Changelog

All notable changes to clue are listed here. The release workflow publishes the section for the version in package.json as that release’s notes.

Unreleased

0.4.1

• The review setting chooses the reviewer, with provider and model. It replaces review_model.
• If a reviewer fails, Claude Haiku at low effort reviews instead, so one provider’s outage or limit no longer holds every action.
• $? no longer sends a command to review.

0.4.0

• Releases are compressed, so a download is about 58% smaller.
• clue upgrade compares versions, so it never replaces a newer build with an older release.
• install.sh and clue upgrade run the new binary before they install it, and keep the old one if it doesn’t start.
• clue upgrade refuses to replace a local build.

0.3.0

• clue "prompt" runs the prompt as the first turn. It also works with -c.
• --name now opens the session with that name in this folder, or creates it. An existing session is resumed and the prompt isn’t sent, so a tool can run the same command every time. It no longer fails on a taken name.
• install.sh and clue upgrade show each step and a download progress bar.

0.2.0

• clue upgrade installs the latest release.
• Sessions can have a name. clue --name <name> starts a named session, /rename changes the name, and clue -r <name> resumes it. The ID stays, and clue ask --resume still takes only an ID.
• In a multiple-choice question, a typed answer now adds to the picked options instead of replacing them.
• The prompt bar, the tool dot and the spinner are blue, and secondary text is one even grey. Prompts are no longer bold.
• Copying a sent prompt no longer picks up the ┃ bar or extra line breaks. The bar shows on its first line only, and the terminal wraps the rest.
• Pasting or dragging an image file adds it to the prompt as [Image #1]. ctrl+v and /paste attach an image from the clipboard on macOS.
• You can type while clue works. Tab or Enter queues one message, and ↑ takes it back to edit.
• A message that starts with a path, such as /tmp/a.png, is no longer read as a slash command.
• auto mode now parses shell commands instead of matching them with a regex. A background &, a newline, or a program that runs other programs, such as env, find or node, now goes to review. The list of routine commands is shorter, so some commands that used to run directly are now reviewed.
• Allow rules, deny rules and “always allow” answers now apply to each command in a chain. Deny rules also apply to read-only tools.
• clue rings the terminal bell when it needs you, and when a turn ends after 30 seconds or more without input from you. Set "bell": false in settings.json to turn it off.
• clue tells the terminal when it is idle, working, waiting for you, finished or failed, through the Program Status Protocol (OSC 7501).
• The footer after each reply shows tokens in and out. Tokens in include cached input.
• clue mcp add <name> -- <command>, clue mcp add --project and clue mcp approve --all now work as they do in /mcp.
• A session that crashed during a tool call can continue. clue records the unfinished calls as failed before the next message.
• A half-written line in a session file no longer hides the records after it.
• A 400 from the API turns off an optional feature only when the error names that feature. A 400 that names none is reported as an error.

0.1.0

The first release.

• An interactive terminal agent on Claude’s Messages API. It starts with Claude Opus 5.5 at medium effort.
• clue ask runs one request without the interactive screen. --json prints the result as JSON.
• Sessions are saved per project. clue -c resumes the last one and clue -r picks one. A resumed session sends its history to the cache while you type.
• Tools: read, find, search, write and edit files, shell commands with a terminal, web search, web fetch, subagents, skills and questions to the user.
• MCP servers over stdio, HTTP and SSE, with OAuth sign-in. Claude finds their tools through capability_search, so the tool list stays the same and stays cached.
• Permission modes ask, auto and full-access. In auto, a reviewer model checks risky actions before they run.
• Sign in with ant auth login, or store an API key with clue login.
• A single binary for macOS and Linux, on arm64 and x86_64, installed with install.sh. It starts in about 7 ms.