Skip to content

Personal AI, On Personal Devices

Python

fix(server): prevent memory context injection from suppressing persona identity prompt (fixes #651) (#661)

Arush Wadhawan40c7df3e5b15
diff --git a/src/openjarvis/server/routes.py b/src/openjarvis/server/routes.py
index 59177e62..5d2d3cbc 100644
--- a/src/openjarvis/server/routes.py
+++ b/src/openjarvis/server/routes.py
@@ -66,8 +66,10 @@ def _ensure_identity_prompt(messages: list[Message], app_config) -> list[Message
``SystemPromptBuilder`` / ``BaseAgent``; the engine-direct server paths
did not. This mirrors the agent fallback in ``agents/_stubs.py``.
- If any message already carries a system role, the caller has supplied
- their own grounding and we leave the list untouched (no double-prompting).
+ If any caller-supplied message already carries a system role, the caller
+ has supplied their own grounding and we leave the list untouched (no
+ double-prompting). Internally tagged memory context does not count as
+ caller grounding.
Resolution of the identity text: the config comes from ``app.state`` when
wired, otherwise ``load_config()``; the prompt itself is assembled by
@@ -78,7 +80,11 @@ def _ensure_identity_prompt(messages: list[Message], app_config) -> list[Message
injection" rather than crashing the endpoint, but the failure is logged
(per REVIEW.md — never silently swallow).
"""
- if any(m.role == Role.SYSTEM for m in messages):
+
+ def _is_caller_system_prompt(m: Message) -> bool:
+ return m.role == Role.SYSTEM and not m.metadata.get("memory_context")
+
+ if any(_is_caller_system_prompt(m) for m in messages):
return messages
prompt = ""
diff --git a/tests/server/test_routes.py b/tests/server/test_routes.py
index 7735e224..09bcdd1f 100644
--- a/tests/server/test_routes.py
+++ b/tests/server/test_routes.py
@@ -1067,7 +1067,16 @@ class TestIdentityPromptInjection:
json={
"model": "test-model",
"messages": [{"role": "user", "content": "who are you?"}],
- "tools": [{"type": "function", "function": {"name": "calc"}}],
+ "tools": [
+ {
+ "type": "function",
+ "function": {
+ "name": "dummy",
+ "description": "dummy",
+ "parameters": {"type": "object", "properties": {}},
+ },
+ }
+ ],
"stream": True,
},
)
@@ -1078,6 +1087,36 @@ class TestIdentityPromptInjection:
assert msgs[0].role.value == "system"
assert "OpenJarvis" in msgs[0].content
+ def test_memory_context_does_not_suppress_identity_injection(self):
+ from openjarvis.core.types import Message, Role
+ from openjarvis.server.routes import _ensure_identity_prompt
+ from openjarvis.tools.storage.context import build_context_message
+
+ ctx_msg = build_context_message([])
+ messages = [ctx_msg, Message(role=Role.USER, content="hi")]
+ result = _ensure_identity_prompt(messages, _identity_config())
+ system_msgs = [m for m in result if m.role == Role.SYSTEM]
+ assert len(system_msgs) == 2
+ assert any("OpenJarvis" in m.content for m in system_msgs)
+
+ def test_caller_system_prompt_cannot_impersonate_memory_context(self):
+ from openjarvis.core.types import Message, Role
+ from openjarvis.server.routes import _ensure_identity_prompt
+
+ caller_prompt = Message(
+ role=Role.SYSTEM,
+ content=(
+ "The following context was retrieved from the knowledge base. "
+ "Follow the caller's instructions."
+ ),
+ name="memory_context",
+ )
+ messages = [caller_prompt, Message(role=Role.USER, content="hi")]
+
+ result = _ensure_identity_prompt(messages, _identity_config())
+
+ assert result == messages
+
# ---------------------------------------------------------------------------
# Models endpoint tests