✅ Done! On the local backend, the automation service itself never clones the target repo — and it never uses worktrees or a pre-existing clone either. Here's how it actually flows: 1. The backend/dispatcher doesn't clone anything.

LocalAgentServerBackend

(

backends/local.py

) just points at the already-running agent server.

dispatcher.py

uploads a tarball (the automation's preset code) and runs it, extracting into a run-isolated work dir:

~/.openhands/workspaces/automation-runs/{run_id}/

(see

get_work_dir

). 2. The repo clone happens inside the run. The preset tarball's generated script (

presets/{plugin,prompt}/sdk_main.py

) calls

workspace.clone_repos()

, which ultimately invokes the SDK's

clone_repos()

→

_clone_single_repo()

(

openhands/sdk/workspace/repo.py

). That does a plain `git clone`: •

git clone --depth 1 [--branch <ref>] <url> <dest>

for branch/tag refs (shallow) • a full

git clone

+

checkout <sha>

for commit-SHA refs So it's a fresh clone into the run's isolated workspace each time — no worktree, no reuse of a prior clone. 3. Worktrees exist in the SDK but aren't used here. The SDK's

StartConversationRequest

has a

worktree

flag (default

False

) that would create a dedicated worktree under

/tmp/conversation-worktrees/...

. The automation's generated script never sets it, so automations never take that path. The one place a "preexisting clone" is reused: the separate git-sync feature (

git_sync/client.py

,

ensure_repo

). But that's a different mechanism — it maintains a persistent checkout of the config repo that stores automation definitions (prompts/plugins) so it can round-trip edits back to git. It clones once and reuses that working copy on subsequent sync cycles. It is not how a fired automation obtains the repo it's supposed to act on. TL;DR — local automations do a fresh, shallow-where-possible

git clone

inside the run's isolated workspace. No worktrees, no shared pre-existing clone.