@zhixuan92/multi-model-agent-mcp 2.7.5 → 2.8.1

package/dist/cli.js

@@ -1,585 +1,7 @@
 #!/usr/bin/env node
-import fs from 'fs';
-import path from 'path';
-import os from 'os';
-import { randomUUID } from 'node:crypto';
-import { createRequire } from 'node:module';
-import { fileURLToPath } from 'url';
-import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
-import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
-import { z } from 'zod';
-import { loadConfigFromFile } from '@zhixuan92/multi-model-agent-core/config/load';
-import { parseConfig } from '@zhixuan92/multi-model-agent-core/config/schema';
-import { runTasks } from '@zhixuan92/multi-model-agent-core/run-tasks';
-import { InMemoryContextBlockStore, createDiagnosticLogger } from '@zhixuan92/multi-model-agent-core';
-import { renderProviderRoutingMatrix } from './routing/render-provider-routing-matrix.js';
-import { computeTimings, computeBatchProgress, computeAggregateCost, } from './tools/batch-response.js';
-import { buildUnifiedResponse, withDiagnostics } from './tools/shared.js';
-import { truncateResults } from './tools/truncation.js';
-import { registerAuditDocument } from './tools/audit-document.js';
-import { registerDebugTask } from './tools/debug-task.js';
-import { registerExecutePlan } from './tools/execute-plan.js';
-import { registerReviewCode } from './tools/review-code.js';
-import { registerVerifyWork } from './tools/verify-work.js';
-import { compileDelegateTasks } from '@zhixuan92/multi-model-agent-core/intake/compilers/delegate';
-import { runIntakePipeline } from '@zhixuan92/multi-model-agent-core/intake/pipeline';
-import { ClarificationStore } from '@zhixuan92/multi-model-agent-core/intake/clarification-store';
-import { registerConfirmClarifications } from './tools/confirm-clarifications.js';
-export { computeTimings, computeBatchProgress, computeAggregateCost } from './tools/batch-response.js';
-export const SERVER_NAME = 'multi-model-agent';
-export const ASSISTANT_MODEL_NAME = 'GPT-5';
-const DEFAULT_LARGE_RESPONSE_THRESHOLD_CHARS = 65_536;
-export function buildCliGreeting() {
-    return `Hi! I'm ${ASSISTANT_MODEL_NAME}, your friendly multi-model agent assistant.`;
-}
-function parsePositiveInt(s) {
-    if (!s)
-        return undefined;
-    const n = Number.parseInt(s, 10);
-    if (Number.isFinite(n) && n > 0 && String(n) === s.trim())
-        return n;
-    return undefined;
-}
-// Read the version from package.json at module load so the MCP server
-// metadata (and tests that assert against it) stays in lockstep with the
-// published npm package version. `createRequire` keeps the JSON read
-// outside tsc's `rootDir: src` constraint and avoids the `with { type:
-// 'json' }` import attribute (which would force us to commit to a
-// specific TS/Node module-resolution combination). The relative path is
-// resolved from the compiled `dist/cli.js` — that sits one level below
-// `packages/mcp/package.json`.
-const packageRequire = createRequire(import.meta.url);
-const pkg = packageRequire('../package.json');
-export const SERVER_VERSION = pkg.version;
-export function buildTaskSchema(availableAgents) {
-    return z.object({
-        prompt: z.string().describe('The task instruction. Required.'),
-        agentType: z.enum(availableAgents).optional().describe('How hard the task is. Default: standard (cost-effective). Set to complex for harder reasoning or ambiguous scope.'),
-        filePaths: z.array(z.string()).optional().describe('Files the sub-agent should focus on. Existing files are pre-verified. Non-existent paths are treated as output targets.'),
-        done: z.string().optional().describe('Acceptance criteria in plain language. The worker works toward this goal. The reviewer verifies it.'),
-        contextBlockIds: z.array(z.string()).optional().describe('IDs from register_context_block to prepend to prompt.'),
-    }).strict();
-}
-/**
- * Batch cache for `retry_tasks`. Every `delegate_tasks` call stashes the
- * original `TaskSpec[]` under a UUID so the caller can later ask us to
- * re-dispatch specific indices without re-transmitting the briefs. Two
- * bounds:
- *
- *   - TTL (30 min from creation): keeps stale batches from lingering
- *     through a long session. TTL is from-creation (not from-last-access),
- *     matching `InMemoryContextBlockStore` — a batch used at minute 29
- *     still dies at minute 30. Access does NOT refresh the expiry.
- *   - LRU cap (100 entries): prevents unbounded growth from a chatty
- *     caller that never retries. Eviction is true LRU (least-recently-
- *     *used*, not least-recently-inserted): a batch that is still being
- *     retried stays hot and a newer but unused batch will be evicted
- *     first. This matters when a caller is iterating on one task while
- *     dispatching unrelated batches in parallel.
- *
- * Eviction on TTL is lazy (checked on `retry_tasks` lookup). Eviction on
- * the LRU cap is eager (runs after every `rememberBatch`).
- *
- * LRU implementation note: we use JavaScript's `Map` which preserves
- * insertion order on iteration. To "touch" an entry on access, we
- * `delete` it and re-`set` it, which moves it to the end of the
- * iteration order. `Map.keys().next().value` is then the oldest-*accessed*
- * entry (not the oldest-inserted entry), giving us O(1) LRU without a
- * separate priority structure. The helpers `touchBatch` (on access) and
- * the eviction loop in `rememberBatch` (on insert) are the only two
- * places that mutate the Map.
- */
-const BATCH_TTL_MS = 30 * 60 * 1000;
-const BATCH_MAX = 100;
-export function buildMcpServer(config, logger, options) {
-    const agentKeys = config.agents ? Object.keys(config.agents) : [];
-    if (agentKeys.length === 0) {
-        throw new Error('buildMcpServer requires at least one configured agent.');
-    }
-    // Resolve the threshold once at server startup
-    const envThreshold = parsePositiveInt(process.env.MULTI_MODEL_LARGE_RESPONSE_THRESHOLD_CHARS);
-    if (process.env.MULTI_MODEL_LARGE_RESPONSE_THRESHOLD_CHARS !== undefined && envThreshold === undefined) {
-        process.stderr.write(`[multi-model-agent] warning: MULTI_MODEL_LARGE_RESPONSE_THRESHOLD_CHARS=${process.env.MULTI_MODEL_LARGE_RESPONSE_THRESHOLD_CHARS} is not a positive integer, ignoring\n`);
-    }
-    const resolvedThreshold = envThreshold
-        ?? config.defaults.largeResponseThresholdChars
-        ?? options?.largeResponseThresholdChars
-        ?? DEFAULT_LARGE_RESPONSE_THRESHOLD_CHARS;
-    const runTasksImpl = options?._testRunTasksOverride ?? runTasks;
-    // Resolve parentModel once: env var > config > undefined
-    const resolvedParentModel = process.env.PARENT_MODEL_NAME || config.defaults.parentModel || undefined;
-    function injectDefaults(tasks) {
-        return tasks.map(t => ({
-            ...t,
-            agentType: t.agentType,
-            tools: config.defaults.tools,
-            timeoutMs: config.defaults.timeoutMs,
-            maxCostUSD: config.defaults.maxCostUSD,
-            sandboxPolicy: config.defaults.sandboxPolicy,
-            cwd: process.cwd(),
-            reviewPolicy: 'full',
-            effort: undefined,
-            parentModel: resolvedParentModel,
-            autoCommit: true,
-        }));
-    }
-    const server = new McpServer({
-        name: SERVER_NAME,
-        version: SERVER_VERSION,
-    });
-    // One context-block store per server instance. Persists across calls
-    // within a single `buildMcpServer(...)` lifetime so `register_context_block`
-    // followed by multiple `delegate_tasks` with `contextBlockIds` works.
-    const contextBlockStore = new InMemoryContextBlockStore();
-    // Clarification store for intake clarification flow
-    const clarificationStore = new ClarificationStore();
-    // Per-server batch cache for `retry_tasks`. See the LRU comment block
-    // above for eviction semantics.
-    const batchCache = new Map();
-    const rememberBatch = (tasks) => {
-        const id = randomUUID();
-        batchCache.set(id, { tasks, expiresAt: Date.now() + BATCH_TTL_MS });
-        // Evict the least-recently-USED entry (not least-recently-inserted).
-        // `touchBatch` below moves accessed entries to the end of insertion
-        // order, so `keys().next().value` is the true LRU head.
-        while (batchCache.size > BATCH_MAX) {
-            const lru = batchCache.keys().next().value;
-            if (lru)
-                batchCache.delete(lru);
-            else
-                break;
-        }
-        return id;
-    };
-    /**
-     * Mark a batch as recently used by reinserting it at the tail of the
-     * Map's iteration order. `touchBatch` is called on every successful
-     * `retry_tasks` lookup so a frequently-retried batch does not get
-     * evicted by `rememberBatch`'s LRU loop. Does NOT refresh the TTL —
-     * expiry stays at the original creation time.
-     */
-    const touchBatch = (id, entry) => {
-        batchCache.delete(id);
-        batchCache.set(id, entry);
-    };
-    const availableAgents = agentKeys;
-    server.tool('delegate_tasks', 'General-purpose task dispatch — use only when no specialized route fits. ' +
-        'Try specialized tools first: audit_document (auditing), review_code (reviewing), verify_work (verifying), debug_task (debugging), execute_plan (implementing from a written plan/spec file on disk). ' +
-        'Use delegate_tasks for ad-hoc implementation, research, or any work that has no plan file and no specialized route.\n\n' +
-        'Minimum: { prompt }. Everything else has good defaults. ' +
-        'Set filePaths whenever the task targets specific files. Set done whenever you have explicit acceptance criteria (required). ' +
-        'Do not invent extra fields such as inputs or done_condition; put extra context in prompt and use only the public schema fields.\n\n' +
-        renderProviderRoutingMatrix(config), {
-        tasks: z.array(buildTaskSchema(availableAgents)).describe('Array of tasks to execute in parallel'),
-    }, withDiagnostics('delegate_tasks', logger, async ({ tasks }, extra) => {
-        const rawToken = extra._meta?.progressToken;
-        const progressToken = typeof rawToken === 'string' || typeof rawToken === 'number'
-            ? rawToken
-            : undefined;
-        let progressCounter = 0;
-        const sendProgress = progressToken !== undefined
-            ? (taskIndex, event) => {
-                progressCounter += 1;
-                const headline = `[task ${taskIndex}] ${event.headline}`;
-                void extra.sendNotification({
-                    method: 'notifications/progress',
-                    params: {
-                        progressToken,
-                        progress: progressCounter,
-                        message: headline,
-                    },
-                });
-            }
-            : undefined;
-        // Intake pipeline: compile → infer → classify → resolve
-        const requestId = randomUUID();
-        const drafts = compileDelegateTasks(tasks, requestId);
-        const intakeResult = runIntakePipeline(drafts, config, contextBlockStore);
-        // Execute ready tasks through normal dispatch
-        let results = [];
-        const readySpecs = intakeResult.ready.map(r => r.task);
-        const batchId = rememberBatch(readySpecs.length > 0 ? readySpecs : tasks);
-        const batchStartMs = Date.now();
-        try {
-            if (readySpecs.length > 0) {
-                const resolvedTasks = injectDefaults(readySpecs);
-                results = await runTasksImpl(resolvedTasks, config, {
-                    onProgress: sendProgress,
-                    runtime: { contextBlockStore },
-                });
-                intakeResult.intakeProgress.executedDrafts = results.length;
-            }
-        }
-        finally {
-            // Always attach results so get_batch_slice/retry_tasks can find the batch
-            const batchEntry = batchCache.get(batchId);
-            if (batchEntry)
-                batchEntry.results = results;
-        }
-        const wallClockMs = Date.now() - batchStartMs;
-        // Create clarification set if needed
-        let clarificationId;
-        if (intakeResult.clarifications.length > 0) {
-            const storedDrafts = intakeResult.clarifications.map(c => ({
-                draft: drafts.find(d => d.draftId === c.draftId),
-                taskIndex: c.taskIndex,
-                roundCount: 0,
-            }));
-            clarificationId = clarificationStore.create(storedDrafts, batchId);
-        }
-        // Apply auto-escape truncation
-        const truncatedResults = truncateResults(results.map(r => ({ status: r.status, output: r.output, filesWritten: r.filesWritten, error: r.error })), batchId, resolvedThreshold);
-        return buildUnifiedResponse({
-            batchId,
-            results: results.map((r, i) => ({ ...r, output: truncatedResults[i].output })),
-            tasks: readySpecs,
-            wallClockMs,
-            parentModel: resolvedParentModel,
-            clarificationId,
-            clarifications: intakeResult.clarifications.length > 0 ? intakeResult.clarifications : undefined,
-        });
-    }));
-    server.tool('register_context_block', 'Store a reusable content block for later delegate_tasks calls. Returns a block id.\n\n' +
-        'When this saves money:\n' +
-        '- You\'re dispatching 3+ tasks that all need the same file or spec as context\n' +
-        '- You\'re doing multiple rounds of review/audit on the same document\n' +
-        '- Your shared context is >2K tokens (below that, duplication cost is negligible)\n\n' +
-        'Common patterns:\n' +
-        '  Delta audit — Register round 1\'s audit report, then dispatch round 2 via\n' +
-        '  delegate_tasks with contextBlockIds + a prompt like "Only report new findings\n' +
-        '  not in the prior report, findings not fixed, and confirm which were fixed."\n' +
-        '  This cuts audit cost roughly in half on subsequent rounds.\n\n' +
-        '  Diff-scoped review — Register the git diff output, then dispatch review via\n' +
-        '  delegate_tasks with contextBlockIds + a prompt like "Review only the changes\n' +
-        '  in the diff, not the entire file." Focuses the reviewer on what changed.\n\n' +
-        '  Shared spec — Register a spec/plan once, reference it from multiple parallel\n' +
-        '  tasks. 3 tasks × 25K tokens = 75K transmitted; with a context block, ~25K total.\n\n' +
-        'Example workflow:\n' +
-        '  1. register_context_block({ content: <spec file contents> })  -> { id: "abc123" }\n' +
-        '  2. delegate_tasks({ tasks: [\n' +
-        '       { prompt: "Review section 1", contextBlockIds: ["abc123"] },\n' +
-        '       { prompt: "Review section 2", contextBlockIds: ["abc123"] },\n' +
-        '       { prompt: "Review section 3", contextBlockIds: ["abc123"] }\n' +
-        '     ]})\n' +
-        '  -> The spec is transmitted once to the server, not three times.\n\n' +
-        'Blocks live in an in-memory store with a 30-minute TTL and 100-entry LRU cap.\n' +
-        'If a block expires before use, delegate_tasks returns an error identifying the missing id.', {
-        id: z.string().optional().describe('Optional id; auto-generated UUID if omitted'),
-        content: z.string().describe('The content to store'),
-    }, async ({ id, content }) => {
-        const result = contextBlockStore.register(content, { id });
-        return {
-            content: [{ type: 'text', text: JSON.stringify({ contextBlockId: result.id }, null, 2) }],
-        };
-    });
-    server.tool('retry_tasks', 'Re-run specific tasks from a previous delegate_tasks batch.\n\n' +
-        'When to use:\n' +
-        '- A task returned \'incomplete\' but you believe a retry will succeed\n' +
-        '  (e.g., after fixing a file the task depends on, or after a parallel conflict is resolved)\n' +
-        '- You want to re-run a subset of a batch without re-transmitting prompts and context blocks\n\n' +
-        'When NOT to use (re-dispatch via delegate_tasks instead):\n' +
-        '- You need to change the task prompt, tools, effort, or limits\n' +
-        '- The original batch is older than 30 minutes (cache TTL)\n' +
-        '- You want to try a different provider or agent type\n\n' +
-        'Pass the batchId returned by delegate_tasks and an array of 0-based task indices.\n' +
-        'Batches live in an in-memory cache with a 30-minute TTL and 100-entry LRU cap.', {
-        batchId: z.string().describe('Batch id returned from a previous delegate_tasks call'),
-        taskIndices: z
-            .array(z.number().int().nonnegative())
-            .describe('Zero-based indices (into the original batch) of the tasks to re-run'),
-    }, async ({ batchId, taskIndices }) => {
-        const batch = batchCache.get(batchId);
-        if (!batch || batch.expiresAt < Date.now()) {
-            // Proactively drop the expired entry so subsequent lookups see
-            // the same "unknown" result and the cache doesn't slowly fill
-            // with stale rows that are never touched again.
-            if (batch)
-                batchCache.delete(batchId);
-            throw new Error(`batch "${batchId}" is unknown or expired — re-dispatch with full task specs via delegate_tasks`);
-        }
-        // Mark this batch as recently used so the LRU eviction in
-        // `rememberBatch` doesn't drop a hot entry when newer batches
-        // arrive. Does NOT refresh TTL — a batch created 29 minutes ago
-        // still dies at minute 30 even if it's retried heavily.
-        touchBatch(batchId, batch);
-        for (const i of taskIndices) {
-            if (i < 0 || i >= batch.tasks.length) {
-                throw new Error(`index ${i} is out of range for batch ${batchId} (size ${batch.tasks.length})`);
-            }
-        }
-        const subset = taskIndices.map((i) => batch.tasks[i]);
-        // Create a fresh batch for the retried tasks so the original batch
-        // entry is preserved and get_batch_slice can still retrieve it.
-        const retryBatchId = rememberBatch(subset);
-        const batchStartMs = Date.now();
-        let results = [];
-        try {
-            results = await runTasksImpl(injectDefaults(subset), config, {
-                runtime: { contextBlockStore },
-            });
-        }
-        finally {
-            const retryEntry = batchCache.get(retryBatchId);
-            if (retryEntry)
-                retryEntry.results = results;
-        }
-        const wallClockMs = Date.now() - batchStartMs;
-        // Apply auto-escape truncation
-        const truncatedResults = truncateResults(results.map(r => ({ status: r.status, output: r.output, filesWritten: r.filesWritten, error: r.error })), retryBatchId, resolvedThreshold);
-        return buildUnifiedResponse({
-            batchId: retryBatchId,
-            results: results.map((r, i) => ({ ...r, output: truncatedResults[i].output })),
-            tasks: subset,
-            wallClockMs,
-            parentModel: resolvedParentModel,
-        });
-    });
-    server.tool('get_batch_slice', `Retrieve full telemetry and output data from a previous delegate_tasks batch.
-
-Returns the complete batch with timings, progress, cost breakdown, and all task results.
-Optionally filter to a single task via taskIndex.
-
-Batches are cached in memory per MCP server instance with a 30-minute TTL from creation
-and a 100-entry LRU cap. Access touches the LRU order but does not refresh TTL. If the
-batch is expired or evicted, re-dispatch via delegate_tasks with the full specs.`, {
-        batchId: z.string().describe('Batch ID from a prior delegate_tasks or retry_tasks response'),
-        taskIndex: z.number().int().min(0).optional().describe('0-based task index. Omit for all tasks.'),
-    }, async ({ batchId, taskIndex }) => {
-        const entry = batchCache.get(batchId);
-        if (!entry || entry.expiresAt < Date.now()) {
-            if (entry)
-                batchCache.delete(batchId);
-            return {
-                content: [{
-                        type: 'text',
-                        text: `Batch "${batchId}" is unknown or expired. Batch results are cached for 30 minutes after completion. Re-dispatch the original task to get fresh results.`,
-                    }],
-            };
-        }
-        touchBatch(batchId, entry);
-        if (!entry.results) {
-            return {
-                content: [{
-                        type: 'text',
-                        text: `Batch "${batchId}" has no results yet — the original dispatch may still be running.`,
-                    }],
-            };
-        }
-        if (taskIndex !== undefined && (taskIndex < 0 || taskIndex >= entry.results.length)) {
-            return {
-                content: [{
-                        type: 'text',
-                        text: `taskIndex ${taskIndex} is out of range. Batch "${batchId}" has ${entry.results.length} tasks (0-based index: 0 to ${entry.results.length - 1}).`,
-                    }],
-            };
-        }
-        const results = taskIndex !== undefined
-            ? [entry.results[taskIndex]]
-            : entry.results;
-        const wallClockMs = Math.max(0, ...entry.results.map((r) => r.durationMs ?? 0));
-        const timings = computeTimings(wallClockMs, entry.results);
-        const batchProgress = computeBatchProgress(entry.results);
-        const aggregateCost = computeAggregateCost(entry.results);
-        return {
-            content: [{
-                    type: 'text',
-                    text: JSON.stringify({
-                        batchId,
-                        timings,
-                        batchProgress,
-                        aggregateCost,
-                        results,
-                    }, null, 2),
-                }],
-        };
-    });
-    registerAuditDocument(server, config, logger, contextBlockStore);
-    registerDebugTask(server, config, logger, contextBlockStore);
-    registerExecutePlan(server, config, logger, contextBlockStore);
-    registerReviewCode(server, config, logger, contextBlockStore);
-    registerVerifyWork(server, config, logger, contextBlockStore);
-    registerConfirmClarifications(server, config, logger, clarificationStore, runTasksImpl, rememberBatch);
-    return server;
-}
-/**
- * MCP CLI config discovery (owned by MCP, not core):
- * 1. --config <path> argument (explicit)
- * 2. MULTI_MODEL_CONFIG environment variable
- * 3. ~/.multi-model/config.json (default home-directory location)
- */
-export async function discoverConfig() {
-    const args = process.argv.slice(2);
-    // 1. Explicit --config
-    const configFlagIdx = args.indexOf('--config');
-    if (configFlagIdx >= 0 && args[configFlagIdx + 1]) {
-        return loadConfigFromFile(args[configFlagIdx + 1]);
-    }
-    // 2. MULTI_MODEL_CONFIG env var (file path)
-    const envPath = process.env.MULTI_MODEL_CONFIG;
-    if (envPath) {
-        return loadConfigFromFile(envPath);
-    }
-    // 3. ~/.multi-model/config.json
-    const defaultPath = path.join(os.homedir(), '.multi-model', 'config.json');
-    if (fs.existsSync(defaultPath)) {
-        return loadConfigFromFile(defaultPath);
-    }
-    // Fallback: empty config with required agents
-    return parseConfig({
-        agents: {
-            standard: { type: 'claude', model: 'claude-sonnet-4-6' },
-            complex: { type: 'claude', model: 'claude-sonnet-4-6' },
-        },
-    });
-}
-let installedLifecycleHandlers = null;
-/**
- * Install safety nets for the stdio transport lifecycle. The MCP SDK's
- * StdioServerTransport writes every JSON-RPC frame to `process.stdout`
- * but never attaches an error handler to it, so when the Claude Code
- * client closes the read end of our stdout (reconnect, /mcp restart,
- * extension reload, client crash, long-running-call abort) the next
- * write emits an `EPIPE` error with no listener, which Node turns into
- * `uncaughtException` and — absent a handler — terminates the process.
- * That is the observed "MCP dies every ~2 calls" failure mode.
- *
- * Single-install contract: calling this more than once in one process is a
- * programmer error. The healthy-server contract ("one stderr line at startup")
- * covers only the first install. A second call writes a warning to stderr
- * and returns — it does not register duplicate handlers. This warning path is
- * outside the healthy-server contract; in normal operation `main()` is the only
- * caller and is invoked exactly once per process.
- */
-export function installStdioLifecycleHandlers(logger) {
-    if (installedLifecycleHandlers !== null) {
-        process.stderr.write('[multi-model-agent] lifecycle handlers already installed; skipping second install\n');
-        return;
-    }
-    const stdoutError = (err) => {
-        if (err.code === 'EPIPE') {
-            logger.shutdown('stdout_epipe');
-            process.exit(0);
-            return;
-        }
-        logger.shutdown('stdout_other_error');
-        process.stderr.write(`[multi-model-agent] stdout error: ${err.message}\n`);
-        process.exit(1);
-    };
-    const stdinEnd = () => {
-        logger.shutdown('stdin_end');
-        process.exit(0);
-    };
-    const uncaught = (err) => {
-        logger.error('uncaughtException', err);
-        logger.shutdown('uncaughtException');
-        process.stderr.write(`[multi-model-agent] uncaughtException: ${err.stack ?? String(err)}\n`);
-        process.exit(1);
-    };
-    const unhandled = (reason) => {
-        logger.error('unhandledRejection', reason);
-        const stack = reason instanceof Error ? (reason.stack ?? reason.message) : String(reason);
-        process.stderr.write(`[multi-model-agent] unhandledRejection: ${stack}\n`);
-        logger.shutdown('unhandledRejection');
-        process.exit(1);
-    };
-    const beforeExit = () => {
-        logger.shutdown('event_loop_empty');
-    };
-    const signals = {
-        SIGTERM: () => {
-            logger.shutdown('SIGTERM');
-            process.exit(0);
-        },
-        SIGINT: () => {
-            logger.shutdown('SIGINT');
-            process.exit(0);
-        },
-        SIGPIPE: () => {
-            logger.shutdown('SIGPIPE');
-            process.exit(1);
-        },
-        SIGHUP: () => {
-            logger.shutdown('SIGHUP');
-            process.exit(0);
-        },
-        SIGABRT: () => {
-            logger.shutdown('SIGABRT');
-            process.exit(1);
-        },
-    };
-    process.stdout.on('error', stdoutError);
-    process.stdin.on('end', stdinEnd);
-    process.on('uncaughtException', uncaught);
-    process.on('unhandledRejection', unhandled);
-    process.on('beforeExit', beforeExit);
-    process.on('SIGTERM', signals.SIGTERM);
-    process.on('SIGINT', signals.SIGINT);
-    process.on('SIGPIPE', signals.SIGPIPE);
-    process.on('SIGHUP', signals.SIGHUP);
-    process.on('SIGABRT', signals.SIGABRT);
-    installedLifecycleHandlers = { stdoutError, stdinEnd, uncaught, unhandled, beforeExit, signals };
-}
-/** Test-only. Not exported from the package public surface. */
-export function __resetStdioLifecycleHandlersForTests() {
-    if (installedLifecycleHandlers === null)
-        return;
-    process.stdout.off('error', installedLifecycleHandlers.stdoutError);
-    process.stdin.off('end', installedLifecycleHandlers.stdinEnd);
-    process.off('uncaughtException', installedLifecycleHandlers.uncaught);
-    process.off('unhandledRejection', installedLifecycleHandlers.unhandled);
-    process.off('beforeExit', installedLifecycleHandlers.beforeExit);
-    process.off('SIGTERM', installedLifecycleHandlers.signals.SIGTERM);
-    process.off('SIGINT', installedLifecycleHandlers.signals.SIGINT);
-    process.off('SIGPIPE', installedLifecycleHandlers.signals.SIGPIPE);
-    process.off('SIGHUP', installedLifecycleHandlers.signals.SIGHUP);
-    process.off('SIGABRT', installedLifecycleHandlers.signals.SIGABRT);
-    installedLifecycleHandlers = null;
-}
-async function main() {
-    const args = process.argv.slice(2);
-    if (args[0] === '--help' || args[0] === '-h') {
-        console.log('Usage: multi-model-agent serve [--config <path>]');
-        process.exit(0);
-    }
-    if (args[0] !== 'serve') {
-        console.error('Usage: multi-model-agent serve [--config <path>]');
-        process.exit(1);
-    }
-    const config = await discoverConfig();
-    const agentNames = config.agents ? Object.keys(config.agents) : [];
-    if (agentNames.length === 0) {
-        console.error('No agents configured. Create ~/.multi-model/config.json or pass --config <path>.');
-        process.exit(1);
-    }
-    const enabled = config.diagnostics?.log ?? false;
-    const logDir = config.diagnostics?.logDir;
-    const logger = createDiagnosticLogger({ enabled, logDir });
-    logger.startup(SERVER_VERSION);
-    const diagnosticLogPath = logger.expectedPath();
-    if (diagnosticLogPath !== undefined) {
-        process.stderr.write(`[multi-model-agent] diagnostic log: ${diagnosticLogPath}\n`);
-    }
-    installStdioLifecycleHandlers(logger);
-    const server = buildMcpServer(config, logger);
-    const transport = new StdioServerTransport();
-    await server.connect(transport);
-}
-// Only run main when executed directly
-const thisFile = fileURLToPath(import.meta.url);
-const isDirectRun = (() => {
-    if (!process.argv[1])
-        return false;
-    try {
-        return fs.realpathSync(process.argv[1]) === fs.realpathSync(thisFile);
-    }
-    catch {
-        return false;
-    }
-})();
-if (isDirectRun) {
-    main().catch((err) => {
-        console.error('Fatal:', err);
-        process.exit(1);
-    });
-}
-//# sourceMappingURL=cli.js.map
+console.error(`
+@zhixuan92/multi-model-agent-mcp has been replaced by @zhixuan92/multi-model-agent in 3.0.0.
+Install the new package: npm i -g @zhixuan92/multi-model-agent
+See: https://github.com/zhixuan312/multi-model-agent/blob/master/CHANGELOG.md#300
+`);
+process.exit(1);

package/package.json

@@ -1,20 +1,12 @@
 {
   "name": "@zhixuan92/multi-model-agent-mcp",
-  "version": "2.7.5",
+  "version": "2.8.1",
   "type": "module",
   "license": "MIT",
-  "description": "MCP server for multi-model-agent. Exposes a delegate_tasks tool that routes work to Claude, Codex, or OpenAI-compatible sub-agents based on capability, quality tier, and cost.",
+  "description": "DEPRECATED. Renamed to @zhixuan92/multi-model-agent in 3.0.0. This package is a stub that prints a migration notice and exits.",
   "author": "Zhang Zhixuan <zhangzhixuan312@gmail.com>",
   "keywords": [
-    "mcp",
-    "model-context-protocol",
-    "llm",
-    "claude",
-    "codex",
-    "openai",
-    "agent",
-    "multi-model",
-    "delegation"
+    "deprecated"
   ],
   "repository": {
     "type": "git",
@@ -26,47 +18,15 @@
   "files": [
     "dist"
   ],
-  "main": "./dist/index.js",
-  "types": "./dist/index.d.ts",
   "bin": {
-    "multi-model-agent": "./dist/cli.js"
-  },
-  "exports": {
-    ".": {
-      "types": "./dist/index.d.ts",
-      "import": "./dist/index.js"
-    },
-    "./routing/render-provider-routing-matrix": {
-      "types": "./dist/routing/render-provider-routing-matrix.d.ts",
-      "import": "./dist/routing/render-provider-routing-matrix.js"
-    },
-    "./tools/audit-document": {
-      "types": "./dist/tools/audit-document.d.ts",
-      "import": "./dist/tools/audit-document.js"
-    },
-    "./tools/debug-task": {
-      "types": "./dist/tools/debug-task.d.ts",
-      "import": "./dist/tools/debug-task.js"
-    },
-    "./tools/review-code": {
-      "types": "./dist/tools/review-code.d.ts",
-      "import": "./dist/tools/review-code.js"
-    },
-    "./tools/verify-work": {
-      "types": "./dist/tools/verify-work.d.ts",
-      "import": "./dist/tools/verify-work.js"
-    }
+    "multi-model-agent": "dist/cli.js",
+    "mmagent": "dist/cli.js"
   },
   "scripts": {
-    "build": "tsc && chmod +x dist/cli.js",
+    "build": "mkdir -p dist && cp src/cli.js dist/cli.js && chmod +x dist/cli.js",
     "prepublishOnly": "npm run build"
   },
   "engines": {
-    "node": ">=22.0.0"
-  },
-  "dependencies": {
-    "@modelcontextprotocol/sdk": "^1.0.0",
-    "@zhixuan92/multi-model-agent-core": "^2.7.5",
-    "zod": "^4.0.0"
+    "node": ">=18.0.0"
   }
-}
+}