diff --git a/src/sre_agent/eval/common/mocks/__init__.py b/src/sre_agent/eval/common/mocks/__init__.py
new file mode 100644
index 00000000..3a0bdd18
--- /dev/null
+++ b/src/sre_agent/eval/common/mocks/__init__.py
@@ -0,0 +1 @@
+"""Shared mocks for evaluation suites."""
diff --git a/src/sre_agent/eval/diagnosis_quality/mocks/cloudwatch.py b/src/sre_agent/eval/common/mocks/cloudwatch.py
similarity index 76%
rename from src/sre_agent/eval/diagnosis_quality/mocks/cloudwatch.py
rename to src/sre_agent/eval/common/mocks/cloudwatch.py
index 86a2b263..bc837cde 100644
--- a/src/sre_agent/eval/diagnosis_quality/mocks/cloudwatch.py
+++ b/src/sre_agent/eval/common/mocks/cloudwatch.py
@@ -1,15 +1,30 @@
-"""Mock CloudWatch tools for diagnosis quality evaluation."""
+"""Mock CloudWatch tools shared across evaluation suites."""
+
+from typing import Protocol
 
 import opik
 
 from sre_agent.core.models import LogEntry, LogQueryResult
-from sre_agent.eval.diagnosis_quality.mocks.runtime import MockToolRuntime
 
 MOCK_TIMESTAMP = "2026-01-01T00:00:00+00:00"
 
 
+class _MockCloudWatchEntryLike(Protocol):
+    message: list[str]
+
+
+class _CaseLike(Protocol):
+    mock_cloudwatch_entries: list[_MockCloudWatchEntryLike]
+
+
+class MockRuntimeLike(Protocol):
+    """Structural type required by the shared CloudWatch mock."""
+
+    case: _CaseLike
+
+
 async def search_error_logs(
-    runtime: MockToolRuntime,
+    runtime: MockRuntimeLike,
     log_group: str,
     service_name: str,
     time_range_minutes: int,
@@ -44,7 +59,7 @@ async def search_error_logs(
         )
 
 
-def _normalise_messages(runtime: MockToolRuntime) -> list[str]:
+def _normalise_messages(runtime: MockRuntimeLike) -> list[str]:
     """Convert multiline fixture entries into non-empty log messages.
 
     Returns:
diff --git a/src/sre_agent/eval/tool_call/mocks/slack.py b/src/sre_agent/eval/common/mocks/slack.py
similarity index 93%
rename from src/sre_agent/eval/tool_call/mocks/slack.py
rename to src/sre_agent/eval/common/mocks/slack.py
index 25dcc73f..bbeb4569 100644
--- a/src/sre_agent/eval/tool_call/mocks/slack.py
+++ b/src/sre_agent/eval/common/mocks/slack.py
@@ -1,4 +1,4 @@
-"""Mock Slack tools for tool call evaluation."""
+"""Mock Slack tools shared across evaluation suites."""
 
 from typing import Any
 
diff --git a/src/sre_agent/eval/diagnosis_quality/mocks/slack.py b/src/sre_agent/eval/diagnosis_quality/mocks/slack.py
deleted file mode 100644
index 0c9aab73..00000000
--- a/src/sre_agent/eval/diagnosis_quality/mocks/slack.py
+++ /dev/null
@@ -1,33 +0,0 @@
-"""Mock Slack tools for diagnosis quality evaluation."""
-
-from typing import Any
-
-import opik
-
-MOCK_THREAD_TS = "1800000000.1000"
-
-
-async def conversations_add_message(
-    channel_id: str,
-    payload: str,
-    thread_ts: str | None,
-) -> dict[str, Any]:
-    """Mock Slack conversations_add_message.
-
-    Returns:
-        A mock Slack API response dict.
-    """
-    span_input: dict[str, Any] = {"channel_id": channel_id, "payload": payload}
-    if thread_ts is not None:
-        span_input["thread_ts"] = thread_ts
-
-    with opik.start_as_current_span(
-        name="conversations_add_message",
-        type="tool",
-        input=span_input,
-        metadata={"mocked": True, "provider": "slack"},
-    ):
-        if thread_ts is None:
-            return {"ok": True, "channel": channel_id, "ts": MOCK_THREAD_TS}
-
-        return {"ok": True, "channel": channel_id, "ts": thread_ts}
diff --git a/src/sre_agent/eval/diagnosis_quality/mocks/toolset.py b/src/sre_agent/eval/diagnosis_quality/mocks/toolset.py
index 21382627..29d9cf51 100644
--- a/src/sre_agent/eval/diagnosis_quality/mocks/toolset.py
+++ b/src/sre_agent/eval/diagnosis_quality/mocks/toolset.py
@@ -5,8 +5,8 @@
 from pydantic_ai import FunctionToolset
 
 from sre_agent.core.models import LogQueryResult
-from sre_agent.eval.diagnosis_quality.mocks import cloudwatch as cloudwatch_mocks
-from sre_agent.eval.diagnosis_quality.mocks import slack as slack_mocks
+from sre_agent.eval.common.mocks import cloudwatch as cloudwatch_mocks
+from sre_agent.eval.common.mocks import slack as slack_mocks
 from sre_agent.eval.diagnosis_quality.mocks.runtime import MockToolRuntime
 
 
diff --git a/src/sre_agent/eval/tool_call/mocks/cloudwatch.py b/src/sre_agent/eval/tool_call/mocks/cloudwatch.py
deleted file mode 100644
index 0706f2a0..00000000
--- a/src/sre_agent/eval/tool_call/mocks/cloudwatch.py
+++ /dev/null
@@ -1,58 +0,0 @@
-"""Mock CloudWatch tools for tool call evaluation."""
-
-import opik
-
-from sre_agent.core.models import LogEntry, LogQueryResult
-from sre_agent.eval.tool_call.mocks.runtime import MockToolRuntime
-
-MOCK_TIMESTAMP = "2026-01-01T00:00:00+00:00"
-
-
-async def search_error_logs(
-    runtime: MockToolRuntime,
-    log_group: str,
-    service_name: str,
-    time_range_minutes: int,
-) -> LogQueryResult:
-    """Mock CloudWatch log lookup using case fixtures.
-
-    Returns:
-        A LogQueryResult populated from case fixtures.
-    """
-    with opik.start_as_current_span(
-        name="search_error_logs",
-        type="tool",
-        input={
-            "log_group": log_group,
-            "service_name": service_name,
-            "time_range_minutes": time_range_minutes,
-        },
-        metadata={"mocked": True, "provider": "cloudwatch"},
-    ):
-        entries = [
-            LogEntry(
-                timestamp=MOCK_TIMESTAMP,
-                message=message,
-                log_stream=None,
-            )
-            for message in _normalise_messages(runtime)
-        ]
-        return LogQueryResult(
-            entries=entries,
-            log_group=log_group,
-            query=f"mock: search_error_logs service={service_name}",
-        )
-
-
-def _normalise_messages(runtime: MockToolRuntime) -> list[str]:
-    """Convert multiline fixture entries into non-empty log messages.
-
-    Returns:
-        A list of non-empty log message strings.
-    """
-    messages: list[str] = []
-    for entry in runtime.case.mock_cloudwatch_entries:
-        message = "\n".join(line.rstrip("\n") for line in entry.message).strip()
-        if message:
-            messages.append(message)
-    return messages
diff --git a/src/sre_agent/eval/tool_call/mocks/toolset.py b/src/sre_agent/eval/tool_call/mocks/toolset.py
index f0efd754..ddaac246 100644
--- a/src/sre_agent/eval/tool_call/mocks/toolset.py
+++ b/src/sre_agent/eval/tool_call/mocks/toolset.py
@@ -5,8 +5,8 @@
 from pydantic_ai import FunctionToolset
 
 from sre_agent.core.models import LogQueryResult
-from sre_agent.eval.tool_call.mocks import cloudwatch as cloudwatch_mocks
-from sre_agent.eval.tool_call.mocks import slack as slack_mocks
+from sre_agent.eval.common.mocks import cloudwatch as cloudwatch_mocks
+from sre_agent.eval.common.mocks import slack as slack_mocks
 from sre_agent.eval.tool_call.mocks.runtime import MockToolRuntime