mcp: service env ({env:...}) not passed to local MCP servers — regression in 2.0.14/2.0.15 #50882

Summary

Since upgrading to opencode 2.0.14/2.0.15, local MCP servers configured with {env:VAR} placeholders in their environment block are spawned without any of the variables stored in the service env (service.json / opencode service get env). Servers that require those vars crash on startup, so opencode mcp list reports Connection closed. Before the upgrade (2.0.13 service) all of them connected with full tool sets.

Environment

• opencode version: 2.0.15 (also reproduced on 2.0.14 service; worked on 2.0.13)
• OS: Darwin 25.6.0 (macOS, arm64)
• Terminal: Apple_Terminal, xterm-256color, truecolor
• Shell: /bin/zsh
• Install/channel: latest; background service launched from the OpenCode desktop app bundled CLI (.../ai.opencode.desktop/cli/2.0.15/opencode-cli serve --service)
• Active plugins: none found

Reproduction

1. Configure a local MCP server in ~/.config/opencode/opencode.jsonc with an environment block using substitutions:

   { "mcp": { "servers": { "demo": {
       "type": "local",
       "command": ["python3", "/tmp/env_dependent_mcp.py"],
       "environment": { "REQUIRED_VAR": "{env:MY_VAR}" }
   }}}}

2. Store MY_VAR in the service env (e.g. ~/.config/opencode/service.json env block).
3. Start the opencode service (2.0.14+) and run opencode mcp list.
4. Observe the child process: the placeholder variable is not set (verified with a debug server that dumps its env — every key from service.json is missing in the child).

Expected Behavior

{env:MY_VAR} should resolve against the service env for spawned MCP processes, as it did on the 2.0.13 service (all servers connected: atlassian tools=98, cedrus tools=8, gp_* tools=9). opencode service get env still returns all configured variables.

Actual Behavior

Every local MCP server whose environment references {env:...} fails to start: the variables are absent in the child environment, so the servers exit with their "missing config" errors and opencode mcp list shows failed: Connection closed. Concrete symptoms:

• postgres-mcp (uvx): exits ValueError: Error: No database URL provided. Please specify via 'DATABASE_URI' environment variable → Connection closed.
• trino-mcp: exits ValueError: TRINO_USER and TRINO_PASSWORD must be set for password authentication → Connection closed.
• mcp-atlassian connects but with tools=0 (tokens absent) while previously it had 98 tools.

Trigger: today 07:43 UTC opencode upgrade (2.0.13 -> 2.0.14/2.0.15) followed by a service restart. Repeated failures on every subsequent connect attempt (consistent, not intermittent).

Additional Context

• Log excerpt (role=server): mcp connect failed ... status.error="Connection closed" for each affected server; mcp connected server=<name> tools=0 for atlassian.
• Direct proof: a minimal MCP server that dumps os.environ was spawned through the service and saw all service.json env keys as missing, while opencode service get env returned them.
• Literal (non-placeholder) values in environment do reach the child: a server with a literal DATABASE_URI started and hung on the DB connection instead of crashing instantly on the "No database URL" check.
• Manual launch of the same commands with the env exported works fine (servers respond to the MCP initialize handshake).
• Workaround applied: replaced {env:...} with literal values in the config — all five servers now connect (atlassian tools=98, cedrus tools=8, gp_smkt/gp_mm/gp_sbl tools=9).