From 73c104ee35656a818898c1cb26cfc49afbdf4294 Mon Sep 17 00:00:00 2001 From: Eva Date: Wed, 9 Sep 2026 16:47:59 +0700 Subject: [PATCH] fix(mcp): preserve uncertainty after retry exit (cherry picked from commit 775dd2c4d7eb28cd82cb8d943c8ea11640474227) --- .../test_mcp_stdio_fastfail_reconnect.py | 47 +++++++++++++++++++ tools/mcp_tool_handlers.py | 6 +++ 2 files changed, 53 insertions(+) diff --git a/tests/tools/test_mcp_stdio_fastfail_reconnect.py b/tests/tools/test_mcp_stdio_fastfail_reconnect.py index dca94a58e6..930a3f389e 100644 --- a/tests/tools/test_mcp_stdio_fastfail_reconnect.py +++ b/tests/tools/test_mcp_stdio_fastfail_reconnect.py @@ -182,6 +182,53 @@ def test_midcall_child_exit_reconnects_without_replay(monkeypatch, tmp_path): _cleanup(mcp_tool, "srv-midcall") +def test_precall_respawn_retry_dying_midcall_is_uncertain_without_replay(monkeypatch, tmp_path): + """A safe pre-call retry becomes uncertain if its replacement dies after dispatch.""" + monkeypatch.setenv("HERMES_HOME", str(tmp_path)) + from tools import mcp_tool + from tools.mcp_tool_handlers import _make_tool_handler + + alive = {"v": False} + effects = {"n": 0} + + async def _never_dispatched(*a, **kw): + raise AssertionError("the original dead child must not receive the call") + + async def _effect_then_die(*a, **kw): + effects["n"] += 1 + alive["v"] = False + await asyncio.sleep(30) + + async def _watch_children(): + while alive["v"]: + await asyncio.sleep(0.05) + + def _respawn(server): + alive["v"] = True + new_session = MagicMock() + new_session.call_tool = _effect_then_die + server.session = new_session + server._ready.set() + + server = _install_stub_server( + mcp_tool, "srv-retry-midcall", _never_dispatched, + children_dead=lambda: not alive["v"], + on_reconnect=_respawn, + ) + server._watch_stdio_children = _watch_children + _mcp_loop._ensure_mcp_loop() + try: + handler = _make_tool_handler("srv-retry-midcall", "tool1", 10.0) + parsed = json.loads(handler({})) + assert parsed["outcome_uncertain"] is True, parsed + assert "may have completed" in parsed["error"], parsed + assert "did not replay" in parsed["error"], parsed + assert server._reconnect_event.set_calls == 1 + assert effects["n"] == 1, "the uncertain retry must not be invoked a third time" + finally: + _cleanup(mcp_tool, "srv-retry-midcall") + + def test_dead_child_never_returning_is_not_reported_as_a_timeout( monkeypatch, tmp_path, ): diff --git a/tools/mcp_tool_handlers.py b/tools/mcp_tool_handlers.py index 4d101b15eb..600bcd7c30 100644 --- a/tools/mcp_tool_handlers.py +++ b/tools/mcp_tool_handlers.py @@ -236,6 +236,12 @@ def _handle_stdio_child_exited_and_retry(server_name: str, exc: Exception, retry # Died again right after respawn: broken server; run()'s budget takes it to the park. logger.warning("MCP server '%s': %s stdio subprocess exited again right after respawn (%s); not retrying " "further.", server_name, op_description, retry_exc) + if retry_exc.in_flight: + return _strike( + server_name, + _STDIO_OUTCOME_UNCERTAIN_MSG.format(s=server_name), + outcome_uncertain=True, + ) return _strike(server_name, _STDIO_DIED_AGAIN_MSG.format(s=server_name)) except Exception as retry_exc: logger.warning("MCP %s/%s retry after stdio respawn failed: %s", server_name, op_description, retry_exc)