Description
After installing or reinstalling nsolid-plugin, the MCP servers may fail to initialize during the harness’s first session. The issue is consistently reproducible in Codex:
MCP startup interrupted. The following servers were not initialized:
ncm, ns-benchmark, nsolid-console
Native plugin installations do not include node_modules, so scripts/mcp-wrapper.js cannot resolve mcp-remote locally. The wrapper then falls back to:
npx -y mcp-remote@0.1.38
This introduces an npm registry lookup or package download during MCP startup. Codex starts all three servers concurrently, and they may exceed startup_timeout_sec, even when configured to 60 seconds.
Subsequent sessions usually work because the first attempt has already warmed the npm/npx cache.
Current behavior
- The first startup depends on npm registry availability and network performance.
- All three MCP servers may time out simultaneously.
- Users receive an initialization error even when the plugin and credentials are correctly installed.
- Behavior differs between a cold session and subsequent sessions.
- Offline environments, or environments with slow or restricted npm access, cannot reliably initialize the MCP servers.
- Claude and Antigravity potentially have the same underlying problem because they use the shared wrapper, although their startup or timeout behavior may hide it.
Expected behavior
After completing the documented installation and setup process, the MCP servers should initialize successfully in the first session without downloading dependencies or contacting npm during harness startup.
Impact
- Poor first-run experience after installing or updating the plugin.
- Misleading configuration or authentication errors.
- Non-deterministic, network-dependent startup.
- Restarting or resuming a session appears to fix the issue without explaining its cause.
- Any harness using the shared STDIO-to-HTTP wrapper may be affected.
Technical evidence
- mcp-remote cannot be resolved from the installed Codex or Claude plugin directories.
- Installed plugin artifacts do not contain node_modules.
- The wrapper fallback executes npx -y mcp-remote@0.1.38.
- A cold execution of that command may take longer than 60 seconds.
- Once the npm/npx cache is warm, subsequent sessions can initialize the same MCP servers successfully.
Description
After installing or reinstalling nsolid-plugin, the MCP servers may fail to initialize during the harness’s first session. The issue is consistently reproducible in Codex:
MCP startup interrupted. The following servers were not initialized:
ncm, ns-benchmark, nsolid-console
Native plugin installations do not include node_modules, so scripts/mcp-wrapper.js cannot resolve mcp-remote locally. The wrapper then falls back to:
npx -y mcp-remote@0.1.38
This introduces an npm registry lookup or package download during MCP startup. Codex starts all three servers concurrently, and they may exceed startup_timeout_sec, even when configured to 60 seconds.
Subsequent sessions usually work because the first attempt has already warmed the npm/npx cache.
Current behavior
Expected behavior
After completing the documented installation and setup process, the MCP servers should initialize successfully in the first session without downloading dependencies or contacting npm during harness startup.
Impact
Technical evidence