mcp_tools() fails with "system error 232" on Windows when `env` is set in MCP server config

[selesnow]

Description:
When using mcp_tools(config = "...") in ellmer with an MCP server configuration that includes an env block, the connection immediately fails on Windows with the following error:

Error in `process_write_input(self, private, str, sep)`:
! Native call to `processx_connection_write_bytes` failed
Caused by error in `chain_call(c_processx_connection_write_bytes, con, str)`:
! Cannot write connection (system error 232, Идет закрытие канала.
) @processx-connection.c:977 (processx_c_connection_write_bytes)
Type .Last.error to see the more details.

If the same configuration is used without the env key, everything works as expected.

---

Steps to reproduce:

1. Create a minimal R MCP server script, e.g. create_pf_mcp_server.R.
2. Add it to the MCP config with env:

{
  "mcpServers": {
    "netpeak_pfworker": {
      "command": "Rscript",
      "args": ["-e", "source('C:/Users/User/Documents/mcp_servers/create_pf_mcp_server.R')"],
      "env": {
        "PFW_TOKEN": "00000000111111111xxxx"
      }
    }
  }
}

3. Call:

library(ellmer)
library(mcptools)

chat <- ellmer::chat_google_gemini(system_prompt = "Test")
chat$set_tools(mcptools::mcp_tools(config = "C:/path/config.json"))

4. Observe the error above.

5. Modify the config to remove env:

{
  "mcpServers": {
    "netpeak_pfworker": {
      "command": "Rscript",
      "args": ["-e", "source('C:/Users/User/Documents/mcp_servers/create_pf_mcp_server.R')"]
    }
  }
}

6. Run again — works fine.

---

Expected behavior:
Passing env should merge environment variables with the current process environment (as is common on other platforms), not completely replace them.

Actual behavior:
On Windows, processx appears to replace the entire environment, which causes the Rscript process to fail almost immediately, breaking the MCP connection.

---

System info:

other attached packages:
[1] ellmer_0.3.0        mcptools_0.1.0.9000

loaded via a namespace (and not attached):
 [1] coro_1.1.0        R6_2.5.1          magrittr_2.0.3    rappdirs_0.3.3    glue_1.8.0        
 [6] lifecycle_1.0.4   ps_1.8.1          cli_3.6.3         S7_0.2.0          processx_3.8.4    
[11] compiler_4.4.2    rstudioapi_0.17.1 tools_4.4.2       httr2_1.2.1       rlang_1.1.4       
[16] jsonlite_1.8.9

Windows 10/11
R 4.x

---

Notes:
This looks like a known processx behavior on Windows where env fully overrides the environment instead of merging it. If this is intentional, mcp_tools() might need to explicitly merge the provided env with the current Sys.getenv() before spawning the process, to ensure required variables like PATH and R_HOME are preserved.

[vladimirobucina]

Experiencing the same issue