feat: migrate LLM backend from OpenAI gpt-5-nano to Claude Sonnet

.env.example

@@ -21,12 +21,10 @@ SLACK_TEST_CHANNEL_ID=
 # =============================================================================
 # LLM Configuration
 # =============================================================================
-# Default: OpenAI API
-# For local Ollama: set LLM_BASE_URL=http://localhost:11434/v1
+# Default: Anthropic Claude (claude-sonnet-4-6)
 
-OPENAI_API_KEY=sk-your-key-here
-LLM_MODEL_NAME=gpt-5-nano
-# LLM_BASE_URL=http://localhost:11434/v1
+ANTHROPIC_API_KEY=sk-ant-your-key-here
+LLM_MODEL_NAME=claude-sonnet-4-6
 #
 # Optional: override max output tokens for all agents
 LLM_MAX_TOKENS=4000

.github/workflows/claude-code-review.yml

@@ -22,7 +22,7 @@ jobs:
     permissions:
       contents: read
       pull-requests: read
-      issues: read
+      issues: write
       id-token: write
 
     steps:
@@ -37,6 +37,9 @@ jobs:
         with:
           anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
           github_token: ${{ github.token }}
+          track_progress: true
+          claude_args: |
+            --model claude-sonnet-4-6
           plugin_marketplaces: 'https://github.com/anthropics/claude-code.git'
           plugins: 'code-review@claude-code-plugins'
           prompt: '/code-review:code-review ${{ github.repository }}/pull/${{ github.event.pull_request.number }}'

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "leads-agent"
-version = "0.1.0"
-description = "Slack leads classifier using Socket Mode + local/OpenAI-compatible LLM."
+version = "0.2.0"
+description = "Slack leads classifier using Socket Mode + Anthropic Claude."
 readme = "README.md"
 requires-python = ">=3.11"
 license = { file = "LICENSE" }
@@ -12,7 +12,7 @@ dependencies = [
   "slack_sdk>=3.27",
   "pydantic>=2.6",
   "pydantic-settings>=2.2",
-  "pydantic-ai-slim[openai,duckduckgo]>=0.0.12",
+  "pydantic-ai-slim[anthropic,duckduckgo]>=0.0.12",
   "python-dotenv>=1.0",
   "typer>=0.12",
   "rich>=13.7",

src/leads_agent/agent.py

@@ -6,8 +6,8 @@
 from pydantic_ai import Agent
 from pydantic_ai.common_tools.duckduckgo import duckduckgo_search_tool
 from pydantic_ai.messages import ModelMessage
-from pydantic_ai.models.openai import OpenAIChatModel, OpenAIChatModelSettings
-from pydantic_ai.providers.openai import OpenAIProvider
+from pydantic_ai.models.anthropic import AnthropicModel, AnthropicModelSettings
+from pydantic_ai.providers.anthropic import AnthropicProvider
 
 from leads_agent.config import Settings
 from leads_agent.models import EnrichedLeadClassification, HubSpotLead, LeadClassification
@@ -77,33 +77,31 @@ def print_debug(self, verbose: bool = False) -> None:
 @overload
 def agent_factory(
     *,
-    llm_base_url: str,
     llm_model_name: str,
-    llm_api_key: str = "ollama",
+    llm_api_key: str,
     instructions: str | None = None,
     output_type: type[TOutput],
-    model_settings: OpenAIChatModelSettings,
+    model_settings: AnthropicModelSettings,
     extra_tools: tuple[Callable, ...] | None = None,
     use_duckduckgo_search: bool = False,
 ) -> Agent[None, TOutput]: ...
 
 
 def agent_factory(
     *,
-    llm_base_url: str,
     llm_model_name: str,
-    llm_api_key: str = "ollama",
+    llm_api_key: str,
     instructions: str | None = None,
     output_type: type[TOutput],
-    model_settings: OpenAIChatModelSettings,
+    model_settings: AnthropicModelSettings,
     extra_tools: tuple[Callable, ...] | None = None,
     use_duckduckgo_search: bool = False,
 ) -> Agent[None, TOutput]:
     """
     Create an agent in a consistent way across triage/research/scoring.
     """
-    provider = OpenAIProvider(base_url=llm_base_url, api_key=llm_api_key)
-    model = OpenAIChatModel(model_name=llm_model_name, provider=provider)
+    provider = AnthropicProvider(api_key=llm_api_key)
+    model = AnthropicModel(model_name=llm_model_name, provider=provider)
 
     tools: list[Any] = list(extra_tools) if extra_tools else []
     if use_duckduckgo_search:
@@ -136,37 +134,34 @@ def _usage_snapshot(result: Any) -> dict[str, Any]:
 def _create_triage_agent(settings: Settings, api_key: str) -> Agent[None, LeadClassification]:
     pm = get_prompt_manager()
     return agent_factory(
-        llm_base_url=settings.llm_base_url,
         llm_model_name=settings.llm_model_name,
         llm_api_key=api_key,
         instructions=pm.build_triage_prompt(),
         output_type=LeadClassification,
-        model_settings=OpenAIChatModelSettings(temperature=0.0, max_tokens=settings.llm_max_tokens),
+        model_settings=AnthropicModelSettings(temperature=0.0, max_tokens=settings.llm_max_tokens),
     )
 
 
 def _create_research_agent(settings: Settings, api_key: str) -> Agent[None, EnrichedLeadClassification]:
     pm = get_prompt_manager()
     return agent_factory(
-        llm_base_url=settings.llm_base_url,
         llm_model_name=settings.llm_model_name,
         llm_api_key=api_key,
         instructions=pm.build_research_prompt(),
         output_type=EnrichedLeadClassification,
-        model_settings=OpenAIChatModelSettings(temperature=0.0, max_tokens=settings.llm_max_tokens),
+        model_settings=AnthropicModelSettings(temperature=0.0, max_tokens=settings.llm_max_tokens),
         use_duckduckgo_search=True,
     )
 
 
 def _create_scoring_agent(settings: Settings, api_key: str) -> Agent[None, EnrichedLeadClassification]:
     pm = get_prompt_manager()
     return agent_factory(
-        llm_base_url=settings.llm_base_url,
         llm_model_name=settings.llm_model_name,
         llm_api_key=api_key,
         instructions=pm.build_scoring_prompt(),
         output_type=EnrichedLeadClassification,
-        model_settings=OpenAIChatModelSettings(temperature=0.0, max_tokens=settings.llm_max_tokens),
+        model_settings=AnthropicModelSettings(temperature=0.0, max_tokens=settings.llm_max_tokens),
     )
 
 
@@ -181,7 +176,7 @@ def classify_lead(
     Classify a HubSpot lead using a multi-stage pipeline:
     triage → (if promising) web research → (if promising) final 1–5 scoring.
     """
-    api_key = settings.openai_api_key.get_secret_value() if settings.openai_api_key else "ollama"
+    api_key = settings.anthropic_api_key.get_secret_value() if settings.anthropic_api_key else ""
 
     triage_agent = _create_triage_agent(settings, api_key)
     prompt = lead.to_prompt_text()
@@ -234,7 +229,7 @@ def _research_lead(
     max_searches: int = 4,
     return_debug: bool = False,
 ) -> EnrichedLeadClassification | tuple[EnrichedLeadClassification, list[ModelMessage], dict[str, Any]]:
-    api_key = settings.openai_api_key.get_secret_value() if settings.openai_api_key else "ollama"
+    api_key = settings.anthropic_api_key.get_secret_value() if settings.anthropic_api_key else ""
     research_agent = _create_research_agent(settings, api_key)
 
     email_domain = ""
@@ -309,7 +304,7 @@ def _score_lead(
     enriched: EnrichedLeadClassification | None,
     return_debug: bool = False,
 ) -> EnrichedLeadClassification | tuple[EnrichedLeadClassification, list[ModelMessage], dict[str, Any]]:
-    api_key = settings.openai_api_key.get_secret_value() if settings.openai_api_key else "ollama"
+    api_key = settings.anthropic_api_key.get_secret_value() if settings.anthropic_api_key else ""
     scoring_agent = _create_scoring_agent(settings, api_key)
 
     name = f"{lead.first_name or ''} {lead.last_name or ''}".strip()
@@ -345,7 +340,7 @@ def _score_lead(
 
 def triage_lead(settings: Settings, lead: HubSpotLead) -> LeadClassification:
     """Run only the triage stage — no research or scoring."""
-    api_key = settings.openai_api_key.get_secret_value() if settings.openai_api_key else "ollama"
+    api_key = settings.anthropic_api_key.get_secret_value() if settings.anthropic_api_key else ""
     agent = _create_triage_agent(settings, api_key)
     run = agent.run_sync(lead.to_prompt_text())
     return run.output

src/leads_agent/config.py

@@ -48,10 +48,9 @@ class Settings(BaseSettings):
     slack_channel_id: str | None = Field(default=None, validation_alias="SLACK_CHANNEL_ID")
     slack_test_channel_id: str | None = Field(default=None, validation_alias="SLACK_TEST_CHANNEL_ID")
 
-    # LLM (OpenAI by default; works with any OpenAI-compatible API)
-    llm_base_url: str = Field(default="https://api.openai.com/v1", validation_alias="LLM_BASE_URL")
-    llm_model_name: str = Field(default="gpt-5-nano", validation_alias="LLM_MODEL_NAME")
-    openai_api_key: SecretStr | None = Field(default=None, validation_alias="OPENAI_API_KEY")
+    # LLM (Anthropic Claude)
+    llm_model_name: str = Field(default="claude-sonnet-4-6", validation_alias="LLM_MODEL_NAME")
+    anthropic_api_key: SecretStr | None = Field(default=None, validation_alias="ANTHROPIC_API_KEY")
     llm_max_tokens: int = Field(default=4000, validation_alias="LLM_MAX_TOKENS")
 
     # Observability
@@ -127,8 +126,7 @@ def display_config():
     table.add_row("SLACK_APP_TOKEN", mask_secret(settings.slack_app_token))
     table.add_row("SLACK_CHANNEL_ID", settings.slack_channel_id or "[not set]")
     table.add_row("SLACK_TEST_CHANNEL_ID", settings.slack_test_channel_id or "[not set]")
-    table.add_row("OPENAI_API_KEY", mask_secret(settings.openai_api_key))
-    table.add_row("LLM_BASE_URL", settings.llm_base_url)
+    table.add_row("ANTHROPIC_API_KEY", mask_secret(settings.anthropic_api_key))
     table.add_row("LLM_MODEL_NAME", settings.llm_model_name)
     table.add_row("LLM_MAX_TOKENS", str(settings.llm_max_tokens))
     table.add_row("LOGFIRE_TOKEN", mask_secret(settings.logfire_token))

src/leads_agent/core/init_wizard.py

@@ -36,15 +36,15 @@ def init_wizard(output: Path, force: bool):
     )
 
     rprint("\n[bold]LLM Configuration[/]")
-    rprint("[dim]Default uses OpenAI; set LLM_BASE_URL for Ollama/other providers[/]\n")
+    rprint("[dim]Default uses Anthropic Claude (claude-sonnet-4-6)[/]\n")
 
-    openai_api_key = Prompt.ask(
-        "  [cyan]OPENAI_API_KEY[/]",
-        default="sk-...",
+    anthropic_api_key = Prompt.ask(
+        "  [cyan]ANTHROPIC_API_KEY[/]",
+        default="sk-ant-...",
     )
     llm_model_name = Prompt.ask(
         "  [cyan]LLM_MODEL_NAME[/]",
-        default="gpt-5-nano",
+        default="claude-sonnet-4-6",
     )
     llm_max_tokens = Prompt.ask(
         "  [cyan]LLM_MAX_TOKENS[/] [dim](max output tokens)[/]",
@@ -122,12 +122,10 @@ def init_wizard(output: Path, force: bool):
     env_lines.extend(
         [
             "",
-            "# LLM configuration (OpenAI by default)",
-            f"OPENAI_API_KEY={openai_api_key}",
+            "# LLM configuration (Anthropic Claude)",
+            f"ANTHROPIC_API_KEY={anthropic_api_key}",
             f"LLM_MODEL_NAME={llm_model_name}",
             f"LLM_MAX_TOKENS={llm_max_tokens}",
-            "# Uncomment for Ollama or other OpenAI-compatible providers:",
-            "# LLM_BASE_URL=http://localhost:11434/v1",
             "",
             "# Runtime",
             f"DRY_RUN={str(dry_run).lower()}",