Files
cleveragents-core/features/overlay_sandbox.feature
Luis Mendes 7f078f75a5
CI / benchmark-publish (pull_request) Has been skipped
CI / lint (pull_request) Successful in 28s
CI / security (pull_request) Successful in 1m8s
CI / build (pull_request) Successful in 20s
CI / helm (pull_request) Successful in 34s
CI / quality (pull_request) Successful in 3m53s
CI / integration_tests (pull_request) Successful in 4m2s
CI / typecheck (pull_request) Successful in 4m9s
CI / unit_tests (pull_request) Successful in 4m48s
CI / docker (pull_request) Successful in 1m19s
CI / e2e_tests (pull_request) Successful in 9m47s
CI / coverage (pull_request) Successful in 11m41s
CI / status-check (pull_request) Successful in 1s
CI / quality (push) Successful in 52s
CI / lint (push) Successful in 3m18s
CI / build (push) Successful in 20s
CI / typecheck (push) Successful in 3m56s
CI / helm (push) Successful in 22s
CI / security (push) Successful in 4m15s
CI / unit_tests (push) Successful in 4m39s
CI / docker (push) Successful in 18s
CI / integration_tests (push) Successful in 6m56s
CI / e2e_tests (push) Successful in 12m34s
CI / coverage (push) Successful in 11m39s
CI / status-check (push) Successful in 1s
CI / benchmark-regression (push) Has been skipped
CI / benchmark-publish (push) Successful in 30m10s
CI / benchmark-regression (pull_request) Successful in 57m33s
fix(sandbox): make commit_all atomic per specification
Changed SandboxManager.commit_all() from partial-commit semantics to
all-or-nothing atomic operation per specification requirement. On
partial failure, already-committed sandboxes are rolled back. Error
reporting indicates which sandbox failed and what was rolled back.
Added Behave scenarios verifying atomicity guarantee.

Hardened atomicity guarantees after code review:

- commit_all and _rollback_committed catch Exception (not just
  SandboxError) so unexpected errors cannot bypass rollback.
  Non-SandboxError exceptions are wrapped in a new AtomicCommitError
  (chaining the original as __cause__) that carries rolled_back_ids
  and failed_rollback_ids attributes so callers can programmatically
  determine rollback outcomes.

- _rollback_committed returns both rolled_back_ids and
  failed_rollback_ids; the error result metadata now carries
  both "rolled_back" and "rollback_failed" keys.

- _rollback_committed iterates in reverse (LIFO) order following
  the standard transaction-log undo pattern.  Clarified in
  docstring that this is distinct from the specification DAG-based
  "top-down" rollback ordering (line 24632).

- TransactionSandbox.rollback() from COMMITTED now raises
  SandboxRollbackError (database commits are irreversible)
  instead of silently transitioning to ROLLED_BACK.

- TransactionSandbox is now classified as non-rollbackable in
  commit_all alongside NoSandbox and committed last in the batch,
  since database COMMIT is irreversible.  Docstring corrected to
  match the raising behavior.

- CopyOnWriteSandbox and OverlaySandbox rollback-from-COMMITTED
  uses rename-based safe_restore() to prevent data loss when
  the copytree step fails after the original was removed.

- CopyOnWriteSandbox and OverlaySandbox rollback() now catches
  Exception (not just OSError), matching the broader catch used
  in _rollback_committed, so non-OSError exceptions from
  safe_restore set the status to ERRORED correctly.

- Extracted shared _fs_utils module (backup_directory,
  safe_restore, compute_diff) with symlink, permission, and
  timestamp preservation (including directory timestamps),
  replacing duplicated per-class _backup_directory and
  _compute_diff methods.

- backup_directory defers directory permissions and timestamps
  to a bottom-up post-walk pass, fixing incorrect mtime
  preservation (POSIX file creation inside a directory overwrites
  its mtime) and preventing restrictive source permissions from
  blocking backup writes.

- backup_directory skips non-regular files (FIFOs, sockets,
  device files) with a warning to prevent hangs on special files.

- CopyOnWriteSandbox and OverlaySandbox commit() now attempts
  to restore the original from the pre-commit backup when the
  file-copy phase fails midway, preventing partial corruption.
  If the restore itself fails the backup is preserved for manual
  recovery (cleanup() still removes it).

- CopyOnWriteSandbox and OverlaySandbox commit() error handler
  now catches Exception (not just OSError) so that unexpected
  errors during the file-copy phase also trigger pre-commit
  backup restoration, preventing partial corruption of the
  original directory.

- Pre-commit backup exception handler catches Exception (not
  just OSError) preventing temp directory leaks on non-OSError
  failures from backup_directory.

- Fixed _pre_commit_backup assignment timing: the backup
  reference is now assigned only AFTER backup_directory()
  succeeds, preventing safe_restore() from corrupting an intact
  original with a partial backup when backup_directory() fails
  (e.g. disk full).

- Rollback from COMMITTED with no pre-commit backup (no changes
  were applied) is now a no-op instead of raising
  SandboxRollbackError, preventing false rollback-failure
  reports in commit_all error metadata.

- commit_all logs a warning when NoSandbox or TransactionSandbox
  instances are present in the batch since their changes cannot
  be rolled back, which breaks the atomicity guarantee.

- Pre-commit backup is skipped when compute_diff returns no
  changes, avoiding a full directory copy for no-op commits.

- GitWorktreeSandbox clears _pre_merge_commit on commit failure
  so the stale value cannot be used by future code.

- commit_all docstring documents Raises clause for AtomicCommitError
  exception wrapping behavior.

- GitWorktreeSandbox.rollback() docstring warns about
  multi-worktree safety when rolling back from COMMITTED.

- Updated SandboxStatus transition diagram in protocol.py to
  clearly show the COMMITTED -> ROLLED_BACK path.

- Added spec-contradiction note (line 45938 vs 19193) in
  commit_all docstring.

- OverlaySandbox rollback from COMMITTED now properly remounts
  OverlayFS for real overlay (unmount, clean upper/work dirs,
  remount) and uses dirs_exist_ok=True for userspace fallback
  to prevent FileExistsError if rmtree silently fails.  The
  merged directory is reset from the restored original,
  preventing stale pre-rollback data from being exposed on
  re-activation via get_path() (which allows ROLLED_BACK status).

- OverlaySandbox rollback from COMMITTED now raises
  SandboxRollbackError if the OverlayFS unmount fails,
  preventing a double-mount attempt that would leave the
  sandbox in an inconsistent state.

- OverlaySandbox rollback from ACTIVE now uses dirs_exist_ok=True
  for userspace fallback to prevent FileExistsError when rmtree
  with ignore_errors=True silently fails.

- CopyOnWriteSandbox rollback from ACTIVE now uses
  dirs_exist_ok=True in copytree to prevent FileExistsError
  when rmtree with ignore_errors=True silently fails, matching
  the fix already applied to OverlaySandbox.

- Non-rollbackable sandboxes (NoSandbox, TransactionSandbox) are
  committed last in the batch so that all rollbackable sandboxes
  commit first; if any rollbackable sandbox fails, none of the
  non-rollbackable sandboxes will have committed yet.

- Moved NoSandbox and TransactionSandbox imports to module level
  in manager.py (no circular dependency exists).

- Pre-commit backups are now created on the same filesystem as
  the original directory (using dir= argument to mkdtemp),
  avoiding cross-device copy overhead and ensuring os.rename
  compatibility.

- safe_restore now renames the target into the mkdtemp directory
  instead of removing the mkdtemp dir first, eliminating the
  residual TOCTOU window between rmdir and rename.

- safe_restore catches BaseException (not just OSError) to
  ensure the original directory is always renamed back on
  unexpected errors, preventing the original from being left
  in the renamed-aside state.

- Added AtomicCommitError exception class to protocol.py
  carrying rolled_back_ids and failed_rollback_ids attributes.

- Exported AtomicCommitError from sandbox package __init__.py
  so callers can import it from the public API.

- Added BDD scenarios: LIFO rollback order, AtomicCommitError
  wrapping with RuntimeError cause and rollback metadata,
  _fs_utils backup/restore coverage, no-change commit rollback
  success, directory timestamp preservation, OverlaySandbox
  merged dir reset after COMMITTED rollback,
  CopyOnWriteSandbox rollback from COMMITTED restores original,
  GitWorktreeSandbox rollback from COMMITTED undoes merge,
  TransactionSandbox rollback from COMMITTED raises
  SandboxRollbackError about irreversible commit.

ISSUES CLOSED: #925

Post-review hardening (PR #1146 review findings):

- OverlaySandbox rollback from COMMITTED with no backup (no-op)
  now skips the merged directory reset entirely, preventing
  unnecessary unmount/remount or re-copy that could fail and
  turn a harmless no-op rollback into a SandboxRollbackError
  during commit_all atomic recovery.

- OverlaySandbox rollback no longer double-wraps
  SandboxRollbackError: the outer except Exception handler
  now has a preceding except SandboxRollbackError clause that
  re-raises directly, avoiding a confusing double-wrapped
  error chain.

- CopyOnWriteSandbox.get_path() now accepts ROLLED_BACK status
  for consistency with OverlaySandbox and the protocol status
  transition table (ROLLED_BACK -> ACTIVE).

- CopyOnWriteSandbox rollback from COMMITTED now resets the
  sandbox copy from the restored original via rmtree+copytree,
  preventing stale pre-rollback modifications from being
  exposed on re-activation.

- rollback_all now catches Exception (not just SandboxError)
  so that unexpected rollback errors do not prevent remaining
  sandboxes from being rolled back, consistent with the pattern
  already used in _rollback_committed.

- commit_all docstring now documents a thread-safety warning:
  the method is not safe for concurrent calls on the same
  plan_id since sandbox commit/rollback runs outside the lock.

- Fixed CHANGELOG.md whitespace inconsistencies (double leading
  spaces on two lines).

Post-review hardening round 2 (PR #1146 automated review):

- safe_restore now uses os.rename (O(1) atomic rename) instead
  of shutil.copytree (O(n) recursive copy) for the main restore
  path, since backup and target are always on the same
  filesystem.  This eliminates the ENOTEMPTY bug where a partial
  copytree failure left target_path partially populated, causing
  the recovery os.rename to fail and strand the original in the
  stale temp directory.

- OverlaySandbox.get_path() now transitions ROLLED_BACK to
  ACTIVE, matching CopyOnWriteSandbox and the protocol
  transition table (ROLLED_BACK -> ACTIVE).

- GitWorktreeSandbox.get_path() now accepts ROLLED_BACK status
  for consistency with all other sandbox implementations and
  the protocol transition table (ROLLED_BACK -> ACTIVE).

- rollback_all now also handles sandboxes in COMMITTED status
  (not just ACTIVE), consistent with the state machine allowing
  COMMITTED -> ROLLED_BACK.

- cleanup_all now catches Exception (not just SandboxError) so
  a single unexpected error does not abort cleanup of remaining
  sandboxes, consistent with _rollback_committed and
  rollback_all.

- Restructured CHANGELOG entry from a single ~90-line paragraph
  into structured sub-bullets for readability.

- Added BDD scenarios: no-op rollback from COMMITTED for
  CopyOnWriteSandbox and OverlaySandbox (zero-change commit),
  commit ordering verification (rollbackable before
  non-rollbackable).

Post-review hardening round 3 (PR #1146 deep automated review):

- cleanup_abandoned now catches Exception (not just SandboxError)
  so that unexpected errors (e.g. raw OSError, PermissionError)
  do not crash the loop and prevent remaining abandoned sandboxes
  from being cleaned up, consistent with cleanup_all,
  rollback_all, and _rollback_committed.

- OverlaySandbox._mount_overlay() now catches
  subprocess.TimeoutExpired (in addition to CalledProcessError
  and OSError), preventing create() from leaving the sandbox in
  PENDING status when mount hangs beyond the timeout.

- OverlaySandbox._unmount_overlay() now catches
  subprocess.TimeoutExpired (in addition to CalledProcessError
  and OSError), preventing cleanup() from leaving the sandbox
  in a zombie state when umount hangs beyond the timeout.

- OverlaySandbox._mount_overlay() validates that overlay paths
  do not contain commas, which would corrupt the OverlayFS mount
  options string (comma is the mount option delimiter).

- GitWorktreeSandbox.commit() now checks git diff return code
  so that a failed diff command raises CalledProcessError instead
  of silently concluding there are no changes and skipping the
  merge.

- safe_restore cleanup of the temporary rollback container now
  runs in a finally block, preventing a temp directory leak
  when the rename fails and the exception is re-raised.

- Fixed misleading BDD step name: "backup path that will cause
  copytree to fail" renamed to "backup path that will cause
  rename to fail" since safe_restore now uses os.rename.
2026-03-29 18:57:15 +01:00

228 lines
10 KiB
Gherkin

Feature: Overlay filesystem sandbox lifecycle
As a developer
I want a sandbox that isolates plan changes using overlay filesystem semantics
So that changes are captured in an upper layer and merged on commit
# --- Creation ---
Scenario: Create an overlay sandbox
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
Then the ovl sandbox should be in the "created" state
And the ovl sandbox context should reference plan "plan-001"
And the ovl sandbox context should have strategy metadata "overlay"
And the ovl sandbox merged path should exist
Scenario: Creating an overlay sandbox with empty plan_id raises ValueError
Given an ovl test directory is initialised
When an ovl sandbox is created with empty plan_id
Then an ovl ValueError should be raised with message "plan_id cannot be empty"
Scenario: Creating an overlay sandbox with empty resource_id raises ValueError
Given an ovl test directory is initialised
When an ovl sandbox is prepared with empty resource_id
Then an ovl ValueError should be raised with message "resource_id cannot be empty"
Scenario: Creating an overlay sandbox with empty original_path raises ValueError
Given an ovl test directory is initialised
When an ovl sandbox is prepared with empty original_path
Then an ovl ValueError should be raised with message "original_path cannot be empty"
Scenario: Creating an overlay sandbox on a non-existent directory raises SandboxCreationError
Given an ovl test directory is initialised
Given an ovl non-existent directory
When an ovl sandbox is created on the non-existent directory for plan "plan-001"
Then an ovl SandboxCreationError should be raised
# --- Directory structure ---
Scenario: Overlay sandbox creates correct directory structure
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
Then the ovl sandbox should have upper, work, and merged directories
# --- Path resolution ---
Scenario: Resolve a path in the overlay merged directory
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl path "data/file.txt" is resolved
Then the ovl resolved path should be inside the merged directory
And the ovl sandbox should be in the "active" state
Scenario: Path traversal is rejected
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl path "../etc/passwd" is resolved
Then an ovl ValueError should be raised with message "Path traversal not allowed"
Scenario: Resolving a path on a cleaned-up sandbox raises SandboxStateError
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl sandbox is cleaned up
And the ovl path "file.txt" is resolved on a cleaned-up sandbox
Then an ovl SandboxStateError should be raised
# --- Commit ---
Scenario: Commit with no changes produces empty result
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl sandbox is committed
Then the ovl commit result should indicate success
And the ovl commit result should have 0 changed files
And the ovl commit result should have 0 added files
And the ovl commit result should have 0 deleted files
Scenario: Commit with a new file syncs it to original
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And an ovl file "new_file.txt" is created in the sandbox with content "hello world"
And the ovl sandbox is committed
Then the ovl commit result should indicate success
And the ovl commit result should have 1 added files
And the ovl file "new_file.txt" should exist in the original directory with content "hello world"
Scenario: Commit with a modified file syncs the change
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl existing file "existing.txt" is modified in the sandbox with content "updated"
And the ovl sandbox is committed
Then the ovl commit result should indicate success
And the ovl commit result should have 1 changed files
And the ovl file "existing.txt" in the original should have content "updated"
Scenario: Commit with a deleted file removes it from original
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl existing file "existing.txt" is deleted from the sandbox
And the ovl sandbox is committed
Then the ovl commit result should indicate success
And the ovl commit result should have 1 deleted files
And the ovl file "existing.txt" should not exist in the original directory
Scenario: Commit on a cleaned-up sandbox raises SandboxStateError
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl sandbox is cleaned up
And the ovl sandbox commit is attempted on cleaned-up sandbox
Then an ovl SandboxStateError should be raised
# --- Rollback ---
Scenario: Rollback restores the sandbox to original state
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And an ovl file "temp.txt" is created in the sandbox with content "temporary"
And the ovl path "temp.txt" is resolved
And the ovl sandbox is rolled back
Then the ovl sandbox should be in the "rolled_back" state
And the ovl file "temp.txt" should not exist in the sandbox
Scenario: Rollback on a non-active sandbox raises SandboxStateError
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl sandbox rollback is attempted on created sandbox
Then an ovl SandboxStateError should be raised
# --- Cleanup ---
Scenario: Cleanup removes the overlay directory
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl sandbox is cleaned up
Then the ovl sandbox should be in the "cleaned_up" state
And the ovl sandbox base path should not exist
Scenario: Cleanup is idempotent
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
And the ovl sandbox is cleaned up
And the ovl sandbox is cleaned up again
Then the ovl sandbox should be in the "cleaned_up" state
# --- Protocol properties ---
Scenario: Overlay sandbox has a unique ULID identifier
Given an ovl test directory is initialised
When an ovl sandbox is instantiated
Then the ovl sandbox_id should be a valid ULID
And the ovl sandbox status should be "pending"
And the ovl sandbox context should be None
# --- Fallback mode ---
Scenario: Fallback mode works when OverlayFS is unavailable
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-fallback"
Then the ovl sandbox should use userspace fallback
# --- Coverage boost: edge cases ---
Scenario: Commit with a message stores it in metadata
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-msg"
And an ovl file "new.txt" is created in the sandbox with content "data"
And the ovl sandbox is committed with message "snapshot before deploy"
Then the ovl commit result should indicate success
And the ovl commit metadata should contain message "snapshot before deploy"
Scenario: get_path raises SandboxStateError when merged_dir is None
Given an ovl sandbox with merged_dir forced to None
When ovl get_path is called with "file.txt" expecting an error
Then an ovl SandboxStateError should be raised with message "merged directory not set"
Scenario: Commit wraps OSError into SandboxCommitError
Given an ovl test directory is initialised
And an ovl sandbox is created and activated for plan "plan-commiterr"
And ovl shutil.copy2 is patched to raise OSError
When the ovl sandbox commit is attempted
Then an ovl SandboxCommitError should be raised
And the ovl sandbox should be in the "errored" state
Scenario: Rollback wraps OSError into SandboxRollbackError
Given an ovl test directory is initialised
And an ovl sandbox is created and activated for plan "plan-rberr"
And ovl shutil.rmtree is patched to raise OSError for rollback
When the ovl sandbox rollback is attempted
Then an ovl SandboxRollbackError should be raised
And the ovl sandbox should be in the "errored" state
Scenario: Create wraps OSError into SandboxCreationError
Given an ovl test directory is initialised
And ovl shutil.copytree is patched to raise OSError
When an ovl sandbox create is attempted for plan "plan-createerr"
Then an ovl SandboxCreationError should be raised
And the ovl sandbox should be in the "errored" state
# --- Atomic rollback from COMMITTED ---
Scenario: Rollback from COMMITTED resets the merged directory
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-rb-committed"
And an ovl file "changed.txt" is created in the sandbox with content "modified"
And the ovl sandbox is committed
Then the ovl sandbox should be in the "committed" state
When the ovl sandbox is rolled back from committed
Then the ovl sandbox should be in the "rolled_back" state
And the ovl original file "existing.txt" should contain "original content"
And the ovl merged directory should not contain stale file "changed.txt"
Scenario: Rollback from COMMITTED with no changes is a no-op
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-rb-noop"
And the ovl sandbox is committed
Then the ovl sandbox should be in the "committed" state
When the ovl sandbox is rolled back from committed
Then the ovl sandbox should be in the "rolled_back" state
And the ovl original file "existing.txt" should contain "original content"
# --- Status transitions ---
Scenario: Status transitions follow protocol
Given an ovl test directory is initialised
When an ovl sandbox is created for plan "plan-001"
Then the ovl sandbox should be in the "created" state
When the ovl path "existing.txt" is resolved
Then the ovl sandbox should be in the "active" state