Enhance reviewer agent loop stop check

.pre-commit-config.yaml

@@ -39,7 +39,7 @@ repos:
     -   id: yamllint
         exclude: pre-commit-config.yaml
 -   repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: "v0.12.3"
+    rev: "v0.12.4"
     hooks:
     -   id: ruff-format
     -   id: ruff-check

packages/manugen-ai/src/manugen_ai/agents/ai_science_writer/sub_agents/reviewer/agent.py

@@ -5,32 +5,36 @@
 import os
 
 from google.adk.agents import Agent, LoopAgent
-from google.adk.tools import FunctionTool
-from manugen_ai.agents.meta_agent import ResilientToolAgent
-from manugen_ai.tools.tools import exit_loop
+from manugen_ai.agents.meta_agent import StopChecker
 from manugen_ai.utils import get_llm
 
 # Environment-driven model names
 MODEL_NAME = os.environ.get("MANUGENAI_MODEL_NAME")
 LLM = get_llm(MODEL_NAME)
-COMPLETION_PHRASE = "THE AGENT HAS COMPLETED THE TASK."
+COMPLETION_PHRASE = "TASK COMPLETED!"
 
 agent_review_loop = Agent(
     model=LLM,
     name="review_and_decide",
-    description="Critique full markdown and decide if it's final.",
-    instruction="""
+    description="Critique full markdown and decide if it's finalized.",
+    instruction=f"""
 You receive content from the user.
 
 Your job:
 1. Provide a concise bullet-list of any changes needed.
-2. If **no** changes are needed (content is publication-ready), call the tool `exit_loop` instead of returning bullets.
+    - Use clear, actionable language.
+    - Focus on clarity, coherence, and scientific rigor.
+    - If changes are needed, return a list of feedback bullets.
+    - Don't go overboard, try to keep it concise and reasonable when it comes to updates.
+    - Do NOT provide guidance about new topics. You are to stick to only the topic areas within the content.
+2. If **no** changes are needed (content is publication-ready), response ONLY with "{COMPLETION_PHRASE}" instead of returning bullets.
 
-Return either:
-- A JSON list of feedback bullets,
-- Or invoke `exit_loop` (via the FunctionTool) to end the loop.
+Return ONLY ONE of the following.
+DO NOT return both!
+1. A list of feedback bullets for use with revision.
+2. Or respond with ONLY "{COMPLETION_PHRASE}" to indicate the content is ready for publication
+If you respond with "{COMPLETION_PHRASE}", do not include any additional text or commentary.
 """,
-    tools=[FunctionTool(func=exit_loop)],
     output_key="feedback",
 )
 
@@ -39,14 +43,20 @@
     model=LLM,
     name="refine_manuscript",
     description="Apply feedback to revise the manuscript.",
-    instruction="""
+    instruction=f"""
 You receive:
 - Content from the user.
-- feedback: a list of bullets in {feedback}
+- feedback: a list of bullets in {{feedback}}
 
 Revise the content accordingly and return *only* updated content.
+You MUST return actual content, not instructions or commentary about the content.
+Do not add additional headers which are not in the original content.
+Do not include "{COMPLETION_PHRASE}" in the output.
+Do not include a header about the feedback.
 Do NOT return additional commentary, JSON, or metadata.
 ONLY return the revised content.
+The revised content should be in markdown format and not include any additional
+wrappers (such as backticks ```) or formatting besides what was provided in the original content.
 """,
     output_key="refined_md",
 )
@@ -55,10 +65,11 @@
 root_agent = LoopAgent(
     name="reviewer_agent",
     description="Iteratively review and refine the manuscript until publication-ready.",
-    # Use ResilientToolAgent so missing-tool errors auto-retry up to N times
     sub_agents=[
-        ResilientToolAgent(agent_review_loop, max_retries=2),
-        ResilientToolAgent(agent_refine, max_retries=2),
+        agent_review_loop,
+        agent_refine,
+        # use stopchecker to check if the work is completed.
+        StopChecker(context_variable="feedback", completion_phrase=COMPLETION_PHRASE),
     ],
     max_iterations=5,  # safety cap
 )

packages/manugen-ai/src/manugen_ai/utils.py

@@ -99,7 +99,7 @@ def get_llm(model_name: str, **kwargs):
     elif model_name.startswith("gemini-"):
         return model_name
 
-    raise ValueError(f"Unknown model name: {model_name}")
+    raise ValueError(f"Unknown model name: {model_name if model_name else '<empty>'}")
 
 
 def prepare_ollama_models_for_adk_state() -> None:

packages/manugen-ai/tests/conftest.py

@@ -2,13 +2,20 @@
 conftest for pytest fixutres and related
 """
 
+import os
 import pathlib
 import tempfile
 from typing import Any, Generator
 
 import pytest
 
 
+@pytest.fixture(autouse=True, scope="session")
+def set_model_name_env():
+    os.environ["MANUGENAI_MODEL_NAME"] = "openai/llama3.2:3b"
+    os.environ["MANUGENAI_FIGURE_MODEL_NAME"] = "openai/llama3.2:3b"
+
+
 @pytest.fixture
 def temp_markdown_dir() -> Generator[
     pathlib.Path,

packages/manugen-ai/tests/test_agents.py

@@ -3,7 +3,7 @@
 """
 
 import pytest
-from manugen_ai.agents.capitalizer.agent import root_agent
+from manugen_ai.agents.capitalizer.agent import root_agent as capitalizer_agent
 from manugen_ai.utils import run_agent_workflow
 
 
@@ -17,7 +17,7 @@ async def test_agent_capitalizer():
     # retry 5 times
     for attempt in range(5):
         _, session_state, _ = await run_agent_workflow(
-            agent=root_agent,
+            agent=capitalizer_agent,
             prompt="""
             this is a sentence to correct
             """,
@@ -32,3 +32,45 @@ async def test_agent_capitalizer():
             # Final attempt failed, raise assertion
             assert "output" in session_state.keys()
             assert session_state["output"] == expected_output
+
+
+@pytest.mark.asyncio
+async def test_agent_reviewer():
+    """
+    Test the reviewer agent to ensure it can loop and improve text
+    """
+
+    # localized imports to help with agent env setting
+    from manugen_ai.agents.ai_science_writer.sub_agents.reviewer.agent import (
+        COMPLETION_PHRASE as review_agent_COMPLETION_PHRASE,
+    )
+    from manugen_ai.agents.ai_science_writer.sub_agents.reviewer.agent import (
+        root_agent as review_agent,
+    )
+
+    # we attempt multiple times to ensure the agent can loop and improve the text
+    for attempt in range(5):
+        _, session_state, output_events = await run_agent_workflow(
+            agent=review_agent,
+            prompt="""
+            Please improve the following text:
+
+            It was a dark and stormy night.
+            """,
+            app_name="app",
+            user_id="user",
+            session_id="0001",
+            # note: use `uv run pytest -s` to view the verbose stdout
+            verbose=True,
+        )
+
+        # check that we received feedback and that the completion phrase was used.
+        if (
+            "feedback" in session_state
+            and session_state["feedback"] == review_agent_COMPLETION_PHRASE
+        ):
+            break
+        if attempt == 4:
+            raise AssertionError(
+                "Final attempt failed, completion phrase not found or does not match expected output."
+            )

packages/manugen-ai/tests/test_tools.py

@@ -1,3 +1,7 @@
+"""
+Tests for custom tools
+"""
+
 import pathlib
 from types import SimpleNamespace