fix(optimal): serialize singleton init, normalize kb_search index_types, surface mentor errors

The OptimalService singleton published the instance before initialize()
finished, so a concurrent caller could observe _initialized=False and hit
"OptimalService not initialized" during RAG indexing. Build the instance,
initialize it, then publish under a lazily-bound asyncio lock so all callers
share a fully-initialized singleton.

roboco_kb_search forwarded the legacy alias index_types=['docs'], which is
not a valid IndexType value (the enum value is 'documentation'), producing a
400 at the route. Normalize the alias in the client before the request is
sent and fix the misleading tool docstring.

The mentor route let exceptions from mentor.ask escape as a bare 500 that
masked the real cause. Catch, log the true upstream error with stack, and
surface it in the response detail so failures are diagnosable.
This commit is contained in:
Renn F
2026-06-03 18:54:07 +02:00
parent 2c96bd09f6
commit b18bdcd41a
5 changed files with 304 additions and 26 deletions
+25 -6
View File
@@ -9,6 +9,7 @@ from datetime import UTC, datetime
from typing import Any
from uuid import uuid4
import structlog
from fastapi import APIRouter, HTTPException, status
from roboco.api.deps import (
@@ -79,6 +80,8 @@ from roboco.services.optimal_brain import (
get_validator_service,
)
logger = structlog.get_logger()
router = APIRouter()
@@ -768,12 +771,28 @@ async def mentor_ask(
detail=f"Mentor service initialization failed: {e}",
) from e
response = await mentor.ask(
question=request.question,
agent_id=str(agent.agent_id),
conversation_id=request.conversation_id,
domain=request.domain,
)
try:
response = await mentor.ask(
question=request.question,
agent_id=str(agent.agent_id),
conversation_id=request.conversation_id,
domain=request.domain,
)
except Exception as e:
# Do NOT let this escape as a bare 500 that masks the real cause.
# Log the true upstream error (with stack) and surface it in the
# detail so the failure is diagnosable end-to-end.
logger.error(
"Mentor ask failed",
agent_id=str(agent.agent_id),
error=str(e),
error_type=type(e).__name__,
exc_info=True,
)
raise HTTPException(
status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
detail=f"Mentor ask failed: {e}",
) from e
return MentorAskResponse(
answer=response.answer,
+24 -3
View File
@@ -44,6 +44,25 @@ class RecordDecisionInput(BaseModel):
tags: list[str] | None = Field(None, description="Tags for categorization")
# Legacy aliases agents commonly pass that are not valid IndexType values.
# The only valid value for documentation is 'documentation' — 'docs' raises
# ValueError at the route's IndexType(...) conversion, yielding a 400.
_INDEX_TYPE_ALIASES = {"docs": "documentation"}
def normalize_index_types(index_types: list[str] | None) -> list[str] | None:
"""Map legacy index-type aliases to valid IndexType values.
Agents (and our own docstrings) historically used 'docs', which is not a
member of the ``IndexType`` enum. Translate it to 'documentation' before
the request leaves the client so the route's ``IndexType(...)`` conversion
succeeds instead of 400-ing.
"""
if index_types is None:
return None
return [_INDEX_TYPE_ALIASES.get(t, t) for t in index_types]
def _register_search_tools(mcp: FastMCP, client: ApiClient) -> None:
"""Register search tools available to all agents."""
@@ -66,7 +85,8 @@ def _register_search_tools(mcp: FastMCP, client: ApiClient) -> None:
top_k: Number of results to return (1-20, default 5)
project: Optional project filter
task_id: Optional task filter
index_types: Index types to search (code, docs, decisions, learnings)
index_types: Index types to search (documentation, decisions,
learnings, standards, errors, reviews, journals, conversations)
Returns:
Search results with relevance scores and source info
@@ -79,8 +99,9 @@ def _register_search_tools(mcp: FastMCP, client: ApiClient) -> None:
payload["project"] = project
if task_id:
payload["task_id"] = task_id
if index_types:
payload["index_types"] = index_types
normalized_index_types = normalize_index_types(index_types)
if normalized_index_types:
payload["index_types"] = normalized_index_types
resp = await client.post("/optimal/kb/search", json=payload)
if not resp.ok:
+33 -17
View File
@@ -11,6 +11,7 @@ The service uses a plugin-based architecture where each index type is handled
by a specialized plugin that implements the BaseIndexPlugin interface.
"""
import asyncio
import re
from dataclasses import dataclass, field
from pathlib import Path
@@ -189,8 +190,6 @@ class OptimalService:
initialized_count = 0
failed_plugins: list[tuple[IndexType, str]] = []
import asyncio
# Per-plugin initialization timeout (embedding validation can be slow)
plugin_init_timeout = 30.0
@@ -439,8 +438,6 @@ class OptimalService:
async def _start_periodic_update(self) -> None:
"""Start periodic update task if enabled in config."""
import asyncio
from roboco.config import get_settings
settings = get_settings()
@@ -459,8 +456,6 @@ class OptimalService:
async def _periodic_update_loop(self, interval: int) -> None:
"""Background loop that checks for file changes periodically."""
import asyncio
while True:
try:
await asyncio.sleep(interval)
@@ -560,7 +555,6 @@ class OptimalService:
async def close(self) -> None:
"""Cleanup resources."""
import asyncio
import contextlib
# Cancel periodic update task
@@ -1249,8 +1243,6 @@ class OptimalService:
(e.g., due to timeouts or errors). Includes retry logic for transient
failures.
"""
import asyncio
import httpx
from roboco.config import settings
@@ -1734,8 +1726,6 @@ class OptimalService:
self, details: dict[str, Any], timeout: float
) -> bool:
"""Test embedding model connectivity; record result in `details`."""
import asyncio
from roboco.config import settings
from roboco.services.optimal_brain.shared_embedder import get_shared_embedder
@@ -1791,8 +1781,6 @@ class OptimalService:
self, details: dict[str, Any], timeout: float
) -> bool:
"""Test vector store connectivity + per-index search; record details."""
import asyncio
try:
async with asyncio.timeout(timeout):
stats = await self.get_stats()
@@ -1837,13 +1825,39 @@ class _OptimalServiceHolder:
"""Holder for singleton OptimalService instance."""
instance: OptimalService | None = None
lock: asyncio.Lock | None = None
def _get_init_lock() -> asyncio.Lock:
"""Return the init lock, lazily bound to the running event loop.
The lock is created on first use (not at import time) so it binds to the
loop that is actually running; a lock created at import time would bind to
the wrong loop and raise "bound to a different event loop".
"""
if _OptimalServiceHolder.lock is None:
_OptimalServiceHolder.lock = asyncio.Lock()
return _OptimalServiceHolder.lock
async def get_optimal_service() -> OptimalService:
"""Get or create the OptimalService instance."""
if _OptimalServiceHolder.instance is None:
_OptimalServiceHolder.instance = OptimalService()
await _OptimalServiceHolder.instance.initialize()
"""Get or create the OptimalService instance.
The instance is only published *after* ``initialize()`` completes. A lock
serializes concurrent first-callers so a second coroutine can never observe
a half-built, ``_initialized == False`` singleton mid-initialization (which
previously surfaced as "OptimalService not initialized" during indexing).
"""
if _OptimalServiceHolder.instance is not None:
return _OptimalServiceHolder.instance
async with _get_init_lock():
# Re-check under the lock: another coroutine may have built it while
# we waited to acquire.
if _OptimalServiceHolder.instance is None:
service = OptimalService()
await service.initialize()
_OptimalServiceHolder.instance = service
return _OptimalServiceHolder.instance
@@ -1852,3 +1866,5 @@ async def close_optimal_service() -> None:
if _OptimalServiceHolder.instance is not None:
await _OptimalServiceHolder.instance.close()
_OptimalServiceHolder.instance = None
# Drop the lock so the next initialization rebinds to the running loop.
_OptimalServiceHolder.lock = None
@@ -0,0 +1,85 @@
"""The mentor route must surface (log) the real upstream error, not mask it.
Previously an exception raised inside ``mentor.ask`` escaped the route as a
bare 500 whose body carried no diagnosable cause agents saw only a generic
mask. The route must now log the true error and return a clean envelope whose
``detail`` names the real cause.
"""
from __future__ import annotations
from http import HTTPStatus
from typing import TYPE_CHECKING
from unittest.mock import AsyncMock, patch
from uuid import uuid4
import pytest
import pytest_asyncio
from fastapi import FastAPI
from httpx import ASGITransport, AsyncClient
from roboco.api.deps import get_agent_context
from roboco.api.routes.optimal import router as optimal_router
from roboco.models import AgentRole
from roboco.models.permissions import AgentContext
if TYPE_CHECKING:
from collections.abc import AsyncIterator
_HDR = {"X-Agent-ID": str(uuid4()), "X-Agent-Role": "ceo"}
@pytest_asyncio.fixture
async def optimal_client() -> AsyncIterator[AsyncClient]:
app = FastAPI()
app.include_router(optimal_router, prefix="/api/optimal")
async def _override_agent() -> AgentContext:
return AgentContext(agent_id=uuid4(), role=AgentRole.CEO, team=None)
app.dependency_overrides[get_agent_context] = _override_agent
transport = ASGITransport(app=app)
async with AsyncClient(transport=transport, base_url="http://test") as client:
yield client
app.dependency_overrides.clear()
@pytest.mark.asyncio
async def test_mentor_ask_logs_and_surfaces_real_upstream_error(
optimal_client: AsyncClient,
) -> None:
mentor = AsyncMock()
mentor._optimal_service = object() # already initialized
mentor.ask = AsyncMock(side_effect=RuntimeError("ollama connection refused"))
with (
patch("roboco.api.routes.optimal.get_mentor_service") as mock_mentor,
patch("roboco.api.routes.optimal.get_optimal_service") as mock_get,
patch("roboco.api.routes.optimal.logger") as mock_logger,
):
mock_mentor.return_value = mentor
mock_get.return_value = AsyncMock()
response = await optimal_client.post(
"/api/optimal/mentor/ask",
json={"question": "What"},
headers=_HDR,
)
# The route must not 200 on a failed ask.
assert response.status_code in (
HTTPStatus.INTERNAL_SERVER_ERROR,
HTTPStatus.SERVICE_UNAVAILABLE,
)
# The real cause must be diagnosable in the response body, not masked.
assert "ollama connection refused" in response.text
# The real error must be logged (so it is diagnosable server-side even if
# the envelope is later sanitized).
logged_text = "".join(
str(call.args) + str(call.kwargs)
for call in (
*mock_logger.error.call_args_list,
*mock_logger.exception.call_args_list,
)
)
assert "ollama connection refused" in logged_text
@@ -0,0 +1,137 @@
"""Grounding-layer regression tests for the Optimal/RAG stack.
Covers three failure modes that surfaced at runtime:
1. ``get_optimal_service`` published a not-yet-initialized singleton while
``initialize()`` was still awaiting, so a concurrent caller hit
"OptimalService not initialized. Call initialize() first." during indexing.
2. ``roboco_kb_search`` forwarded the legacy alias ``index_types=['docs']``,
which is not a valid ``IndexType`` value ('documentation' is), producing a
400 at the route.
3. The mentor route let exceptions from ``mentor.ask`` escape as a bare 500
with no log of the true upstream cause.
"""
from __future__ import annotations
import asyncio
import pytest
from roboco.mcp.optimal_server import normalize_index_types
from roboco.models.optimal import IndexType
from roboco.services import optimal as optimal_module
from roboco.services.optimal import (
OptimalService,
close_optimal_service,
get_optimal_service,
)
# ---------------------------------------------------------------------------
# Sub-issue 2: kb_search must not forward the invalid 'docs' alias
# ---------------------------------------------------------------------------
def test_normalize_index_types_maps_docs_alias_to_documentation() -> None:
"""The legacy 'docs' alias must become the valid 'documentation' value.
``IndexType('docs')`` raises ``ValueError`` the only valid value is
``IndexType.DOCUMENTATION`` whose string value is 'documentation'.
"""
assert normalize_index_types(["docs"]) == ["documentation"]
# Every normalized value must be a constructible IndexType.
for value in normalize_index_types(["docs"]):
IndexType(value)
def test_normalize_index_types_passes_valid_values_through() -> None:
assert normalize_index_types(["documentation", "decisions"]) == [
"documentation",
"decisions",
]
def test_normalize_index_types_none_returns_none() -> None:
assert normalize_index_types(None) is None
# ---------------------------------------------------------------------------
# Sub-issue 1: the init entrypoint must never expose an uninitialized singleton
# ---------------------------------------------------------------------------
class _SlowInitService(OptimalService):
"""OptimalService whose initialize() yields control mid-flight.
This reproduces the publish-before-initialize race: while one coroutine
is awaiting inside ``initialize()``, a second coroutine calls
``get_optimal_service()``. With the old code the second caller received
the instance with ``_initialized == False``.
"""
init_calls = 0
async def initialize(self) -> None:
type(self).init_calls += 1
# Cooperatively yield so a concurrent get_optimal_service() can run
# during the window the old code left the instance unpublished/uninit.
await asyncio.sleep(0)
self._initialized = True
@pytest.mark.asyncio
async def test_get_optimal_service_never_returns_uninitialized(
monkeypatch: pytest.MonkeyPatch,
) -> None:
"""Concurrent callers must all receive a fully-initialized singleton.
The indexing entrypoint calls ``_get_plugin`` which raises
"OptimalService not initialized" when ``_initialized`` is False. This test
asserts no concurrent caller can observe that state.
"""
await close_optimal_service()
_SlowInitService.init_calls = 0
monkeypatch.setattr(optimal_module, "OptimalService", _SlowInitService)
try:
results: list[OptimalService] = await asyncio.gather(
get_optimal_service(),
get_optimal_service(),
get_optimal_service(),
)
for svc in results:
assert svc._initialized is True
# All callers share the one singleton, initialized exactly once.
assert len({id(s) for s in results}) == 1
assert _SlowInitService.init_calls == 1
finally:
await close_optimal_service()
@pytest.mark.asyncio
async def test_indexing_entrypoint_does_not_raise_not_initialized(
monkeypatch: pytest.MonkeyPatch,
) -> None:
"""A caller obtaining the service mid-init must not hit _get_plugin's guard."""
await close_optimal_service()
_SlowInitService.init_calls = 0
monkeypatch.setattr(optimal_module, "OptimalService", _SlowInitService)
async def _use_service() -> None:
svc = await get_optimal_service()
# Mirror what index_documentation does first: resolve the plugin,
# which raises RuntimeError("OptimalService not initialized...") if
# the singleton was published before initialize() completed.
svc._plugins[IndexType.DOCUMENTATION] = _FakePlugin()
svc._get_plugin(IndexType.DOCUMENTATION)
try:
await asyncio.gather(_use_service(), _use_service())
finally:
await close_optimal_service()
class _FakePlugin:
"""Minimal stand-in so _get_plugin returns without a real plugin."""
async def close(self) -> None: # pragma: no cover - never awaited here
return None