@agentvalet/mcp-server 1.1.1 → 1.3.0

package/dist/index.js

@@ -18,9 +18,12 @@ process.on("unhandledRejection", (err) => {
 import { Server } from "@modelcontextprotocol/sdk/server/index.js";
 import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
 import { CallToolRequestSchema, ListToolsRequestSchema, } from "@modelcontextprotocol/sdk/types.js";
-import { SignJWT } from "jose";
 import { validateConfig } from "./config.js";
 import { renderInstructions } from "./instructions.js";
+import { signJWT } from "./auth.js";
+import { ALLOWED_METHODS, ALL_TOOLS } from "./tools/schemas.js";
+import { handleListPlatforms, handleUsePlatform, handleAgentRegister, handleAgentStatus, handleAuthzenEvaluate, handleListMyPendingActions, handleReportSelfDiagnostic, } from "./tools/handlers.js";
+import { errorContent } from "./net.js";
 // ---------------------------------------------------------------------------
 // Startup env validation
 // ---------------------------------------------------------------------------
@@ -35,376 +38,50 @@ catch (err) {
 const { agentId: AGENT_ID, ownerId: OWNER_ID, proxyUrl: PROXY_URL, privateKeyPem: AGENT_PRIVATE_KEY_RAW, privateKey } = configResult;
 process.stderr.write(`[mcp-server] config ok | agent=${AGENT_ID} | owner=${OWNER_ID} | proxy=${PROXY_URL} | has_key=${!!privateKey}\n`);
 // ---------------------------------------------------------------------------
-// JWT signing
-// ---------------------------------------------------------------------------
-async function signJWT() {
-    if (!privateKey)
-        throw new Error("Private key not loaded");
-    return new SignJWT({ agent_id: AGENT_ID, owner_id: OWNER_ID })
-        .setProtectedHeader({ alg: "RS256" })
-        .setIssuedAt()
-        .setExpirationTime("60s")
-        .sign(privateKey);
-}
-async function notifyBindSecret() {
-    try {
-        await fetchWithTimeout(`${PROXY_URL}/v1/bind-secret`, {
-            method: "POST",
-            headers: { "Content-Type": "application/json" },
-            body: JSON.stringify({ agent_id: AGENT_ID, owner_id: OWNER_ID }),
-        }, 8_000);
-    }
-    catch {
-        // best-effort — don't block the error response
-    }
-}
-function pendingFirstCallResponse() {
-    return {
-        content: [{
-                type: "text",
-                text: JSON.stringify({
-                    error: "Agent not yet activated — owner confirmation pending. Retry in 60 seconds.",
-                }),
-            }],
-        isError: true,
-    };
-}
-// State C — no credentials at all (empty env / Glama-style sandbox). The
-// server still boots and answers introspection; an authed tool call lands
-// here and gets an actionable message instead of a crash. Distinct from the
-// state-B "owner confirmation pending" response above, which means an identity
-// IS configured but its key hasn't arrived yet.
-function credentialsNotConfiguredResponse() {
-    return {
-        content: [{
-                type: "text",
-                text: "AgentValet credentials are not configured. Set AGENTVALET_AGENT_ID, " +
-                    "AGENTVALET_OWNER_ID, and the agent private key (and optionally " +
-                    "AGENTVALET_PROXY_URL). Run npx @agentvalet/register to create an agent. " +
-                    "Docs: https://github.com/AgentValet/AgentValet#quickstart",
-            }],
-        isError: true,
-    };
-}
-// Credential gate for authed tools/call. Returns null when the call may
-// proceed (state A — a JWT can be signed), or a ready-to-return MCP tool
-// result for the two no-key states:
-//   • state B (identity present, key pending) → existing invite-bind pending
-//     response, preserving the MCPB first-run flow.
-//   • state C (no identity at all)            → credentials-not-configured.
-// The no-auth tools (agent_register / agent_status) intentionally never call
-// this — you must be able to register in order to OBTAIN credentials.
-async function requireCredentials() {
-    if (AGENT_PRIVATE_KEY_RAW !== null)
-        return null;
-    if (AGENT_ID && OWNER_ID) {
-        await notifyBindSecret();
-        return pendingFirstCallResponse();
-    }
-    return credentialsNotConfiguredResponse();
-}
-// ---------------------------------------------------------------------------
-// Tool definitions
+// MCP server setup
 // ---------------------------------------------------------------------------
-const ALLOWED_METHODS = ["GET", "POST", "PUT", "PATCH", "DELETE"];
-const LIST_PLATFORMS_TOOL = {
-    name: "list_platforms",
-    description: "list_platforms: List the platforms and permission scopes this agent has access to.\nInput: None.\nReturns: { platforms: [{ platformId, platformName, scopes, requireApproval }], version: \"<hex>\" }.\nVersion: a deterministic hash that only changes when the platform set or scopes change. Cache the value across calls in the same session — only refresh when you suspect platforms have changed (e.g. user mentions a new connection).\nAuth: Bearer JWT.",
-    inputSchema: { type: "object", properties: {} },
-    outputSchema: {
-        type: "object",
-        properties: {
-            platforms: {
-                type: "array",
-                items: {
-                    type: "object",
-                    properties: {
-                        platformId: { type: "string" },
-                        platformName: { type: "string" },
-                        scopes: { type: "array", items: { type: "string" } },
-                        requireApproval: { type: "boolean" },
-                    },
-                    required: ["platformId", "platformName", "scopes"],
-                },
-            },
-            version: { type: "string" },
-        },
-        required: ["platforms"],
-    },
-};
-const USE_PLATFORM_TOOL = {
-    name: "use_platform",
-    description: "use_platform: Call an external platform API (Airtable, GitHub, Slack, Metabase, etc.) through the AgentValet proxy.\nInput: platform (string), endpoint (string), method (GET|POST|PUT|PATCH|DELETE), scope (string), body (object, optional — JSON request body for POST/PUT/PATCH/DELETE).\nReturns: upstream API response body. May take up to 50 seconds when the action requires owner approval — the call will block while we wait, then return the approved result transparently. If approval doesn't land in time, returns a `pending_approval` envelope and the action runs asynchronously; the user is notified when it completes.\nAuth: Bearer JWT.\nNote: legacy clients passing `data` instead of `body` are still accepted for backwards compatibility, but `body` is the canonical name.",
-    inputSchema: {
-        type: "object",
-        properties: {
-            platform: {
-                type: "string",
-                description: "Platform ID (e.g. airtable, github, slack, metabase)",
-            },
-            endpoint: {
-                type: "string",
-                description: "API path on the target platform (e.g. /v0/meta/bases or /api/dataset)",
-            },
-            method: {
-                type: "string",
-                enum: ["GET", "POST", "PUT", "PATCH", "DELETE"],
-                description: "HTTP method to use",
-            },
-            scope: {
-                type: "string",
-                description: "Permission scope required for this action (e.g. records:read)",
-            },
-            body: {
-                type: "object",
-                description: "JSON request body for POST/PUT/PATCH/DELETE. Optional. Forwarded verbatim to the upstream API.",
-            },
-            data: {
-                type: "object",
-                description: "Deprecated alias for `body` — prefer `body`. Kept for backwards compatibility.",
-            },
-        },
-        required: ["platform", "endpoint", "method", "scope"],
-    },
-};
-const AGENT_REGISTER_TOOL = {
-    name: "agent_register",
-    description: "agent_register: Self-register this agent with an owner. No auth required.\nInput: owner_id (string), agent_name (string), requested_scopes (array of {platformId, scopes}).\nReturns: registration_token, poll_url, client_id, scope, expires_in.\nAuth: None.",
-    inputSchema: {
-        type: "object",
-        properties: {
-            owner_id: {
-                type: "string",
-                description: "The owner ID to register this agent under",
-            },
-            agent_name: {
-                type: "string",
-                description: "Human-readable name for this agent",
-            },
-            requested_scopes: {
-                type: "array",
-                description: "Array of platform scope requests",
-                items: {
-                    type: "object",
-                    properties: {
-                        platformId: { type: "string" },
-                        scopes: { type: "array", items: { type: "string" } },
-                    },
-                    required: ["platformId", "scopes"],
-                },
-            },
-        },
-        required: ["owner_id", "agent_name", "requested_scopes"],
-    },
-    outputSchema: {
-        type: "object",
-        properties: {
-            registration_token: { type: "string" },
-            poll_url: { type: "string" },
-            client_id: { type: "string" },
-            scope: { type: "string" },
-            expires_in: { type: "number" },
-        },
-        required: ["registration_token"],
-    },
-};
-const AGENT_STATUS_TOOL = {
-    name: "agent_status",
-    description: "agent_status: Poll registration status using the token from agent_register.\nInput: token (string, required).\nReturns: status (\"pending_approval\"|\"approved\"|\"rejected\"), agent_id (if approved), mcp_config (if approved).\nAuth: None.",
-    inputSchema: {
-        type: "object",
-        properties: {
-            token: {
-                type: "string",
-                description: "Registration token returned by agent_register",
-            },
-        },
-        required: ["token"],
-    },
-    outputSchema: {
-        type: "object",
-        properties: {
-            status: { type: "string", enum: ["pending_approval", "approved", "rejected"] },
-            agent_id: { type: "string" },
-            mcp_config: { type: "object" },
-        },
-        required: ["status"],
-    },
-};
-const AUTHZEN_EVALUATE_TOOL = {
-    name: "authzen_evaluate",
-    description: "authzen_evaluate: Evaluate whether this agent has access to a specific platform scope. Call this BEFORE use_platform when you want to pre-check without making the upstream call.\nInput: platform_id (string), scope (string).\nReturns: decision (boolean), reason (\"approved\"|\"denied\"|\"revoked\"|\"scope_not_granted\").\nAuth: Bearer agent JWT (sent automatically by this MCP server).",
-    inputSchema: {
-        type: "object",
-        properties: {
-            platform_id: {
-                type: "string",
-                description: "The platform identifier (e.g. airtable, github)",
-            },
-            scope: {
-                type: "string",
-                description: "The permission scope to evaluate (e.g. records:read)",
-            },
-        },
-        required: ["platform_id", "scope"],
-    },
-    outputSchema: {
-        type: "object",
-        properties: {
-            decision: { type: "boolean" },
-            context: { type: "object", properties: { reason: { type: "string" } } },
-        },
-        required: ["decision"],
-    },
-};
-const REPORT_SELF_DIAGNOSTIC_TOOL = {
-    name: "report_self_diagnostic",
-    description: "report_self_diagnostic: Lodge a self-report (error/warning/info) with the AgentValet owner. Use after a use_platform error returns a report_hint, OR proactively when you encounter a problem the user should know about.\nInput: severity (debug|info|warn|error|critical), message (string, required, max 4096 bytes), code (string, optional, max 128 chars), platform (string, optional), endpoint (string, optional), correlation_id (uuid string, optional — copy from the failing call's report_hint to stitch this report to the broker-side audit row), context (object, optional, JSON-serialised must be < 16 KiB).\nReturns: { id, received_at } on success.\nAuth: Bearer agent JWT (sent automatically).",
-    inputSchema: {
-        type: "object",
-        properties: {
-            severity: {
-                type: "string",
-                enum: ["debug", "info", "warn", "error", "critical"],
-                description: "Severity level. error/critical trigger an owner notification.",
-            },
-            message: {
-                type: "string",
-                description: "One-sentence agent narrative describing what happened.",
-            },
-            code: {
-                type: "string",
-                description: "Optional short machine code (e.g. 'permission_denied').",
-            },
-            platform: {
-                type: "string",
-                description: "Optional platform id this report relates to.",
-            },
-            endpoint: {
-                type: "string",
-                description: "Optional endpoint that failed.",
-            },
-            correlation_id: {
-                type: "string",
-                description: "Optional UUID — copy from a use_platform error's report_hint to stitch this report to the audit row.",
-            },
-            context: {
-                type: "object",
-                description: "Optional structured context (request params, error details). Avoid secrets.",
-            },
-        },
-        required: ["severity", "message"],
-    },
-    outputSchema: {
-        type: "object",
-        properties: {
-            id: { type: "string" },
-            received_at: { type: "string" },
-        },
-        required: ["id"],
-    },
-};
-const LIST_MY_PENDING_ACTIONS_TOOL = {
-    name: "list_my_pending_actions",
-    description: "list_my_pending_actions: Returns this agent's currently-pending approval requests AND any that completed in the last 24 hours. Use this at session start when the user mentions an earlier action, or when use_platform's long-poll timed out and the user comes back asking what happened.\nInput: None.\nReturns: { pending: [{approval_id, platform_id, scope, created_at, expires_at}], recently_completed: [{approval_id, platform_id, scope, status, executed_at, result_summary, execution_error}] }.\nAuth: Bearer agent JWT (sent automatically).",
-    inputSchema: { type: "object", properties: {} },
-    outputSchema: {
-        type: "object",
-        properties: {
-            pending: {
-                type: "array",
-                items: {
-                    type: "object",
-                    properties: {
-                        approval_id: { type: "string" },
-                        platform_id: { type: "string" },
-                        scope: { type: "string" },
-                        created_at: { type: "string" },
-                        expires_at: { type: "string" },
-                    },
-                    required: ["approval_id", "platform_id", "scope"],
-                },
-            },
-            recently_completed: {
-                type: "array",
-                items: {
-                    type: "object",
-                    properties: {
-                        approval_id: { type: "string" },
-                        platform_id: { type: "string" },
-                        scope: { type: "string" },
-                        status: { type: "string" },
-                        executed_at: { type: "string" },
-                        result_summary: { type: "string" },
-                        execution_error: { type: "string" },
-                    },
-                    required: ["approval_id", "status"],
-                },
-            },
-        },
-        required: ["pending", "recently_completed"],
-    },
+const server = new Server({ name: "agentvalet", version: "1.0.0" }, { capabilities: { tools: {} }, instructions: renderInstructions(undefined) });
+// The config + server bundle threaded into auth + handlers (see context.ts) —
+// keeps those modules free of globals.
+const ctx = {
+    AGENT_ID,
+    OWNER_ID,
+    PROXY_URL,
+    AGENT_PRIVATE_KEY_RAW,
+    privateKey,
+    server,
+    OBSERVE_PLATFORM: process.env.OBSERVE_PLATFORM ?? "",
+    OBSERVE_CREDENTIAL: process.env.OBSERVE_CREDENTIAL ?? "",
 };
-// TODO: intent_resolve tool — planned for future release
-// ---------------------------------------------------------------------------
-// Boot-time platform fetch — primes the instruction block with the live
-// platform list so the host LLM sees an exact catalogue. Best-effort: if the
-// fetch fails (agent not activated, network down), we fall back to the static
-// instructions and rely on the agent calling list_platforms at runtime.
-// ---------------------------------------------------------------------------
-async function fetchPlatformNamesForInstructions() {
+// Boot-time platform fetch — primes the proxy connection and surfaces auth
+// failures in the stderr boot diagnostics. Best-effort and fire-and-forget so
+// it can NEVER delay the `initialize` response (a top-level await here used to
+// block Claude Desktop for seconds on cold Azure CA and cause host timeouts).
+// The host LLM learns the catalogue from list_platforms at runtime.
+void (async () => {
     if (AGENT_PRIVATE_KEY_RAW === null)
-        return undefined;
+        return;
     try {
-        const token = await signJWT();
+        const token = await signJWT(ctx);
         const ac = new AbortController();
         const timer = setTimeout(() => ac.abort(), 4_000);
-        const res = await fetch(`${PROXY_URL}/v1/agent/permissions`, {
+        await fetch(`${PROXY_URL}/v1/agent/permissions`, {
             method: "GET",
             headers: { Authorization: `Bearer ${token}`, "Content-Type": "application/json" },
             signal: ac.signal,
         }).finally(() => clearTimeout(timer));
-        if (!res.ok)
-            return undefined;
-        const body = (await res.json());
-        const names = (body.platforms ?? [])
-            .map((p) => p.platformName ?? p.platformId)
-            .filter((n) => typeof n === "string" && n.length > 0);
-        return names.length > 0 ? names : undefined;
     }
     catch {
-        return undefined;
+        // best-effort warmup only
     }
-}
-// NOTE: We intentionally do NOT prefetch platform names at boot. Doing so
-// added a top-level await on the proxy and blocked the `initialize` response
-// to Claude Desktop for several seconds (worse on cold Azure CA), causing
-// host-side timeouts. The LLM learns the catalogue from list_platforms at
-// runtime — boot-time enrichment was nice-to-have, not load-bearing.
-// ---------------------------------------------------------------------------
-// MCP server setup
-// ---------------------------------------------------------------------------
-const server = new Server({ name: "agentvalet", version: "1.0.0" }, { capabilities: { tools: {} }, instructions: renderInstructions(undefined) });
-// Fire the platform-names fetch in the background after the server is up so
-// it can't delay initialize. The result isn't surfaced to the host (MCP has
-// no instructions-update message), but the network warmup primes the proxy
-// and surfaces auth failures in the stderr boot diagnostics path.
-void fetchPlatformNamesForInstructions().catch(() => { });
+})().catch(() => { });
 server.setRequestHandler(ListToolsRequestSchema, async () => ({
-    tools: [
-        LIST_PLATFORMS_TOOL,
-        USE_PLATFORM_TOOL,
-        AGENT_REGISTER_TOOL,
-        AGENT_STATUS_TOOL,
-        AUTHZEN_EVALUATE_TOOL,
-        REPORT_SELF_DIAGNOSTIC_TOOL,
-        LIST_MY_PENDING_ACTIONS_TOOL,
-    ],
+    tools: ALL_TOOLS,
 }));
 server.setRequestHandler(CallToolRequestSchema, async (request) => {
     const { name, arguments: args } = request.params;
     if (name === "list_platforms") {
-        return await handleListPlatforms();
+        return await handleListPlatforms(ctx);
     }
     if (name === "use_platform") {
         if (!args ||
@@ -430,40 +107,41 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
         // schema only declared `data` which made every POST get silently
         // dropped to an empty body. See the use_platform tool description.
         const bodyArg = (args.body ?? args.data);
-        return await handleUsePlatform({
+        return await handleUsePlatform(ctx, {
             platform: args.platform,
             endpoint: args.endpoint,
             method: args.method,
             scope: args.scope,
             data: bodyArg,
+            ...(typeof args.connection_id === "string" ? { connection_id: args.connection_id } : {}),
         }, progressToken);
     }
     if (name === "agent_register") {
         if (!args || typeof args.owner_id !== "string" || typeof args.agent_name !== "string" || !Array.isArray(args.requested_scopes)) {
             return errorContent("Invalid or missing arguments: owner_id, agent_name, requested_scopes are required");
         }
-        return await handleAgentRegister(args);
+        return await handleAgentRegister(ctx, args);
     }
     if (name === "agent_status") {
         if (!args || typeof args.token !== "string") {
             return errorContent("Invalid or missing argument: token is required");
         }
-        return await handleAgentStatus(args.token);
+        return await handleAgentStatus(ctx, args.token);
     }
     if (name === "authzen_evaluate") {
         if (!args || typeof args.platform_id !== "string" || typeof args.scope !== "string") {
             return errorContent("Invalid or missing arguments: platform_id and scope are required");
         }
-        return await handleAuthzenEvaluate(args.platform_id, args.scope);
+        return await handleAuthzenEvaluate(ctx, args.platform_id, args.scope);
     }
     if (name === "list_my_pending_actions") {
-        return await handleListMyPendingActions();
+        return await handleListMyPendingActions(ctx);
     }
     if (name === "report_self_diagnostic") {
         if (!args || typeof args.severity !== "string" || typeof args.message !== "string") {
             return errorContent("Invalid or missing arguments: severity and message are required");
         }
-        return await handleReportSelfDiagnostic(args);
+        return await handleReportSelfDiagnostic(ctx, args);
     }
     return {
         content: [{ type: "text", text: `Unknown tool: ${name}` }],
@@ -471,365 +149,6 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
     };
 });
 // ---------------------------------------------------------------------------
-// Helpers
-// ---------------------------------------------------------------------------
-function fetchWithTimeout(url, init, timeoutMs = 15_000) {
-    const ac = new AbortController();
-    const timer = setTimeout(() => ac.abort(), timeoutMs);
-    return fetch(url, { ...init, signal: ac.signal }).finally(() => clearTimeout(timer));
-}
-async function fetchWithAuth(url, init) {
-    const makeRequest = async () => {
-        const token = await signJWT();
-        return fetchWithTimeout(url, {
-            ...init,
-            headers: { Authorization: `Bearer ${token}`, "Content-Type": "application/json", ...init.headers },
-        });
-    };
-    const response = await makeRequest();
-    // Retry once on 401 — the JWT may have been issued just before clock skew threshold
-    if (response.status === 401) {
-        return makeRequest();
-    }
-    return response;
-}
-function errorContent(message) {
-    return { content: [{ type: "text", text: message }], isError: true };
-}
-// Return helper that gives an MCP-aware host the structured content directly
-// (no parse step), while still emitting the text-content envelope every MCP
-// client understands. Avoids the "list of length 1 with empty fields → let me
-// parse the wrapper" round-trip Claude does on tool results that are raw
-// JSON strings — every read-heavy use_platform call pays that cost otherwise.
-//
-// If `body` isn't valid JSON, we just return the text envelope unchanged —
-// callers that emit prose (summaries, error strings) get the old behaviour.
-function jsonContent(body) {
-    const parsed = tryParseJson(body);
-    if (parsed === undefined) {
-        return { content: [{ type: "text", text: body }] };
-    }
-    return {
-        content: [{ type: "text", text: body }],
-        structuredContent: parsed,
-    };
-}
-function tryParseJson(text) {
-    const trimmed = text.trim();
-    // Only attempt parse if it looks structured. Avoids parsing a stray "true"
-    // / number / etc. into structuredContent and confusing the host.
-    if (!trimmed.startsWith("{") && !trimmed.startsWith("["))
-        return undefined;
-    try {
-        const v = JSON.parse(trimmed);
-        if (v && typeof v === "object")
-            return v;
-        return undefined;
-    }
-    catch {
-        return undefined;
-    }
-}
-// ---------------------------------------------------------------------------
-// Tool handlers
-// ---------------------------------------------------------------------------
-async function handleListPlatforms() {
-    const gate = await requireCredentials();
-    if (gate)
-        return gate;
-    let response;
-    try {
-        response = await fetchWithAuth(`${PROXY_URL}/v1/agent/permissions`, { method: "GET", headers: {} });
-    }
-    catch (err) {
-        return errorContent(diagnoseNetworkError(err, PROXY_URL));
-    }
-    const body = await response.text();
-    if (!response.ok)
-        return errorContent(`Proxy error ${response.status}: ${body}`);
-    // Proxy wraps the payload as { data: { platforms, version }, _meta: {...} }
-    // but outputSchema declares the flat { platforms, version } shape, so the
-    // wrapped envelope fails strict structuredContent validation and the host
-    // surfaces "tool execution failed (no upstream body)". Unwrap .data.
-    const parsed = tryParseJson(body);
-    const inner = parsed && typeof parsed === "object" && parsed !== null && "data" in parsed
-        ? parsed.data
-        : parsed;
-    if (inner && typeof inner === "object") {
-        return {
-            content: [{ type: "text", text: body }],
-            structuredContent: inner,
-        };
-    }
-    return jsonContent(body);
-}
-// Translates a fetch() failure into something an end-user can actually act on.
-// The default "Network error: fetch failed" message tells a non-developer
-// nothing. Look at the underlying cause keyword and map to a concrete fix.
-function diagnoseNetworkError(err, proxyUrl) {
-    const raw = err instanceof Error ? err.message : String(err);
-    const lower = raw.toLowerCase();
-    // Node's undici surfaces DNS failures as "getaddrinfo ENOTFOUND <host>".
-    if (lower.includes("enotfound") || lower.includes("getaddrinfo")) {
-        return `Network error: cannot resolve ${proxyUrl}. Check your DNS / corporate proxy / VPN, or confirm the PROXY_URL setting is correct. Raw: ${raw}`;
-    }
-    // Connection refused / unreachable / TLS handshake failure.
-    if (lower.includes("econnrefused") || lower.includes("econnreset")) {
-        return `Network error: connection to ${proxyUrl} was refused or reset. The proxy may be down — check https://status.agentvalet.ai — or a firewall is blocking the request. Raw: ${raw}`;
-    }
-    if (lower.includes("etimedout") || lower.includes("timeout") || lower.includes("aborterror")) {
-        return `Network error: request to ${proxyUrl} timed out. Likely causes: VPN routing, corporate proxy buffering, or slow network. Try again or confirm api.agentvalet.ai is reachable from this machine. Raw: ${raw}`;
-    }
-    if (lower.includes("self signed") || lower.includes("cert") || lower.includes("ssl") || lower.includes("tls")) {
-        return `Network error: TLS / certificate problem talking to ${proxyUrl}. A corporate MITM proxy may be intercepting traffic. Raw: ${raw}`;
-    }
-    return `Network error reaching ${proxyUrl}: ${raw}. Check VPN, corporate proxy, and firewall rules for api.agentvalet.ai. If the proxy itself is down, see https://status.agentvalet.ai.`;
-}
-// Long-poll budget — under Claude Desktop's hardcoded 60s tool timeout
-// (with 10s of safety). After this we return a graceful "queued" message and
-// the action lands in the user's pending-actions list (Layer 4).
-const APPROVAL_POLL_BUDGET_MS = 50_000;
-const APPROVAL_POLL_INTERVAL_MS = 2_000;
-const APPROVAL_PROGRESS_INTERVAL_MS = 5_000;
-async function handleUsePlatform(params, progressToken) {
-    const gate = await requireCredentials();
-    if (gate)
-        return gate;
-    const requestBody = {
-        platform: params.platform,
-        endpoint: params.endpoint,
-        method: params.method,
-        scope: params.scope,
-        ...(params.data !== undefined && { data: params.data }),
-    };
-    let response;
-    try {
-        response = await fetchWithAuth(`${PROXY_URL}/v1/actions`, {
-            method: "POST",
-            body: JSON.stringify(requestBody),
-        });
-    }
-    catch (err) {
-        return errorContent(`Network error: ${err instanceof Error ? err.message : err}`);
-    }
-    const body = await response.text();
-    // 202 + approval_id → enter long-poll. The proxy returns this when the
-    // owner needs to approve the call. We poll /v1/approvals/:id every ~2s
-    // for up to 50s; if the owner approves in time the proxy re-runs the
-    // upstream call and we return its result to the agent transparently.
-    if (response.status === 202) {
-        const parsed = safeJsonParse(body);
-        const approvalId = parsed?.approval_id;
-        if (approvalId) {
-            return await waitForApproval(approvalId, params, progressToken);
-        }
-    }
-    if (!response.ok)
-        return errorContent(`Proxy error ${response.status}: ${body}`);
-    return jsonContent(body);
-}
-async function waitForApproval(approvalId, originalCall, progressToken) {
-    const startedAt = Date.now();
-    let lastProgressAt = 0;
-    // Initial progress so the user sees activity immediately.
-    await sendProgress(progressToken, 0, APPROVAL_POLL_BUDGET_MS, `Owner approval required for ${originalCall.platform}:${originalCall.scope} — waiting…`);
-    while (Date.now() - startedAt < APPROVAL_POLL_BUDGET_MS) {
-        await sleep(APPROVAL_POLL_INTERVAL_MS);
-        const elapsed = Date.now() - startedAt;
-        let res;
-        try {
-            res = await fetchWithAuth(`${PROXY_URL}/v1/approvals/${approvalId}`, { method: "GET" });
-        }
-        catch {
-            // Transient network error — try again next tick.
-            continue;
-        }
-        if (!res.ok) {
-            // 404/410 etc. — treat as terminal. Surface back to the agent.
-            const text = await res.text();
-            return errorContent(`Approval status error ${res.status}: ${text}`);
-        }
-        const status = (await res.json());
-        if (status.status === "approved" && status.executed_at && status.result) {
-            // Re-execution complete. Return the upstream response transparently —
-            // the agent sees this as if the original use_platform call just succeeded.
-            const r = status.result;
-            // Mirror the non-2xx-from-upstream behaviour of the regular path.
-            if (r.status < 200 || r.status >= 300) {
-                return errorContent(`Upstream returned ${r.status}: ${JSON.stringify(r.data)}`);
-            }
-            // Two-channel return: the text envelope carries the human-readable
-            // "[Owner approved after Ns]" prefix (visible to non-structured-aware
-            // clients), and structuredContent carries the parsed upstream JSON
-            // directly so MCP-aware hosts (Claude Desktop ≥ MCP spec 2025-06-18)
-            // never have to strip-the-prefix-and-reparse. Without the second
-            // channel the prefix forces Claude into the same wrapper-parse
-            // round-trip the bare /v1/actions path used to.
-            const data = r.data ?? null;
-            const summary = `[Owner approved after ${Math.round(elapsed / 1000)}s]\n` +
-                JSON.stringify(data);
-            const result = {
-                content: [{ type: "text", text: summary }],
-            };
-            if (data && typeof data === "object" && !Array.isArray(data)) {
-                result.structuredContent = data;
-            }
-            return result;
-        }
-        if (status.status === "denied") {
-            return errorContent(`Owner denied this action${status.execution_error ? `: ${status.execution_error}` : "."}`);
-        }
-        if (status.status === "expired") {
-            return errorContent("Approval request expired before the owner responded.");
-        }
-        if (status.execution_error) {
-            // Approved but re-execution failed (network / upstream / agent revoked).
-            return errorContent(`Approved but re-execution failed: ${status.execution_error}`);
-        }
-        // Still pending — emit a progress update at most once per ~5s so the
-        // user can see the wait advancing in the Claude Desktop tool UI.
-        if (Date.now() - lastProgressAt >= APPROVAL_PROGRESS_INTERVAL_MS) {
-            await sendProgress(progressToken, elapsed, APPROVAL_POLL_BUDGET_MS, `Waiting for owner approval — ${Math.round(elapsed / 1000)}s elapsed`);
-            lastProgressAt = Date.now();
-        }
-    }
-    // Timed out — fall through to async-recap path. The action is still
-    // queued and will run when the owner approves; the user is notified via
-    // push/email at that point. Layer 4's list_my_pending_actions surfaces it.
-    return jsonContent(JSON.stringify({
-        status: "pending_approval",
-        approval_id: approvalId,
-        message: `Owner hasn't approved within ${APPROVAL_POLL_BUDGET_MS / 1000}s. ` +
-            `The action is queued — your owner will be notified, and you'll see it ` +
-            `complete next time we chat (or you can ask me to check pending actions).`,
-    }));
-}
-function safeJsonParse(text) {
-    try {
-        return JSON.parse(text);
-    }
-    catch {
-        return null;
-    }
-}
-function sleep(ms) {
-    return new Promise((resolve) => setTimeout(resolve, ms));
-}
-async function sendProgress(token, progress, total, message) {
-    if (token === undefined)
-        return;
-    try {
-        await server.notification({
-            method: "notifications/progress",
-            params: { progressToken: token, progress, total, message },
-        });
-    }
-    catch {
-        // Best-effort — never let a failed notification break the call.
-    }
-}
-async function handleAgentRegister(args) {
-    let response;
-    try {
-        response = await fetchWithTimeout(`${PROXY_URL}/v1/register`, {
-            method: "POST",
-            headers: { "Content-Type": "application/json" },
-            body: JSON.stringify(args),
-        });
-    }
-    catch (err) {
-        return errorContent(`Network error: ${err instanceof Error ? err.message : err}`);
-    }
-    const body = await response.text();
-    if (!response.ok)
-        return errorContent(`Proxy error ${response.status}: ${body}`);
-    return jsonContent(body);
-}
-async function handleAgentStatus(token) {
-    let response;
-    try {
-        response = await fetchWithTimeout(`${PROXY_URL}/v1/register/status/${encodeURIComponent(token)}`, {
-            method: "GET",
-            headers: { "Content-Type": "application/json" },
-        });
-    }
-    catch (err) {
-        return errorContent(`Network error: ${err instanceof Error ? err.message : err}`);
-    }
-    const body = await response.text();
-    if (!response.ok)
-        return errorContent(`Proxy error ${response.status}: ${body}`);
-    return jsonContent(body);
-}
-async function handleAuthzenEvaluate(platformId, scope) {
-    const gate = await requireCredentials();
-    if (gate)
-        return gate;
-    const authzenBody = {
-        subject: { type: "agent", id: AGENT_ID },
-        action: { name: "tool_call" },
-        resource: { type: "platform_scope", id: `${platformId}:${scope}` },
-    };
-    let response;
-    try {
-        response = await fetchWithAuth(`${PROXY_URL}/v1/authzen/access`, {
-            method: "POST",
-            body: JSON.stringify(authzenBody),
-        });
-    }
-    catch (err) {
-        return errorContent(`Network error: ${err instanceof Error ? err.message : err}`);
-    }
-    const body = await response.text();
-    if (!response.ok)
-        return errorContent(`Proxy error ${response.status}: ${body}`);
-    return jsonContent(body);
-}
-async function handleListMyPendingActions() {
-    const gate = await requireCredentials();
-    if (gate)
-        return gate;
-    let response;
-    try {
-        response = await fetchWithAuth(`${PROXY_URL}/v1/agents/me/pending-actions`, { method: "GET" });
-    }
-    catch (err) {
-        return errorContent(`Network error: ${err instanceof Error ? err.message : err}`);
-    }
-    const text = await response.text();
-    if (!response.ok)
-        return errorContent(`Proxy error ${response.status}: ${text}`);
-    return jsonContent(text);
-}
-async function handleReportSelfDiagnostic(args) {
-    const gate = await requireCredentials();
-    if (gate)
-        return gate;
-    // Whitelist body fields — never forward owner_id/agent_id (proxy derives those from JWT).
-    const body = {
-        severity: args.severity,
-        message: args.message,
-    };
-    for (const k of ["code", "platform", "endpoint", "correlation_id", "context"]) {
-        if (args[k] !== undefined)
-            body[k] = args[k];
-    }
-    let response;
-    try {
-        response = await fetchWithAuth(`${PROXY_URL}/v1/agents/self/diagnostics`, {
-            method: "POST",
-            body: JSON.stringify(body),
-        });
-    }
-    catch (err) {
-        return errorContent(`Network error: ${err instanceof Error ? err.message : err}`);
-    }
-    const text = await response.text();
-    if (!response.ok)
-        return errorContent(`Proxy error ${response.status}: ${text}`);
-    return jsonContent(text);
-}
-// ---------------------------------------------------------------------------
 // Connect transport
 // ---------------------------------------------------------------------------
 if (process.env.MCP_TRANSPORT === "http") {