AgentRunState

AgentRunState is the semantic step-level run state of the built-in step-driven ReAct strategy. It lives outside the workflow and is shared between AgentLoopState.run_state and the strategy (one instance, bridged by AGENT_ENTRY).

from amrita_core.builtins.agent.state import AgentRunState

state = AgentRunState()
state.begin_step("execute")

Fields

Field	Type	Meaning
step_index	int	Global step counter (1-based)
current_phase	str | None	The active phase — a DAG node id (or "execute" in simple mode)
plan	list[DAGNode] | None	The task DAG (None = no decomposition)
simple_mode	bool	True when the LLM decided to run directly
current_step_id	str | None	Id of the DAG node being executed
completed_step_ids	list[str]	Finished DAG node ids (dependency check)
plan_revision	int	update_step revision counter
step_tool_signatures	list[str]	Tool-call signatures in the current Step (stall window)
stall_injected	bool	Give-up prompt injected (once per Step)
last_summary	StepSummary | None	Subject-predicate summary of the previous Step
tokens	TokenBudget	Real API token accounting
exec_finished	bool	Strategy done calling tools → iteration loop ends

Methods

Method	Purpose
begin_step(phase)	Enter a new Step: advance counter, reset per-Step state
begin_node(node)	begin_step(node.id) + track current_step_id
record_tool_call(signature)	Record a tool signature in the current Step
is_stalled(threshold)	True when the last N signatures are identical
would_stall(signature, threshold)	True if recording this signature would trip the detector (pre-execution cancel)
next_ready_node()	Next DAG node in topological order (graphlib.TopologicalSorter; cyclic plans degrade gracefully)
complete_current_node()	Mark the current node done
all_plan_done()	True when every DAG node is completed

Related

• DAGNode — a plan sub-step
• StepSummary — subject-predicate summary
• TokenBudget — token accounting
• See Advanced → Step Loop for the full picture