soe-ai 0.1.1__py3-none-any.whl → 0.1.3__py3-none-any.whl

soe/docs/primitives/signals.md

@@ -0,0 +1,865 @@
+
+# Appendix C: Signals Reference
+
+This appendix covers everything about signals in SOE: how they trigger nodes, how they're emitted, and the condition evaluation rules for each node type.
+
+All examples are tested. Run them with:
+
+```bash
+uv run pytest tests/test_cases/appendix/c_signals/ -v
+```
+
+---
+
+## What Are Signals?
+
+Signals are the **communication mechanism** between nodes in SOE. They:
+
+- **Trigger** nodes via `event_triggers`
+- **Route** execution flow between nodes
+- **Propagate** to parent workflows in sub-orchestration
+
+Think of signals like events in a pub/sub system. Nodes subscribe to signals they care about, and emit signals to notify other nodes.
+
+---
+
+## Signal Naming Best Practices
+
+### Use Descriptive, Action-Oriented Names
+
+```yaml
+# ✅ Good - clear intent
+event_emissions:
+  - signal_name: ANALYSIS_COMPLETE
+  - signal_name: VALIDATION_FAILED
+  - signal_name: USER_VERIFIED
+
+# ❌ Bad - vague or generic
+event_emissions:
+  - signal_name: DONE
+  - signal_name: NEXT
+  - signal_name: STEP_2
+```
+
+### Prefix Workflow-Specific Signals
+
+When using sub-orchestration, prefix signals with the workflow name for clarity:
+
+```yaml
+# Parent workflow
+signals_to_parent: [ANALYSIS_COMPLETE]  # Clear origin
+
+# Instead of
+signals_to_parent: [DONE]  # What finished? Unclear to parent
+```
+
+### Use Consistent Conventions
+
+| Convention | Example | Use Case |
+|------------|---------|----------|
+| `*_COMPLETE` | `ANALYSIS_COMPLETE` | Successful completion |
+| `*_FAILED` | `VALIDATION_FAILED` | Error states |
+| `*_READY` | `DATA_READY` | Readiness signals |
+| `*_REQUIRED` | `REVIEW_REQUIRED` | Action needed |
+
+---
+
+## The `event_emissions` Field
+
+Every node type can emit signals via `event_emissions`:
+
+```yaml
+event_emissions:
+  - signal_name: SUCCESS
+  - signal_name: NEEDS_REVIEW
+    condition: "{{ result.confidence < 0.8 }}"
+```
+
+### Fields
+
+| Field | Type | Required | Description |
+|-------|------|----------|-------------|
+| `signal_name` | `str` | Yes | The signal to emit |
+| `condition` | `str` | No | Controls when the signal is emitted |
+
+### No Signals (Terminal Node)
+
+A node with empty or missing `event_emissions` is a **terminal node**—it executes but emits nothing:
+
+```yaml
+LogAndFinish:
+  node_type: tool
+  event_triggers: [COMPLETE]
+  tool_name: log_result
+  input_fields: [result]
+  # No event_emissions = terminal node
+```
+
+Use this pattern for:
+- Final logging/cleanup nodes
+- Fire-and-forget operations
+- Workflow endpoints
+
+---
+
+## Condition Types: The Three Modes
+
+The `condition` field has **three modes** that determine when and how signals are emitted:
+
+### Mode 1: No Condition (Unconditional)
+
+Signals without conditions **always emit** after node execution:
+
+```yaml
+example_workflow:
+  TriggerMultiple:
+    node_type: router
+    event_triggers: [START]
+    event_emissions:
+      - signal_name: PROCESSING_DONE
+      - signal_name: LOG_EVENT
+```
+
+Both `PROCESSING_DONE` and `LOG_EVENT` emit every time the node runs.
+
+> **Note**: For Router nodes, multiple unconditional signals all emit simultaneously (fan-out pattern). For LLM/Agent nodes with multiple signals, the LLM must select one - use Jinja conditions like `{{ true }}` if you want all signals to emit.
+
+### Mode 2: Jinja Template (Programmatic)
+
+Conditions containing `{{ }}` are evaluated programmatically:
+
+```yaml
+example_workflow:
+  AnalyzeAndRoute:
+    node_type: llm
+    event_triggers: [START]
+    prompt: "Analyze: &#123;&#123; context.text &#125;&#125;"
+    output_field: analysis
+    event_emissions:
+      - signal_name: HIGH_PRIORITY
+        condition: "&#123;&#123; context.priority > 5 &#125;&#125;"
+      - signal_name: NORMAL_PRIORITY
+        condition: "&#123;&#123; context.priority <= 5 &#125;&#125;"
+```
+
+SOE evaluates `context.priority > 5` and emits the matching signal. No LLM involvement.
+
+### Mode 3: Plain Text (Semantic/LLM Selection)
+
+Plain text conditions trigger LLM signal selection:
+
+```yaml
+example_workflow:
+  SentimentRouter:
+    node_type: llm
+    event_triggers: [START]
+    prompt: "Analyze the sentiment of this message: &#123;&#123; context.message &#125;&#125;"
+    output_field: sentiment_analysis
+    event_emissions:
+      - signal_name: POSITIVE_SENTIMENT
+        condition: "The message expresses happiness, satisfaction, or positive emotions"
+      - signal_name: NEGATIVE_SENTIMENT
+        condition: "The message expresses anger, frustration, or negative emotions"
+      - signal_name: NEUTRAL_SENTIMENT
+        condition: "The message is factual, neutral, or emotionally ambiguous"
+```
+
+SOE asks the LLM: "Based on your analysis, which signal should be emitted?" The LLM chooses based on semantic understanding.
+
+---
+
+## How SOE Decides: The Decision Tree
+
+The behavior depends on the node type:
+
+### Router Node
+
+```
+
+┌─────────────────────────────────────────────────────────────┐
+│                  Router Signal Emission                      │
+├─────────────────────────────────────────────────────────────┤
+│                                                              │
+│  For EACH signal in event_emissions:                         │
+│     └─ No condition? → Emit                                  │
+│     └─ Jinja condition? → Evaluate, emit if truthy           │
+│                                                              │
+│  → Multiple signals can emit (fan-out pattern)               │
+│                                                              │
+└─────────────────────────────────────────────────────────────┘
+
+```
+
+### LLM/Agent Node
+
+```
+
+┌─────────────────────────────────────────────────────────────┐
+│               LLM/Agent Signal Emission                      │
+├─────────────────────────────────────────────────────────────┤
+│                                                              │
+│  1. Any condition contains {{ }}?                            │
+│     └─ YES → Evaluate ALL conditions programmatically        │
+│              Emit signals where condition is truthy          │
+│     └─ NO → Continue to step 2                               │
+│                                                              │
+│  2. Count signals                                            │
+│     └─ Zero signals? → Nothing emitted                       │
+│     └─ Single signal? → Emit unconditionally                 │
+│     └─ Multiple signals?                                     │
+│         └─ LLM selects ONE signal                            │
+│            (uses conditions as semantic descriptions)        │
+│                                                              │
+└─────────────────────────────────────────────────────────────┘
+
+```
+
+### Critical Rule: Jinja Takes Over
+
+If **any** condition contains `{{ }}`, ALL conditions are evaluated programmatically. The LLM never selects signals when Jinja is present.
+
+---
+
+## Signal Behavior by Node Type
+
+### Router Node
+
+**Purpose**: Conditional branching based on context.
+
+**Condition Context**: `context` only.
+
+**Evaluation**: Always programmatic (Jinja). Plain text conditions are not LLM-selected.
+
+```yaml
+example_workflow:
+  ValidateInput:
+    node_type: router
+    event_triggers: [START]
+    event_emissions:
+      - signal_name: HAS_DATA
+        condition: "&#123;&#123; context.data is defined and context.data &#125;&#125;"
+      - signal_name: NO_DATA
+        condition: "&#123;&#123; context.data is not defined or not context.data &#125;&#125;"
+
+  ProcessData:
+    node_type: router
+    event_triggers: [HAS_DATA]
+    event_emissions:
+      - signal_name: DONE
+
+  HandleMissing:
+    node_type: router
+    event_triggers: [NO_DATA]
+    event_emissions:
+      - signal_name: DONE
+```
+
+**Key Point**: Router conditions must use Jinja. Plain text won't work as expected.
+
+---
+
+### LLM Node
+
+**Purpose**: Single LLM call with optional signal selection.
+
+**Condition Context**: `context` only (LLM output stored in `output_field`).
+
+**Evaluation**: Jinja → programmatic. Plain text → LLM selection.
+
+**Jinja Example** (SOE evaluates):
+
+```yaml
+example_workflow:
+  AnalyzeAndRoute:
+    node_type: llm
+    event_triggers: [START]
+    prompt: "Analyze: &#123;&#123; context.text &#125;&#125;"
+    output_field: analysis
+    event_emissions:
+      - signal_name: HIGH_PRIORITY
+        condition: "&#123;&#123; context.priority > 5 &#125;&#125;"
+      - signal_name: NORMAL_PRIORITY
+        condition: "&#123;&#123; context.priority <= 5 &#125;&#125;"
+```
+
+**Plain Text Example** (LLM selects):
+
+```yaml
+example_workflow:
+  SentimentRouter:
+    node_type: llm
+    event_triggers: [START]
+    prompt: "Analyze the sentiment of this message: &#123;&#123; context.message &#125;&#125;"
+    output_field: sentiment_analysis
+    event_emissions:
+      - signal_name: POSITIVE_SENTIMENT
+        condition: "The message expresses happiness, satisfaction, or positive emotions"
+      - signal_name: NEGATIVE_SENTIMENT
+        condition: "The message expresses anger, frustration, or negative emotions"
+      - signal_name: NEUTRAL_SENTIMENT
+        condition: "The message is factual, neutral, or emotionally ambiguous"
+```
+
+**LLM Selection Mechanism**: SOE adds a `selected_signal` field to the response model, forcing the LLM to choose from the options. The condition text serves as the description.
+
+---
+
+### Agent Node
+
+**Purpose**: ReAct loop with tool access.
+
+**Condition Context**: `context` only.
+
+**Evaluation**: Same as LLM node—Jinja → programmatic, plain text → LLM selection.
+
+```yaml
+example_workflow:
+  TaskAgent:
+    node_type: agent
+    event_triggers: [START]
+    prompt: "Help the user with their request: &#123;&#123; context.request &#125;&#125;"
+    available_tools: [search, calculate]
+    output_field: result
+    event_emissions:
+      - signal_name: TASK_COMPLETED
+        condition: "The task was successfully completed with a satisfactory result"
+      - signal_name: TASK_FAILED
+        condition: "The task could not be completed due to limitations or errors"
+      - signal_name: TASK_NEEDS_CLARIFICATION
+        condition: "The request is ambiguous and needs more information from the user"
+```
+
+**Note**: Agent tools are selected by the LLM within the ReAct loop. Signal selection is a separate decision made after the agent loop completes.
+
+---
+
+### Tool Node
+
+**Purpose**: Direct function execution.
+
+**Condition Context**: `result` AND `context`.
+
+**Evaluation**: Always programmatic. No LLM selection (tools don't call LLMs).
+
+```yaml
+example_workflow:
+  ProcessPayment:
+    node_type: tool
+    event_triggers: [START]
+    tool_name: process_payment
+    context_parameter_field: payment_data
+    output_field: payment_result
+    event_emissions:
+      - signal_name: PAYMENT_APPROVED
+        condition: "&#123;&#123; result.status == 'approved' &#125;&#125;"
+      - signal_name: PAYMENT_DECLINED
+        condition: "&#123;&#123; result.status == 'declined' &#125;&#125;"
+      - signal_name: PAYMENT_PENDING
+        condition: "&#123;&#123; result.status == 'pending' &#125;&#125;"
+
+  OnApproved:
+    node_type: router
+    event_triggers: [PAYMENT_APPROVED]
+    event_emissions:
+      - signal_name: DONE
+
+  OnDeclined:
+    node_type: router
+    event_triggers: [PAYMENT_DECLINED]
+    event_emissions:
+      - signal_name: DONE
+
+  OnPending:
+    node_type: router
+    event_triggers: [PAYMENT_PENDING]
+    event_emissions:
+      - signal_name: DONE
+```
+
+#### The `result` Keyword
+
+Tool nodes have a special `result` variable in condition evaluation:
+
+| Variable | Description | Example |
+|----------|-------------|---------|
+| `result` | The return value of the tool function | `{{ result.status }}` |
+| `context` | The execution context | `{{ context.user_id }}` |
+
+**Combining result and context**:
+
+```yaml
+example_workflow:
+  CheckOrder:
+    node_type: tool
+    event_triggers: [START]
+    tool_name: validate_order
+    context_parameter_field: order
+    output_field: validation
+    event_emissions:
+      - signal_name: VIP_LARGE_ORDER
+        condition: "&#123;&#123; result.valid and context.customer.is_vip and context.order.total > 1000 &#125;&#125;"
+      - signal_name: VIP_ORDER
+        condition: "&#123;&#123; result.valid and context.customer.is_vip &#125;&#125;"
+      - signal_name: LARGE_ORDER
+        condition: "&#123;&#123; result.valid and context.order.total > 1000 &#125;&#125;"
+      - signal_name: STANDARD_ORDER
+        condition: "&#123;&#123; result.valid &#125;&#125;"
+      - signal_name: INVALID_ORDER
+        condition: "&#123;&#123; not result.valid &#125;&#125;"
+```
+
+**Why `result`?**: Tools return values that aren't stored in context until after condition evaluation. The `result` keyword provides access to the raw return value.
+
+---
+
+### Child Node
+
+**Purpose**: Sub-orchestration.
+
+**Condition Context**: `context` only.
+
+**Evaluation**: Always programmatic.
+
+**Note**: Child node `event_emissions` fire after the child workflow **starts**, not when it completes. Use `signals_to_parent` to get completion signals.
+
+---
+
+## Failure Signals
+
+LLM, Agent, and Tool nodes can emit **failure signals** when they fail after exhausting retries. This enables graceful error handling.
+
+### LLM Failure Signal
+
+```yaml
+example_workflow:
+  RiskyLLMCall:
+    node_type: llm
+    event_triggers: [START]
+    prompt: "Generate a complex response for: &#123;&#123; context.input &#125;&#125;"
+    output_field: response
+    retries: 2
+    llm_failure_signal: LLM_FAILED
+    event_emissions:
+      - signal_name: SUCCESS
+
+  HandleSuccess:
+    node_type: router
+    event_triggers: [SUCCESS]
+    event_emissions:
+      - signal_name: WORKFLOW_COMPLETE
+
+  HandleFailure:
+    node_type: router
+    event_triggers: [LLM_FAILED]
+    event_emissions:
+      - signal_name: WORKFLOW_COMPLETE
+```
+
+When the LLM fails after 2 retries (3 total attempts), `LLM_FAILED` is emitted instead of `SUCCESS`.
+
+### Agent Failure Signal
+
+```yaml
+example_workflow:
+  ComplexAgent:
+    node_type: agent
+    event_triggers: [START]
+    prompt: "Complete this complex task: &#123;&#123; context.task &#125;&#125;"
+    available_tools: [search]
+    output_field: result
+    retries: 1
+    llm_failure_signal: AGENT_EXHAUSTED
+    event_emissions:
+      - signal_name: TASK_DONE
+
+  OnSuccess:
+    node_type: router
+    event_triggers: [TASK_DONE]
+    event_emissions:
+      - signal_name: DONE
+
+  OnFailure:
+    node_type: router
+    event_triggers: [AGENT_EXHAUSTED]
+    event_emissions:
+      - signal_name: DONE
+```
+
+### Tool Failure Signal (Registry-Based)
+
+Unlike LLM/Agent nodes, tool failure signals are configured in the **tools registry**, not in the YAML workflow:
+
+```yaml
+example_workflow:
+  CallExternalAPI:
+    node_type: tool
+    event_triggers: [START]
+    tool_name: flaky_api
+    context_parameter_field: api_params
+    output_field: api_result
+    event_emissions:
+      - signal_name: API_SUCCESS
+
+  OnSuccess:
+    node_type: router
+    event_triggers: [API_SUCCESS]
+    event_emissions:
+      - signal_name: DONE
+
+  OnFailure:
+    node_type: router
+    event_triggers: [API_FAILED]
+    event_emissions:
+      - signal_name: DONE
+```
+
+The `failure_signal` is configured when creating the tools registry:
+
+```python
+tools_registry = {
+    "flaky_api": {
+        "function": flaky_api,
+        "max_retries": 2,           # Retry up to 2 times after initial failure
+        "failure_signal": "API_FAILED",  # Emit when all retries exhausted
+    }
+}
+```
+
+When `flaky_api` throws an exception and exhausts all retries, `API_FAILED` is emitted.
+
+### Failure Signal Behavior
+
+| Field | Node Type | Location | Description |
+|-------|-----------|----------|-------------|
+| `llm_failure_signal` | LLM, Agent | YAML workflow | Signal emitted when LLM call fails after retries |
+| `retries` | LLM, Agent | YAML workflow | Number of retry attempts (default: 3) |
+| `failure_signal` | Tool | Tools registry | Signal emitted when tool fails after retries |
+| `max_retries` | Tool | Tools registry | Number of retry attempts (default: 1) |
+
+**Key Points**:
+- Failure signals are **only emitted** if configured
+- They replace normal `event_emissions` on failure
+- Use them to create error handling branches in your workflow
+- Without a failure signal, failures raise exceptions
+- **Tool failure signals** are in the registry because tools are Python functions with no YAML config
+
+---
+
+## Sub-Orchestration Signal Propagation
+
+### The `signals_to_parent` Field
+
+Controls which child signals propagate to the parent workflow:
+
+```yaml
+parent_workflow:
+  StartAnalysis:
+    node_type: child
+    event_triggers: [START]
+    child_workflow_name: analysis_child
+    child_initial_signals: [BEGIN]
+    input_fields: [data_to_analyze]
+    signals_to_parent: [ANALYSIS_SUCCESS, ANALYSIS_FAILED]
+    context_updates_to_parent: [analysis_result]
+
+  OnSuccess:
+    node_type: router
+    event_triggers: [ANALYSIS_SUCCESS]
+    event_emissions:
+      - signal_name: PARENT_DONE
+
+  OnFailure:
+    node_type: router
+    event_triggers: [ANALYSIS_FAILED]
+    event_emissions:
+      - signal_name: PARENT_DONE
+
+analysis_child:
+  Analyze:
+    node_type: llm
+    event_triggers: [BEGIN]
+    prompt: "Analyze this data: &#123;&#123; context.data_to_analyze &#125;&#125;"
+    output_field: analysis_result
+    event_emissions:
+      - signal_name: ANALYSIS_SUCCESS
+        condition: "Analysis completed successfully"
+      - signal_name: ANALYSIS_FAILED
+        condition: "Analysis could not be completed"
+```
+
+Only `ANALYSIS_SUCCESS` and `ANALYSIS_FAILED` reach the parent. Other child signals stay internal.
+
+### Best Practices for Sub-Orchestration Signals
+
+#### 1. Use Specific, Workflow-Prefixed Names
+
+```yaml
+# ✅ Good - parent knows exactly what completed
+signals_to_parent: [ANALYZER_COMPLETE, ANALYZER_FAILED]
+
+# ❌ Bad - ambiguous in parent context
+signals_to_parent: [DONE, ERROR]
+```
+
+#### 2. Keep Internal Signals Internal
+
+Don't propagate signals that only matter within the child.
+
+#### 3. Consider the Parent's Perspective
+
+The parent workflow should receive signals that are actionable.
+
+---
+
+## The `context_updates_to_parent` Field
+
+Controls which context keys are synced to the parent:
+
+```yaml
+SpawnAnalyzer:
+  node_type: child
+  context_updates_to_parent: [analysis_result, confidence_score]
+```
+
+When the child updates these keys, they're automatically copied to the parent's context.
+
+---
+
+## LLM Signal Selection: Under the Hood
+
+When the LLM selects a signal, SOE:
+
+1. **Builds a response model** with a `selected_signal` field:
+   ```python
+   class Response(BaseModel):
+       response: str
+       selected_signal: Literal["POSITIVE", "NEGATIVE", "NEUTRAL"]
+   ```
+
+2. **Provides descriptions** from the `condition` field:
+   ```
+   Select one of these signals based on your response:
+   - POSITIVE: The message expresses positive sentiment
+   - NEGATIVE: The message expresses negative sentiment
+   - NEUTRAL: The message is neutral
+   ```
+
+3. **Extracts the selection** and emits that signal.
+
+This is why plain-text conditions are called "semantic"—the LLM understands the description and makes a judgment call.
+
+---
+
+## Common Patterns
+
+### Exclusive Routing
+
+Only one path taken based on mutually exclusive conditions:
+
+```yaml
+example_workflow:
+  RouteByType:
+    node_type: router
+    event_triggers: [START]
+    event_emissions:
+      - signal_name: TYPE_A
+        condition: "&#123;&#123; context.type == 'a' &#125;&#125;"
+      - signal_name: TYPE_B
+        condition: "&#123;&#123; context.type == 'b' &#125;&#125;"
+      - signal_name: TYPE_DEFAULT
+        condition: "&#123;&#123; context.type not in ['a', 'b'] &#125;&#125;"
+
+  HandleA:
+    node_type: router
+    event_triggers: [TYPE_A]
+    event_emissions:
+      - signal_name: DONE
+
+  HandleB:
+    node_type: router
+    event_triggers: [TYPE_B]
+    event_emissions:
+      - signal_name: DONE
+
+  HandleDefault:
+    node_type: router
+    event_triggers: [TYPE_DEFAULT]
+    event_emissions:
+      - signal_name: DONE
+```
+
+### Fan-Out (Multiple Signals)
+
+Multiple signals can emit simultaneously, triggering parallel processing:
+
+```yaml
+example_workflow:
+  TriggerMultiple:
+    node_type: router
+    event_triggers: [START]
+    event_emissions:
+      - signal_name: NOTIFY_USER
+        condition: "&#123;&#123; context.notify_user &#125;&#125;"
+      - signal_name: LOG_EVENT
+        condition: "&#123;&#123; context.log_enabled &#125;&#125;"
+      - signal_name: UPDATE_METRICS
+
+  NotifyUser:
+    node_type: router
+    event_triggers: [NOTIFY_USER]
+    event_emissions:
+      - signal_name: NOTIFICATION_SENT
+
+  LogEvent:
+    node_type: router
+    event_triggers: [LOG_EVENT]
+    event_emissions:
+      - signal_name: EVENT_LOGGED
+
+  UpdateMetrics:
+    node_type: router
+    event_triggers: [UPDATE_METRICS]
+    event_emissions:
+      - signal_name: METRICS_UPDATED
+```
+
+---
+
+## Complete Example: Order Processing
+
+This workflow demonstrates multiple signal patterns working together:
+
+```yaml
+order_processing:
+  # 1. Router with Jinja conditions (programmatic)
+  ValidateOrder:
+    node_type: router
+    event_triggers: [START]
+    event_emissions:
+      - signal_name: ORDER_VALID
+        condition: "&#123;&#123; context.order['line_items']|length > 0 and context.order.total > 0 &#125;&#125;"
+      - signal_name: ORDER_INVALID
+        condition: "&#123;&#123; context.order['line_items']|length == 0 or context.order.total <= 0 &#125;&#125;"
+
+  # 2. Tool with result conditions
+  ProcessPayment:
+    node_type: tool
+    event_triggers: [ORDER_VALID]
+    tool_name: charge_card
+    context_parameter_field: payment_info
+    output_field: payment_result
+    event_emissions:
+      - signal_name: PAYMENT_SUCCESS
+        condition: "&#123;&#123; result.charged == true &#125;&#125;"
+      - signal_name: PAYMENT_FAILED
+        condition: "&#123;&#123; result.charged == false &#125;&#125;"
+
+  # 3. LLM with failure signal and unconditional emission
+  GenerateConfirmation:
+    node_type: llm
+    event_triggers: [PAYMENT_SUCCESS]
+    prompt: "Generate a friendly order confirmation for order #&#123;&#123; context.order.id &#125;&#125;"
+    output_field: confirmation_message
+    retries: 2
+    llm_failure_signal: CONFIRMATION_FAILED
+    event_emissions:
+      - signal_name: ORDER_COMPLETE
+
+  # 4. Fan-out: notify multiple systems
+  NotifySystems:
+    node_type: router
+    event_triggers: [ORDER_COMPLETE]
+    event_emissions:
+      - signal_name: NOTIFY_CUSTOMER
+      - signal_name: UPDATE_INVENTORY
+      - signal_name: LOG_ORDER
+
+  # Handle failures
+  HandleInvalidOrder:
+    node_type: router
+    event_triggers: [ORDER_INVALID]
+    event_emissions:
+      - signal_name: WORKFLOW_ERROR
+
+  HandlePaymentFailed:
+    node_type: router
+    event_triggers: [PAYMENT_FAILED]
+    event_emissions:
+      - signal_name: WORKFLOW_ERROR
+
+  HandleConfirmationFailed:
+    node_type: router
+    event_triggers: [CONFIRMATION_FAILED]
+    event_emissions:
+      - signal_name: ORDER_COMPLETE
+```
+
+**Patterns demonstrated**:
+1. **Jinja conditions** (Router): `ORDER_VALID` vs `ORDER_INVALID`
+2. **Tool result conditions**: `PAYMENT_SUCCESS` vs `PAYMENT_FAILED`
+3. **Failure signal**: `CONFIRMATION_FAILED` on LLM error
+4. **Fan-out**: `NOTIFY_CUSTOMER`, `UPDATE_INVENTORY`, `LOG_ORDER` emit together
+
+---
+
+## Debugging Signal Issues
+
+### Signal Not Emitting
+
+1. **Check condition syntax**: Is the Jinja valid?
+2. **Check condition logic**: Does it evaluate to truthy?
+3. **Check for typos**: Signal names are case-sensitive.
+
+### Wrong Signal Emitting
+
+1. **Jinja vs plain text**: Did you mean LLM selection but used Jinja?
+2. **Missing condition**: Signals without conditions always emit.
+3. **Multiple matches**: Multiple conditions can be truthy simultaneously.
+
+### Jinja Attribute Access Gotcha
+
+When accessing dict keys in Jinja, be careful with keys that conflict with dict methods:
+
+```yaml
+# ❌ BAD: 'items' conflicts with dict.items() method
+condition: "{{ context.order.items|length > 0 }}"
+
+# ✅ GOOD: Use bracket notation for conflicting keys
+condition: "{{ context.order['items']|length > 0 }}"
+
+# ✅ GOOD: Rename to avoid conflicts
+condition: "{{ context.order.line_items|length > 0 }}"
+```
+
+**Conflicting dict method names to avoid**: `items`, `keys`, `values`, `get`, `pop`, `update`
+
+### Parent Not Receiving Signal
+
+1. **Check `signals_to_parent`**: Is the signal listed?
+2. **Check signal name match**: Exact string match required.
+3. **Verify child emitted**: Debug the child workflow first.
+
+---
+
+## Summary Table
+
+| Node Type | Condition Context | LLM Selection? | Jinja Support | Failure Signal |
+|-----------|-------------------|----------------|---------------|----------------|
+| Router | `context` | No | Yes | No |
+| LLM | `context` | Yes (plain text) | Yes | `llm_failure_signal` |
+| Agent | `context` | Yes (plain text) | Yes | `llm_failure_signal` |
+| Tool | `result`, `context` | No | Yes | Registry-based |
+| Child | `context` | No | Yes | No |
+
+---
+
+## Key Takeaways
+
+1. **No condition** = signal always emits
+2. **Jinja condition** = programmatic evaluation by SOE
+3. **Plain text condition** = semantic selection by LLM (LLM/Agent nodes only)
+4. **`result` keyword** = tool return value (Tool nodes only)
+5. **`signals_to_parent`** = controls sub-orchestration signal propagation
+6. **Failure signals** = error handling for LLM/Agent nodes
+7. Use **descriptive signal names** for maintainability
+8. **Jinja takes over**: If any condition has `{{ }}`, all are evaluated programmatically