Codex install omits CBM_CACHE_DIR from env_vars, causing MCP handshake failure #1562

Version

codebase-memory-mcp 0.10.2

Platform

Linux (arm64)

Install channel

GitHub release archive / install.sh

Binary variant

standard

What happened, and what did you expect?

When CBM_CACHE_DIR points to a non-default cache directory, codebase-memory-mcp install -y --force configures the Codex MCP server without forwarding that variable:

[mcp_servers.codebase-memory-mcp]
command = "/root/.local/bin/codebase-memory-mcp"
args = []

Codex sanitizes the environment of stdio MCP subprocesses unless variables are explicitly listed in env_vars. Consequently, the CBM subprocess spawned by Codex falls back to the default cache directory while the active account daemon uses the custom CBM_CACHE_DIR.

The cache mismatch closes the MCP connection during initialization. Codex then exposes none of the CBM tools.

I expected the CBM installer to configure Codex to forward CBM_CACHE_DIR whenever that variable is set:

[mcp_servers.codebase-memory-mcp]
command = "/root/.local/bin/codebase-memory-mcp"
args = []
env_vars = ["CBM_CACHE_DIR"]

Reproduction

1. Set a custom cache directory:

   export CBM_CACHE_DIR=/workspaces/example/.cache/codebase-memory-mcp

2. Install/configure CBM:

   codebase-memory-mcp install -y --force

3. Start the daemon and index a repository:

   codebase-memory-mcp daemon start
   codebase-memory-mcp cli index_repository --repo-path /workspaces/example

4. Inspect ~/.codex/config.toml. The generated MCP entry has command and args, but no env_vars.

5. Start Codex from an environment containing CBM_CACHE_DIR and request a CBM tool such as list_projects.

6. Observe that CBM tools are unavailable and the MCP handshake fails.

For comparison, add this manually:

env_vars = ["CBM_CACHE_DIR"]

Start a new Codex session and call list_projects again. The MCP connection succeeds and the indexed project is returned.

No repository source code is required to reproduce the problem.

Logs

error: CBM could not start because the active account daemon uses a different cache

Codex also reports:

MCP startup failed: handshaking with MCP server failed: connection closed: initialize response

The two cache locations in the reproduced case were:

Active daemon: /workspaces/example/.cache/codebase-memory-mcp
Codex-spawned CBM fallback: /root/.cache/codebase-memory-mcp

Proposed solution

When CBM_CACHE_DIR is set during installation, add it to the Codex server’s forwarded environment:

[mcp_servers.codebase-memory-mcp]
env_vars = ["CBM_CACHE_DIR"]

The installer should preserve existing env_vars entries and add CBM_CACHE_DIR only if it is absent.

The same setting may also be needed in installer-generated Codex agent profiles that declare their own [mcp_servers.codebase-memory-mcp] transport.

A regression test could:

1. Run configuration generation with CBM_CACHE_DIR set.
2. Verify that the resulting Codex server table contains env_vars = ["CBM_CACHE_DIR"].
3. Verify that existing forwarded variables are preserved.
4. Verify that the generated TOML remains idempotent across repeated installs.

Project scale

Not relevant; the failure occurs during MCP initialization before graph access.

Confirmations

• I searched existing issues and did not find a duplicate mentioning Codex env_vars or CBM_CACHE_DIR.
• The reproduction does not require proprietary code.

Related but distinct: #1383 discusses account-daemon cache conflicts, but not the Codex installer’s missing environment forwarding.