fix(cli): retry Windows worktree cleanup locks

[alex-alecu]

Summary

Followup for #9600

• Make Windows worktree and temp-directory cleanup tolerant of transient file-lock errors.
• Keep the implementation in Kilo-specific helpers, with only small annotated hook points in shared OpenCode files.

Why

This is both a real cleanup bug and a test reliability bug: Worktree.remove can hit the same slow Windows handle-release race in normal worktree removal, while CI also exposed it during test temp-directory teardown after otherwise-passing tests.

We know that from the failure split: some failures came from the production worktree removal path (git worktree remove / directory cleanup), and others came only from global test cleanup (EACCES/EBUSY deleting opencode-test-data-*).

Upstream

I checked anomalyco/opencode dev, recent PRs/issues, and code for an equivalent EBUSY/EACCES/EPERM worktree cleanup fix. I did not find an upstream fix to port.

Upstream currently has similar cleanup assumptions, but its recent Windows unit runs are passing; this appears to be a latent Windows race that Kilo hits more often because we run extra worktree/session tests and branch CI combinations.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (7 files)

• .changeset/windows-worktree-cleanup.md
• packages/opencode/src/kilocode/worktree-cleanup.ts
• packages/opencode/src/worktree/index.ts
• packages/opencode/test/fixture/fixture.ts
• packages/opencode/test/kilocode/cleanup.ts
• packages/opencode/test/kilocode/worktree-remove-lock.test.ts
• packages/opencode/test/preload.ts

---

_{Reviewed by gpt-5.5-2026-04-23 · 365,523 tokens}