Add lcode acp: lcode inside editors over the Agent Client Protocol

[nasser1941]

Closes #49.

What

lcode acp runs lcode as the agent inside an editor, over the Agent Client Protocol, version 1 (the stable one).

• Transport: newline-delimited JSON-RPC 2.0 on stdin/stdout. Requests run in their own threads, so cancellations and the editor's answers keep flowing during a request. Nothing else reaches stdout: print goes to stderr, and stdin is replaced so nothing can wait for typed input.
• Mapping:
  • Streaming: the model's text and thinking arrive as agent_message_chunk / agent_thought_chunk.
  • Tool calls: tool_call → in_progress → completed / failed, with a kind (read, search, edit, execute, fetch, …), the files touched, and a whole-file diff for edits.
  • Permission questions go to session/request_permission (allow once, always allow for this session, reject).
  • Plan mode: plan approval goes through the same dialog, and current_mode_update follows when it switches the mode.
  • lcode's state: the permission modes are the session modes, the todo list becomes plan updates, and usage_update and session_info_update follow each request.
  • Sessions: session/load replays a saved lcode session.
  • Commands: /review and the user's own commands go in available_commands_update.
  • The editor's buffers: with fs.readTextFile/writeTextFile, the model reads what's in the editor (unsaved changes included), and edits go through the editor. The file on disk is kept in step for checkpoints.
  • MCP: servers the editor passes (stdio and http) join lcode's own.
  • Prompts: text, resource_link and embedded resource (inlined), and images (saved and attached, so lcode's vision handles them).
• Agent hooks this needed: on_delta (streamed text and thinking), current_call, plan_review (an editor approves plans), and Toolbox.reader/writer. Permissions.approve may now return a message, and gets the permission key.
• Fix: cancelling now also ends a request whose answer was cut off mid-stream. Before, it ended as if complete.
• Docs: a new Editors page covering Zed (agent_servers with "type": "custom"), JetBrains (~/.jetbrains/acp.json), other ACP clients, what maps to what, and settings.

Tests

• tests/test_acp.py (8 tests), with a pipe-connected editor stand-in and a scripted model. It covers:
  • initializing, sessions and unknown methods;
  • edit permissions with diffs and writes through the editor;
  • reject and always-allow;
  • reading unsaved buffers;
  • plan approval with plan and mode updates;
  • cancelling;
  • files, images and /review in prompts;
  • loading and replaying a session.
• Real runs with the official ACP Python SDK (agent-client-protocol 0.12.1) as the client, which validates every message against its models, and Qwen3.6 35B through Ollama:
  1. Ask mode: fixing two bugs asked twice in the editor's dialog (allow_once/allow_always/reject_once) with whole-file diffs, and both edits went through fs/write_text_file.
  2. Unsaved buffer: a new session read calc.py and quoted a line that existed only in the editor's buffer.
  3. Plan mode: the plan was approved in the editor; the mode switched to auto-edit (current_mode_update), and the todo list arrived as plan updates.
  4. Cancel: session/cancel during a request returned stopReason: cancelled.
  5. Load: session/load in a new connection replayed the conversation.
• Not run inside Zed or a JetBrains IDE. Those need a desktop session; the docs ask for the ACP log if something looks off.

🤖 Generated with Claude Code