Struct RunConfig 

pub struct RunConfig {

    pub streaming_mode: StreamingMode,
    pub tool_confirmation_decisions: HashMap<String, ToolConfirmationDecision>,
    pub tool_confirmation_fingerprints: HashMap<String, String>,
    pub tool_confirmation_handler: Option<Arc<dyn ToolConfirmationHandler>>,
    pub runtime_toolsets: Vec<RuntimeToolset>,
    pub cached_content: Option<String>,
    pub transfer_targets: Vec<String>,
    pub parent_agent: Option<String>,
    pub auto_cache: bool,
    pub history_max_events: Option<usize>,
    pub tool_concurrency: ToolConcurrencyConfig,
    pub record_payloads: bool,
    pub trace_payload_max_bytes: usize,
    pub max_transfer_depth: Option<u32>,
}

Configuration for a single agent run.

Controls streaming behavior, tool confirmation, caching, transfer targets, and concurrency settings. Use RunConfig::builder() to construct from

Fields

streaming_mode: StreamingMode

The streaming mode for agent responses.

tool_confirmation_decisions: HashMap<String, ToolConfirmationDecision>

Static confirmation decisions for the current run, keyed by function call ID.

The ID is the one reported on ToolConfirmationRequest::function_call_id, so a decision authorizes the exact call it was requested for. A decision under a tool name is not consulted, because one name can cover materially different calls — a delete_file approval for a scratch path must not authorize a call that targets a different path.

Use tool_confirmation_fingerprints to additionally bind a decision to the arguments it was granted for. For name-wide or policy-driven decisions, supply a tool_confirmation_handler instead.

tool_confirmation_fingerprints: HashMap<String, String>

Optional argument binding for entries in tool_confirmation_decisions, keyed by the same function call ID.

The value is the fingerprint produced by tool_call_fingerprint for the call the decision was granted for. When an entry is present and the actual call does not match it, the decision is ignored and the call is treated as unconfirmed — the safe direction. Use this when a decision travels through an untrusted round trip, such as a browser, where the arguments could be changed while the call ID is replayed.

tool_confirmation_handler: Option<Arc<dyn ToolConfirmationHandler>>

Optional live decision source for confirmations that have no static entry in tool_confirmation_decisions.

runtime_toolsets: Vec<RuntimeToolset>

Toolsets made available only for this invocation.

cached_content: Option<String>

Optional cached content name for automatic prompt caching. When set by the runner’s cache lifecycle manager, agents should attach this name to their GenerateContentConfig so the LLM provider can reuse cached system instructions and tool definitions.

transfer_targets: Vec<String>

Valid agent names this agent can transfer to (parent, peers, children). Set by the runner when invoking agents in a multi-agent tree. When non-empty, the transfer_to_agent tool is injected and validation uses this list instead of only checking sub_agents.

parent_agent: Option<String>

The name of the parent agent, if this agent was invoked via transfer. Used by the agent to apply disallow_transfer_to_parent filtering.

auto_cache: bool

Enable automatic prompt caching for all providers that support it.

When true (the default), the runner enables provider-level caching:

• Anthropic: sets prompt_caching = true on the config
• Bedrock: sets prompt_caching = Some(BedrockCacheConfig::default())
• OpenAI / DeepSeek: no action needed (caching is automatic)
• Gemini: handled separately via ContextCacheConfig

history_max_events: Option<usize>

Maximum number of recent persisted events to load at the start of a run.

None preserves the previous behavior and loads the full session history. Set this for chat surfaces that already summarize older turns and need predictable startup latency.

tool_concurrency: ToolConcurrencyConfig

Tool concurrency configuration controlling parallel tool dispatch limits, per-tool overrides, and backpressure behavior.

The default (ToolConcurrencyConfig::default()) imposes no limits, preserving backward compatibility with the previous max_tool_concurrency: None.

record_payloads: bool

Whether tracing spans may include full request, response, and tool payloads when the record-payloads crate feature is enabled.

trace_payload_max_bytes: usize

Maximum serialized bytes recorded for tracing payload fields when full payload recording is disabled.

max_transfer_depth: Option<u32>

Maximum number of agent-to-agent transfers allowed in a single run.

Prevents infinite transfer loops when agents transfer back and forth. Defaults to 10 when None.

Implementations

impl RunConfig

pub fn builder() -> RunConfigBuilder

Creates a new RunConfigBuilder initialized with default values.

Use the builder to construct a RunConfig when struct literal syntax

Example

use adk_core::{RunConfig, StreamingMode};

let config = RunConfig::builder()
    .streaming_mode(StreamingMode::None)
    .auto_cache(false)
    .build();

assert_eq!(config.streaming_mode, StreamingMode::None);
assert!(!config.auto_cache);

Trait Implementations

impl Clone for RunConfig

fn clone(&self) -> RunConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for RunConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for RunConfig

fn default() -> Self

Returns the “default value” for a type.