crewmaster

MODULE	DESCRIPTION
agents	Agent models for CrewMaster v2.0.0.
api	Public API for CrewMaster v2.0.0.
api_integration_test	Integration tests for CrewMaster v2.0.0 — full tracer bullet verification.
api_test	Tests for the public API: crewmaster.execute() and execute_stream().
cli	CrewMaster CLI — top-level command group.
collaboration	Collaboration protocol models for CrewMaster v2.0.0.
conversation	Conversation module for CrewMaster v2.0.0.
core
evaluation
execution	Execution runtime models for CrewMaster v2.0.0.
init_test
operations	Operations models for CrewMaster v2.0.0.
sandbox	CrewMaster v2.0.0 Sandbox — FastAPI demo application.
tools	Tool registry models for CrewMaster v2.0.0.

CLASS	DESCRIPTION
ExecutionError	Raised when execution of an ExecutionPlan fails.

FUNCTION	DESCRIPTION
execute	Execute an Operation (single-node or multi-node) and return the final output.
execute_stream	Execute an Operation with streaming output.

__all__ module-attribute

__all__ = ['core', 'execute', 'execute_stream', 'ExecutionError']

ExecutionError

Raised when execution of an ExecutionPlan fails.

execute async

execute(operation: Operation, context_store: ContextStore, default_runtime: RuntimeDriver, block_store: BlockStore | None = None, tool_registry: ToolRegistry | None = None, plan: ExecutionPlan | None = None, prompt_engine: PromptEngine | None = None, observer: ExecutionObserver | None = None, context_overrides: dict[str, Any] | None = None) -> Any

Execute an Operation (single-node or multi-node) and return the final output.

If no plan is provided, one is built from operation via :func:resolve_plan. The plan is then executed node by node in topological order. Each node is dispatched to its assigned runtime driver with :func:secure_retry_execution wrapping the call. The output of each node is passed as deps context to downstream nodes.

When a node has max_retrieval_rounds > 0, the request_capability system tool is injected into the toolset. If the agent invokes it, execution is paused, the ToolRegistry is queried via :meth:ToolRegistry.retrieve_by_description, matching capabilities are added to the toolset, and the agent is re-dispatched with the expanded toolset.

PARAMETER	DESCRIPTION
operation	The root Operation to execute. TYPE: Operation
context_store	ContextStore with domain instances for resolution. TYPE: ContextStore
default_runtime	Fallback RuntimeDriver when an AgentConfig does not specify its own. TYPE: RuntimeDriver
block_store	BlockStore for resolving block URIs. Required if plan is not pre-built. TYPE: BlockStore | None DEFAULT: None
tool_registry	Optional ToolRegistry for tool retrieval. TYPE: ToolRegistry | None DEFAULT: None
plan	An optional pre-built ExecutionPlan. If omitted, one is built from operation. TYPE: ExecutionPlan | None DEFAULT: None
prompt_engine	Optional PromptEngine for rendering prompts at execution time. When provided, each node's prompt is rendered via prompt_engine.compose() using the current per-node context. If None, falls back to disclosure.rendered_prompt. TYPE: PromptEngine | None DEFAULT: None
observer	Optional ExecutionObserver for hooking into the execution lifecycle (on_plan_start, on_node_start, etc.). TYPE: ExecutionObserver | None DEFAULT: None

RETURNS	DESCRIPTION
Any	The structured output of the final node in the plan (matches
Any	operation.produces).

RAISES	DESCRIPTION
PlanBuildError	If the plan cannot be built due to unresolved dependencies.
ExecutionError	If a node fails after exhausting retry attempts.

execute_stream async

execute_stream(operation: Operation, context_store: ContextStore, default_runtime: RuntimeDriver, block_store: BlockStore | None = None, tool_registry: ToolRegistry | None = None, plan: ExecutionPlan | None = None, prompt_engine: PromptEngine | None = None, observer: ExecutionObserver | None = None, context_overrides: dict[str, Any] | None = None) -> AsyncIterator[RuntimeStreamChunk]

Execute an Operation with streaming output.

Like :func:execute, but yields :class:RuntimeStreamChunk events as the plan progresses. Each node yields text_delta, tool_call, and tool_result events during its execution, followed by a node_complete chunk carrying the node's artifact. The final final chunk carries the last node's output.

When a node has max_retrieval_rounds > 0, the request_capability system tool is injected. If a tool_call chunk targets request_capability, it is intercepted and handled exactly as in :func:execute.

PARAMETER	DESCRIPTION
operation	The root Operation to execute. TYPE: Operation
context_store	ContextStore with domain instances for resolution. TYPE: ContextStore
default_runtime	Fallback RuntimeDriver. TYPE: RuntimeDriver
block_store	BlockStore for resolving block URIs. TYPE: BlockStore | None DEFAULT: None
tool_registry	Optional ToolRegistry for tool retrieval. TYPE: ToolRegistry | None DEFAULT: None
plan	Optional pre-built ExecutionPlan. TYPE: ExecutionPlan | None DEFAULT: None
prompt_engine	Optional PromptEngine for rendering prompts at execution time. Falls back to disclosure.rendered_prompt. TYPE: PromptEngine | None DEFAULT: None
observer	Optional ExecutionObserver for hooking into the execution lifecycle. TYPE: ExecutionObserver | None DEFAULT: None

YIELDS	DESCRIPTION
AsyncIterator[RuntimeStreamChunk]	RuntimeStreamChunk events during plan execution.