fix(cursor): detect Cursor by its project state dir, not just .cursor/

[Soph]

https://entire.io/gh/entireio/cli/trails/1132

Every agent's DetectPresence asks "does <repo>/.<agent>/ exist". That's a fair test for Claude Code, Codex, and Gemini, which all create repo-local config as a matter of course. It's structurally unfair to Cursor: project rules are optional, and Cursor's real per-project state lives in ~/.cursor/projects/<sanitized-repo-path>.

So a repo driven daily from Cursor reports "not present". entire enable enables only detected agents when non-interactive, and pre-selects only detected ones on first run (selectAgents, setup.go:1771-1795). The observed result is worse than Cursor merely being skipped — enable falls back to the default agent, wires Claude Code, and prints "Ready.":

	before	after
enable says	Agent: Claude Code (use --agent to change)	Detected agent: Cursor
.cursor/hooks.json	absent	7 hooks
wired instead	.claude/settings.json	—

A/B'd with real binaries in a fresh repo with ~/.cursor/projects/<repo>/ present and no repo-local .cursor/.

Why this is load-bearing: the missing counterpart to #1262

#1262 reported this exact repo state — Claude Code hooks installed, .cursor/hooks.json absent, Cursor in use — and its consequence: the Cursor session was claimed by Claude's forwarded hooks and misattributed. fadc4a06a fixed that with hook_guard.go, which inspects the transcript path and makes the forwarded hook no-op when the session belongs to another agent.

That fix is correct, and it changes the failure mode for anyone in that state:

• before — Claude's forwarded hook claimed the session → sessions appeared under the wrong agent. Wrong, but visible.
• after — the guard correctly declines → nothing claims the session → zero capture, silently.

#1262 shipped one half. The guard stops the wrong agent from claiming a Cursor session; nothing ensured the right agent is wired to claim it. Users in that state went from misattributed sessions to correct attribution of nothing at all. This is that second half.

The user-visible symptom has been reported from outside more than once — #574 ("Unable to see agent steps on dashboard") and #898 ("empty Cursor session panel").

What changed

DetectPresence keeps the repo-local check first and adds the state dir as a second signal. GetSessionDir and DetectPresence both resolve it through a new projectStateDir helper, so the two cannot disagree about where Cursor keeps a repo's state.

The second signal deliberately tolerates a false positive — a repo opened in Cursor once and never used with its agent. That installs hooks which never fire, which is harmless; a false negative captures nothing and says nothing.

Verification

Cursor capture itself is healthy once hooks exist — TestSingleSessionManualCommit with cursor-cli passes on Cursor CLI 2026.08.11, full lifecycle (SessionStart → TurnStart → TurnEnd → checkpoint saved → condensed), commit trailered Entire-Checkpoint, v1 checkpoint carrying Entire-Agent: Cursor. Detection was the only thing standing between a Cursor user and that path.

Four tests, two RED-verified (they fail with the second signal stubbed, pass with it). Tests set HOME/USERPROFILE to a temp dir so they never read the developer's real ~/.cursor.

initGitRepo now runs a real git init: the hand-written .git/HEAD it created before is not a repo the git CLI accepts, so paths.WorktreeRoot failed with exit 128 on it — existing tests never noticed because DetectPresence falls back to CWD on that error. This package can't use testutil.InitRepo; architecture_test.go forbids agent packages from importing cli internals.

mise run check green.

Not addressed here

Found while investigating, each worth its own issue: headless (cursor-agent -p) captures nothing at all because only sessionStart/sessionEnd fire so no session state is created — though the transcript is written to disk, so it's fixable from sessionEnd; Cursor has no ContextInjector, and sessionStart's additional_context is verified working, so that's now viable; and nothing anywhere detects "hooks installed, never fired", which is the state every one of these leaves you in.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes which agent entire enable auto-selects and installs hooks for. False positives can wire unused Cursor hooks; false negatives previously caused silent missed capture.

Overview
Cursor is now treated as present when ~/.cursor/projects/<sanitized-repo-path> exists, even if the repo has no .cursor/ directory. That was the common case (project rules are optional), so entire enable skipped Cursor, installed the default agent instead, and captured nothing.

DetectPresence still checks repo-local .cursor/ first, then the home-dir project state dir. GetSessionDir and detection share a new projectStateDir helper so they cannot disagree on the path. An empty project dir still counts as present (hooks that never fire beat silent data loss).

Tests isolate HOME/USERPROFILE and use a real git init so WorktreeRoot resolves the same path Cursor would sanitize.

^{Reviewed by Cursor Bugbot for commit 42d4a7b. Configure here.}

[Copilot]

Pull request overview

This PR improves Cursor agent detection so entire enable can correctly auto-select Cursor in repos that don’t have a repo-local .cursor/ directory, by also treating Cursor’s home-directory per-project state (~/.cursor/projects/<sanitized-repo-path>) as a presence signal.

Changes:

• Extend CursorAgent.DetectPresence to check Cursor’s per-project state dir as a second signal after repo-local .cursor/.
• Refactor Cursor path resolution via a shared projectStateDir helper so DetectPresence and GetSessionDir can’t diverge.
• Add regression tests covering the new Cursor presence signal and update git repo test setup to use a real git init.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
cmd/entire/cli/agent/cursor/cursor.go	Adds home-dir Cursor project-state detection and shares path resolution via projectStateDir.
cmd/entire/cli/agent/cursor/cursor_test.go	Adds tests for the new detection signal and switches test repo setup to run git init.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want fixes drafted automatically? Bugbot Autofix can create code changes for findings. A team admin can enable Autofix in the Cursor dashboard.

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 42d4a7b. Configure here.}

[cursor]

Windows-broken git config isolation

Medium Severity

initGitRepo sets GIT_CONFIG_GLOBAL and GIT_CONFIG_SYSTEM to /dev/null. That path is not a usable git config on Windows (NUL also fails), so git init can fatal and these new DetectPresence tests fail on a full Windows run. The repo already documents this pitfall and isolates via a real empty config file instead.

 

^{Reviewed by Cursor Bugbot for commit 42d4a7b. Configure here.}