Struct ForkedAgentConfig 

pub struct ForkedAgentConfig {
    pub prompt_messages: Vec<Message>,
    pub cache_safe_params: CacheSafeParams,
    pub can_use_tool: Arc<CanUseToolFn>,
    pub query_source: QuerySource,
    pub fork_label: String,
    pub overrides: Option<SubagentContextOverrides>,
    pub max_output_tokens: Option<u64>,
    pub max_turns: Option<u32>,
    pub on_message: Option<Arc<dyn Fn(Message) + Send + Sync>>,
    pub skip_transcript: bool,
    pub skip_cache_write: bool,
}

Configuration for a forked agent query.

Fields

prompt_messages: Vec<Message>

Messages to start the forked query loop with

cache_safe_params: CacheSafeParams

Cache-safe parameters that must match the parent query

can_use_tool: Arc<CanUseToolFn>

Permission check function for the forked agent

query_source: QuerySource

Source identifier for tracking

fork_label: String

Label for analytics (e.g., ‘session_memory’, ‘supervisor’)

overrides: Option<SubagentContextOverrides>

Optional overrides for the subagent context

max_output_tokens: Option<u64>

Optional cap on output tokens. CAUTION: setting this changes both max_tokens AND budget_tokens (via clamping in claude.ts). If the fork uses cache_safe_params to share the parent’s prompt cache, a different budget_tokens will invalidate the cache — thinking config is part of the cache key. Only set this when cache sharing is not a goal (e.g., compact summaries).

max_turns: Option<u32>

Optional cap on number of turns (API round-trips)

on_message: Option<Arc<dyn Fn(Message) + Send + Sync>>

Optional callback invoked for each message as it arrives (for streaming UI)

skip_transcript: bool

Skip sidechain transcript recording (e.g., for ephemeral work like speculation)

skip_cache_write: bool

Skip writing new prompt cache entries on the last message. For fire-and-forget forks where no future request will read from this prefix.