Handling Model Selection in ACP

[iamthebot]

Currently, ACP adapters work a little differently than traditional remote LLM adapters in that if, for example, you're using ACP to connect to claude code you only see a "default" adapter. This is because codecompanion doesn't yet support SetSessionModel*. Also, /model is not supported as a slash command in the ACP world (because SetSessionModel is in the ACP spec).

*In the ACP docs this is considered an unstable API, but alas it is the only way to select a model within a session and is already supported by the claude code ACP.

My thinking is in the same way that we update slash commands after they're broadcasted by the ACP host we can update the available models for the adapter after we receive the model list after session initialization. It does mean that, initially, you'd still only be able to select "default" as the model until a few seconds after you open the chat buffer (this limitation already exists for slash commands though).

Happy to throw up a PR again to add this. Any thoughts?

[olimorris]

The Zed guys dropped me a message about this a couple of weeks ago. It might be marked as unstable but my experience is that they're normally pretty full baked features by the time they let me know about them so happy to accept one of your awesome PRs to add it in.

From a usability perspective, this poses an interesting dilemma. Right now when users change model in the chat buffer with ga, it displays commands in place of models. I think this should change to display the available models and I propose that an ACP adapter's commands becomes a flat array, like:

commands = {
  "npx",
  "--silent",
  "--yes",
  "@zed-industries/claude-code-acp",
},

[stuckinsnow]

I was able to get it working by updating to the latest acp (pnpm list -g @zed-industries/claude-code-acp = 0.12.6) and then adding a small bit of code to the config -

        keymaps = {
          change_model = {
            modes = { n = "gm" },
            description = "Change Model",
            callback = function(chat)
              if chat.adapter.type == "acp" then
                local models = { "sonnet", "opus", "haiku" }
                vim.ui.select(models, { prompt = "Select Model" }, function(selected)
                  if not selected or not chat.acp_connection then
                    return
                  end
                  local result = chat.acp_connection:send_rpc_request("session/set_model", {
                    sessionId = chat.acp_connection.session_id,
                    modelId = selected,
                  })
                  if result then
                    vim.notify("Model changed to " .. selected, vim.log.levels.INFO)
                  else
                    vim.notify("Failed to change model (unstable API)", vim.log.levels.WARN)
                  end
                end)
              else
                require("codecompanion.interactions.chat.keymaps.change_adapter").select_model(chat)
              end
            end,
          },
        },

[serranomorante]

For gemini-cli you could use: --model=gemini-2.5-pro

        gemini_cli = function()
          return require("codecompanion.adapters").extend("gemini_cli", {
            commands = {
              pro_2_5 = {
                "gemini",
                "--model=gemini-2.5-pro",
                "--experimental-acp",
              },
            },
   ...

[olimorris]

#2641 will add support for this. It's fully working right now and behaves the same way as changing models for a http adapter. Just need to allow users to select the command via the command line or slash command.

[JPFrancoia]

Thank you, I was just looking for that 🚀