acp: session/new ignores configured default model (always falls back to opencode/deepseek-v4.1-flash) #50378

Summary

When OpenCode runs as an ACP agent (opencode acp), session/new ignores the model set in opencode.json(c) and always reports a built-in fallback model (opencode/deepseek-v4.1-flash) as currentValue. ACP clients (e.g. Neovim plugins using OpenCode as their provider) therefore start sessions with the wrong model, while the TUI/server resolve the configured default correctly.

Environment

• opencode version: 2.0.11
• OS: Linux 7.0.0-31-generic (Ubuntu, x86_64)
• Terminal: alacritty (TERM=alacritty, COLORTERM=truecolor)
• Shell: /bin/bash
• Install/channel: latest (release)
• Active plugins: none

Config in use (~/.config/opencode/opencode.jsonc):

{
  "$schema": "https://opencode.ai/config.json",
  "model": "github-copilot/kimi-k3"
}

Reproduction

1. Set a default model in the global config, e.g. "model": "github-copilot/kimi-k3" as above.
2. Verify the server honors it: opencode api get /api/model/default → returns github-copilot/kimi-k3 (works as expected).
3. Start OpenCode as an ACP agent: opencode acp (this is what editor integrations spawn).
4. Send initialize (protocolVersion 1), then session/new with any cwd and mcpServers: [] over newline-delimited JSON-RPC.
5. Inspect the configOptions in the session/new result.

Minimal probe script (Python):

import json, subprocess, threading, time

proc = subprocess.Popen(["opencode", "acp"], stdin=subprocess.PIPE,
                        stdout=subprocess.PIPE, text=True, cwd="/tmp")
lines = []
threading.Thread(target=lambda: [lines.append(l) for l in proc.stdout], daemon=True).start()

def send(obj):
    proc.stdin.write(json.dumps(obj) + "\n"); proc.stdin.flush()

send({"jsonrpc": "2.0", "id": 1, "method": "initialize", "params": {
    "protocolVersion": 1,
    "clientCapabilities": {"fs": {"readTextFile": False, "writeTextFile": False}, "terminal": False},
    "clientInfo": {"name": "probe", "version": "0.0.1"}}})
time.sleep(2)
send({"jsonrpc": "2.0", "id": 2, "method": "session/new",
      "params": {"cwd": "/tmp", "mcpServers": []}})
time.sleep(4)

for l in lines:
    try:
        res = json.loads(l).get("result") or {}
    except ValueError:
        continue
    for opt in res.get("configOptions") or []:
        if opt.get("category") == "model":
            print("currentValue:", opt.get("currentValue"))
proc.kill()

Expected Behavior

The model config option's currentValue in the session/new response reflects the configured default model (github-copilot/kimi-k3), matching what the TUI and the HTTP API (/api/model/default) resolve.

Actual Behavior

currentValue is always opencode/deepseek-v4.1-flash, regardless of the model set in global or project config, and regardless of the cwd passed to session/new (tested from a config-free temp dir, from the home directory, and from ~/.config/opencode itself — same result everywhere).

Notes:

• The configured model is present in the option's options list, so this is not an availability problem.
• The model can be corrected per session afterwards: session/set_config_option with configId: "model", value: "github-copilot/kimi-k3" succeeds and sticks. So only the initial default is wrong.
• Impact: any ACP client that relies on the agent's default (e.g. agentic.nvim with provider = "opencode-acp") silently runs sessions on the fallback model instead of the user's configured one. Users must either switch models manually per session or configure the client to force the model after session/new (agentic.nvim: acp_providers["opencode-acp"].initial_model).
• Reproducible: consistently, on every session/new.