@bluefly/openstandardagents 0.2.8 → 0.2.9

package/spec/v0.2.9/runtime-semantics.md

@@ -0,0 +1,464 @@
+# OSSA Runtime Semantics Specification
+
+**Version**: 0.2.9
+**Status**: Draft
+**Last Updated**: 2025-12-04
+
+This document defines the formal runtime semantics for OSSA-compliant agent execution. All OSSA runtime implementations MUST conform to these semantics.
+
+## 1. Turn Lifecycle
+
+An OSSA agent turn consists of 7 sequential phases:
+
+```
+┌─────────────────────────────────────────────────────────────────┐
+│                        TURN LIFECYCLE                            │
+├─────────────────────────────────────────────────────────────────┤
+│                                                                  │
+│  ┌──────────┐   ┌──────────┐   ┌──────────┐   ┌──────────┐     │
+│  │ 1. INIT  │──▶│ 2. NORM  │──▶│ 3. RESOL │──▶│ 4. INFER │     │
+│  │          │   │          │   │          │   │          │     │
+│  │ Generate │   │ Normalize│   │ Resolve  │   │ LLM Call │     │
+│  │ IDs      │   │ Message  │   │ Tools    │   │          │     │
+│  └──────────┘   └──────────┘   └──────────┘   └──────────┘     │
+│                                                      │          │
+│  ┌──────────┐   ┌──────────┐   ┌──────────┐         │          │
+│  │ 7. EMIT  │◀──│ 6. PERSIST│◀──│ 5. EXEC  │◀────────┘          │
+│  │          │   │          │   │          │                     │
+│  │ Observe  │   │ State    │   │ Tools    │                     │
+│  │          │   │          │   │          │                     │
+│  └──────────┘   └──────────┘   └──────────┘                     │
+│                                                                  │
+└─────────────────────────────────────────────────────────────────┘
+```
+
+### Phase 1: Initialization (INIT)
+
+**Purpose**: Generate runtime identifiers and establish context.
+
+```typescript
+interface InitPhase {
+  // MUST generate per spec.identity configuration
+  interaction_id: string;  // New UUID per turn
+  session_id: string;      // From context or new
+  instance_id: string;     // Stable per pod/process
+
+  // MUST extract from context
+  trace_context: W3CTraceContext | null;
+  baggage: W3CBaggage | null;
+
+  // MUST record
+  turn_number: number;
+  started_at: ISO8601Timestamp;
+}
+```
+
+**Invariants**:
+- `interaction_id` MUST be unique across all turns
+- `session_id` MUST be consistent within a conversation
+- `instance_id` MUST be stable for the agent instance lifetime
+
+### Phase 2: Message Normalization (NORM)
+
+**Purpose**: Transform input into canonical message format.
+
+```typescript
+interface NormalizedMessage {
+  role: 'user' | 'system' | 'assistant' | 'tool';
+  content: string | ContentPart[];
+  metadata: {
+    source: 'user' | 'api' | 'delegation';
+    original_format: string;
+    normalized_at: ISO8601Timestamp;
+  };
+}
+
+type ContentPart = TextPart | ImagePart | FilePart;
+```
+
+**Rules**:
+1. All inputs MUST be converted to `NormalizedMessage`
+2. Content MUST be sanitized per `spec.safety.content_filtering`
+3. Multi-modal content MUST be split into `ContentPart[]`
+4. Original format MUST be preserved in metadata
+
+### Phase 3: Tool Resolution (RESOL)
+
+**Purpose**: Resolve available tools from manifest and MCP servers.
+
+```typescript
+interface ResolvedTool {
+  name: string;
+  description: string;
+  input_schema: JSONSchema;
+  source: ToolSource;
+  transport: 'function' | 'http' | 'mcp' | 'grpc';
+  timeout_ms: number;
+  retry_policy: RetryPolicy;
+}
+
+type ToolSource =
+  | { type: 'manifest'; tool_index: number }
+  | { type: 'mcp'; server_uri: string; capability: string }
+  | { type: 'delegation'; agent_id: string };
+```
+
+**Rules**:
+1. Tools MUST be resolved from `spec.tools[]` array
+2. MCP tools MUST be discovered via `tools/list` call
+3. Tool schemas MUST be validated against JSON Schema
+4. Unavailable tools MUST be excluded with warning
+
+### Phase 4: LLM Inference (INFER)
+
+**Purpose**: Execute LLM call with resolved context.
+
+```typescript
+interface InferenceRequest {
+  model: string;                    // From spec.llm.model
+  messages: NormalizedMessage[];    // Conversation history
+  tools: ResolvedTool[];            // Available tools
+  system_prompt: string;            // From spec.role or spec.prompts
+  parameters: LLMParameters;        // From spec.llm.parameters
+}
+
+interface InferenceResponse {
+  content: string | null;
+  tool_calls: ToolCallRequest[];
+  finish_reason: 'stop' | 'length' | 'tool_use' | 'content_filter';
+  usage: TokenUsage;
+  model: string;                    // Actual model used
+}
+```
+
+**Rules**:
+1. System prompt MUST include `spec.role`
+2. Few-shot examples from `spec.prompts.few_shot_examples` MUST be prepended
+3. Token limits from `spec.constraints` MUST be enforced
+4. Timeout from `spec.constraints.timeout_seconds` MUST be enforced
+
+### Phase 5: Tool Execution (EXEC)
+
+**Purpose**: Execute requested tool calls.
+
+```typescript
+interface ToolExecution {
+  call_id: string;
+  tool_name: string;
+  input: unknown;
+  started_at: ISO8601Timestamp;
+  completed_at: ISO8601Timestamp;
+  result: ToolResult;
+  retries: number;
+}
+
+type ToolResult =
+  | { status: 'success'; output: unknown }
+  | { status: 'error'; error: ToolError }
+  | { status: 'timeout'; partial_output?: unknown };
+```
+
+**Execution Order**:
+1. Tool calls MAY execute in parallel if `spec.tools[].parallel: true`
+2. By default, tools execute sequentially in request order
+3. Failed tools MUST be retried per `spec.reliability.retry`
+4. Circuit breaker MUST trip after threshold failures
+
+**Rules**:
+1. Each tool call MUST have unique `call_id`
+2. Tool input MUST validate against tool's `input_schema`
+3. Tool output MUST be serializable to JSON
+4. Sensitive outputs MUST be redacted per `spec.security.output_filtering`
+
+### Phase 6: State Persistence (PERSIST)
+
+**Purpose**: Persist state changes atomically.
+
+```typescript
+interface StatePersistence {
+  mode: 'none' | 'session' | 'persistent';
+  storage: StateStorage;
+  changes: StateChange[];
+  transaction_id: string;
+}
+
+interface StateChange {
+  key: string;
+  previous_value: unknown | null;
+  new_value: unknown;
+  operation: 'set' | 'delete' | 'append';
+}
+```
+
+**Transaction Semantics**:
+1. All state changes within a turn MUST be atomic
+2. On failure, all changes MUST be rolled back
+3. State MUST be isolated per `spec.state.isolation_level`:
+   - `read_uncommitted`: No isolation
+   - `read_committed`: See only committed state
+   - `serializable`: Full isolation (default)
+
+### Phase 7: Observability Emission (EMIT)
+
+**Purpose**: Emit telemetry data.
+
+```typescript
+interface ObservabilityEmission {
+  // Tracing
+  span: OTelSpan;
+
+  // Metrics
+  metrics: {
+    tokens_input: number;
+    tokens_output: number;
+    latency_ms: number;
+    tool_calls: number;
+    errors: number;
+  };
+
+  // Logging
+  log_entries: LogEntry[];
+
+  // Activity Stream (if enabled)
+  event?: CloudEvent;
+}
+```
+
+**Rules**:
+1. Span MUST include all attributes from semantic-conventions.md
+2. Metrics MUST NOT include high-cardinality attributes
+3. Logs MUST be correlated via `trace_id` and `span_id`
+4. CloudEvents MUST conform to spec/v0.2.9/cloudevents.md
+
+---
+
+## 2. Error Handling
+
+### Error Code Taxonomy
+
+```typescript
+type ErrorCode =
+  // Input Errors (4xx equivalent)
+  | 'VALIDATION_ERROR'      // Invalid input format
+  | 'SCHEMA_VIOLATION'      // Input doesn't match schema
+  | 'CONTENT_FILTERED'      // Content blocked by safety
+  | 'POLICY_VIOLATION'      // Violates agent policy
+
+  // Execution Errors (5xx equivalent)
+  | 'TOOL_ERROR'            // Tool execution failed
+  | 'TOOL_TIMEOUT'          // Tool exceeded timeout
+  | 'LLM_ERROR'             // LLM provider error
+  | 'LLM_TIMEOUT'           // LLM exceeded timeout
+  | 'STATE_ERROR'           // State persistence failed
+
+  // Resource Errors
+  | 'RATE_LIMITED'          // Rate limit exceeded
+  | 'QUOTA_EXCEEDED'        // Token/request quota exceeded
+  | 'CIRCUIT_OPEN'          // Circuit breaker tripped
+
+  // Agent Errors
+  | 'MAX_TURNS_EXCEEDED'    // Hit spec.constraints.max_turns
+  | 'MAX_TOKENS_EXCEEDED'   // Hit spec.constraints.max_tokens
+  | 'DELEGATION_FAILED'     // Child agent failed
+  | 'ESCALATION_REQUIRED';  // Needs human intervention
+
+interface OSSAError {
+  code: ErrorCode;
+  message: string;
+  details?: Record<string, unknown>;
+  recoverable: boolean;
+  retry_after_ms?: number;
+  escalation?: EscalationAction;
+}
+```
+
+### Recovery Strategies
+
+```typescript
+type RecoveryStrategy =
+  | 'retry'           // Retry with backoff
+  | 'fallback'        // Use fallback value/tool
+  | 'skip'            // Skip and continue
+  | 'escalate'        // Escalate to human/supervisor
+  | 'abort';          // Abort turn/session
+
+interface ErrorHandler {
+  error_codes: ErrorCode[];
+  strategy: RecoveryStrategy;
+  max_retries?: number;
+  fallback_value?: unknown;
+  escalation_target?: string;
+}
+```
+
+### Default Error Handling
+
+| Error Code | Default Strategy | Retries |
+|------------|------------------|---------|
+| `VALIDATION_ERROR` | abort | 0 |
+| `TOOL_ERROR` | retry | 3 |
+| `TOOL_TIMEOUT` | retry | 2 |
+| `LLM_ERROR` | retry | 3 |
+| `LLM_TIMEOUT` | retry | 2 |
+| `RATE_LIMITED` | retry | 3 (with backoff) |
+| `CIRCUIT_OPEN` | fallback | 0 |
+| `MAX_TURNS_EXCEEDED` | escalate | 0 |
+
+---
+
+## 3. State Semantics
+
+### Isolation Levels
+
+```typescript
+type IsolationLevel =
+  | 'read_uncommitted'  // Can see uncommitted changes
+  | 'read_committed'    // Only see committed changes
+  | 'repeatable_read'   // Consistent reads within turn
+  | 'serializable';     // Full isolation (default)
+```
+
+### Transaction Boundaries
+
+```
+Turn Start ─────────────────────────────────────────── Turn End
+     │                                                      │
+     │  ┌─────────────────────────────────────────────┐    │
+     │  │           TRANSACTION BOUNDARY               │    │
+     │  │                                             │    │
+     │  │  state.get()  ──▶  compute  ──▶  state.set()│    │
+     │  │                                             │    │
+     │  │  ALL changes atomic                         │    │
+     │  │  ROLLBACK on any failure                    │    │
+     │  │                                             │    │
+     │  └─────────────────────────────────────────────┘    │
+     │                                                      │
+     └──────────────────────────────────────────────────────┘
+```
+
+### State Operations
+
+```typescript
+interface StateOperations {
+  // Read operations
+  get<T>(key: string): Promise<T | null>;
+  getMany<T>(keys: string[]): Promise<Map<string, T>>;
+  exists(key: string): Promise<boolean>;
+
+  // Write operations (within transaction)
+  set<T>(key: string, value: T, ttl?: number): Promise<void>;
+  delete(key: string): Promise<void>;
+  append<T>(key: string, value: T): Promise<void>;
+
+  // Transaction control
+  checkpoint(): Promise<string>;  // Save point
+  rollback(checkpoint?: string): Promise<void>;
+}
+```
+
+### Rollback Behavior
+
+1. **Automatic Rollback**: On unhandled error, all state changes roll back
+2. **Partial Rollback**: Can rollback to named checkpoint
+3. **No Rollback**: `isolation_level: read_uncommitted` disables rollback
+
+---
+
+## 4. Multi-Turn Behavior
+
+### Turn Chaining
+
+```typescript
+interface TurnChain {
+  // Continue if LLM requests tool use
+  continue_on_tool_use: boolean;  // default: true
+
+  // Max consecutive tool turns before response required
+  max_tool_turns: number;  // default: 10
+
+  // Require user confirmation between turns
+  confirm_between_turns: boolean;  // default: false
+}
+```
+
+### Conversation Memory
+
+```typescript
+interface ConversationMemory {
+  // How many messages to include in context
+  window_size: number;  // default: 20
+
+  // Token budget for history
+  max_history_tokens: number;  // default: 4000
+
+  // Summarization strategy when window exceeded
+  summarization: 'truncate' | 'summarize' | 'sliding';
+}
+```
+
+---
+
+## 5. Delegation Semantics
+
+### Agent-to-Agent Handoff
+
+```typescript
+interface Delegation {
+  type: 'handoff' | 'parallel' | 'supervisor';
+  target_agent: string;
+  input: unknown;
+  context_propagation: ContextPropagation;
+  timeout_ms: number;
+  on_failure: 'retry' | 'fallback' | 'escalate';
+}
+
+interface ContextPropagation {
+  // What to pass to child agent
+  include_session_id: boolean;   // default: true
+  include_history: boolean;      // default: false
+  include_state: boolean;        // default: false
+  custom_context: Record<string, unknown>;
+}
+```
+
+### Delegation Patterns
+
+1. **Handoff**: Transfer control completely to target agent
+2. **Parallel**: Execute multiple agents concurrently
+3. **Supervisor**: Parent monitors child execution
+
+---
+
+## 6. Compliance Requirements
+
+### MUST Requirements
+
+1. Runtime MUST generate unique `interaction_id` per turn
+2. Runtime MUST enforce `spec.constraints.max_turns`
+3. Runtime MUST enforce `spec.constraints.timeout_seconds`
+4. Runtime MUST emit OpenTelemetry spans per semantic-conventions.md
+5. Runtime MUST rollback state on unhandled errors
+6. Runtime MUST validate tool inputs against schemas
+
+### SHOULD Requirements
+
+1. Runtime SHOULD support all isolation levels
+2. Runtime SHOULD implement circuit breaker pattern
+3. Runtime SHOULD support W3C trace context propagation
+4. Runtime SHOULD emit CloudEvents for activity streams
+
+### MAY Requirements
+
+1. Runtime MAY implement custom error handlers
+2. Runtime MAY support additional state backends
+3. Runtime MAY implement conversation summarization
+
+---
+
+## References
+
+- [OSSA Schema v0.2.9](./ossa-0.2.9.schema.json)
+- [Semantic Conventions](./semantic-conventions.md)
+- [OpenTelemetry Specification](https://opentelemetry.io/docs/specs/)
+- [CloudEvents Specification](https://cloudevents.io/)
+- [W3C Trace Context](https://www.w3.org/TR/trace-context/)