fix: Docker API hook manager crashes for all user-provided hooks (#1878)

[hafezparast]

Summary

• Fixes [Bug]: Docker API hook manager crashes for all kind of user-provided hooks #1878: PR Release v0.8.0: Crash Recovery, Prefetch Mode & Security Fixes #1712 removed __import__ from safe builtins for security, but the hook compiler still used exec("import asyncio", namespace) which requires __import__ — breaking ALL user-provided hooks
• Replace exec-based imports with direct module injection into the namespace (namespace['asyncio'] = asyncio, etc.)
• Add standard exception classes (ValueError, TypeError, KeyError, etc.) to safe builtins so hooks can use try/except properly
• __import__ remains blocked — users still cannot import arbitrary modules at runtime

Changes

• deploy/docker/hook_manager.py: Fix module injection + add exception classes to allowed builtins
• tests/docker/test_hook_manager_unit.py: 36 unit tests covering:
  • Issue [Bug]: Docker API hook manager crashes for all kind of user-provided hooks #1878 exact repro + all 8 hook points
  • Security: import/__import__/exec/eval/open all blocked
  • Validation: empty code, sync func, missing params, syntax errors
  • Adversarial: infinite loop timeout, exception capture, 10k-line code, unicode, namespace isolation
  • Exception classes: raise/catch, isinstance checks, all 14 classes verified

Test plan

• 36 unit tests passing locally (pytest tests/docker/test_hook_manager_unit.py -v)
• Verified exact curl from issue [Bug]: Docker API hook manager crashes for all kind of user-provided hooks #1878 scenario compiles and runs
• Verified __import__ still blocked at runtime (security preserved)
• Integration test with Docker server (needs running container)

🤖 Generated with Claude Code

[ntohidi]

@hafezparast , thanks for the solid fix — the approach of replacing exec("import ...") with direct module injection is clean and correctly preserves the security intent of PR #1712. Tests are thorough.

A few things to address before the merge:

Must fix (critical):

• Remove BaseException from allowed_builtins. Since asyncio.CancelledError inherits from BaseException (Python 3.9+), a hook with except BaseException: pass can defeat the asyncio.wait_for timeout in execute_hook_safely, creating a DoS vector. Users should only be able to catch Exception and its subclasses.

Should fix:

• Inject the typing module itself into the namespace (namespace['typing'] = typing) alongside the individual names (Dict, List, Optional). Currently, if a user needs Union, Any, Tuple, etc., they have no way to access them since import is blocked.
• Replace asyncio.get_event_loop().run_until_complete() in tests — it's deprecated since Python 3.10 and noisy on 3.12+. Use asyncio.run() or pytest-asyncio with @pytest.mark.asyncio.

[hafezparast]

Docker verification ✅

Tested against a locally built image from this branch with CRAWL4AI_HOOKS_ENABLED=true:

Test	Result
Simplest hook: async def hook(page, context, **kwargs): return page	✅ success: True
Hook using asyncio.sleep()	✅ success: True
Hook calling page.set_extra_http_headers()	✅ success: True

Docker logs show Hooks attachment status: success for all three — no __import__ not found errors. The module injection approach works cleanly.

[ntohidi]

@hafezparast Thanks for the Docker verification — good to know it works end-to-end. However, the review items I listed still need to be addressed before we can merge:

1. Remove BaseException from allowed_builtins — this is critical (DoS vector via except BaseException: pass defeating asyncio.wait_for timeout)
2. Inject the typing module itself into the namespace (namespace['typing'] = typing) so users can access Union, Any, Tuple, etc.
3. Replace asyncio.get_event_loop().run_until_complete() in tests with asyncio.run() or pytest-asyncio

Could you push those changes? Thanks!

[hafezparast]

All three review items addressed in the latest commit:

1. BaseException removed — left a comment explaining why (defeats asyncio.wait_for timeout via CancelledError)
2. typing module injected — plus Union, Any, Tuple, Set alongside the existing Dict, List, Optional
3. get_event_loop() replaced — asyncio.get_running_loop().time() in hook_manager.py (3 calls), asyncio.run() in test_hook_manager_unit.py (20 calls)

[ntohidi]

@hafezparast Thanks for the work on this! The core bug has already been fixed on develop as part of the security remediation commit (e326da9) — exec("import ...") was replaced with direct module injection, which is the same approach your PR takes.

Since the fix is already in and this PR has merge conflicts with the current develop, we're closing it. If you'd like to contribute the test file (test_hook_manager_unit.py) separately, feel free to open a new PR rebased on current develop. The test coverage would be welcome!