feat: Phase 9 Agentic Layer - Release v0.3.0

.env.example

@@ -53,5 +53,51 @@ RAG_INDEX_TYPE=hnsw
 RAG_HNSW_M=16
 RAG_HNSW_EF_CONSTRUCTION=64
 
+# =============================================================================
+# Agentic Layer Configuration (PydanticAI v1.48.0)
+# =============================================================================
+
+# Model Configuration
+# Model identifier format: "provider:model-name"
+# Supported providers:
+#   - anthropic: Claude models (claude-sonnet-4-5, claude-opus-4-5, etc.)
+#   - openai: GPT models (gpt-4o, gpt-4o-mini, etc.)
+#   - google-gla: Gemini models via Google AI Studio (gemini-2-5-flash, gemini-3-flash, gemini-3-pro)
+#   - google-vertex: Gemini models via Vertex AI (gemini-*) [requires GCP auth]
+AGENT_DEFAULT_MODEL=anthropic:claude-sonnet-4-5
+AGENT_FALLBACK_MODEL=openai:gpt-4o
+
+# API Keys (only one needed based on your chosen provider)
+ANTHROPIC_API_KEY=sk-ant-your-anthropic-api-key-here
+# OPENAI_API_KEY=sk-your-openai-api-key-here
+# GOOGLE_API_KEY=your-google-api-key-here  # For google-gla:* models
+
+# Gemini Extended Reasoning (optional)
+# Thinking mode for Gemini 2.5+ models (requires additional tokens)
+# Set a token budget (e.g., 2000-8000) or leave unset to disable
+# Recommended: 4000 tokens for complex agent planning tasks
+# AGENT_THINKING_BUDGET=4000
+
+# Model parameters
+AGENT_TEMPERATURE=0.1
+AGENT_MAX_TOKENS=4096
+
+# Execution settings
+AGENT_MAX_TOOL_CALLS=10
+AGENT_TIMEOUT_SECONDS=120
+AGENT_RETRY_ATTEMPTS=3
+AGENT_RETRY_DELAY_SECONDS=1.0
+
+# Session settings
+AGENT_SESSION_TTL_MINUTES=120
+AGENT_MAX_SESSIONS_PER_USER=5
+
+# Human-in-the-loop actions (JSON array format required for safe parsing)
+AGENT_REQUIRE_APPROVAL=["create_alias","archive_run"]
+AGENT_APPROVAL_TIMEOUT_MINUTES=60
+
+# Streaming
+AGENT_ENABLE_STREAMING=true
+
 # Frontend (Vite)
 VITE_API_BASE_URL=http://localhost:8123

.github/workflows/cd-release.yml

@@ -31,7 +31,10 @@ jobs:
       - uses: googleapis/release-please-action@v4
         id: release
         with:
-          token: ${{ secrets.GITHUB_TOKEN }}
+          # Use PAT to trigger CI workflows on release PRs
+          # GITHUB_TOKEN won't trigger workflows (GitHub security feature)
+          # If RELEASE_PAT is not set, falls back to GITHUB_TOKEN (CI won't auto-trigger)
+          token: ${{ secrets.RELEASE_PAT || secrets.GITHUB_TOKEN }}
           config-file: release-please-config.json
           manifest-file: .release-please-manifest.json
 

PRPs/PRP-10-agentic-layer.md

@@ -2,7 +2,8 @@
 
 **Feature**: INITIAL-10.md — Agentic Layer
 **Status**: Ready for Implementation
-**Confidence Score**: 7.5/10
+**Confidence Score**: 8.0/10
+**Last Updated**: 2026-02-01 (Post Phase-9 RAG Review)
 
 ---
 
@@ -65,7 +66,7 @@ This is the "Brain" layer that orchestrates tools from INITIAL-9 (RAG), Phase 5
   why: "Official PydanticAI docs - main reference"
 
 - url: https://ai.pydantic.dev/agents/
-  why: "Agent constructor, result_type, system_prompt, run/run_stream methods"
+  why: "Agent constructor, output_type, system_prompt, run/run_stream methods"
 
 - url: https://ai.pydantic.dev/tools/
   why: "@agent.tool decorator, RunContext, deps_type, tool parameters"
@@ -165,13 +166,14 @@ examples/agents/
 ### Known Gotchas & Library Quirks
 
 ```python
-# CRITICAL: PydanticAI model identifier format
-# Use "anthropic:claude-sonnet-4-20250514" NOT "claude-sonnet-4-20250514"
-agent = Agent(model="anthropic:claude-sonnet-4-20250514")
+# CRITICAL: PydanticAI model identifier format (updated Jan 2026)
+# Use "anthropic:claude-sonnet-4-5" NOT "claude-sonnet-4-5"
+# For production, pin specific version: "anthropic:claude-sonnet-4-5-20250929"
+agent = Agent(model="anthropic:claude-sonnet-4-5")
 
 # CRITICAL: deps_type must match RunContext generic parameter
 agent = Agent(
-    model="anthropic:claude-sonnet-4-20250514",
+    model="anthropic:claude-sonnet-4-5",
     deps_type=AgentDeps,  # Your dependency dataclass
 )
 
@@ -502,9 +504,12 @@ class WSEvent(BaseModel):
 ```yaml
 MODIFY: pyproject.toml
 ADD to dependencies:
-  - "pydantic-ai>=0.1.0"    # PydanticAI agent framework
-  - "anthropic>=0.40.0"      # Anthropic SDK for Claude
+  - "pydantic-ai>=1.48.0"    # PydanticAI agent framework (v1 stable, API guaranteed)
+  - "anthropic>=0.50.0"      # Anthropic SDK for Claude
   - "websockets>=13.0"       # WebSocket support (already in uvicorn[standard])
+
+NOTE: PydanticAI v1.0 was released Sept 2025 with API stability guarantee.
+      Current version is 1.48.0 (Jan 2026). Do NOT use 0.x versions.
 ```
 
 ### Task 2: Add Agent Settings to config.py
@@ -514,7 +519,7 @@ MODIFY: app/core/config.py
 ADD after RAG settings:
 
   # Agent LLM Configuration
-  agent_default_model: str = "anthropic:claude-sonnet-4-20250514"
+  agent_default_model: str = "anthropic:claude-sonnet-4-5"
   agent_fallback_model: str = "openai:gpt-4o"
   agent_temperature: float = 0.1
   agent_max_tokens: int = 4096
@@ -596,28 +601,41 @@ INCLUDE:
 CREATE: app/features/agents/tools/registry_tools.py
 TOOLS:
   - list_runs(ctx, filters) -> list[RunSummary]
+      # Wraps: RegistryService.list_runs(db, page, page_size, model_type, status, store_id, product_id)
   - compare_runs(ctx, run_id_a, run_id_b) -> CompareResult
+      # Wraps: RegistryService.compare_runs(db, run_id_a, run_id_b)
   - create_alias(ctx, alias_name, run_id) -> AliasResult
+      # Wraps: RegistryService.create_alias(db, AliasCreate(...))
+      # REQUIRES HUMAN APPROVAL
   - archive_run(ctx, run_id) -> ArchiveResult
+      # Wraps: RegistryService.update_run(db, run_id, RunUpdate(status=RunStatus.ARCHIVED))
+      # NOTE: No direct archive method - use update_run with ARCHIVED status
+      # REQUIRES HUMAN APPROVAL
 
 CREATE: app/features/agents/tools/backtesting_tools.py
 TOOLS:
   - run_backtest(ctx, model_type, config, store_id, product_id, n_splits) -> BacktestResult
+      # Wraps: BacktestingService.run_backtest(db, store_id, product_id, start_date, end_date, config)
 
 CREATE: app/features/agents/tools/forecasting_tools.py
 TOOLS:
   - list_models(ctx) -> list[ModelInfo]
+      # Returns available model types: naive, seasonal_naive, moving_average, lightgbm (if enabled)
 
 CREATE: app/features/agents/tools/rag_tools.py
 TOOLS:
-  - retrieve_context(ctx, query, top_k) -> list[RetrievedChunk]
+  - retrieve_context(ctx, query, top_k) -> list[ChunkResult]
+      # Wraps: RAGService.retrieve(db, RetrieveRequest(query=query, top_k=top_k))
+      # NOTE: RAG service uses retrieve() not retrieve_context()
   - format_citation(ctx, chunk) -> Citation
+      # Transforms ChunkResult to Citation schema
 
 CRITICAL for all tools:
   - Use @agent.tool decorator (not @agent.tool_plain) for db access
   - First param is RunContext[AgentDeps]
-  - Detailed docstrings for LLM schema
+  - Detailed docstrings for LLM schema (Google/numpy style supported)
   - Structured logging with timing
+  - Match actual service method signatures from Phase 5-9 implementations
 ```
 
 ### Task 8: Create Agent Definitions
@@ -736,11 +754,58 @@ ADD websocket: app.add_api_websocket_route("/agents/stream", websocket_stream)
 ```yaml
 CREATE: app/features/agents/tests/conftest.py
 FIXTURES:
-  - db_session: Async session with cleanup
+  - db_session: Async session with cleanup (follow registry/tests/conftest.py pattern)
   - client: AsyncClient with db override
-  - mock_anthropic: Mock Anthropic API responses
-  - sample_experiment_request: Test request
-  - sample_rag_request: Test request
+  - mock_pydantic_ai_agent: Mock PydanticAI Agent (see pattern below)
+  - sample_experiment_request: ExperimentRequest fixture
+  - sample_rag_request: RAGQueryRequest fixture
+  - sample_agent_session: AgentSession ORM fixture
+
+MOCK PATTERN (following rag/tests/conftest.py mock_embedding_service):
+```
+
+```python
+@pytest.fixture
+def mock_pydantic_ai_agent():
+    """Mock PydanticAI Agent for unit tests without LLM calls.
+
+    Follows the mock_embedding_service pattern from RAG tests.
+    Returns deterministic responses without API calls.
+    """
+    from unittest.mock import AsyncMock, MagicMock
+    from app.features.agents.schemas import ExperimentReport, RunSummary
+
+    # Create mock structured output
+    mock_report = ExperimentReport(
+        objective="Test objective",
+        methodology="Tested naive and seasonal_naive models",
+        experiments_run=2,
+        best_run=RunSummary(
+            run_id="test123",
+            model_type="seasonal_naive",
+            config={"season_length": 7},
+            metrics={"mae": 5.0, "smape": 10.0},
+        ),
+        baseline_comparison=None,
+        recommendation="Deploy seasonal_naive model",
+        approval_required=False,
+    )
+
+    # Mock result object
+    mock_result = MagicMock()
+    mock_result.output = mock_report
+    mock_result.usage.return_value = MagicMock(
+        input_tokens=100,
+        output_tokens=50,
+    )
+    mock_result.messages = []
+
+    # Mock agent
+    agent = MagicMock()
+    agent.run = AsyncMock(return_value=mock_result)
+    agent.run_stream = AsyncMock()
+
+    return agent
 ```
 
 ### Task 14: Create Unit Tests
@@ -792,9 +857,11 @@ MODIFY: .env.example
 ADD:
   # Agent Configuration
   ANTHROPIC_API_KEY=sk-ant-...
-  AGENT_DEFAULT_MODEL=anthropic:claude-sonnet-4-20250514
+  AGENT_DEFAULT_MODEL=anthropic:claude-sonnet-4-5
+  AGENT_FALLBACK_MODEL=openai:gpt-4o
   AGENT_MAX_TOOL_CALLS=10
   AGENT_TIMEOUT_SECONDS=120
+  AGENT_TEMPERATURE=0.1
 ```
 
 ---
@@ -899,22 +966,31 @@ python examples/agents/websocket_client.py
 
 ---
 
-## Confidence Score: 7.5/10
+## Confidence Score: 8.0/10
 
 **Strengths:**
-- PydanticAI has excellent documentation
-- Clear FastAPI integration patterns
-- Existing service patterns to follow
-- Tool integrations with existing modules
+- PydanticAI v1.x provides API stability guarantee (released Sept 2025)
+- Clear FastAPI integration patterns with excellent documentation
+- Existing service patterns from Registry/RAG/Backtesting to follow
+- Tool integrations with existing modules well-defined
+- Mock patterns established in RAG tests (mock_embedding_service)
 
 **Risks:**
-- PydanticAI is relatively new (versioning may change)
 - WebSocket streaming with tools is complex
-- LLM rate limits may affect tests
+- LLM rate limits may affect integration tests
 - Message history serialization edge cases
+- Tool execution ordering in multi-step workflows
 
 **Mitigations:**
-- Pin PydanticAI version in pyproject.toml
-- Comprehensive mocking for unit tests
-- Rate-limited integration tests
+- Pin PydanticAI version >=1.48.0 in pyproject.toml
+- Comprehensive mocking following RAG test patterns
+- Rate-limited integration tests with retry logic
 - JSONB for flexible message storage
+- Timeout handling with asyncio.wait_for
+
+**Changes Since Initial Review (2026-02-01):**
+- Updated PydanticAI from 0.1.0 to 1.48.0 (v1 stable)
+- Updated Claude model identifier to claude-sonnet-4-5 format
+- Added service method mapping notes to Task 7
+- Added mock_pydantic_ai_agent fixture pattern
+- Verified tool wrappers match actual service APIs

README.md

@@ -9,6 +9,7 @@ Portfolio-grade end-to-end retail demand forecasting system.
 - **Serving Layer**: Typed FastAPI endpoints (Pydantic v2 validation)
 - **Model Registry**: Run configs, metrics, artifacts, and data windows for reproducibility
 - **RAG Knowledge Base**: Postgres pgvector embeddings + evidence-grounded answers with citations
+- **Agentic Layer**: PydanticAI agents for autonomous experimentation and evidence-grounded Q&A with human-in-the-loop approval
 
 ## Quick Start
 
@@ -120,6 +121,8 @@ app/
 │   ├── forecasting/    # Model training, prediction, persistence
 │   ├── backtesting/    # Time-series CV, metrics, baseline comparisons
 │   ├── registry/       # Model run tracking, artifacts, deployment aliases
+│   ├── rag/            # pgvector embeddings, semantic search, citations
+│   ├── agents/         # PydanticAI agents (experiment, RAG assistant)
 │   ├── dimensions/     # Store/product discovery for LLM tool-calling
 │   ├── analytics/      # KPI aggregations and drilldown analysis
 │   └── jobs/           # Async-ready task orchestration
@@ -507,6 +510,85 @@ curl -X POST http://localhost:8123/rag/retrieve \
 - Markdown and OpenAPI chunking strategies
 - Configurable embedding dimensions
 
+### Agentic Layer
+
+- `POST /agents/sessions` - Create a new agent session
+- `GET /agents/sessions/{session_id}` - Get session status and details
+- `POST /agents/sessions/{session_id}/chat` - Send a message to the agent
+- `POST /agents/sessions/{session_id}/approve` - Approve or reject a pending action
+- `DELETE /agents/sessions/{session_id}` - Close a session
+- `WS /agents/stream` - WebSocket streaming endpoint for real-time responses
+
+**Agent Types:**
+
+1. **Experiment Orchestrator** (`agent_type: "experiment"`):
+   - Autonomous model experimentation workflow
+   - Runs backtests and compares configurations
+   - Recommends best model with human-in-the-loop approval
+
+2. **RAG Assistant** (`agent_type: "rag_assistant"`):
+   - Evidence-grounded documentation Q&A
+   - Citation-backed responses with confidence scoring
+   - "Insufficient evidence" detection to prevent hallucination
+
+**Example Create Session Request:**
+```bash
+curl -X POST http://localhost:8123/agents/sessions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "agent_type": "rag_assistant",
+    "initial_context": null
+  }'
+```
+
+**Example Chat Request:**
+```bash
+curl -X POST http://localhost:8123/agents/sessions/{session_id}/chat \
+  -H "Content-Type: application/json" \
+  -d '{
+    "message": "How does backtesting prevent data leakage?"
+  }'
+```
+
+**Features:**
+- PydanticAI v1.48.0 for structured, type-safe agent orchestration
+- Session management with PostgreSQL JSONB message history
+- Human-in-the-loop approval for sensitive actions (create_alias, archive_run)
+- WebSocket streaming for real-time token delivery
+- Token usage tracking and tool call auditing
+
+**Configuration:**
+```bash
+# Agent LLM Configuration
+# Model format: "provider:model-name" (e.g., anthropic:claude-sonnet-4-5)
+AGENT_DEFAULT_MODEL=anthropic:claude-sonnet-4-5
+AGENT_FALLBACK_MODEL=openai:gpt-4o
+AGENT_TEMPERATURE=0.1
+AGENT_MAX_TOKENS=4096
+
+# API Keys (set based on your chosen provider)
+ANTHROPIC_API_KEY=sk-ant-your-key
+# OPENAI_API_KEY=sk-your-key
+# GOOGLE_API_KEY=your-google-api-key  # For Gemini models
+
+# Execution Configuration
+AGENT_MAX_TOOL_CALLS=10
+AGENT_TIMEOUT_SECONDS=120
+AGENT_RETRY_ATTEMPTS=3
+AGENT_RETRY_DELAY_SECONDS=1.0
+
+# Session Configuration
+AGENT_SESSION_TTL_MINUTES=120
+AGENT_MAX_SESSIONS_PER_USER=5
+
+# Human-in-the-loop Configuration (JSON array format)
+AGENT_REQUIRE_APPROVAL=["create_alias","archive_run"]
+AGENT_APPROVAL_TIMEOUT_MINUTES=60
+
+# Streaming Configuration
+AGENT_ENABLE_STREAMING=true
+```
+
 ### Error Responses (RFC 7807)
 
 All error responses follow RFC 7807 Problem Details format with `Content-Type: application/problem+json`:

alembic/env.py

@@ -12,6 +12,7 @@
 from app.core.database import Base
 
 # Import all models for Alembic autogenerate detection
+from app.features.agents import models as agents_models  # noqa: F401
 from app.features.data_platform import models as data_platform_models  # noqa: F401
 from app.features.jobs import models as jobs_models  # noqa: F401
 from app.features.rag import models as rag_models  # noqa: F401