fix(tools): route bound task and batch tools' sync path through the bound runtime - #5861

Why

get_available_tools wraps the process-wide tool singletons in place with a sync func (backend/packages/harness/deerflow/tools/tools.py:72, inside _ensure_sync_invocable_tool, reached from lines 172 and 305) so synchronous agent callers can invoke them. bind_task_tool (backend/packages/harness/deerflow/tools/builtins/task_tool.py:457 on main) and _bind_batch_tool (backend/packages/harness/deerflow/tools/builtins/batch_task_tool.py:84 on main) rebind only the coroutine of their copies, so each copy inherits the singleton's func — a sync wrapper around the unbound coroutine. A synchronous invocation of a bound copy therefore bypassed the explicit SDK submitter / execution capacity and fell through to the process-global one — the exact fallback bind_batch_tools' docstring rules out ("never fall through to another application's process-global submitter") — or raised NotImplementedError in a fresh process where the singleton had never been wrapped.

What changed

• bind_task_tool and _bind_batch_tool now also rebind the copy's func to make_sync_tool_wrapper(bound_coroutine, ...) — the same helper _ensure_sync_invocable_tool uses — so both invocation paths of a bound copy go through the bound coroutine and its runtime's own submitter/capacity.
• Bound copies now always expose a usable sync path, even when created before any sync wrap has happened.

Surface area

• Agents / LangGraph — tool binding under backend/packages/harness/deerflow/tools/ (no prompt changes, no new data sources in text)
• Frontend UI
• Backend API
• Sandbox
• Skills
• Dependencies
• Default behavior change
• Docs / tests / CI only

Bug fix verification

• Test path that reproduces the bug:
  • backend/tests/test_batch_task_tool.py::test_bound_batch_tools_sync_path_uses_the_explicit_submitter
  • backend/tests/test_batch_task_tool.py::test_bound_batch_tools_expose_a_sync_invocation_path
  • backend/tests/test_task_tool_core_logic.py::test_bound_task_tool_sync_path_uses_the_explicit_capacity
• Did it go red on main and green on this branch? yes — with the two source changes reverted (tests kept), all three fail: the batch sync path returns a fallback-submitter result instead of the explicit submitter's, bound copies have func is None, and the task sync path raises KeyError: 'execution_capacity'. All three pass on this branch.

Validation

All commands run in backend/:

• make lint → All checks passed! / 1633 files already formatted
• New regression tests → 3 passed
PYTHONPATH=. uv run pytest tests/test_batch_task_tool.py::test_bound_batch_tools_sync_path_uses_the_explicit_submitter tests/test_batch_task_tool.py::test_bound_batch_tools_expose_a_sync_invocation_path tests/test_task_tool_core_logic.py::test_bound_task_tool_sync_path_uses_the_explicit_capacity
• Module tests → 103 passed
PYTHONPATH=. uv run pytest tests/test_batch_task_tool.py tests/test_task_tool_core_logic.py
• Related suites → 32 passed (tests/test_batch_acceptance.py), 14 passed (tests/test_mcp_sync_wrapper.py tests/test_batch_extension_snapshot.py)
• Full default suite (PYTHONPATH=. uv run pytest -m "not live" --ignore=tests/blocking_io tests/ -q): 19474 passed, 17 failed on this branch vs 19471 passed, 20 failed with the fix reverted — the 3 extra failures without the fix are the new regression tests; the 17 remaining failures are pre-existing on main in this environment (browser automation / web-fetch / extension-install tests, untouched by this change).
• Also verified that LangGraph's injected-arg detection (_get_all_injected_args) still resolves runtime and tool_call_id on the bound copies, since the new wrapper's __wrapped__ chain points at the bound coroutine: runtime and tool_call_id are detected for batch_task/task, runtime for batch_status/cancel_batch, matching the singletons.

AI assistance

Tool(s) used: Claude Code (autonomous coding agent, submitted from the BlueX888 account)

How you used it: The agent diagnosed the root cause, wrote the fix and the regression tests, and ran every validation command listed above; this PR text was also drafted by the agent.

• I've read and understand every line of this change and take responsibility for it — it's not unreviewed AI output.