Inferpal 1.7.5

Inferpal now reads the setup your repository already wrote for other coding agents — its instructions, its commands, its skills and its MCP servers — and connects to MCP servers of the latest protocol revision. Plus eighteen fixes and two security changes.

Your repository's agent setup, read as it is

• The instructions your repository gives coding agents go with your questions. AGENTS.md, Copilot's instruction files, CLAUDE.md and the rule folders of Claude Code, Cursor, Cline, Roo and Continue — each read the way its own tool reads it: a rule written for some files only when you are on one of them, AGENTS.md from the repository's root down to your file. Inferpal's own .inferpal/ files still win a conflict, and a file too long for the model's window is shortened and says so.
• You can see what each question carries. The Context settings page and the new /instructions command list every instruction file found, what the next question sends of it, and why when it sends nothing. A new setting turns off the files of a tool your team does not want applied.
• The commands written for Copilot, Claude Code or Continue are slash commands in Inferpal (.github/prompts, .claude/commands, .continue/prompts), and the words you type after one go where it asks for them.
• Skills. /skill lists the SKILL.md skills of your repository and your own, and /skill <name> <request> asks with one. A new setting, off by default, lets the model pick the skill a request needs. A skill's scripts only ever run through the usual command approval.
• The MCP servers your repository declares start beside yours — once you agree, even with approvals switched off. The question shows everything that changes what runs — the command, its environment, its folder, its headers — and your answer is kept until the definition changes. A repository file Inferpal cannot read is named, and your own servers keep running.

MCP

• Servers written for the 2026-07-28 revision of the protocol connect, on stdio and over HTTP. They refused Inferpal at once; servers of the earlier revisions keep receiving exactly what they did.
• A tool's structured argument reaches the model with its shape. With Ollama, an argument described as a nested object arrived with no type at all.
• A tool that answers with structured content only no longer reads "(no output)", and a tool that asks for something Inferpal cannot give says so instead of answering nothing.

Safety

• Signing in to an MCP server checks which authorization server answered before the sign-in goes on, and a client registered with one authorization server is no longer offered to another.
• The agent always asks before editing the instructions a repository gives coding agents (AGENTS.md, CLAUDE.md, Copilot's and the other tools' rule files), even with an allow rule, "Always" or security alerts switched off.
• A permission rule removed in one window stops approving in the others. It went on approving commands and edits without asking until those windows were restarted.

Nothing lost

• What you type is no longer lost when the agent edits the same file. In Visual Studio, reloading a file the agent had changed erased your unsaved typing; the agent now refuses to write over unsaved changes, and checks again after you approve.
• Nothing Inferpal started keeps running after a crash — background commands, their shells, the processes an MCP server started — on Windows, Linux and macOS.
• The first index of a large solution no longer starts over when the editor closes: it saves its progress as it goes and picks up where it stopped.
• Two solutions open side by side each get their last conversation back, instead of one overwriting the other's.
• A conversation reopened in the other editor comes back whole, including the messages and the plan card only the other editor shows.

Figures that tell the truth

• A count agrees with what it counts, in every language — no more "1 files"; French, Russian and the other languages each count their own way.
• /task, the run line, /history, Smart Fix and the code added to each question count what they say they count: every step a task took, the files a run read, the messages of a conversation, compilation errors only where the compiler gives them, and each snippet left out for its own reason.
• The project map sees positional records and primary constructors, and so do /doc and the impact analysis — in a modern C# codebase, most records were missing.

Smaller

• A multi-line value in the front matter of a rule, a check or a prompt file is read whole, and the words you type after a custom command are no longer lost.

Full changelog: v1.7.4...v1.7.5