copilot.el v0.7.0

New Features

Agent mode tooling:

• Add MCP (Model Context Protocol) server support for agent mode via the new copilot-mcp-servers option. Configured servers are forwarded to the language server and their tools become available in copilot-chat, each prompting for confirmation like the built-in tools.
• Add copilot-chat-list-mcp-tools to see the connected MCP servers, their status, and their tools, and warn when an MCP server fails to start (previously silent).
• Preview file changes in a temporary buffer before confirming an agent-mode edit tool (create_file, insert_edit_into_file, replace_string_in_file), so you can see what will be written before approving. Controlled by copilot-chat-preview-tool-edits (default on).
• Offer an always choice at agent-mode tool confirmation prompts, alongside yes and no, to approve a tool for the rest of the conversation instead of confirming each call. The choice is remembered until a new conversation starts.
• Run the agent-mode run_in_terminal tool asynchronously instead of blocking Emacs for the whole command. The editor stays responsive, the language-server connection keeps flowing, C-g aborts a running command, and copilot-chat-terminal-timeout (default 30s) kills runaways. The command's exit status is now reported back to the model.
• Show a status line in the chat buffer when Copilot runs one of its own built-in or MCP tools that asks for confirmation (read_file, the edit tools, etc.), so those tool calls leave a trace, not just the ones copilot.el executes locally.

Workspace access for agent mode:

• Answer the server's workspace/findFiles and workspace/findTextInFiles requests so agent mode can search the project by file glob and by text/regex (via ripgrep), a step toward whole-workspace questions.
• Answer the server's workspace/readFile and workspace/readDirectory requests so agent mode can read files and list directories across the project.
• Add opt-in semantic search (copilot-chat-enable-semantic-search): declare the watchedFiles capability and answer the server's copilot/watchedFiles request with the project's file list, so the server can build a workspace index for whole-codebase questions.

Chat usability:

• Add copilot-chat-slash-command (C-c /) to pick and send a chat slash command (/explain, /fix, /tests, /doc, etc.) fetched from the server, with optional arguments.
• Act on chat code blocks: copilot-chat-insert-code-block (C-c C-i) inserts the block at point into the source buffer, and copilot-chat-copy-code-block (C-c M-w) copies it to the kill ring, instead of leaving chat output as read-only text.
• Attach extra context to a chat message: copilot-chat-add-file-reference (C-c C-f) and copilot-chat-add-region-reference send specific files or selections along with the next message, and copilot-chat-clear-references drops anything pending.

Account and usage:

• Track Copilot usage quota: surface the server's quota warnings and add copilot-quota to show how much of your chat, completion, and premium-request allowance is left.
• Announce when a suggestion matches public code, and collect the matches (with licenses and reference URLs) in a buffer shown by copilot-list-code-citations. Controlled by copilot-show-code-citations (default on). (#471)

Changes

• Show readable agent-mode confirmation prompts for the server's own tools (read_file, insert_edit_into_file, replace_string_in_file, etc.) instead of a raw plist dump, falling back to the server-provided message, then the raw input, for any tool we don't recognize. copilot-chat-auto-approve-tools is matched exactly so it never auto-approves a same-named tool from a different namespace.

Bug Fixes

• Fix agent-mode tool confirmations always failing (so Copilot Chat couldn't read files, edit, etc.) by returning the (:result "accept")/(:result "dismiss") shape the server expects instead of a bare string. (#483)
• Resolve a concrete default chat model from the server when copilot-chat-model is nil (preferring the server's designated chat default, then an auto model) instead of leaving the model unset, which some servers answer with an empty reply. The lookup runs once per session, only against an already-running server. (#473)
• Surface a turn-level error reported by the server at the end of a chat turn, instead of leaving only an empty reply. (#473)
• Stop the copilot--infer-indentation-offset warning from firing while generating chat context, so chatting from a buffer whose mode has no configured indentation offset no longer nags. (#473)
• Report Flycheck diagnostics for the agent-mode get_errors tool, not just Flymake, so Flycheck users get real diagnostics instead of "no diagnostics available".