fix(#245): codex-sdk MCP injection — commhub_send_task + per-node alias

Closes the codex-sdk side of #245 ("新成员都用不了 send_task"). Root cause
locked + verified by 通信龙 + TMCode 负责人 diagnostic (msg 459d1b6c) +
my own audit (cli.ts:1337-1340 had ZERO mcp_servers in CODEX_CONFIG;
createCommhubSdkMcpServer at agent-node/src/cli.ts:1126 is wired ONLY
into the claude-agent-sdk processWithSdk path).

Background
==========

Vincent's "new agents can't use send_task" was: new nodes default to
codex-sdk runtime (per [[feedback_new_node_codex_default]]); codex
inherits MCP servers from `~/.codex/config.toml [mcp_servers.*]`
(TOML — codex does NOT read cwd `.mcp.json`); for months the only
[mcp_servers.commhub-proxy] block was a global stale proxy with a
hardcoded `COMMHUB_ALIAS = "codex-硅谷"` pointing at a separate
proxy script (`proxy/commhub-proxy.ts`) that exposed only 4 tools
(get_task / report_result / get_status / send_message — NO send_task,
NO get_all_status), and that script got rm-rf'd in the 06-16 incident
so codex-sdk nodes started crashing "No such file or directory" on
top of being functionally limited. 通信龙 removed the broken block
this morning (~/.codex/config.toml.bak-20260617-245), exposing the
real underlying gap: agent-node never injected commhub MCP for
codex-sdk in the first place.

Fix (two-package commit)
========================

agent-node/src/cli.ts:
  - Replace the bare CODEX_CONFIG constant with `buildCodexConfig(workdir)`
    that adds `mcp_servers.commhub = { command: "bun", args: [<cwd>/.anet/node-server.js] }`.
  - env intentionally OMITTED from the mcp_servers.commhub block — codex
    CLI subprocess inherits this agent-node process's env, which anet's
    launchAgent already sets per-node (COMMHUB_ALIAS = displayName,
    COMMHUB_TOKEN, COMMHUB_URL). Hard-coding env here would re-introduce
    the global-alias bug the 06-17 incident exposed.
  - The MCP server points at agent-network's `.anet/node-server.js` (a
    bun stdio MCP server bundled in agent-network's dist/), which exposes
    the full commhub_send_task / commhub_get_all_status / commhub_get_task
    / commhub_send_message / commhub_send_reply / commhub_report_status /
    commhub_get_session_status / commhub_list_tasks set.

agent-network/bin/cli.ts:
  - ensureMcpJson runtime gate widened from `claude-code-cli only` to
    `claude-code-cli OR codex-sdk`. For codex-sdk the function still
    refreshes `.anet/node-server.js` from the npm package + self-heals
    `@modelcontextprotocol/sdk` in `.anet/node_modules` + writes
    `.anet/.env` (hub URL + token), but SKIPS the `.mcp.json` write
    (codex does not read cwd `.mcp.json`; that file would be a silent
    no-op + a debugging-trap).

Why this fix is per-node-safe
=============================

`Codex({ config })` (the SDK constructor) accepts the JSON object and
flattens it into `--config key=value` TOML literal overrides on the
codex CLI command line. The override is per-Codex-instance, in-memory,
and does NOT mutate `~/.codex/config.toml`. Two codex-sdk nodes running
simultaneously each get their own codex CLI subprocess with its own
inherited env from its own agent-node parent — so their commhub MCP
servers register against their own per-node alias.

Verified
========

Docker node:24-alpine, fresh `npm i -g @preview` of both packages:
  - agent-node v2.4.11-preview.2 installed cleanly ✓
  - agent-node bundle markers: `mcp_servers` ×1, `node-server.js` path ×1,
    `commhub` ×36 ✓
  - agent-network 2.2.14-preview.0 installed cleanly ✓
  - agent-network bundle marker: `codex-sdk` ×14 (the new gate
    addition + cross-references in existing #237 + #133 code) ✓
  - `anet doctor` clean output ✓

Live verification (TMCode 副责人 codex-sdk node restart, agent self-
report) is the next gate before promote latest — fast-follow patch
flow per 通信龙 dispatch 478de9fe.

Co-changes:
  - agent-network 2.2.13 → 2.2.14-preview.0
  - agent-node    2.4.10 → 2.4.11-preview.2 (.0 and .1 occupied by
                  prior #211 / #168 shipping cycles)

Branch: fix/245-codex-mcp (forked from fix/245-connectivity-reliability
which carries the just-released 2.2.13 telegram series + bump commit
e409aff)

Refs: #245 (umbrella), 通信龙 dispatches d2814e67 + 7c7c7e1f + 478de9fe
+ 7507d1fb + 71675e69, TMCode 负责人 diagnostic 459d1b6c, agent-node
in-process commhub for claude-agent-sdk at cli.ts:1126
(createCommhubSdkMcpServer), agent-network ensureMcpJson at cli.ts:2233.

Author-Agent: 通信工程马

Author: SmartFlowAITeam

agent-network/bin/cli.ts

@@ -2231,7 +2231,24 @@ async function interactiveCreateProfile(id: string): Promise<Profile> {
 // ── ensure .mcp.json has commhub server ──
 
 function ensureMcpJson(profile: Profile) {
-  if (normalizeRuntime(profile) !== "claude-code-cli") return;
+  // #245 codex-sdk fix — widened gate (was claude-code-cli only).
+  //
+  // Both claude-code-cli and codex-sdk runtimes need `.anet/node-server.js`
+  // (the in-process commhub MCP stdio server) refreshed + the @modelcontextprotocol/sdk
+  // dependency self-healed. The difference is the discovery mechanism:
+  //   * claude-code-cli reads cwd `.mcp.json` and finds commhub there
+  //   * codex-sdk reads `~/.codex/config.toml [mcp_servers.*]` and CANNOT use
+  //     `.mcp.json` (TMCode负责人 459d1b6c diagnostic confirmed). For codex-sdk
+  //     anet-node passes a `mcp_servers.commhub` override via the Codex SDK's
+  //     `CodexOptions.config` (per-instance, in-memory) — see agent-node/src/cli.ts
+  //     `CODEX_CONFIG.mcp_servers` block. That override points at the same
+  //     `.anet/node-server.js`, so we still need to keep it fresh on this side.
+  //
+  // claude-agent-sdk and grok-build-acp do NOT use cwd .anet/node-server.js
+  // (they inject in-process via createCommhubSdkMcpServer at agent-node), so
+  // they keep the early-return.
+  const runtime = normalizeRuntime(profile);
+  if (runtime !== "claude-code-cli" && runtime !== "codex-sdk") return;
   if (!profile.channels?.some(ch => ch.includes("commhub"))) return;
 
   const mcpJsonPath = join(process.cwd(), ".mcp.json");
@@ -2322,28 +2339,25 @@ function ensureMcpJson(profile: Profile) {
     }
   }
 
-  // 只在没有 commhub 配置时才写 .mcp.json
-  // 用户可能手动配了指向开发源码(commhub-channel.ts)，不能覆盖
-  mcpConfig.mcpServers = mcpConfig.mcpServers || {};
-  const hasCommhub = Object.keys(mcpConfig.mcpServers).some(k => k.includes("commhub"));
-  if (!hasCommhub) {
-    mcpConfig.mcpServers.commhub = { type: "stdio", command: "bun", args: [".anet/node-server.js"] };
-    writeFileSync(mcpJsonPath, JSON.stringify(mcpConfig, null, 2) + "\n");
-    console.log(`[anet] .mcp.json: added commhub`);
-  }
-
-  // Write .anet/.env (hub URL + token)
+  // Write .anet/.env (hub URL + token) — both runtimes need this; node-server.js
+  // reads COMMHUB_URL / COMMHUB_TOKEN from this file when spawned as a stdio MCP.
   const anetEnvPath = join(anetDir, ".env");
   const token = profile.token || "";
   let envContent = `COMMHUB_URL=${profile.hub || "http://127.0.0.1:9200"}\n`;
   if (token) envContent += `COMMHUB_TOKEN=${token}\n`;
   writeFileSync(anetEnvPath, envContent);
 
-  // Write .mcp.json
-  mcpConfig.mcpServers = mcpConfig.mcpServers || {};
-  mcpConfig.mcpServers.commhub = { type: "stdio", command: "bun", args: [".anet/node-server.js"] };
-  writeFileSync(mcpJsonPath, JSON.stringify(mcpConfig, null, 2) + "\n");
-  console.log(`[anet] .mcp.json: added commhub channel server`);
+  // #245 codex-sdk fix — only write `.mcp.json` for claude-code-cli. codex-sdk
+  // does not read cwd `.mcp.json`; it reads `~/.codex/config.toml [mcp_servers.*]`
+  // (or accepts a `CodexOptions.config` override from agent-node, which is the
+  // path this fix uses). Writing `.mcp.json` for codex-sdk would be a silent
+  // no-op + confuse anyone reading the file expecting it to work.
+  if (runtime === "claude-code-cli") {
+    mcpConfig.mcpServers = mcpConfig.mcpServers || {};
+    mcpConfig.mcpServers.commhub = { type: "stdio", command: "bun", args: [".anet/node-server.js"] };
+    writeFileSync(mcpJsonPath, JSON.stringify(mcpConfig, null, 2) + "\n");
+    console.log(`[anet] .mcp.json: added commhub channel server`);
+  }
 }
 
 // ── launch helper (shared by start + resume) ──

agent-network/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sleep2agi/agent-network",
-  "version": "2.2.13",
+  "version": "2.2.14-preview.0",
   "description": "AI Agent Network CLI — Local-first multi-agent orchestration across 4 runtimes (Claude Code CLI / Claude Agent SDK / Codex SDK / Grok Build ACP) and 8+ LLM providers (Anthropic / OpenAI / xAI Grok / MiniMax / DeepSeek / GLM / Kimi / InternLM / Xiaomi MiMo / OpenRouter). Apache 2.0.",
   "type": "module",
   "main": "dist/src/client.js",

agent-node/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sleep2agi/agent-node",
-  "version": "2.4.10",
+  "version": "2.4.11-preview.2",
   "description": "AI Agent runtime for CommHub networks. Supports 4 runtimes: Claude Code CLI, Claude Agent SDK, Codex SDK, and Grok Build ACP.",
   "bin": {
     "agent-node": "dist/cli.js"

agent-node/src/cli.ts

@@ -1334,10 +1334,47 @@ const CODEX_INSTRUCTIONS = SYSTEM_PROMPT || [
   `你的最终回复会被系统自动 send_reply 给任务发起者。`,
 ].join("\n");
 
-const CODEX_CONFIG = {
-  model_auto_compact_token_limit: 200000,
-  developer_instructions: CODEX_INSTRUCTIONS,
-};
+// #245 codex-sdk fix — inject commhub MCP server config so codex-sdk nodes
+// get the full commhub_send_task / get_all_status / etc. tool set, with
+// per-node identity (alias / token) inherited from this agent-node process.
+//
+// Vincent retro 2026-06-17: "新成员都用不了 send_task" — new nodes default to
+// codex-sdk runtime (per [[feedback_new_node_codex_default]]); for claude-
+// agent-sdk runtime agent-node injects commhub via an in-process McpServer
+// (createCommhubSdkMcpServer at line 1126); for codex-sdk no equivalent
+// in-process channel exists, and CODEX_CONFIG had ZERO mcp_servers field —
+// so codex inherited only whatever was in `~/.codex/config.toml` (a stale,
+// limited "commhub-proxy" pointing at a separate script that got rm-rf'd
+// in the 06-16 incident).
+//
+// Fix: the Codex SDK's `CodexOptions.config` accepts a JSON object that gets
+// flattened to `--config key=value` TOML literal overrides — per-Codex-
+// instance, in-memory, no mutation of `~/.codex/config.toml`. We point
+// `mcp_servers.commhub` at `.anet/node-server.js` (the bun stdio commhub
+// MCP server that agent-network's ensureMcpJson refreshes on every start —
+// also patched in this fix to widen its runtime gate to codex-sdk). codex
+// CLI subprocess inherits parent agent-node's env (COMMHUB_ALIAS / TOKEN /
+// URL set by anet launchAgent per-node), so node-server.js runs with the
+// correct per-node identity.
+function buildCodexConfig(workdir: string): Record<string, any> {
+  const nodeServerPath = join(workdir, ".anet", "node-server.js");
+  return {
+    model_auto_compact_token_limit: 200000,
+    developer_instructions: CODEX_INSTRUCTIONS,
+    mcp_servers: {
+      commhub: {
+        command: "bun",
+        args: [nodeServerPath],
+        // env intentionally omitted: codex CLI subprocess inherits this
+        // agent-node parent process's env, which includes the per-node
+        // COMMHUB_ALIAS / COMMHUB_TOKEN / COMMHUB_URL that anet's
+        // launchAgent already sets. Hard-coding env here would re-introduce
+        // the global-alias bug the 06-17 incident exposed.
+      },
+    },
+  };
+}
+const CODEX_CONFIG = buildCodexConfig(process.cwd());
 
 async function processWithCodex(task: string, from: string, images?: string[]): Promise<string> {
   // Ensure system-installed codex binary is found (npm global bin)