Repository navigation
Agent Trail now reads GitHub Copilot, OpenCode, and Cursor next to Claude, Codex, and Grok. Their projects share the same cards, and none of them is added until you say so.
- GitHub Copilot: one profile reads Copilot Chat in VS Code and the Copilot CLI, along with the Copilot app and VS Code's Copilot CLI sessions, which write the same files.
- Copilot Chat in VS Code has history only. An edit tool call counts once VS Code recorded applying it to a file. Edits undone by restoring a checkpoint, rejected, or undone in the next request don't count. Chats in Remote-SSH windows, windows without a folder, and chats with no requests are skipped. VS Code saves these records about once a minute while idle, so the newest edits can show up to a minute late.
- Copilot CLI has history and live status. A successful edit counts, and subagent edits count as delegated. An open session is Working during a turn or tool call and Needs input while a permission request or a question to you is unanswered. Jump brings its terminal forward.
configDirectoryis the Copilot CLI home (~/.copilot), anduserDataDirectoryis VS Code's user-data folder; set it for Insiders.
- OpenCode: one profile reads
opencode.db, the database the terminal app, the desktop app, andopencode serveshare, opened read-only.- History: a completed
edit,write, orpatchcall counts (multieditandapply_patchin OpenCode 1.x). Calls that failed, were denied, or still wait for approval don't count, and neither do shell commands. Subagent edits count as delegated,/undohides the edits it rolled back until/redo, and OpenCode 2's import of 1.x sessions counts each session once. - Live status, OpenCode 2 only: a session executing in the background service is Working, and Needs input while it asks you a question or waits for you to approve a file edit or read. OpenCode keeps approval prompts only in memory and stores a waiting call like a running one, so an
edit,write,patch, orreadcall still running after 5 seconds counts as waiting; a slow code formatter can hold one too. A session waiting to approve a shell command or a subagent shows as Working. Whileopencode acp,--standalone, or a hand-runopencode serveshares the database, a turn it could be running is Unknown. - Jump brings the OpenCode desktop app forward when it runs (it has no link to one session), and otherwise shows the session's folder in your editor or Finder. A turn you end by denying an approval or dismissing a question leaves the live list.
- History: a completed
- Cursor: one profile reads the Cursor app's agent chats and Cursor CLI sessions.
- Cursor app history comes from Cursor's chat database, read-only, in short, time-limited queries that leave Cursor's own saves alone. A completed edit or delete call counts once Cursor recorded the result. Copies in forked chats count once, subagent edits count as delegated (a side chat's edits are yours), and worktree chats count for the worktree. Chats in remote windows, windows without a folder, and cloud agents are skipped. In Cursor 3.29, restoring a checkpoint keeps the conversation, so the edits it undid still count.
- Live status, Cursor 3.29 and later: a chat is Working from the moment you send until the turn ends or you press Stop. It needs input while its agent's question waits for your answer or Cursor flags it as blocked on you; for an approval, a subagent's included, that is about 8 seconds after the approval opens. A subagent sent to the background keeps its chat Working in background until it finishes, and a side chat shows on its own. Only runs since the running Cursor app started count, so a quit mid-turn leaves nothing behind. Jump opens the chat in Cursor's Agents window.
- Cursor CLI: sessions only, counted on their folder. Their edits are not counted yet.
configDirectoryis the Cursor CLI home (~/.cursor), anduserDataDirectoryis the Cursor app's data folder.
- Found on this Mac: a saved
profiles.jsonnever picks up new default profiles, and Copilot, OpenCode, and Cursor aren't defaults. When an assistant has sessions on this Mac but no profile reads them, the Projects panel offers Add profile. Not now hides the offer for good, and nothing is added without your click. - Unknown assistants in
profiles.json: a file that names an assistant this version doesn't know now says which one, instead of a generic JSON error. It still fails closed, so saving can't drop that profile. - Under the hood: each assistant is now described once, and its readers are registered in one place, so a new assistant can't be left out of scans or cache flushes. Existing caches keep their names and format, so nothing reindexes.
- Jump brings apps forward more reliably: a host app that isn't a terminal is brought forward through LaunchServices, as a Dock click does, when one instance of it runs (Cursor ignored the usual activation request). This also covers Claude Desktop's terminal pane, an editor's terminal pane running
claudeor Grok, and the Codex app when its link can't be opened. It never launches an app that isn't running. - Multi-root windows: in a VS Code or Cursor window with several folders, an edit in a folder beside the checkout the session started in now counts for that folder.
Validation: 477 tests pass, and CI passed on Apple Silicon and Intel for each pull request. Copilot Chat in VS Code was checked against the maintainer's VS Code data: 66 chats and 627 edits across 14 projects, matching an independent replay of the same rules except 3 edits to /tmp files outside any project. On an OpenCode 2.2.7 desktop session, Agent Trail's count matched the database exactly: 5 edits in 2 sessions, including a delegated subagent edit. An open approval showed Needs input about 5–8 seconds after it opened, and Jump and a denied approval leaving the live list were confirmed in the app. On Cursor 3.29.15, these were each checked on real sessions:
- edits
- Working within about a second of a send
- questions and approvals, a subagent's and a side chat's included
- a background subagent
- a quit mid-turn
- history after quitting Cursor
- Jump to the chat
Reviews found 20 issues in the Copilot reader and 3 in OpenCode's. Four separate reviews covered the Cursor reader and its live status. All confirmed findings were fixed, with tests.
Not yet checked on real sessions:
- the Copilot CLI
- Copilot Chat's reject, Undo, and checkpoint rules one by one
- OpenCode's
/redo - OpenCode history reads after every client and the service quit
- a Cursor worktree agent's edits
A Cursor run orphaned by a window crash or reload while Cursor keeps running stays Working until Cursor saves that chat again.
The universal2 ZIP supports Apple Silicon and Intel Macs on macOS 14+. It is ad-hoc signed and not Apple-notarized, so macOS stops the first launch; Install has the steps (Open Anyway in System Settings › Privacy & Security) and the checksum check. A local build opens without the warning. Assistant runtime formats are private and version-dependent.
Gemini CLI and Qwen Code are not in this release; they are being finished separately (#28).