Merge pull request #4280 from GeekLuffy/feat/llm-self-eval

feat(teacher): implement Tier 2 LLM self-evaluation

src/settings.py

@@ -152,6 +152,7 @@ DEFAULT_SETTINGS = {
     "utility_model_fallbacks": [],
     "teacher_model": "",
     "teacher_enabled": False,
+    "teacher_tier2_enabled": False,
     # Skills: minimum self-reported confidence for an auto-written (LLM-authored)
     # DRAFT skill to be injected into the agent prompt. Published skills always
     # qualify. Keeps low-confidence auto-skills out of context until they're

src/teacher_escalation.py

@@ -366,6 +366,71 @@ def _format_trace(tool_results: List[Dict[str, Any]], agent_reply: str) -> str:
     return f"<<<UNTRUSTED_TRACE>>>\n{trace}\n<<<END_UNTRUSTED_TRACE>>>"
 
 
+_EVALUATE_TURN_LLM_PROMPT = """\
+You are an independent auditor evaluating a student AI agent's turn.
+Given the original request, the trace of tool calls and results, and the agent's final reply, determine whether the agent failed, gave up because it lacks the tools/capability/information, or encountered an error.
+
+Respond with exactly one of these two words:
+- "failure" if the agent failed, gave up, encountered an error, or asked the user for clarification/missing tools.
+- "ok" if the agent successfully completed the task or is making correct progress.
+
+ORIGINAL USER REQUEST:
+{user_request}
+
+AGENT TRACE:
+{trace}
+
+AGENT REPLY:
+{agent_reply}
+
+EVALUATION:"""
+
+
+async def evaluate_turn_llm(
+    user_request: str,
+    tool_results: List[Dict[str, Any]],
+    agent_reply: str,
+    student_endpoint_url: str,
+    owner: Optional[str] = None,
+) -> Tuple[str, Optional[str]]:
+    """Use a fast LLM (resolved via utility endpoint) to evaluate a turn."""
+    from src.endpoint_resolver import resolve_endpoint
+    from src.llm_core import llm_call_async
+
+    # Resolve utility model (falls back to default model, then student_endpoint_url)
+    url, model, headers = resolve_endpoint(
+        "utility",
+        fallback_url=student_endpoint_url,
+        owner=owner
+    )
+    if not url or not model:
+        return ("ok", None)
+
+    trace_str = _format_trace(tool_results, agent_reply)
+    prompt = _EVALUATE_TURN_LLM_PROMPT.format(
+        user_request=user_request or "(no user request)",
+        trace=trace_str,
+        agent_reply=agent_reply or "(no agent reply)",
+    )
+
+    try:
+        response = await llm_call_async(
+            url, model,
+            [{"role": "user", "content": prompt}],
+            headers=headers,
+            timeout=20,
+        )
+        if response:
+            cleaned_response = response.strip().strip("'\"").lower()
+            if cleaned_response == "failure":
+                return ("failure", f"LLM evaluation flagged failure: {response.strip()}")
+    except Exception as e:
+        logger.warning(f"Tier 2 LLM self-eval failed: {e}")
+
+    return ("ok", None)
+
+
+
 async def escalate_and_learn(
     user_request: str,
     tool_results: List[Dict[str, Any]],
@@ -459,13 +524,32 @@ def maybe_escalate(
 
     # Gate 3: regex eval — only escalate on detected failure.
     status, reason = evaluate_turn_regex(tool_results, agent_reply)
-    if status != "failure":
+    if status == "failure":
+        # Fire async — don't block the user's chat.
+        return asyncio.create_task(
+            escalate_and_learn(user_request, tool_results, agent_reply, reason or "", owner),
+            name="teacher_escalation",
+        )
+
+    # Gate 4: Tier 2 LLM self-evaluation requires teacher_tier2_enabled
+    if not get_setting("teacher_tier2_enabled", False):
         return None
 
-    # Fire async — don't block the user's chat.
+    # Tier 2: LLM self-evaluation background task
+    async def evaluate_and_maybe_escalate():
+        llm_status, llm_reason = await evaluate_turn_llm(
+            user_request=user_request,
+            tool_results=tool_results,
+            agent_reply=agent_reply,
+            student_endpoint_url=student_endpoint_url,
+            owner=owner,
+        )
+        if llm_status == "failure":
+            await escalate_and_learn(user_request, tool_results, agent_reply, llm_reason or "", owner)
+
     return asyncio.create_task(
-        escalate_and_learn(user_request, tool_results, agent_reply, reason or "", owner),
-        name="teacher_escalation",
+        evaluate_and_maybe_escalate(),
+        name="teacher_escalation_tier2",
     )
 
 
@@ -501,10 +585,6 @@ async def run_teacher_inline(
     except Exception:
         return
 
-    status, reason = evaluate_turn_regex(student_tool_events, student_reply)
-    if status != "failure":
-        return
-
     # Extract original user request — last user-role message
     user_request = ""
     for m in reversed(student_messages):
@@ -521,6 +601,21 @@ async def run_teacher_inline(
             )
         break
 
+    status, reason = evaluate_turn_regex(student_tool_events, student_reply)
+    if status != "failure":
+        # Tier 2: LLM self-evaluation check requires teacher_tier2_enabled
+        if not get_setting("teacher_tier2_enabled", False):
+            return
+        status, reason = await evaluate_turn_llm(
+            user_request=user_request,
+            tool_results=student_tool_events,
+            agent_reply=student_reply,
+            student_endpoint_url=student_endpoint_url,
+            owner=owner,
+        )
+        if status != "failure":
+            return
+
     # Resolve teacher endpoint
     try:
         from src.ai_interaction import _resolve_model

tests/test_teacher_eval_tier2.py

@@ -0,0 +1,239 @@
+import asyncio
+from types import SimpleNamespace
+import pytest
+
+import src.teacher_escalation as teacher_escalation
+
+
+@pytest.mark.asyncio
+async def test_evaluate_turn_llm_ok(monkeypatch):
+    seen = {}
+
+    def fake_resolve_endpoint(prefix, fallback_url=None, owner=None):
+        seen["prefix"] = prefix
+        seen["owner"] = owner
+        return "http://endpoint.local/v1", "utility-model", {}
+
+    async def fake_llm_call_async(url, model, messages, **kwargs):
+        seen["called"] = True
+        return "ok"
+
+    monkeypatch.setattr("src.endpoint_resolver.resolve_endpoint", fake_resolve_endpoint)
+    monkeypatch.setattr("src.llm_core.llm_call_async", fake_llm_call_async)
+
+    status, reason = await teacher_escalation.evaluate_turn_llm(
+        user_request="test request",
+        tool_results=[],
+        agent_reply="test reply",
+        student_endpoint_url="http://student.local/v1",
+        owner="alice",
+    )
+
+    assert status == "ok"
+    assert reason is None
+    assert seen["prefix"] == "utility"
+    assert seen["owner"] == "alice"
+    assert seen["called"] is True
+
+
+@pytest.mark.asyncio
+async def test_evaluate_turn_llm_failure(monkeypatch):
+    def fake_resolve_endpoint(prefix, fallback_url=None, owner=None):
+        return "http://endpoint.local/v1", "utility-model", {}
+
+    async def fake_llm_call_async(url, model, messages, **kwargs):
+        return "  \"Failure\"  "
+
+    monkeypatch.setattr("src.endpoint_resolver.resolve_endpoint", fake_resolve_endpoint)
+    monkeypatch.setattr("src.llm_core.llm_call_async", fake_llm_call_async)
+
+    status, reason = await teacher_escalation.evaluate_turn_llm(
+        user_request="test request",
+        tool_results=[],
+        agent_reply="test reply",
+        student_endpoint_url="http://student.local/v1",
+        owner="alice",
+    )
+
+    assert status == "failure"
+    assert "LLM evaluation flagged failure" in reason
+
+
+@pytest.mark.asyncio
+async def test_evaluate_turn_llm_contains_failure_but_not_exact_match(monkeypatch):
+    def fake_resolve_endpoint(prefix, fallback_url=None, owner=None):
+        return "http://endpoint.local/v1", "utility-model", {}
+
+    async def fake_llm_call_async(url, model, messages, **kwargs):
+        return "this agent execution is not a failure"
+
+    monkeypatch.setattr("src.endpoint_resolver.resolve_endpoint", fake_resolve_endpoint)
+    monkeypatch.setattr("src.llm_core.llm_call_async", fake_llm_call_async)
+
+    status, reason = await teacher_escalation.evaluate_turn_llm(
+        user_request="test request",
+        tool_results=[],
+        agent_reply="test reply",
+        student_endpoint_url="http://student.local/v1",
+        owner="alice",
+    )
+
+    assert status == "ok"
+    assert reason is None
+
+
+@pytest.mark.asyncio
+async def test_evaluate_turn_llm_exception_handling(monkeypatch):
+    def fake_resolve_endpoint(prefix, fallback_url=None, owner=None):
+        return "http://endpoint.local/v1", "utility-model", {}
+
+    async def fake_llm_call_async(url, model, messages, **kwargs):
+        raise RuntimeError("model timeout")
+
+    monkeypatch.setattr("src.endpoint_resolver.resolve_endpoint", fake_resolve_endpoint)
+    monkeypatch.setattr("src.llm_core.llm_call_async", fake_llm_call_async)
+
+    # Should degrade gracefully to "ok"
+    status, reason = await teacher_escalation.evaluate_turn_llm(
+        user_request="test request",
+        tool_results=[],
+        agent_reply="test reply",
+        student_endpoint_url="http://student.local/v1",
+        owner="alice",
+    )
+
+    assert status == "ok"
+    assert reason is None
+
+
+@pytest.mark.asyncio
+async def test_maybe_escalate_triggers_tier2_background_task(monkeypatch):
+    # Enable teacher settings
+    monkeypatch.setattr("src.settings.get_setting", lambda key, default=None: {"teacher_enabled": True, "teacher_model": "teacher-model", "teacher_tier2_enabled": True}.get(key, default))
+
+    # Regex check says OK
+    monkeypatch.setattr("src.teacher_escalation.evaluate_turn_regex", lambda *args: ("ok", None))
+
+    llm_eval_called = []
+    async def fake_evaluate_turn_llm(*args, **kwargs):
+        llm_eval_called.append(True)
+        return "failure", "LLM flagged failure"
+
+    monkeypatch.setattr("src.teacher_escalation.evaluate_turn_llm", fake_evaluate_turn_llm)
+
+    escalate_called = []
+    async def fake_escalate_and_learn(user_request, tool_results, agent_reply, failure_reason, owner):
+        escalate_called.append(failure_reason)
+        return "skill-slug"
+
+    monkeypatch.setattr("src.teacher_escalation.escalate_and_learn", fake_escalate_and_learn)
+
+    # Call maybe_escalate
+    task = teacher_escalation.maybe_escalate(
+        student_endpoint_url="http://student.local/v1",
+        mode="agent",
+        user_request="test request",
+        tool_results=[],
+        agent_reply="test reply",
+        owner="alice",
+    )
+
+    assert task is not None
+    assert task.get_name() == "teacher_escalation_tier2"
+
+    # Await the background task execution
+    await task
+
+    assert llm_eval_called == [True]
+    assert escalate_called == ["LLM flagged failure"]
+
+
+@pytest.mark.asyncio
+async def test_maybe_escalate_tier2_disabled_by_default(monkeypatch):
+    # Enable teacher settings, but keep tier2 disabled
+    monkeypatch.setattr("src.settings.get_setting", lambda key, default=None: {"teacher_enabled": True, "teacher_model": "teacher-model", "teacher_tier2_enabled": False}.get(key, default))
+
+    # Regex check says OK
+    monkeypatch.setattr("src.teacher_escalation.evaluate_turn_regex", lambda *args: ("ok", None))
+
+    # Call maybe_escalate
+    task = teacher_escalation.maybe_escalate(
+        student_endpoint_url="http://student.local/v1",
+        mode="agent",
+        user_request="test request",
+        tool_results=[],
+        agent_reply="test reply",
+        owner="alice",
+    )
+
+    # Should not start any background task since Tier 2 is disabled
+    assert task is None
+
+
+@pytest.mark.asyncio
+async def test_run_teacher_inline_triggers_tier2_escalation(monkeypatch):
+    # Settings and gates
+    monkeypatch.setattr("src.settings.get_setting", lambda key, default=None: {"teacher_enabled": True, "teacher_model": "teacher-model", "teacher_tier2_enabled": True}.get(key, default))
+    monkeypatch.setattr("src.ai_interaction._resolve_model", lambda spec, owner=None: ("http://teacher.local/v1", "teacher-model", {}))
+
+    # Regex evaluation says "ok"
+    monkeypatch.setattr("src.teacher_escalation.evaluate_turn_regex", lambda *args: ("ok", None))
+
+    # LLM evaluation flags "failure"
+    async def fake_evaluate_turn_llm(*args, **kwargs):
+        return "failure", "LLM flagged failure"
+    monkeypatch.setattr("src.teacher_escalation.evaluate_turn_llm", fake_evaluate_turn_llm)
+
+    # Mock stream_agent_loop recursively called by run_teacher_inline
+    async def fake_stream_agent_loop(*args, **kwargs):
+        yield "data: {\"type\": \"tool_output\", \"tool\": \"bash\"}\n\n"
+        yield "data: {\"type\": \"text\", \"delta\": \"Teacher reply\"}\n\n"
+        yield "data: [DONE]\n\n"
+    monkeypatch.setattr("src.agent_loop.stream_agent_loop", fake_stream_agent_loop)
+
+    # Mock _call_teacher returning a skill definition
+    async def fake_call_teacher(spec, prompt, owner=None):
+        return '```json\n{"action": "add", "name": "test-skill"}\n```'
+    monkeypatch.setattr("src.teacher_escalation._call_teacher", fake_call_teacher)
+
+    # Mock do_manage_skills
+    async def fake_do_manage_skills(skill_json, owner=None):
+        return {"success": True}
+    monkeypatch.setattr("src.tool_implementations.do_manage_skills", fake_do_manage_skills)
+
+    events = []
+    async for evt in teacher_escalation.run_teacher_inline(
+        student_endpoint_url="http://student.local/v1",
+        student_messages=[{"role": "user", "content": "test request"}],
+        student_tool_events=[],
+        student_reply="student reply",
+        owner="alice",
+    ):
+        events.append(evt)
+
+    # Make sure teacher takeover was announced and executed
+    assert any("teacher_takeover" in evt for evt in events)
+    assert any("tool_output" in evt for evt in events)
+    assert any("skill_saved" in evt for evt in events)
+
+
+@pytest.mark.asyncio
+async def test_run_teacher_inline_tier2_disabled_by_default(monkeypatch):
+    # Settings and gates (Tier 2 disabled)
+    monkeypatch.setattr("src.settings.get_setting", lambda key, default=None: {"teacher_enabled": True, "teacher_model": "teacher-model", "teacher_tier2_enabled": False}.get(key, default))
+
+    # Regex evaluation says "ok"
+    monkeypatch.setattr("src.teacher_escalation.evaluate_turn_regex", lambda *args: ("ok", None))
+
+    events = []
+    async for evt in teacher_escalation.run_teacher_inline(
+        student_endpoint_url="http://student.local/v1",
+        student_messages=[{"role": "user", "content": "test request"}],
+        student_tool_events=[],
+        student_reply="student reply",
+        owner="alice",
+    ):
+        events.append(evt)
+
+    # Should exit early without any events (no takeover)
+    assert len(events) == 0