fix(acp): add bounded wait + late-refresh for configured MCP servers

ACP entry.py fires MCP discovery in a background daemon thread, but
_make_agent snapshots tools once at build and never re-reads the registry.
Unlike CLI/TUI, ACP had no bounded wait before the snapshot and no
late-refresh for configured (config.yaml) MCP servers — a reachable-but-
slow server that finished after agent build was invisible for the whole
session.

Changes:
- acp_adapter/session.py (_make_agent): call wait_for_mcp_discovery()
  before AIAgent construction, bounded by mcp_discovery_timeout (default
  ~1.5s). A dead server can't block; servers that miss the bound are
  picked up by the late-refresh below.
- acp_adapter/server.py (_schedule_mcp_late_refresh): new method on
  HermesACPAgent — if discovery is still in flight after session creation,
  spawns an off-critical-path daemon that joins it (bounded 30s), then
  rebuilds the tool snapshot via the shared refresh_agent_mcp_tools helper.
  Cache-safe: only runs pre-first-turn (_user_turn_count/_api_call_count
  both 0); once the user has sent a message the snapshot is frozen, exactly
  as TUI PR #48403 does.
- Called from new_session, load_session, resume_session.
- Mirrors the TUI pattern (tui_gateway _schedule_mcp_late_refresh, PR
  #48403) and the CLI pattern (get_tool_definitions → wait_for_mcp_discovery).

Tests:
- Replace the AST-based test (source-text inspection) with three
  behavioral regression tests in tests/acp_adapter/test_acp_mcp_discovery.py:
  1. Blocked discovery does not block startup (non-blocking contract)
  2. Delayed discovery lands tools via late-refresh (pre-first-turn)
  3. Late-refresh is cache-safe: skips rebuild after first turn

Addresses teknium1 review on PR #32811.
This commit is contained in:
Sergey Prontsevich 2026-07-13 21:19:30 +03:00 committed by kshitij
parent 89f0b63da4
commit 640de6562e
3 changed files with 362 additions and 0 deletions

View File

@ -1037,6 +1037,84 @@ class HermesACPAgent(acp.Agent):
exc_info=True,
)
def _schedule_mcp_late_refresh(self, state: SessionState) -> None:
"""Refresh the agent's tool snapshot when background MCP discovery lands late.
ACP entry.py starts MCP tool discovery in a background daemon thread so a
slow/dead configured server can't block ``asyncio.run()``. ``_make_agent``
briefly joins that thread (``wait_for_mcp_discovery``, bounded ~1.5s) so
already-spawning fast servers land in the snapshot but a server slower
than the bound lands *after* the agent is built, leaving its tools absent
for the whole session.
This schedules an off-critical-path daemon that waits for discovery to
finish (bounded 30s), then rebuilds the snapshot via the shared
``refresh_agent_mcp_tools`` helper the same rebuild ``/reload-mcp``
performs, but automatic. Mirrors the TUI late-refresh (PR #48403).
Cache safety: the rebuild only runs while the session is still
pre-first-turn (no API call made yet nothing cached to invalidate).
Once the user has sent a message we leave the snapshot frozen rather
than break the cached prompt prefix mid-conversation; late tools then
require an explicit ``/reload-mcp`` (user-consented), exactly as today.
No-op when discovery already finished, when the join times out, when the
registry was unchanged, or when the session was closed while waiting.
"""
try:
from hermes_cli.mcp_startup import mcp_discovery_in_flight
except Exception:
return
if not mcp_discovery_in_flight():
return
import threading
agent = state.agent
session_id = state.session_id
def _wait_then_refresh() -> None:
try:
from hermes_cli.mcp_startup import join_mcp_discovery
if not join_mcp_discovery(timeout=30.0):
return
# Session may have been closed while we waited.
current = self.session_manager.get_session(session_id)
if current is None or current.agent is not agent:
return
# Cache safety: never rebuild the tool list once the conversation
# has started — that would invalidate the cached prompt prefix.
if (
int(getattr(agent, "_user_turn_count", 0) or 0) > 0
or int(getattr(agent, "_api_call_count", 0) or 0) > 0
):
return
from tools.mcp_tool import refresh_agent_mcp_tools
added = refresh_agent_mcp_tools(agent)
if added:
logger.info(
"Session %s: late MCP refresh added %d tools: %s",
session_id,
len(added),
", ".join(sorted(added)),
)
except Exception:
logger.debug(
"Session %s: late MCP refresh failed",
session_id,
exc_info=True,
)
threading.Thread(
target=_wait_then_refresh,
name=f"acp-mcp-late-refresh-{session_id}",
daemon=True,
).start()
# ---- ACP lifecycle ------------------------------------------------------
async def initialize(
@ -1343,6 +1421,7 @@ class HermesACPAgent(acp.Agent):
) -> NewSessionResponse:
state = self.session_manager.create_session(cwd=cwd)
await self._register_session_mcp_servers(state, mcp_servers)
self._schedule_mcp_late_refresh(state)
logger.info("New session %s (cwd=%s)", state.session_id, cwd)
self._schedule_available_commands_update(state.session_id)
self._schedule_usage_update(state)
@ -1367,6 +1446,7 @@ class HermesACPAgent(acp.Agent):
logger.warning("load_session: session %s not found", session_id)
return None
await self._register_session_mcp_servers(state, mcp_servers)
self._schedule_mcp_late_refresh(state)
logger.info("Loaded session %s", session_id)
# Per ACP spec, `session/load` must stream the prior conversation back
# to the client via `session/update` notifications BEFORE responding,
@ -1414,6 +1494,7 @@ class HermesACPAgent(acp.Agent):
logger.warning("resume_session: session %s not found, creating new", session_id)
state = self.session_manager.create_session(cwd=cwd)
await self._register_session_mcp_servers(state, mcp_servers)
self._schedule_mcp_late_refresh(state)
logger.info("Resumed session %s", state.session_id)
# See `load_session` above for the spec rationale — replay must
# complete before the response so clients receive the full transcript

View File

@ -648,6 +648,23 @@ class SessionManager:
logger.debug("ACP session falling back to default provider resolution", exc_info=True)
_register_task_cwd(session_id, cwd)
# Bounded wait for background MCP discovery so already-spawning fast
# servers land in the agent's tool snapshot. ACP entry.py fires
# discovery in a background daemon thread (start_background_mcp_discovery);
# the agent snapshots tools once at build (run_agent/agent_init) and
# never re-reads the registry, so without this join a reachable-but-
# slow configured server would be invisible for the whole session.
# Bounded by ``mcp_discovery_timeout`` (config.yaml, default ~1.5s) so a
# dead server can't block — servers that miss the bound are picked up
# by the automatic late-refresh (see HermesACPAgent._schedule_mcp_late_refresh).
try:
from hermes_cli.mcp_startup import wait_for_mcp_discovery
wait_for_mcp_discovery()
except Exception:
logger.debug("ACP: bounded MCP discovery wait failed", exc_info=True)
agent = AIAgent(**kwargs)
# Codex app-server sessions are spawned lazily on the first turn. Stamp
# the ACP workspace onto the agent so the Codex runtime starts from the

View File

@ -0,0 +1,264 @@
"""Behavioral regression tests for ACP background MCP discovery + late-refresh.
These replace the previous AST-based test that only inspected source text.
They verify the *behavior*: (1) a blocked discovery doesn't block startup, and
(2) a delayed-but-reachable MCP server's tools land in the agent's snapshot
via the automatic late-refresh, cache-safely (pre-first-turn only).
"""
from __future__ import annotations
import sys
import threading
import time
import types
from contextlib import nullcontext
from types import ModuleType, SimpleNamespace
import pytest
from acp_adapter.server import HermesACPAgent
from acp_adapter.session import SessionManager, SessionState
from hermes_cli import mcp_startup
# ---------------------------------------------------------------------------
# Shared helpers
# ---------------------------------------------------------------------------
class FakeAgent:
"""Minimal stand-in for AIAgent with the attributes late-refresh touches."""
def __init__(self):
self.model = "fake-model"
self.provider = "fake-provider"
self.enabled_toolsets = ["hermes-acp"]
self.disabled_toolsets = []
self.tools = []
self.valid_tool_names = set()
self._user_turn_count = 0
self._api_call_count = 0
class NoopDb:
def get_session(self, *_a, **_k):
return None
def create_session(self, *_a, **_k):
return None
def update_session(self, *_a, **_k):
return None
def _mod(name: str, **attrs) -> ModuleType:
module = ModuleType(name)
for key, value in attrs.items():
setattr(module, key, value)
return module
@pytest.fixture(autouse=True)
def _reset_mcp_startup_state():
"""Ensure each test starts with a clean discovery thread state."""
saved_started = mcp_startup._mcp_discovery_started
saved_thread = mcp_startup._mcp_discovery_thread
mcp_startup._mcp_discovery_started = False
mcp_startup._mcp_discovery_thread = None
yield
thread = mcp_startup._mcp_discovery_thread
if thread is not None and thread.is_alive():
thread.join(timeout=2.0)
mcp_startup._mcp_discovery_started = saved_started
mcp_startup._mcp_discovery_thread = saved_thread
# ---------------------------------------------------------------------------
# Test 1 — blocked discovery does not block startup
# ---------------------------------------------------------------------------
def test_acp_background_discovery_does_not_block_startup(monkeypatch):
"""start_background_mcp_discovery must return immediately even if discovery hangs."""
block = threading.Event()
def _blocking_discover():
block.wait(timeout=5.0)
monkeypatch.setitem(
sys.modules,
"hermes_cli.config",
_mod(
"hermes_cli.config",
read_raw_config=lambda: {"mcp_servers": {"slow": {"url": "https://mcp.example.test"}}},
),
)
monkeypatch.setitem(
sys.modules,
"tools.mcp_oauth",
_mod("tools.mcp_oauth", suppress_interactive_oauth=lambda: nullcontext()),
)
monkeypatch.setitem(
sys.modules,
"tools.mcp_tool",
_mod("tools.mcp_tool", discover_mcp_tools=_blocking_discover),
)
start = time.monotonic()
mcp_startup.start_background_mcp_discovery(
logger=SimpleNamespace(debug=lambda *_a, **_k: None),
thread_name="test-acp-discovery",
)
elapsed = time.monotonic() - start
assert elapsed < 0.2, "start_background_mcp_discovery blocked for {:.3f}s".format(elapsed)
assert mcp_startup._mcp_discovery_thread is not None
assert mcp_startup._mcp_discovery_thread.is_alive()
block.set()
mcp_startup._mcp_discovery_thread.join(timeout=2.0)
# ---------------------------------------------------------------------------
# Test 2 — delayed discovery lands tools via late-refresh (pre-first-turn)
# ---------------------------------------------------------------------------
def test_acp_late_refresh_adds_tools_when_discovery_lands_after_build(monkeypatch):
"""A slow MCP server that finishes after agent build must still appear in tools."""
discovery_block = threading.Event()
discovery_done = threading.Event()
def _slow_discover():
discovery_block.wait(timeout=5.0)
discovery_done.set()
monkeypatch.setitem(
sys.modules,
"hermes_cli.config",
_mod(
"hermes_cli.config",
read_raw_config=lambda: {"mcp_servers": {"slow": {"url": "https://mcp.example.test"}}},
),
)
monkeypatch.setitem(
sys.modules,
"tools.mcp_oauth",
_mod("tools.mcp_oauth", suppress_interactive_oauth=lambda: nullcontext()),
)
monkeypatch.setitem(
sys.modules,
"tools.mcp_tool",
_mod("tools.mcp_tool", discover_mcp_tools=_slow_discover),
)
mcp_startup.start_background_mcp_discovery(
logger=SimpleNamespace(debug=lambda *_a, **_k: None),
thread_name="test-acp-late",
)
# Build the session immediately — discovery is still in flight.
fake = FakeAgent()
manager = SessionManager(agent_factory=lambda **_k: fake, db=NoopDb())
acp_agent = HermesACPAgent(session_manager=manager)
state = manager.create_session(cwd=".")
# Discovery is blocked, so it must still be in flight.
assert not discovery_done.is_set(), "discovery finished too early for this test"
# Track refresh_agent_mcp_tools calls.
refreshed = []
def _fake_refresh(agent, **_kw):
agent.tools = [{"function": {"name": "mcp_slow_tool"}}]
agent.valid_tool_names = {"mcp_slow_tool"}
refreshed.append(agent)
return {"mcp_slow_tool"}
monkeypatch.setitem(
sys.modules,
"tools.mcp_tool",
_mod("tools.mcp_tool", refresh_agent_mcp_tools=_fake_refresh),
)
# Trigger late-refresh.
acp_agent._schedule_mcp_late_refresh(state)
# Release discovery so the late-refresh daemon can proceed.
discovery_block.set()
# Wait for the late-refresh daemon to finish.
deadline = time.monotonic() + 5.0
while not refreshed and time.monotonic() < deadline:
time.sleep(0.01)
assert refreshed, "late-refresh daemon did not call refresh_agent_mcp_tools"
assert refreshed[0] is fake
assert "mcp_slow_tool" in fake.valid_tool_names
# ---------------------------------------------------------------------------
# Test 3 — late-refresh is cache-safe: skips after first turn
# ---------------------------------------------------------------------------
def test_acp_late_refresh_skips_after_first_turn(monkeypatch):
"""Once the user has sent a message, late-refresh must NOT rebuild tools."""
discovery_block = threading.Event()
def _slow_discover():
discovery_block.wait(timeout=5.0)
monkeypatch.setitem(
sys.modules,
"hermes_cli.config",
_mod(
"hermes_cli.config",
read_raw_config=lambda: {"mcp_servers": {"slow": {"url": "https://mcp.example.test"}}},
),
)
monkeypatch.setitem(
sys.modules,
"tools.mcp_oauth",
_mod("tools.mcp_oauth", suppress_interactive_oauth=lambda: nullcontext()),
)
monkeypatch.setitem(
sys.modules,
"tools.mcp_tool",
_mod("tools.mcp_tool", discover_mcp_tools=_slow_discover),
)
mcp_startup.start_background_mcp_discovery(
logger=SimpleNamespace(debug=lambda *_a, **_k: None),
thread_name="test-acp-cache",
)
fake = FakeAgent()
fake._api_call_count = 1 # simulate: user already sent a message
manager = SessionManager(agent_factory=lambda **_k: fake, db=NoopDb())
acp_agent = HermesACPAgent(session_manager=manager)
state = manager.create_session(cwd=".")
refreshed = []
def _fake_refresh(agent, **_kw):
refreshed.append(agent)
return set()
monkeypatch.setitem(
sys.modules,
"tools.mcp_tool",
_mod("tools.mcp_tool", refresh_agent_mcp_tools=_fake_refresh),
)
acp_agent._schedule_mcp_late_refresh(state)
# Release discovery so the daemon can proceed (if it were going to).
discovery_block.set()
# Give the daemon time to run (if it were going to).
time.sleep(0.5)
assert not refreshed, "late-refresh rebuilt tools after the first turn — cache broken!"