feat(llm): add profile-based configuration system

Implement centralized LLM configuration via ~/.workflows/llm-config.yml with provider reuse and hierarchical parameter resolution.

- Add LLMConfigLoader for profile management with built-in defaults
- Update LLMCallInput to support profile field with backward compatibility
- Integrate LLMConfigLoader into ExecutionContext and AppContext
- Update schema.json to reflect new profile-based configuration
- Document profile-based configuration in README.md
- Add layer1 and agent tags to git-analyze-changes workflow

Profiles enable "right-sized" LLM selection similar to AWS instance types, supporting both cloud and local providers with inline parameter overrides.

Author: ibacalu

README.md

@@ -117,45 +117,86 @@ blocks:
 
 ### 🤖 LLM Integration
 
-Call LLM APIs directly from workflows with automatic retry logic and JSON schema validation:
+Call LLM APIs directly from workflows with automatic retry logic and JSON schema validation.
+
+**Profile-Based Configuration (Recommended):**
+
+Centralized LLM configuration via `~/.workflows/llm-config.yml`:
 
 ```yaml
-blocks:
-  - id: extract_data
-    type: LLMCall
-    inputs:
-      provider: openai
-      model: gpt-4o-mini
-      api_key: "{{secrets.OPENAI_API_KEY}}"
-      prompt: "Extract key information from: {{inputs.text}}"
-      response_schema:
-        type: object
-        required: [summary, confidence]
-        properties:
-          summary: {type: string}
-          confidence: {type: number, minimum: 0, maximum: 1}
-      max_retries: 3
+version: "1.0"
+
+providers:
+  cloud:
+    type: openai
+    api_key_secret: "OPENAI_API_KEY"
+
+  local:
+    type: openai
+    api_url: "http://127.0.0.1:1234/v1/chat/completions"
+
+profiles:
+  default:
+    provider: cloud
+    model: gpt-4o-mini
+    temperature: 0.7
+    max_tokens: 4000
+
+  local:
+    provider: local
+    model: ""
+    temperature: 0.8
+    max_tokens: 4000
+
+default_profile: default
 ```
 
-**Supported Providers:**
+Usage:
+
+```yaml
+# Profile-based
+- id: analyze
+  type: LLMCall
+  inputs:
+    profile: default
+    prompt: "Analyze: {{inputs.text}}"
+
+# Profile with overrides
+- id: summary
+  type: LLMCall
+  inputs:
+    profile: default
+    temperature: 0.1
+    prompt: "Summarize: {{inputs.text}}"
+```
 
-- **OpenAI** - API access with native Structured Outputs
-- **Anthropic** - API access for Claude models
-- **Gemini** - API access for Gemini models
-- **Ollama** - Local Ollama server API
-- **OpenAI-compatible** - LM Studio, vLLM, and other compatible endpoints
+**Direct Specification (Backward Compatible):**
+
+```yaml
+- id: extract_data
+  type: LLMCall
+  inputs:
+    provider: openai
+    model: gpt-4o-mini
+    api_key: "{{secrets.OPENAI_API_KEY}}"
+    prompt: "Extract: {{inputs.text}}"
+    response_schema:
+      type: object
+      required: [summary, confidence]
+      properties:
+        summary: {type: string}
+        confidence: {type: number, minimum: 0, maximum: 1}
+```
+
+**Supported Providers:** OpenAI, Anthropic, Gemini, Ollama, OpenAI-compatible (LM Studio, vLLM)
 
 **Key Features:**
 
-- ✅ **Native schema validation** - Send JSON Schema to API for guaranteed structure
-- ✅ **Automatic retry** - Exponential backoff with validation feedback loop
+- ✅ **Profile-based config** - Centralized provider/model management
+- ✅ **Schema validation** - Guaranteed JSON structure with automatic retry
 - ✅ **Token tracking** - Monitor usage and costs
 - ✅ **Secrets integration** - Secure API key management
 
-**For Local CLI Tools:**
-
-To use local Claude CLI or Gemini CLI instead of APIs, use the `llm-process` workflow template which wraps CLI commands.
-
 ## What Can You Do With It?
 
 ### Built-in Workflows

schema.json

@@ -615,23 +615,47 @@
                   "additionalProperties": false,
                   "description": "Inputs for LLMCall block",
                   "properties": {
+                    "profile": {
+                      "anyOf": [
+                        {
+                          "type": "string"
+                        },
+                        {
+                          "type": "null"
+                        }
+                      ],
+                      "default": null,
+                      "description": "Profile name from ~/.workflows/llm-config.yml (e.g., 'cloud', 'local', 'default'). If specified, provider/model are loaded from config. Mutually exclusive with direct provider/model specification.",
+                      "title": "Profile"
+                    },
                     "provider": {
                       "anyOf": [
                         {
                           "$ref": "#/$defs/LLMProvider"
                         },
                         {
                           "type": "string"
+                        },
+                        {
+                          "type": "null"
                         }
                       ],
-                      "default": "openai",
-                      "description": "LLM provider (enum or interpolation string)",
+                      "default": null,
+                      "description": "LLM provider (enum or interpolation string). Required if profile not specified. Ignored if profile specified.",
                       "title": "Provider"
                     },
                     "model": {
-                      "description": "Model name (e.g., gpt-4o, claude-3-5-sonnet-20241022, gemini-2.0-flash-exp)",
-                      "title": "Model",
-                      "type": "string"
+                      "anyOf": [
+                        {
+                          "type": "string"
+                        },
+                        {
+                          "type": "null"
+                        }
+                      ],
+                      "default": null,
+                      "description": "Model name (e.g., gpt-4o, claude-3-5-sonnet-20241022, gemini-2.0-flash-exp). Required if profile not specified. Can override profile model if both specified.",
+                      "title": "Model"
                     },
                     "prompt": {
                       "description": "User prompt to send to the LLM",
@@ -773,7 +797,6 @@
                     }
                   },
                   "required": [
-                    "model",
                     "prompt"
                   ],
                   "title": "LLMCallInput",
@@ -1357,25 +1380,49 @@
     },
     "LLMCallInput": {
       "additionalProperties": false,
-      "description": "Input model for LLMCall executor.\n\nAll inputs are pre-resolved by VariableResolver:\n- {{secrets.OPENAI_API_KEY}} \u2192 actual key value\n- {{inputs.prompt}} \u2192 actual prompt text\n- {{blocks.previous.outputs.data}} \u2192 actual data",
+      "description": "Input model for LLMCall executor.\n\nAll inputs are pre-resolved by VariableResolver:\n- {{secrets.OPENAI_API_KEY}} \u2192 actual key value\n- {{inputs.prompt}} \u2192 actual prompt text\n- {{blocks.previous.outputs.data}} \u2192 actual data\n\nConfiguration Modes:\n1. Profile-based (NEW): Specify profile name, optionally override parameters\n2. Direct specification (backward compatible): Specify provider + model directly\n\nExamples:\n    # Profile-based (recommended)\n    profile: quick\n    prompt: \"...\"\n\n    # Profile with overrides\n    profile: standard\n    temperature: 1.0\n    max_tokens: 8000\n    prompt: \"...\"\n\n    # Direct specification (backward compatible)\n    provider: openai\n    model: gpt-4o\n    api_key: \"{{secrets.OPENAI_API_KEY}}\"\n    prompt: \"...\"",
       "properties": {
+        "profile": {
+          "anyOf": [
+            {
+              "type": "string"
+            },
+            {
+              "type": "null"
+            }
+          ],
+          "default": null,
+          "description": "Profile name from ~/.workflows/llm-config.yml (e.g., 'cloud', 'local', 'default'). If specified, provider/model are loaded from config. Mutually exclusive with direct provider/model specification.",
+          "title": "Profile"
+        },
         "provider": {
           "anyOf": [
             {
               "$ref": "#/$defs/LLMProvider"
             },
             {
               "type": "string"
+            },
+            {
+              "type": "null"
             }
           ],
-          "default": "openai",
-          "description": "LLM provider (enum or interpolation string)",
+          "default": null,
+          "description": "LLM provider (enum or interpolation string). Required if profile not specified. Ignored if profile specified.",
           "title": "Provider"
         },
         "model": {
-          "description": "Model name (e.g., gpt-4o, claude-3-5-sonnet-20241022, gemini-2.0-flash-exp)",
-          "title": "Model",
-          "type": "string"
+          "anyOf": [
+            {
+              "type": "string"
+            },
+            {
+              "type": "null"
+            }
+          ],
+          "default": null,
+          "description": "Model name (e.g., gpt-4o, claude-3-5-sonnet-20241022, gemini-2.0-flash-exp). Required if profile not specified. Can override profile model if both specified.",
+          "title": "Model"
         },
         "prompt": {
           "description": "User prompt to send to the LLM",
@@ -1517,7 +1564,6 @@
         }
       },
       "required": [
-        "model",
         "prompt"
       ],
       "title": "LLMCallInput",

src/src/main.py

@@ -0,0 +1 @@
+print('hello world')

src/src/utils.py

@@ -0,0 +1 @@
+# Utility functions

src/workflows_mcp/context.py

@@ -12,6 +12,7 @@
 from .engine import ExecutionContext, WorkflowRegistry
 from .engine.checkpoint_store import CheckpointStore
 from .engine.executor_base import ExecutorRegistry
+from .engine.llm_config import LLMConfigLoader
 
 
 @dataclass
@@ -27,6 +28,7 @@ class AppContext:
     registry: WorkflowRegistry
     executor_registry: ExecutorRegistry
     checkpoint_store: CheckpointStore
+    llm_config_loader: LLMConfigLoader
     max_recursion_depth: int = 50  # Default recursion depth limit
 
     def create_execution_context(self) -> ExecutionContext:
@@ -39,6 +41,7 @@ def create_execution_context(self) -> ExecutionContext:
             workflow_registry=self.registry,
             executor_registry=self.executor_registry,
             checkpoint_store=self.checkpoint_store,
+            llm_config_loader=self.llm_config_loader,
             parent=None,
             workflow_stack=[],
             max_recursion_depth=self.max_recursion_depth,

src/workflows_mcp/engine/__init__.py

@@ -37,6 +37,7 @@
     executors_file,  # File operation executors
     executors_http,  # HTTP/REST API executor
     executors_interactive,  # Interactive executors
+    executors_llm,  # LLM call executor
     executors_state,  # JSON state executors
     executors_workflow,  # Workflow executor (ADR-008)
 )

src/workflows_mcp/engine/execution_context.py

@@ -21,6 +21,7 @@
     from .checkpoint_store import CheckpointStore
     from .execution import Execution
     from .executor_base import ExecutorRegistry
+    from .llm_config import LLMConfigLoader
     from .registry import WorkflowRegistry
     from .schema import WorkflowSchema
 
@@ -48,6 +49,7 @@ def __init__(
         workflow_registry: WorkflowRegistry,
         executor_registry: ExecutorRegistry,
         checkpoint_store: CheckpointStore,
+        llm_config_loader: LLMConfigLoader,
         parent: Execution | None = None,
         workflow_stack: list[str] | None = None,
         max_recursion_depth: int = 50,
@@ -59,13 +61,15 @@ def __init__(
             workflow_registry: Registry of workflow definitions
             executor_registry: Registry of block executors
             checkpoint_store: Store for checkpoints (pause/resume)
+            llm_config_loader: Loader for LLM profile configuration
             parent: Parent execution (for nested workflows)
             workflow_stack: Workflow execution stack (depth tracking)
             max_recursion_depth: Maximum allowed recursion depth (default: 50)
         """
         self.workflow_registry = workflow_registry
         self.executor_registry = executor_registry
         self.checkpoint_store = checkpoint_store
+        self.llm_config_loader = llm_config_loader
         self.parent = parent
         self.workflow_stack = workflow_stack or []
         self.max_recursion_depth = max_recursion_depth
@@ -107,6 +111,7 @@ def create_child_context(
             workflow_registry=self.workflow_registry,
             executor_registry=self.executor_registry,
             checkpoint_store=self.checkpoint_store,
+            llm_config_loader=self.llm_config_loader,
             parent=parent_execution,
             workflow_stack=self.workflow_stack + [workflow_name],
             max_recursion_depth=self.max_recursion_depth,