Cleanup: single source of truth for backend, fresh agent per review

app.py

@@ -9,7 +9,6 @@ Deployed demo uses `groq` (free tier); local dev uses `ollama`.
 
 from __future__ import annotations
 
-import os
 import time
 
 import gradio as gr
@@ -19,7 +18,7 @@ load_dotenv()  # load .env locally; a no-op in hosted envs that use real env var
 
 from pr_review_agent.agent import (
     assemble_context,
-    build_agent,
+    current_model_label,
     reason_over,
 )
 from pr_review_agent.tools import (
@@ -42,17 +41,9 @@ try:  # pragma: no cover - only runs on Hugging Face Spaces
 except Exception:
     pass
 
-# Build the agent once at startup and reuse it across requests.
-_agent = build_agent()
-
 # Which model is answering — shown in the header as a small technical detail.
-_BACKEND = os.getenv("MODEL_BACKEND", "groq").lower()
-_MODEL_LABEL = {
-    "groq": f"{os.getenv('GROQ_MODEL', 'openai/gpt-oss-120b')} · Groq",
-    "gemini": f"{os.getenv('GEMINI_MODEL', 'gemini-2.0-flash')} · Gemini",
-    "bedrock": "Claude · Bedrock",
-    "ollama": f"{os.getenv('OLLAMA_MODEL', 'qwen2.5:7b')} · local",
-}.get(_BACKEND, _BACKEND)
+# Resolved from the agent module so the label can't drift from what actually runs.
+_MODEL_LABEL = current_model_label()
 
 CSS = """
 @import url('https://fonts.googleapis.com/css2?family=Space+Grotesk:wght@500;600;700&family=IBM+Plex+Sans:wght@400;500;600&family=IBM+Plex+Mono:wght@400;500;600&display=swap');
@@ -220,7 +211,9 @@ def review(pr_ref: str):
         yield _loader(2)
         conventions = get_repo_conventions(pr_ref)
         yield _loader(3)
-        report = reason_over(assemble_context(pr_ref, metadata, diff, conventions), agent=_agent)
+        # No shared agent: reason_over builds a fresh one per review, so concurrent
+        # reviews never share mutable conversation state.
+        report = reason_over(assemble_context(pr_ref, metadata, diff, conventions))
         # The model call does reasoning + writing in one shot; briefly show the final
         # "Writing the briefing" step so the loader visibly completes before the report.
         yield _loader(4)

pr_review_agent/agent.py

@@ -54,15 +54,41 @@ Rules:
 """
 
 
+DEFAULT_BACKEND = "groq"
+
+# Single source of truth for each backend: env var holding the model id, its default,
+# and a display name. Used both to build the model and to label it in the UI, so the two
+# can never drift apart.
+_BACKENDS = {
+    "groq": ("GROQ_MODEL", "openai/gpt-oss-120b", "Groq"),
+    "gemini": ("GEMINI_MODEL", "gemini-2.0-flash", "Gemini"),
+    "bedrock": ("BEDROCK_MODEL_ID", "us.anthropic.claude-sonnet-4-5-20250929-v1:0", "Bedrock"),
+    "ollama": ("OLLAMA_MODEL", "qwen2.5:7b", "local"),
+}
+
+
+def current_backend() -> str:
+    """The active model backend, from MODEL_BACKEND (defaulting to DEFAULT_BACKEND)."""
+    return os.getenv("MODEL_BACKEND", DEFAULT_BACKEND).lower()
+
+
+def current_model_label() -> str:
+    """A 'model-id · Backend' label for the active backend, for the UI header."""
+    backend = current_backend()
+    env_var, default_id, nice = _BACKENDS.get(backend, (None, "", backend))
+    model_id = os.getenv(env_var, default_id) if env_var else backend
+    return f"{model_id} · {nice}"
+
+
 def _build_model():
-    backend = os.getenv("MODEL_BACKEND", "bedrock").lower()
+    backend = current_backend()
 
     if backend == "ollama":
         from strands.models.ollama import OllamaModel
 
         return OllamaModel(
             host=os.getenv("OLLAMA_HOST", "http://localhost:11434"),
-            model_id=os.getenv("OLLAMA_MODEL", "llama3.1"),
+            model_id=os.getenv("OLLAMA_MODEL", _BACKENDS["ollama"][1]),
             temperature=0.1,
         )
 
@@ -77,7 +103,7 @@ def _build_model():
                 "api_key": os.getenv("GROQ_API_KEY"),
                 "base_url": "https://api.groq.com/openai/v1",
             },
-            model_id=os.getenv("GROQ_MODEL", "openai/gpt-oss-120b"),
+            model_id=os.getenv("GROQ_MODEL", _BACKENDS["groq"][1]),
             params={"temperature": 0.2},
         )
 
@@ -90,7 +116,7 @@ def _build_model():
             client_args={
                 "api_key": os.getenv("GEMINI_API_KEY") or os.getenv("GOOGLE_API_KEY")
             },
-            model_id=os.getenv("GEMINI_MODEL", "gemini-2.0-flash"),
+            model_id=os.getenv("GEMINI_MODEL", _BACKENDS["gemini"][1]),
             params={"temperature": 0.2},
         )
 
@@ -99,10 +125,7 @@ def _build_model():
     from strands.models import BedrockModel
 
     return BedrockModel(
-        model_id=os.getenv(
-            "BEDROCK_MODEL_ID",
-            "us.anthropic.claude-sonnet-4-5-20250929-v1:0",
-        ),
+        model_id=os.getenv("BEDROCK_MODEL_ID", _BACKENDS["bedrock"][1]),
         region_name=os.getenv("AWS_REGION", "us-west-2"),
         temperature=0.2,
     )
@@ -155,15 +178,15 @@ def gather_context(pr_ref: str) -> str:
 
 
 def reason_over(context: str, agent: Agent | None = None) -> str:
-    """Run the agent over an assembled context blob and return the cleaned report."""
+    """Run the agent over an assembled context blob and return the cleaned report.
+
+    Each review is an independent single-turn call. When no agent is passed, a fresh
+    one is built — so there is no prior turn to replay (a reused agent would send its
+    previous assistant turn back, and gpt-oss reasoning content is rejected by Groq in
+    multi-turn conversations). Callers that serve concurrent reviews should NOT share
+    one agent; pass None and let each review build its own.
+    """
     agent = agent or build_agent()
-    # Each review is an independent single-turn call. Clear any prior history so we
-    # never replay a previous assistant turn — gpt-oss reasoning content is rejected
-    # by Groq in multi-turn conversations, which would break the 2nd+ review.
-    try:
-        agent.messages.clear()
-    except Exception:
-        agent.messages = []
     instruction = (
         f"{context}\n"
         "Using ONLY the context above, write the review report now, with all five "
@@ -189,5 +212,4 @@ def prepare_review(pr_ref: str, agent: Agent | None = None) -> str:
     Returns:
         The five-section markdown review report.
     """
-    agent = agent or build_agent()
     return reason_over(gather_context(pr_ref), agent)