Verification: SDK defects KI-001 to KI-007 and KI-010 fixed (python-sdk client-sdk)

Date: 2026-10-12
Change: fixes of the defects pinned by the real-cloud E2E suite (e2e/KNOWN-ISSUES.md); CHANGELOG Unreleased > Fixed
Result: PASS for the eight fixed defects; one throughput-bound test is environment-flaky (KI-008, see below)
Package under test: agent-sandbox (client-sdk) installed editable from the working tree
Environment: real Alibaba Cloud FC agent sandbox platform (cn-hangzhou, envd 0.5.2), Python 3.11, Linux. Credentials from the gitignored .env (values not recorded)

Protocol evidence (captured from a real sandbox)

ListDir entry: {"name","path","type":2,"size":4096,"mode","permissions","owner","group","modifiedTime"} (type 1 = file, 2 = directory, number); Stat -> {"entry": {...}}; Stat of a missing path -> HTTP 404 with an empty body; ListDir of a missing path -> HTTP 200 {"entries": []}.
Process/List -> {"processes":[{"config":{"cmd","args","cwd","envs"},"pid":1,"tag":""}]}: envd pids are its own counters (1, 2, 3, ...), not OS pids; the config key is envs (confirms KI-001).
SendInput with a nested selector and an unknown pid -> HTTP 404 process not found; SendSignal {"process":{"pid":N},"signal":"SIGNAL_SIGKILL"} on a running sleep ends the stream with an exit event and empties List.

Commands run

cd client-sdk && python -m pytest tests -m "not integration"      # 2222 passed, 1 skipped, 1 failed (see below)
cd client-sdk && python -m pytest tests/test_known_issues          # 81 new regression tests, all pass
cd client-sdk && ruff check src tests && mypy src/agent_sandbox/    # clean
cd e2e && python -m pytest -m "e2e and not e2e_docker and not e2e_slow" -ra     # fast path (make e2e)
python scripts/e2e-cleanup.py --all-tagged                         # dry run: 0 leftovers

Results

Run	Selection	Result
unit	`client-sdk` `-m "not integration"`	2222 passed, 1 skipped, 1 failed: `test_agent/test_mcp_stdio.py::...ping_is_answered_while_a_tool_call_is_in_flight` fails identically before any of these changes (baseline: 2141 passed, same single failure)
unit	`client-sdk/tests/test_known_issues/` (new)	81 passed (KI-001 ... KI-008, KI-010)
real cloud	`test_03_commands.py` (17 tests)	17 passed, including the formerly xfailed `per_call_env`, `command_timeout_raises_command_timeout_error`, `kill_and_signal_terminate_a_process[kill/signal_term/signal_kill]`, `send_input_reaches_stdin`, plus the new `timed_out_command_is_not_left_running`
real cloud	`test_01_platform.py`	all passed, including `wrong_api_key_raises_authentication_error` and `connect_to_missing_sandbox_raises_sdk_error` (now `SandboxNotFoundError`)
real cloud	`test_04_files_envd.py`	`list_and_stat_report_directory_type`, `read_missing_file_raises_sdk_file_error` (`FileNotFoundError_`), `list_missing_directory_raises` (`FileNotFoundError_`) pass; see the throughput note
real cloud	`test_07_declarative.py`	`async_decorated_function` passes (`42`)
real cloud	fast path `make e2e` selection	79 passed, 1 xfailed (KI-009, platform), 1 failed (`test_medium_file_roundtrip_matches_remote_checksum`, see below)

Throughput-bound test (not a regression)

test_medium_file_roundtrip_matches_remote_checksum uploads 400 KB. From this host the upload rate measured 20-30 s for 400 KB in the later runs (it passed in the first run after the fixes), i.e. right at the 30 s default HTTP timeout (KI-008). It now fails with FileOperationError: Upload to <path> (400000 bytes) timed out; the transfer took longer than the HTTP timeout plus the ABX_HTTP_TIMEOUT hint instead of a bare httpx.ReadTimeout. The upload code path is otherwise unchanged.

Test-suite fixes made on the way

e2e/test_03_commands.py: the helpers started processes and picked procs[-1] of a list shared with earlier tests of the module, and force-killed with kill -9 <pid> although envd pids are not OS pids; they now pick the process whose pid was not listed before and force-kill through commands.send_signal(pid, 9). This is what made the first post-fix run flaky for the process-control tests.

Cleanup

Every sandbox was tracked and deleted by the session finalizer (leaked_after_cleanup=0 in every run); python scripts/e2e-cleanup.py --all-tagged reports no tagged sandbox. Probe sandboxes created while investigating were killed in finally blocks.

Not fixed / out of scope

KI-008 (gateway throughput), KI-009 (Jupyter service 404), KI-011 (ACR push bandwidth): platform/environment, unchanged by design; KI-008 is only reported more clearly.
commands.run(timeout=...) is an inactivity limit (no output for timeout + 5 s), not a wall-clock deadline: a chatty command is never timed out. Pre-existing, outside KI-003.
ProcessInfo.command stays empty (envd nests it under config.cmd); not part of the recorded issues.