Something went wrong. Try again.
MCP based ACP agent
Something went wrong. Try again.
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576"""The execute kernel's timeout path — regression for EXECUTE_TODO B2/B3/B4/B5.
A cell that outlives its timeout returns an honest "(kernel busy: …)" noticeinstead of a bare, traceback-less ``Error:`` (an unhandled ``queue.Empty``whose ``str()`` is empty), the cell is NOT killed, and — the subtle one — theNEXT cell re-syncs via ``parent_header.msg_id`` filtering instead ofinheriting the stale cell's output. Before the fix, a timed-out cell left itsiopub messages (including its final ``status:idle``) queued, so the nextexecute drained those and broke on the STALE idle, returning the PREVIOUScell's output and desyncing attribution by one until a kernel reset.
Plus the MCP tool's ``timeout`` argument (B2): the caller can raise theceiling for a long-running process, and it actually reaches the kernel."""
import sys
import pytestfrom fastmcp import Client
from crow_cli.mcp.execute.kernel import CrowKernel
@pytest.fixturedef kernel(): k = CrowKernel(python_path=sys.executable) try: yield k finally: k.shutdown()
def test_long_cell_returns_busy_not_bare_error(kernel): """B3/B5: a cell that outlives the timeout yields the honest busy notice, not an empty ``Error:``/exception — and the kernel survives it (still usable on the next call, not wedged, not killed).""" out = kernel.execute("import time; time.sleep(3); print('slept')", timeout=1) assert "kernel busy" in out assert kernel.execute("print('alive')", timeout=15).strip() == "alive"
def test_next_cell_resyncs_after_a_timeout(kernel): """B4: the cell after a timeout returns ITS OWN output, not the stale output of the cell that was still running. msg_id filtering re-syncs the stream with no kernel reset — the old code returned 'LATE' here.""" busy = kernel.execute("import time; time.sleep(3); print('LATE')", timeout=1) assert "kernel busy" in busy # Waits for the kernel to free up, then returns its OWN output; the stale # 'LATE' (parent msg_id of the abandoned cell) must NOT be misattributed. out = kernel.execute("print('SECOND')", timeout=15) assert "SECOND" in out assert "LATE" not in out
@pytest.mark.asyncioasync def test_execute_tool_forwards_timeout(): """B2: the MCP execute tool exposes ``timeout`` and forwards it to the kernel — a cell that outlives the PASSED timeout comes back busy, which the hardcoded 30s default could never have produced.""" from crow_cli.mcp.server.app import mcp import crow_cli.mcp.execute.main # noqa: F401 — registers the tool from crow_cli.mcp.execute.main import shutdown_all
try: async with Client(mcp) as client: result = await client.call_tool( "execute", {"code": "import time; time.sleep(3); print('done')", "timeout": 1}, meta={"session_id": "timeout-test"}, ) text = result.content[0].text assert "kernel busy" in text assert "done" not in text finally: shutdown_all()