refactor(phase0): replace Anthropic SDK with LLMBackend abstraction (default Qwen)

viral_script_engine/agents/critic.py

@@ -1,12 +1,9 @@
 import json
-import os
 from typing import List
 
-import anthropic
-from dotenv import load_dotenv
 from pydantic import BaseModel
 
-load_dotenv()
+from viral_script_engine.agents.llm_backend import LLMBackend
 
 SYSTEM_PROMPT = """You are an expert social media content critic specialising in short-form video scripts for Reels and YouTube Shorts. Your job is to find specific, real problems in creator scripts — not vague feedback.
 
@@ -70,27 +67,17 @@ class CritiqueOutput(BaseModel):
 
 
 class CriticAgent:
-    def __init__(self, model_name: str = "claude-sonnet-4-20250514"):
-        self.model_name = model_name
-        self.client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])
-
-    def _call_api(self, user_content: str) -> str:
-        message = self.client.messages.create(
-            model=self.model_name,
-            max_tokens=2048,
-            system=SYSTEM_PROMPT,
-            messages=[{"role": "user", "content": user_content}],
-        )
-        return message.content[0].text
+    def __init__(self, backend: str = "qwen", model_name: str = "Qwen/Qwen2.5-7B-Instruct"):
+        self.llm = LLMBackend(backend=backend, model_name=model_name)
 
-    def _parse_response(self, raw: str, user_content: str) -> CritiqueOutput:
+    def _parse_response(self, raw: str, user_prompt: str) -> CritiqueOutput:
         try:
             data = json.loads(raw)
             data["raw_response"] = raw
             return CritiqueOutput(**data)
         except Exception:
-            strict_content = user_content + STRICT_RETRY_SUFFIX
-            raw2 = self._call_api(strict_content)
+            strict_prompt = user_prompt + STRICT_RETRY_SUFFIX
+            raw2 = self.llm.generate(SYSTEM_PROMPT, strict_prompt, max_tokens=2048)
             try:
                 data = json.loads(raw2)
                 data["raw_response"] = raw2
@@ -99,8 +86,8 @@ class CriticAgent:
                 raise CriticParseError(f"Failed to parse critique after 2 attempts: {e}")
 
     def critique(self, script: str, region: str, platform: str, niche: str) -> CritiqueOutput:
-        user_content = USER_PROMPT_TEMPLATE.format(
+        user_prompt = USER_PROMPT_TEMPLATE.format(
             script=script, region=region, platform=platform, niche=niche
         )
-        raw = self._call_api(user_content)
-        return self._parse_response(raw, user_content)
+        raw = self.llm.generate(SYSTEM_PROMPT, user_prompt, max_tokens=2048)
+        return self._parse_response(raw, user_prompt)

viral_script_engine/agents/llm_backend.py

@@ -0,0 +1,59 @@
+class LLMBackend:
+    def __init__(self, backend: str = "qwen", model_name: str = "Qwen/Qwen2.5-7B-Instruct"):
+        """
+        backend: "qwen" | "anthropic" | "openai"
+        Default is local Qwen — no API key needed.
+        Pipeline is lazy-loaded on first generate() call.
+        """
+        self.backend = backend
+        self.model_name = model_name
+        self._pipe = None
+        self._client = None
+
+        if backend not in ("qwen", "anthropic", "openai"):
+            raise ValueError(f"Unknown backend: {backend!r}. Choose qwen | anthropic | openai")
+
+    def _get_pipe(self):
+        if self._pipe is None:
+            from transformers import pipeline
+            self._pipe = pipeline("text-generation", model=self.model_name, device_map="auto")
+        return self._pipe
+
+    def _get_client(self):
+        if self._client is None:
+            if self.backend == "anthropic":
+                import anthropic
+                self._client = anthropic.Anthropic()
+            elif self.backend == "openai":
+                from openai import OpenAI
+                self._client = OpenAI()
+        return self._client
+
+    def generate(self, system_prompt: str, user_prompt: str, max_tokens: int = 512) -> str:
+        if self.backend == "qwen":
+            messages = [
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_prompt},
+            ]
+            out = self._get_pipe()(messages, max_new_tokens=max_tokens, return_full_text=False)
+            return out[0]["generated_text"]
+
+        elif self.backend == "anthropic":
+            msg = self._get_client().messages.create(
+                model=self.model_name,
+                max_tokens=max_tokens,
+                system=system_prompt,
+                messages=[{"role": "user", "content": user_prompt}],
+            )
+            return msg.content[0].text
+
+        elif self.backend == "openai":
+            resp = self._get_client().chat.completions.create(
+                model=self.model_name,
+                max_tokens=max_tokens,
+                messages=[
+                    {"role": "system", "content": system_prompt},
+                    {"role": "user", "content": user_prompt},
+                ],
+            )
+            return resp.choices[0].message.content

viral_script_engine/requirements.txt

@@ -1,7 +1,20 @@
-anthropic>=0.40.0
+# Core — required
+transformers>=4.40.0
+torch>=2.2.0
+accelerate>=0.28.0
+unsloth
+trl>=0.12.0
 sentence-transformers>=2.7.0
-numpy>=1.26.0
 pydantic>=2.0.0
+numpy>=1.26.0
 python-dotenv>=1.0.0
 rich>=13.0.0
+fastapi>=0.110.0
+uvicorn>=0.29.0
 pytest>=8.0.0
+matplotlib>=3.8.0
+openenv
+
+# Optional — only needed if using non-Qwen backends
+anthropic>=0.40.0      # only if backend="anthropic"
+openai>=1.0.0          # only if backend="openai"

viral_script_engine/scripts/run_critic_gate.py

@@ -30,8 +30,8 @@ def load_scripts(dry_run: bool) -> list:
     return scripts
 
 
-def run_gate(max_retries: int = 3, dry_run: bool = False) -> bool:
-    agent = CriticAgent()
+def run_gate(max_retries: int = 3, dry_run: bool = False, backend: str = "qwen", model_name: str = "Qwen/Qwen2.5-7B-Instruct") -> bool:
+    agent = CriticAgent(backend=backend, model_name=model_name)
     evaluator = CriticEvaluator()
     scripts = load_scripts(dry_run)
 
@@ -126,9 +126,11 @@ def main():
     parser = argparse.ArgumentParser(description="Run Critic quality gate")
     parser.add_argument("--max-retries", type=int, default=3)
     parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--backend", default="qwen", choices=["qwen", "anthropic", "openai"])
+    parser.add_argument("--model-name", default="Qwen/Qwen2.5-7B-Instruct")
     args = parser.parse_args()
 
-    passed = run_gate(max_retries=args.max_retries, dry_run=args.dry_run)
+    passed = run_gate(max_retries=args.max_retries, dry_run=args.dry_run, backend=args.backend, model_name=args.model_name)
     sys.exit(0 if passed else 1)
 
 

viral_script_engine/tests/conftest.py

@@ -0,0 +1,44 @@
+import json
+import pytest
+
+MOCK_VALID_RESPONSE = json.dumps({
+    "claims": [
+        {
+            "claim_id": "C1",
+            "critique_class": "hook_weakness",
+            "claim_text": "Weak hook.",
+            "timestamp_range": "0:00-0:03",
+            "evidence": "Let me tell you a secret",
+            "is_falsifiable": True,
+            "severity": "high",
+        },
+        {
+            "claim_id": "C2",
+            "critique_class": "cta_buried",
+            "claim_text": "CTA at end.",
+            "timestamp_range": "0:45-0:50",
+            "evidence": "Like and save this video",
+            "is_falsifiable": True,
+            "severity": "medium",
+        },
+        {
+            "claim_id": "C3",
+            "critique_class": "pacing_issue",
+            "claim_text": "Pacing issue.",
+            "timestamp_range": "N/A",
+            "evidence": "save twenty percent",
+            "is_falsifiable": True,
+            "severity": "low",
+        },
+    ],
+    "overall_severity": "high",
+})
+
+
+@pytest.fixture
+def mock_llm(monkeypatch):
+    monkeypatch.setattr(
+        "viral_script_engine.agents.llm_backend.LLMBackend.generate",
+        lambda self, sys_prompt, usr_prompt, **kw: MOCK_VALID_RESPONSE,
+    )
+    return MOCK_VALID_RESPONSE

viral_script_engine/tests/test_critic.py

@@ -1,9 +1,16 @@
 import json
+import os
+import subprocess
+import sys
+
 import pytest
-from unittest.mock import MagicMock, patch
+from unittest.mock import patch
 
-from viral_script_engine.agents.critic import CritiqueClaim, CritiqueOutput
+from viral_script_engine.agents.critic import CritiqueClaim, CritiqueOutput, CriticAgent, CriticParseError
 from viral_script_engine.evaluation.critic_evaluator import CriticEvaluator, EvaluationResult
+from tests.conftest import MOCK_VALID_RESPONSE
+
+MOCK_INVALID_RESPONSE = "Here is my feedback: The hook is weak and the CTA is missing."
 
 
 # ── Task 2: Model parsing tests ───────────────────────────────────────────────
@@ -91,11 +98,6 @@ def test_evaluator_fails_low_specificity():
 
 # ── Task 5: CLI exit-code test ────────────────────────────────────────────────
 
-import os
-import subprocess
-import sys
-
-
 def test_cli_dry_run_exits_zero_or_one():
     """CLI must exit 0 (pass) or 1 (fail) — never crash with unhandled exception."""
     result = subprocess.run(
@@ -103,7 +105,7 @@ def test_cli_dry_run_exits_zero_or_one():
         capture_output=True,
         text=True,
         cwd=str(__import__("pathlib").Path(__file__).parent.parent),
-        env={**os.environ, "ANTHROPIC_API_KEY": "sk-fake-key-for-test"},
+        env={**os.environ},
     )
     assert result.returncode in (0, 1), (
         f"Unexpected exit code: {result.returncode}\nSTDERR: {result.stderr}"
@@ -112,65 +114,37 @@ def test_cli_dry_run_exits_zero_or_one():
 
 # ── Task 6: Mocked CriticAgent tests ─────────────────────────────────────────
 
-from viral_script_engine.agents.critic import CriticAgent, CriticParseError
-
-MOCK_VALID_JSON = json.dumps({
-    "claims": [
-        {"claim_id": "C1", "critique_class": "hook_weakness", "claim_text": "Weak hook.", "timestamp_range": "0:00-0:03", "evidence": "Let me tell you a secret", "is_falsifiable": True, "severity": "high"},
-        {"claim_id": "C2", "critique_class": "cta_buried", "claim_text": "CTA at end.", "timestamp_range": "0:45-0:50", "evidence": "Like and save this video", "is_falsifiable": True, "severity": "medium"},
-        {"claim_id": "C3", "critique_class": "pacing_issue", "claim_text": "Pacing issue.", "timestamp_range": "N/A", "evidence": "save twenty percent", "is_falsifiable": True, "severity": "low"},
-    ],
-    "overall_severity": "high",
-})
-
-MOCK_INVALID_JSON = "Here is my feedback: The hook is weak and the CTA is missing."
-
-
-@patch.dict(os.environ, {"ANTHROPIC_API_KEY": "sk-test"})
-@patch("viral_script_engine.agents.critic.anthropic.Anthropic")
-def test_critic_agent_returns_critique_output(mock_anthropic_cls):
-    mock_client = MagicMock()
-    mock_anthropic_cls.return_value = mock_client
-    mock_msg = MagicMock()
-    mock_msg.content = [MagicMock(text=MOCK_VALID_JSON)]
-    mock_client.messages.create.return_value = mock_msg
-
-    agent = CriticAgent()
+def test_critic_agent_returns_critique_output(mock_llm):
+    agent = CriticAgent(backend="qwen")
     result = agent.critique("Some script text here", "Mumbai", "Reels", "finance")
-
     assert len(result.claims) == 3
     assert result.claims[0].claim_id == "C1"
     assert result.overall_severity == "high"
-    assert mock_client.messages.create.call_count == 1
 
 
-@patch.dict(os.environ, {"ANTHROPIC_API_KEY": "sk-test"})
-@patch("viral_script_engine.agents.critic.anthropic.Anthropic")
-def test_critic_agent_retries_on_bad_json(mock_anthropic_cls):
-    mock_client = MagicMock()
-    mock_anthropic_cls.return_value = mock_client
-    bad_msg = MagicMock()
-    bad_msg.content = [MagicMock(text=MOCK_INVALID_JSON)]
-    good_msg = MagicMock()
-    good_msg.content = [MagicMock(text=MOCK_VALID_JSON)]
-    mock_client.messages.create.side_effect = [bad_msg, good_msg]
+def test_critic_agent_retries_on_bad_json(monkeypatch):
+    calls = []
+
+    def fake_generate(self, sys_prompt, usr_prompt, **kw):
+        calls.append(usr_prompt)
+        if len(calls) == 1:
+            return MOCK_INVALID_RESPONSE
+        return MOCK_VALID_RESPONSE
 
-    agent = CriticAgent()
+    monkeypatch.setattr(
+        "viral_script_engine.agents.llm_backend.LLMBackend.generate", fake_generate
+    )
+    agent = CriticAgent(backend="qwen")
     result = agent.critique("Some script text here", "Mumbai", "Reels", "finance")
     assert len(result.claims) == 3
-    assert mock_client.messages.create.call_count == 2
+    assert len(calls) == 2
 
 
-@patch.dict(os.environ, {"ANTHROPIC_API_KEY": "sk-test"})
-@patch("viral_script_engine.agents.critic.anthropic.Anthropic")
-def test_critic_agent_raises_parse_error_after_two_failures(mock_anthropic_cls):
-    mock_client = MagicMock()
-    mock_anthropic_cls.return_value = mock_client
-    bad_msg = MagicMock()
-    bad_msg.content = [MagicMock(text=MOCK_INVALID_JSON)]
-    mock_client.messages.create.return_value = bad_msg
-
-    agent = CriticAgent()
+def test_critic_agent_raises_parse_error_after_two_failures(monkeypatch):
+    monkeypatch.setattr(
+        "viral_script_engine.agents.llm_backend.LLMBackend.generate",
+        lambda self, sys_prompt, usr_prompt, **kw: MOCK_INVALID_RESPONSE,
+    )
+    agent = CriticAgent(backend="qwen")
     with pytest.raises(CriticParseError):
         agent.critique("Some script text here", "Mumbai", "Reels", "finance")
-    assert mock_client.messages.create.call_count == 2