fix(api): mcp error with remote server

[cameronr]

Fixes #93

[sudo-tee]

@cameronr

I did some tests on this and according to the opencode docs custom command can take arguments.

---
description: Say hello to bob
agent: plan
model: github-copilot/gpt-4.1
---

Say hello to bob $ARGUMENTS and ask how he is doing today.

This dummy command used to let me add arguments after completion

What it did was inserting the text /bob in the input window and let me add arguments.

Hope it helps

[cameronr]

Ah, that's very helpful. I'll take a look!

[cameronr]

@sudo-tee I fixed arguments for users commands so now it works for both :Opencode command cmd arg and /cmd arg.

In my testing, I did uncover two race conditions. I've already fixed the first (we were triggering a full session load when sending a user command but we don't need to do that so I just removed the full re-render).

The second can still happen via the following sequence:

1. use something that sets a display_route, e.g. /help
2. enter and submit an input prompt

The issue is that api.submit_input_prompt doesn't wait for the session to finish loading before sending the input prompt:

opencode.nvim/lua/opencode/api.lua

Lines 272 to 281 in 1941a34

	function M.submit_input_prompt()
	if state.display_route then
	-- we're displaying /help or something similar, need to clear that and refresh
	-- the session data before sending the command
	state.display_route = nil
	ui.render_output()
	end
	input_window.handle_submit()
	end

Because renderer.render_full_session doesn't :wait():

opencode.nvim/lua/opencode/ui/renderer.lua

Lines 117 to 123 in 1941a34

	function M.render_full_session()
	if not output_window.mounted() or not state.api_client then
	return
	end
	fetch_session():and_then(M._render_full_session_data)
	end

I think that means it's possible we can start processing events before we've reloaded the session, which means I think would lose events if the sequence went:

1. request session
2. send prompt
3. get prompt responses
4. render session response which resets state.messages (but the session response may not include most recent prompt responses that happened while it was inflight)

render_full_session is only called by ui.render_output and that's called in 3 places (now):

1. api.close when closing a display_route:
   https://github.com/cameronr/opencode.nvim/blob/c18d8f55bf892f103876a9220df5e52b25e8fa03/lua/opencode/api.lua#L31-L38

2. core.send_message if there's an error:
   https://github.com/cameronr/opencode.nvim/blob/c18d8f55bf892f103876a9220df5e52b25e8fa03/lua/opencode/core.lua#L129-L136

3. core.open when we're not creating a new session. This happens regularly when opening
   https://github.com/cameronr/opencode.nvim/blob/c18d8f55bf892f103876a9220df5e52b25e8fa03/lua/opencode/core.lua#L67-L83

For 1, we need to force the session load to be synchronous. My proposal is to change ui.render_output to be:

---Force a full rerender of the output buffer. Should be done synchronously if
---called before submitting input or doing something that might generate events
---from opencode
---@param synchronous boolean If true, waits until session is full rendered
function M.render_output(synchronous)
  local ret = renderer.render_full_session()

  if ret and synchronous then
    ret:wait()
  end
end

And then I think we only need to set synchronous in case 1.

How's all of that sound?

[sudo-tee]

Great catch.

I think the proposed solution is good enough.

[cameronr]

Ok, I think this should be all set now

[sudo-tee]

Great job on this :)

Thanks a lot