Module agent_loop 

Agent loop — the core execution engine for phi-core agents.

This module is split into focused sub-modules:

• [config] — Hook type aliases and AgentLoopConfig
• [core] — Entry points: agent_loop, agent_loop_continue
• [run] — Core turn engine (run_loop)
• [streaming] — LLM response streaming
• [tools] — Tool execution pipeline
• [parallel] — Evaluational parallelism (agent_loop_parallel)
• evaluation — Pluggable evaluation strategies for parallel branch selection
• [helpers] — Utilities (input filters, config derivation, etc.)

Re-exports

pub use evaluation::ElaborateEvaluation;

pub use evaluation::LlmJudgeEvaluation;

pub use evaluation::PickFirstEvaluation;

pub use evaluation::TokenEfficientEvaluation;

pub use evaluation::TransparentEvaluation;

Modules

evaluation

Pluggable evaluation strategies for agent_loop_parallel.

script_callback

Script-based callbacks — run shell or Python scripts as lifecycle hooks.

Structs

AgentLoopConfig

All static settings for a single [agent_loop] / [agent_loop_continue] call.

Functions

agent_loop

Start an agent loop with new prompt messages.

agent_loop_continue

Resume an agent loop from existing context without new prompts.

agent_loop_parallel

Run multiple agent loop configurations concurrently from a shared base context, evaluate the results with the supplied strategy, and return the selected outcome.

Type Aliases

AfterCompactionEndFn

Called after compaction completes.

AfterLoopFn

Called once after the entire agent loop ends (after AgentEnd is emitted).

AfterToolExecutionFn

Called after each tool call (after ToolExecutionEnd is emitted).

AfterToolExecutionUpdateFn

Called after each incremental tool update (after ToolExecutionUpdate is emitted).

AfterTurnFn

Called after each LLM turn (after TurnEnd is emitted).

BeforeCompactionStartFn

Called before compaction starts.

BeforeLoopFn

Called once before the entire agent loop begins (before AgentStart is emitted).

BeforeToolExecutionFn

Called before each tool call (before ToolExecutionStart is emitted).

BeforeToolExecutionUpdateFn

Called before each incremental tool update (before ToolExecutionUpdate is emitted).

BeforeTurnFn

Called before each LLM turn (before TurnStart is emitted).

ConvertToLlmFn

All hook types use Arc (shared ownership) so they can be cloned into closures and stored without lifetime complications. Box<dyn Fn> would suffice for single-owner cases but Arc makes it trivially cheap to share across async tasks. Converts AgentMessage[] → Message[] before each LLM call.

GetMessagesFn

Returns pending messages (steering interrupts or follow-up work) when polled.

HookFuture

Boxed-future return type used by all 0.9.0 async lifecycle hooks. T is the hook’s logical return value (often bool for veto-returning hooks or () for fire-and-forget hooks).

OnErrorFn

Called when the LLM returns StopReason::Error. Argument: the error message string.

TransformContextFn

Transforms the full context before convert_to_llm (for pruning, reordering, injection).