refactor: migrate all LLM calls to Claude Agent SDK

.github/workflows/test.yaml

@@ -23,7 +23,6 @@ jobs:
           python-version: "3.11"
           cache: "pip"
           cache-dependency-path: |
-            libs/openant-core/requirements.txt
             libs/openant-core/pyproject.toml
 
       - name: Set up Node.js
@@ -35,7 +34,7 @@ jobs:
 
       - name: Install Python dependencies
         working-directory: libs/openant-core
-        run: pip install -r requirements.txt && pip install ".[dev]"
+        run: pip install -e ".[dev]"
 
       - name: Cache JS parser node_modules
         id: cache-node-modules
@@ -51,7 +50,7 @@ jobs:
 
       - name: Run Python and parser tests
         working-directory: libs/openant-core
-        run: python -m pytest tests/test_token_tracker.py tests/test_parser_adapter.py tests/test_python_parser.py tests/test_js_parser.py -v
+        run: python -m pytest tests/test_token_tracker.py tests/test_parser_adapter.py tests/test_python_parser.py tests/test_js_parser.py tests/test_declared_dependencies.py -v
 
   go-tests:
     name: Go build + integration (${{ matrix.os }})
@@ -79,7 +78,6 @@ jobs:
           python-version: "3.11"
           cache: "pip"
           cache-dependency-path: |
-            libs/openant-core/requirements.txt
             libs/openant-core/pyproject.toml
 
       - name: Set up Node.js
@@ -115,7 +113,7 @@ jobs:
 
       - name: Install Python dependencies
         working-directory: libs/openant-core
-        run: pip install -r requirements.txt && pip install ".[dev]"
+        run: pip install -e ".[dev]"
 
       - name: Cache JS parser node_modules
         id: cache-node-modules

libs/openant-core/README.md

@@ -24,7 +24,7 @@ git clone https://github.com/your-org/openant.git
 cd openant
 
 # Install Python dependencies
-pip install -r requirements.txt
+pip install -e .
 
 # Set API key
 echo "ANTHROPIC_API_KEY=your-key-here" > .env

libs/openant-core/context/application_context.py

@@ -29,9 +29,17 @@
 from pathlib import Path
 from typing import Any
 
-from anthropic import Anthropic
 from dotenv import load_dotenv
 
+# Ensure libs/openant-core is on sys.path so `utilities.*` imports resolve
+# regardless of how this module is loaded.
+_OPENANT_CORE_ROOT = str(Path(__file__).parent.parent)
+if _OPENANT_CORE_ROOT not in sys.path:
+    sys.path.insert(0, _OPENANT_CORE_ROOT)
+
+from utilities.model_config import MODEL_AUXILIARY  # noqa: E402
+from utilities.llm_client import AnthropicClient  # noqa: E402
+
 # Load environment variables
 load_dotenv()
 
@@ -462,7 +470,7 @@ def _build_type_descriptions() -> str:
 
 def generate_application_context(
     repo_path: Path,
-    model: str = "claude-sonnet-4-20250514",
+    model: str = MODEL_AUXILIARY,
     force_regenerate: bool = False,
 ) -> ApplicationContext:
     """Generate application context using LLM analysis.
@@ -503,19 +511,14 @@ def generate_application_context(
 
     # Call LLM
     print(f"Generating context with {model}...", file=sys.stderr)
-    client = Anthropic()
-    response = client.messages.create(
-        model=model,
+    # AnthropicClient is the SDK-backed wrapper; routes through the Claude
+    # Agent SDK so this works with both API keys and local Claude Code sessions.
+    client = AnthropicClient(model=model)
+    response_text = client.analyze_sync(
+        CONTEXT_GENERATION_PROMPT.format(sources=sources_text),
         max_tokens=2000,
-        messages=[{
-            "role": "user",
-            "content": CONTEXT_GENERATION_PROMPT.format(sources=sources_text)
-        }]
     )
 
-    # Parse response
-    response_text = response.content[0].text
-
     # Extract JSON from response
     json_match = re.search(r'```json\s*(.*?)\s*```', response_text, re.DOTALL)
     if json_match:

libs/openant-core/context/generate_context.py

@@ -14,6 +14,7 @@
 # Add parent directory to path for imports
 sys.path.insert(0, str(Path(__file__).parent.parent))
 
+from utilities.model_config import MODEL_AUXILIARY
 from context.application_context import (
     ApplicationType,
     APPLICATION_TYPE_INFO,
@@ -78,8 +79,8 @@ def main():
 
     parser.add_argument(
         "--model", "-m",
-        default="claude-sonnet-4-20250514",
-        help="Anthropic model to use (default: claude-sonnet-4-20250514)",
+        default=MODEL_AUXILIARY,
+        help=f"Anthropic model to use (default: {MODEL_AUXILIARY})",
     )
 
     parser.add_argument(

libs/openant-core/core/analyzer.py

@@ -313,7 +313,8 @@ def run_analysis(
     checkpoint.dir = checkpoint_path
 
     # Select model
-    model_id = "claude-opus-4-6" if model == "opus" else "claude-sonnet-4-20250514"
+    from utilities.model_config import MODEL_AUXILIARY, MODEL_PRIMARY
+    model_id = MODEL_PRIMARY if model == "opus" else MODEL_AUXILIARY
     print(f"[Analyze] Model: {model_id}", file=sys.stderr)
 
     # Initialize client

libs/openant-core/core/enhancer.py

@@ -50,7 +50,8 @@ def enhance_dataset(
     # Configure global rate limiter
     configure_rate_limiter(backoff_seconds=float(backoff_seconds))
 
-    model_id = "claude-sonnet-4-20250514" if model == "sonnet" else "claude-opus-4-6"
+    from utilities.model_config import MODEL_AUXILIARY, MODEL_PRIMARY
+    model_id = MODEL_AUXILIARY if model == "sonnet" else MODEL_PRIMARY
     print(f"[Enhance] Mode: {mode}", file=sys.stderr)
     print(f"[Enhance] Model: {model_id}", file=sys.stderr)
 

libs/openant-core/core/reporter.py

@@ -587,11 +587,12 @@ def _record_usage_in_tracker(usage: dict):
     """Record usage in the global TokenTracker so step_context captures it."""
     try:
         from utilities.llm_client import get_global_tracker
+        from utilities.model_config import MODEL_PRIMARY
         tracker = get_global_tracker()
         # Record as a single aggregated call
         if usage.get("total_tokens", 0) > 0:
             tracker.record_call(
-                model="claude-opus-4-6",
+                model=MODEL_PRIMARY,
                 input_tokens=usage["input_tokens"],
                 output_tokens=usage["output_tokens"],
             )

libs/openant-core/experiment.py

@@ -474,7 +474,8 @@ def run_experiment(
         Experiment results with metrics
     """
     # Select model
-    model_id = "claude-opus-4-20250514" if model == "opus" else "claude-sonnet-4-20250514"
+    from utilities.model_config import MODEL_AUXILIARY, MODEL_PRIMARY
+    model_id = MODEL_PRIMARY if model == "opus" else MODEL_AUXILIARY
     print(f"Using model: {model_id}")
     print(f"Enhanced context: {enhanced}")
     print(f"Context correction: {correct_context}")

libs/openant-core/generate_report.py

@@ -27,16 +27,26 @@
 import json
 import html
 import os
+import sys
 from datetime import datetime
+from pathlib import Path
 
-import anthropic
 from dotenv import load_dotenv
 
+# Ensure libs/openant-core is on sys.path so `utilities.*` imports resolve
+# regardless of the caller's working directory.
+_OPENANT_CORE_ROOT = str(Path(__file__).parent)
+if _OPENANT_CORE_ROOT not in sys.path:
+    sys.path.insert(0, _OPENANT_CORE_ROOT)
+
+from utilities.model_config import MODEL_AUXILIARY  # noqa: E402
+from utilities.llm_client import AnthropicClient  # noqa: E402
+
 # Load environment variables from .env file
 load_dotenv()
 
 
-REPORT_MODEL = "claude-sonnet-4-20250514"
+REPORT_MODEL = MODEL_AUXILIARY
 MAX_TOKENS = 4096
 
 
@@ -198,18 +208,10 @@ def generate_remediation_guidance(findings: list) -> str:
 {findings_text}
 """
 
-    api_key = os.getenv("ANTHROPIC_API_KEY")
-    if not api_key:
-        raise ValueError("ANTHROPIC_API_KEY not found in environment")
-
-    client = anthropic.Anthropic(api_key=api_key)
-    response = client.messages.create(
-        model=REPORT_MODEL,
-        max_tokens=MAX_TOKENS,
-        messages=[{"role": "user", "content": prompt}]
-    )
-
-    return response.content[0].text
+    # AnthropicClient is the SDK-backed wrapper; it handles auth (API key
+    # or local Claude Code session) and surfaces structured SDK errors.
+    client = AnthropicClient(model=REPORT_MODEL)
+    return client.analyze_sync(prompt, max_tokens=MAX_TOKENS)
 
 
 def _build_pipeline_costs_html(step_reports: list[dict]) -> str:

libs/openant-core/openant/cli.py

@@ -587,10 +587,9 @@ def cmd_report_data(args):
     and step reports — everything display-ready.
     """
     import html as html_mod
-    import anthropic
     from core.schemas import success, error
     from core.step_report import step_context
-    from utilities.llm_client import get_global_tracker
+    from utilities.llm_client import AnthropicClient, get_global_tracker
 
     results_path = args.results
     dataset_path = args.dataset
@@ -810,13 +809,10 @@ def cmd_report_data(args):
 {findings_text}
 """
                 print("[Report] Generating remediation guidance (LLM)...", file=sys.stderr)
-                client = anthropic.Anthropic()
-                response = client.messages.create(
-                    model="claude-sonnet-4-20250514",
-                    max_tokens=4096,
-                    messages=[{"role": "user", "content": prompt}],
-                )
-                remediation_html = response.content[0].text
+                from utilities.model_config import MODEL_AUXILIARY
+                # AnthropicClient handles usage tracking via the global TokenTracker.
+                remediation_client = AnthropicClient(model=MODEL_AUXILIARY)
+                remediation_html = remediation_client.analyze_sync(prompt, max_tokens=4096)
 
                 # Post-process: linkify finding references like #4, #12-#14
                 import re
@@ -825,15 +821,11 @@ def _linkify_finding(m):
                     return f'<a href="#finding-{num}" class="finding-ref">#{num}</a>'
                 remediation_html = re.sub(r'#(\d+)', _linkify_finding, remediation_html)
 
-                # Track usage
-                usage = response.usage
-                tracker = get_global_tracker()
-                tracker.record_call(
-                    model="claude-sonnet-4-20250514",
-                    input_tokens=usage.input_tokens,
-                    output_tokens=usage.output_tokens,
+                last = remediation_client.get_last_call() or {}
+                print(
+                    f"  Remediation cost: ${last.get('cost_usd', 0.0):.4f}",
+                    file=sys.stderr,
                 )
-                print(f"  Remediation cost: ${(usage.input_tokens / 1e6) * 3.0 + (usage.output_tokens / 1e6) * 15.0:.4f}", file=sys.stderr)
 
             # --- Step reports ---
             step_reports_data = []