Thanks for the thorough investigation and detailed PR description, @AnganSamadder! These are real issues (hierarchical branch crashes, FD exhaustion, adapter hangs) and I appreciate you digging into all of them. A few things need fixing before we can merge:

1. Build failure: tests not updated for sessionIndex removal

The opencode adapter_test.go references sessionIndex in 7 places (lines 22-23, 401, 527, 555, 588, 644, 687). Removing the field from the struct without updating tests will break compilation:

if a.sessionIndex == nil {  // line 22
sessionIndex: make(map[string]string),  // lines 401, 527, 555, 588, 644, 687

Please remove these references from the tests as well.

2. Concurrency bug in plugin_loading.go — shared channel across loop iterations

The timeout pattern has a channel-mixing race condition:

resultChan := make(chan result, len(worktreePaths))
for _, wtPath := range worktreePaths {
    go func(path string) {
        sess, err := adpt.Sessions(path)
        resultChan <- result{sessions: sess, err: err}
    }(wtPath)

    select {
    case res := <-resultChan:
        // ...
    case <-time.After(5 * time.Second):
        continue
    }
}

If iteration 1 times out, its goroutine is still running. When iteration 2's select fires, it may receive the stale result from iteration 1's goroutine instead of iteration 2's. This means results get attributed to the wrong worktree (wtName/wtPath mismatch) and some results are silently dropped.

Fix: use a per-iteration channel instead of a shared one:

for _, wtPath := range worktreePaths {
    ch := make(chan result, 1)
    go func(path string) {
        sess, err := adpt.Sessions(path)
        ch <- result{sessions: sess, err: err}
    }(wtPath)

    select {
    case res := <-ch:
        // ...
    case <-time.After(5 * time.Second):
        continue
    }
}

3. Data race on defaultBranchCache

defaultBranchCache is a package-level map[string]string with no synchronization. detectDefaultBranch is called from multiple goroutines (each adapter goroutine in loadSessions → Sessions() → diff paths). Concurrent map read+write in Go is a fatal crash — which is ironic given this PR is fixing crashes.

Please protect it with a sync.RWMutex or use sync.Map.

The worktree MkdirAll fix, git status timeout, and cursor adapter guard all look good. Just need these three items addressed. Thanks again!

Review — Round 2

Nice work addressing the feedback, @AnganSamadder. The per-worktree channel fix, sync.RWMutex on the branch cache, and the in-flight dedup guard with token validation all look solid. Tests are clean and cover the important edge cases (stale token, reset).

Recommendation: Approve with one minor note.

The three blocking issues from the first review are all resolved correctly. The in-flight guard is a good addition beyond what was asked — it prevents duplicate timeout goroutines from piling up, which is a real concern in the refresh-heavy path.

One thing to address

The PR description mentions adding a 10s timeout to git status in internal/plugins/gitstatus/tree.go, but that file isn't in the diff. Was this dropped intentionally, or did it get lost in a rebase? If it's still needed, it can go in a follow-up — not a blocker for this PR.

Verdict

Ready to merge once Marcus gives the final 👍. The crash fixes (worktree MkdirAll, concurrent map write removal), the FD exhaustion cache, and the adapter timeout hardening are all well-implemented. Good contribution. 🪶