Summary

Sessions created by the Desktop app ignore the configured default_agent and root model, opening as build with a carried-over model instead. Sessions created by the CLI resolve both correctly from the identical config. This matters because agent-specific permission rules (including subagent approval gates) belong to the configured default agent, so they are silently not applied in Desktop windows.

Environment

• opencode version: 2.0.25
• OS: Windows_NT 10.0.19045 (win32 x64)
• Terminal: Unavailable: OpenCode Desktop app (Electron); terminal environment variables are not set in this context
• Shell: C:\Windows\system32\cmd.exe (diagnostics run via Windows PowerShell 5.1)
• Install/channel: latest (OpenCode Desktop app with bundled CLI v2.0.25)
• Active plugins: None found in config

Reproduction

1. Create ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "default_agent": "coordinator",
  "model": "openai/gpt-6-sol",
  "agents": {
    "coordinator": {
      "mode": "primary",
      "model": "openai/gpt-6-sol",
      "system": "...",
      "permissions": [
        { "action": "edit", "resource": "*", "effect": "deny" },
        { "action": "shell", "resource": "*", "effect": "deny" },
        { "action": "subagent", "resource": "*", "effect": "deny" },
        { "action": "subagent", "resource": "specialist", "effect": "ask" }
      ]
    },
    "assistant": { "mode": "subagent", "model": "opencode/mimo-v2.6-flash-free" },
    "maker": { "mode": "subagent", "model": "openai/gpt-6.1-sol" },
    "specialist": { "mode": "subagent", "model": "openai/gpt-6-astra" }
  }
}

2. Verify via API that the config loaded correctly:

GET /api/config
-> "default_agent": "coordinator"
-> "model": { "providerID": "openai", "model": "gpt-6-sol" }

3. From the Desktop app, open a new window / start a new session.
4. Inspect the resulting session via GET /api/session.
5. Repeat from a terminal with a bare opencode run "Reply: OK" (no --model) in the same project.

Expected Behavior

Newly created sessions use default_agent (coordinator) and the root model (openai/gpt-6-sol), regardless of whether they were created by the Desktop app or the CLI.

Actual Behavior

CLI resolves both correctly:

$ opencode run "Reply with exactly: DEFAULT_OK"
> coordinator · gpt-6-sol
DEFAULT_OK

Desktop-created sessions do not. Root sessions (parentID null), newest first:

Two distinct deviations:

1. Agent: every Desktop-created session carries an explicit agent: "build", overriding default_agent. The CLI-created session correctly leaves it unset and resolves to coordinator.
2. Model: the Desktop model is inconsistent (mimo-v2.6-flash-free vs gpt-6-sol across sessions created minutes apart under identical config) rather than tracking the configured root model.

A session created directly through POST /api/session with no agent/model fields returns both as empty, suggesting the Desktop client is supplying them explicitly rather than leaving them to be resolved.

Additional Context

Impact. While running as build, the configured agent's permission rules are not applied. In this config the coordinator rules include edit: deny, shell: deny, and a subagent deny-* with explicit allow/ask exceptions — notably specialist → ask, which is a deliberate cost gate. Under build's default permissive policy those gates are bypassed.

Frequency. Reproduced consistently on every Desktop-created session observed (5 of 5 root sessions with an explicit agent).

Workaround tried. Running /agents → coordinator and /models → openai/gpt-6-sol in each new window. Works, but must be repeated per session. Using the CLI instead also works, but loses the Desktop UI.

Related. GET /api/agent confirms all four custom agents register correctly with the right models and permission rules, and GET /api/config confirms default_agent is present — so config parsing and agent registration appear unaffected. The deviation happens only at Desktop session-creation time.