memory/index.ts: 3-tier memory (context/daily/core)
- Extend MCP client: mcp-config.ts env-var substitution
- Update schema.sql: agent_sessions, tasks, pending_changes extensions
- Update API types: MessageMetadata, ErrorReason, AgentSessionConfig
- Update routes: chats, messages, sessions — column renames and agent_session_id
- Update inference: error handler, payload builder, stream phase, turn orchestrator

apps/server/src/services/inference/supervisor.ts

@@ -0,0 +1,75 @@
+// Supervisor agent: routes user requests to the best agent via a cheap LLM
+// classification call. Activated when session.agent_id === 'supervisor'.
+
+import type { Agent } from '../../types/api.js';
+import { taskModelCompletion } from '../task-model.js';
+
+export interface SupervisorRoute {
+  agent_id: string;
+  confidence: number;
+  reasoning: string;
+}
+
+const SUPERVISOR_SYSTEM_PROMPT = `You are a router. Given the user's request and the available agents, choose the best agent to handle the request.
+
+Rules:
+- Match the request to the agent whose description and toolset best fits the task.
+- For code review / bug finding requests → code-reviewer
+- For debugging / diagnosing failures → debugger
+- For refactoring / simplifying code → refactorer
+- For architecture / design / planning → architect or planner
+- For security audits → security-auditor
+- For building prompts for other agents → prompt-builder
+- For exploring / understanding unfamiliar code → recon
+- For implementing / writing code changes → builder
+- Respond with ONLY the agent id (e.g. "builder") or "none" if no agent fits.
+- Do not include any other text, punctuation, or explanation.`;
+
+const MAX_ROUTING_TOKENS = 30;
+
+/**
+ * Given the user's latest message and available agents, classifies which agent
+ * should handle this turn. Returns null to fall through to default (no agent).
+ */
+export async function resolveSupervisorTurn(
+  latestUserMessage: string,
+  agents: Agent[],
+  fallbackModel?: string,
+): Promise<SupervisorRoute | null> {
+  // Build agent listing — skip the supervisor itself to avoid self-routing.
+  const agentList = agents
+    .filter((a) => a.id !== 'supervisor')
+    .map((a) => `- ${a.id}: ${a.description} (${a.tools.length} tools)`)
+    .join('\n');
+
+  if (!agentList) {
+    return null;
+  }
+
+  const userPrompt = `Available agents:\n${agentList}\n\nUser request: ${latestUserMessage.slice(0, 2000)}`;
+
+  const response = await taskModelCompletion({
+    system: SUPERVISOR_SYSTEM_PROMPT,
+    user: userPrompt,
+    maxTokens: MAX_ROUTING_TOKENS,
+    temperature: 0.1,
+    fallbackModel,
+  });
+
+  const agentId = response.trim().toLowerCase();
+  if (!agentId || agentId === 'none') {
+    return null;
+  }
+
+  // Map back to a real agent to validate the id.
+  const matched = agents.find((a) => a.id === agentId);
+  if (!matched) {
+    return null;
+  }
+
+  return {
+    agent_id: matched.id,
+    confidence: 1,
+    reasoning: `supervisor routed to "${matched.name}" based on request classification`,
+  };
+}