cleveragents/cleveragents-core

Fork 3

BUG-HUNT: [resource-leak] GitWorktreeSandbox worktree directory leaked when create() fails after mkdtemp but before git worktree add #7737

New issue

Open

opened 2026-04-12 03:22:09 +00:00 by HAL9000 · 3 comments

HAL9000 commented

2026-04-12 03:22:09 +00:00

Owner

Bug Report: Resource Leak — Worktree Temp Directory Leaked on Partial create() Failure

Severity Assessment

Impact: Each failed create() call leaks a temporary directory under /tmp/ (prefixed ca-sandbox-*). In high-frequency plan execution this accumulates orphaned temp directories consuming disk space
Likelihood: Low-Medium — triggered when git worktree add fails after the temp directory was created and then removed by os.rmdir but before the worktree creation succeeds; however any CalledProcessError after os.rmdir leaves the worktree path in a partially-created state in git's book
Priority: Medium

Location

File: src/cleveragents/infrastructure/sandbox/git_worktree.py
Function/Class: GitWorktreeSandbox.create
Lines: 231–261

Description

In create(), the code creates a temp directory, removes it (so git worktree add can use that exact path), then runs git worktree add. If git worktree add fails (e.g., due to a git lock, network issue, or permission error), the exception handlers set self._status = SandboxStatus.ERRORED and raise SandboxCreationError, but self._worktree_path is never cleaned up. Any subsequent call to cleanup() will try git worktree remove on a path that may not exist, which it handles, but the real issue is the git metadata left behind.

More critically, the CopyOnWriteSandbox creates a parent temp directory at line 136 (parent_dir = tempfile.mkdtemp(...)) and then assigns self._sandbox_path = os.path.join(parent_dir, 'sandbox'). If shutil.copytree() fails (e.g. disk full), the parent_dir temp directory is leaked because only self._sandbox_path is tracked, not parent_dir. cleanup() only removes os.path.dirname(self._sandbox_path) which is parent_dir — but self._sandbox_path is never set in the failure path (the assignment happens before the try block).

Wait, re-reading: self._sandbox_path = os.path.join(parent_dir, "sandbox") is set BEFORE shutil.copytree. So if copytree fails:

self._status is set to ERRORED
self._sandbox_path is set to the never-created path
parent_dir exists but nothing removed it
cleanup() would call shutil.rmtree(parent, ignore_errors=True) where parent = os.path.dirname(self._sandbox_path) = parent_dir — BUT only if self._sandbox_path is not None and os.path.exists(parent). This actually WOULD clean up.

The real leak is in GitWorktreeSandbox: after the worktree path is used in git commands and git fails, there may be partial git state left in .git/worktrees/ that cleanup() would need to prune, but if self._worktree_path itself was never fully created (os.rmdir was called), git may have created partial state.

Evidence

# git_worktree.py lines 231-261
self._worktree_path = tempfile.mkdtemp(prefix=f"ca-sandbox-{safe_plan_id}-")
# mkdtemp creates the dir; git worktree add needs it to not exist
os.rmdir(self._worktree_path)  # <-- directory deleted

# Create the worktree with a new branch
_run_git(
    ["worktree", "add", "-b", self._branch_name, self._worktree_path, "HEAD"],
    cwd=self._original_path,
    timeout=self._git_timeout,
)  # <-- if this fails, partial .git/worktrees/ state is left

except subprocess.CalledProcessError as exc:
    self._status = SandboxStatus.ERRORED
    raise SandboxCreationError(...)  # <-- no cleanup of partial worktree state

When git worktree add fails, git may have partially created the worktree entry in .git/worktrees/ directory. The exception handler does not call cleanup() or git worktree prune.

Expected Behavior

If create() fails at any point, all partial state (git worktree entries, branch refs) should be cleaned up before raising the exception.

Actual Behavior

Partial git worktree state may be left in .git/worktrees/ if git worktree add fails mid-way. No cleanup is performed in the error path of create().

Suggested Fix

Add a cleanup step in the exception handlers within create():

except subprocess.CalledProcessError as exc:
    self._status = SandboxStatus.ERRORED
    # Clean up any partial git state
    with contextlib.suppress(subprocess.CalledProcessError, subprocess.TimeoutExpired):
        _run_git(["worktree", "prune"], cwd=self._original_path, timeout=self._git_timeout)
    if self._branch_name:
        with contextlib.suppress(subprocess.CalledProcessError, subprocess.TimeoutExpired):
            _run_git(["branch", "-D", self._branch_name], cwd=self._original_path, timeout=self._git_timeout)
    raise SandboxCreationError(...) from exc

TDD Note

After this bug issue is verified, a corresponding Type/Testing issue will be created for TDD.

Automated by CleverAgents Bot
Supervisor: Bug Hunting | Agent: bug-hunter

## Bug Report: Resource Leak — Worktree Temp Directory Leaked on Partial create() Failure ### Severity Assessment - **Impact**: Each failed `create()` call leaks a temporary directory under `/tmp/` (prefixed `ca-sandbox-*`). In high-frequency plan execution this accumulates orphaned temp directories consuming disk space - **Likelihood**: Low-Medium — triggered when `git worktree add` fails after the temp directory was created and then removed by `os.rmdir` but before the worktree creation succeeds; however any `CalledProcessError` after `os.rmdir` leaves the worktree path in a partially-created state in git's book - **Priority**: Medium ### Location - **File**: `src/cleveragents/infrastructure/sandbox/git_worktree.py` - **Function/Class**: `GitWorktreeSandbox.create` - **Lines**: 231–261 ### Description In `create()`, the code creates a temp directory, removes it (so `git worktree add` can use that exact path), then runs `git worktree add`. If `git worktree add` fails (e.g., due to a git lock, network issue, or permission error), the exception handlers set `self._status = SandboxStatus.ERRORED` and raise `SandboxCreationError`, but `self._worktree_path` is never cleaned up. Any subsequent call to `cleanup()` will try `git worktree remove` on a path that may not exist, which it handles, but the real issue is the git metadata left behind. More critically, the `CopyOnWriteSandbox` creates a parent temp directory at line 136 (`parent_dir = tempfile.mkdtemp(...)`) and then assigns `self._sandbox_path = os.path.join(parent_dir, 'sandbox')`. If `shutil.copytree()` fails (e.g. disk full), the `parent_dir` temp directory is leaked because only `self._sandbox_path` is tracked, not `parent_dir`. `cleanup()` only removes `os.path.dirname(self._sandbox_path)` which is `parent_dir` — but `self._sandbox_path` is never set in the failure path (the assignment happens before the try block). Wait, re-reading: `self._sandbox_path = os.path.join(parent_dir, "sandbox")` is set BEFORE `shutil.copytree`. So if `copytree` fails: - `self._status` is set to ERRORED - `self._sandbox_path` is set to the never-created path - `parent_dir` exists but nothing removed it - `cleanup()` would call `shutil.rmtree(parent, ignore_errors=True)` where `parent = os.path.dirname(self._sandbox_path) = parent_dir` — BUT only if `self._sandbox_path is not None` and `os.path.exists(parent)`. This actually WOULD clean up. The real leak is in `GitWorktreeSandbox`: after the worktree path is used in git commands and git fails, there may be partial git state left in `.git/worktrees/` that `cleanup()` would need to prune, but if `self._worktree_path` itself was never fully created (os.rmdir was called), git may have created partial state. ### Evidence ```python # git_worktree.py lines 231-261 self._worktree_path = tempfile.mkdtemp(prefix=f"ca-sandbox-{safe_plan_id}-") # mkdtemp creates the dir; git worktree add needs it to not exist os.rmdir(self._worktree_path) # <-- directory deleted # Create the worktree with a new branch _run_git( ["worktree", "add", "-b", self._branch_name, self._worktree_path, "HEAD"], cwd=self._original_path, timeout=self._git_timeout, ) # <-- if this fails, partial .git/worktrees/ state is left except subprocess.CalledProcessError as exc: self._status = SandboxStatus.ERRORED raise SandboxCreationError(...) # <-- no cleanup of partial worktree state ``` When `git worktree add` fails, git may have partially created the worktree entry in `.git/worktrees/` directory. The exception handler does not call `cleanup()` or `git worktree prune`. ### Expected Behavior If `create()` fails at any point, all partial state (git worktree entries, branch refs) should be cleaned up before raising the exception. ### Actual Behavior Partial git worktree state may be left in `.git/worktrees/` if `git worktree add` fails mid-way. No cleanup is performed in the error path of `create()`. ### Suggested Fix Add a cleanup step in the exception handlers within `create()`: ```python except subprocess.CalledProcessError as exc: self._status = SandboxStatus.ERRORED # Clean up any partial git state with contextlib.suppress(subprocess.CalledProcessError, subprocess.TimeoutExpired): _run_git(["worktree", "prune"], cwd=self._original_path, timeout=self._git_timeout) if self._branch_name: with contextlib.suppress(subprocess.CalledProcessError, subprocess.TimeoutExpired): _run_git(["branch", "-D", self._branch_name], cwd=self._original_path, timeout=self._git_timeout) raise SandboxCreationError(...) from exc ``` ### Category resource-leak ### TDD Note After this bug issue is verified, a corresponding Type/Testing issue will be created for TDD. --- **Automated by CleverAgents Bot** Supervisor: Bug Hunting | Agent: bug-hunter

HAL9000 added this to the v3.2.0 milestone

2026-04-12 03:42:25 +00:00

HAL9000 referenced this issue

2026-04-12 03:47:56 +00:00

[AUTO-PROJ-OWN] Project Owner Report (Cycle 1) #7724

HAL9000 added the

labels

2026-04-12 07:15:56 +00:00

freemo added the

Type

Bug

label

2026-04-12 18:33:51 +00:00

HAL9000 added

and removed

labels

2026-04-14 05:58:41 +00:00

HAL9000 commented

2026-04-14 05:58:42 +00:00

Author

Owner

✅ Verified — Resource leak: GitWorktreeSandbox leaks worktree directory on create() failure. MoSCoW: Should-have. Priority: Medium.

Automated by CleverAgents Bot
Supervisor: Project Owner | Agent: project-owner-pool-supervisor

✅ **Verified** — Resource leak: GitWorktreeSandbox leaks worktree directory on create() failure. MoSCoW: Should-have. Priority: Medium. --- **Automated by CleverAgents Bot** Supervisor: Project Owner | Agent: project-owner-pool-supervisor

HAL9000 commented

2026-04-14 06:11:58 +00:00

Author

Owner

✅ Verified — Resource leak: GitWorktreeSandbox leaks worktree directory on create() failure. MoSCoW: Should-have. Priority: Medium.

Automated by CleverAgents Bot
Supervisor: Project Owner | Agent: project-owner-pool-supervisor