Modifications to the code buffer seem to happen twice

[bronger]

Checklist

• I agree to communicate with the author myself (not AI-generated).
• I've read the README's Filing issues section.
• I'm running the latest versions (fill in below).
  • agent-shell version: v0.51.1
  • acp.el version: 0.11.3
  • ACP package (e.g. claude-agent-acp) version: ~/.emacs.d/elpa/agent-shell-20260512.1559/agent-shell-opencode.el
  • Agent CLI (e.g. claude, gemini) version: I don’t know
• For requesting new agent support, I'm including a link to the ACP-capable agent or related ACP package.
• For issues, I'm including ACP traffic (as per README).

I use opencode with locally running LLMs. Here is a fraction of my opencode configuration:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["@tarquinen/opencode-dcp@latest"],
  "model": "blablador/alias-fast",
  "small_model": "blablador/alias-fast",
  "permission": {
    "edit": "ask",
    "bash": "ask",
    "webfetch": "allow"
  },

When the agent is supposed to modify a code buffer, the following happens: (Taken from *Messages*.)

Press RET or y to Allow once
Allow once
Reverting buffer ‘controllers.go’
controllers.go changed on disk; really edit the buffer? (y, n, r or C-h) r
LSP :: Disconnected
Press RET or y to Allow once
Allow once

(I use gopls as my Go LSP server.)

This seems to mean that the source code file is changed twice: Once directly in the file system, and then by patching the buffer contents. Funny enough, there is a second edit which happened without issues.

The traffic is here: https://wilson.bronger.org/traffic

[xenodium]

Thanks for filing.

What's likely happening here is that the agent modified the file outside of Emacs, accessing the file directly through the filesystem. When the agent requested another file edit, it went through Emacs, and you got what you saw. ACP enables agents (servers) to modify through the clients (Emacs/agent-shell).

Agents can't always route their edits through the client, especially if batching changes through a script. Having said that, the agent can always grep and subsequently ask the client to make the actual changes.

If you'd like to turn off client (Emacs) file handling as a workaround, it can be turned off via agent-shell-text-file-capabilities.

[bronger]

I set agent-shell-text-file-capabilities to nil, restarted Emacs, and started a new session, still, sometimes Emacs wants me to resolve the above conflict.

[xenodium]

sometimes Emacs wants me to resolve the above conflict

Emacs is going to want to do this anyway if a buffer is visiting a file and something outside Emacs modified such file. It's whether or not it happens when an agent asks Emacs to modify a file on its behalf that we can control with that variable. If it did happen at this time, then we'd need traffic to see whats going on after unsetting the variable.