Struct NativeTurnInput 

pub struct NativeTurnInput {
    pub prompt_text: String,
    pub system_prompt: Option<String>,
    pub attachments: Vec<UserAttachment>,
    pub cancel_token: Option<CancellationToken>,
    pub prior_messages: Vec<ChatMessage>,
    pub context_path: Option<PathBuf>,
    pub turn_end_grace: Option<Duration>,
}

Fields

prompt_text: String

The user-facing prompt that triggered this turn. Lands as the first ChatMessage::User.content the model sees.

system_prompt: Option<String>

Pre-composed system prompt (e.g. spec_snapshot.system_prompt + driver.append_system_prompt). None ⇒ no system message sent.

attachments: Vec<UserAttachment>

Non-text attachments lifted from the inbound UserMessagePayload content blocks. Empty Vec ⇒ pure-text prompt. Lands on the first ChatMessage::User.attachments so each provider’s projection can render them as image / file content blocks.

cancel_token: Option<CancellationToken>

Optional cancellation handle. When fired, the harness loop short-circuits at the next stream-chunk await (or before the next step starts) and emits TurnEnd { stop_reason: "interrupt" }. None ⇒ harness cannot be cancelled mid-flight (fine for tests and for fire-and-forget turns); production wires this through from RD’s active_native_cancel.

prior_messages: Vec<ChatMessage>

Prior messages history for in-memory mode (context_path = None). Ignored when context_path is Some — harness loads history from the JSONL file instead.

Empty Vec + context_path = None = fresh in-memory conversation.

context_path: Option<PathBuf>

Absolute path to harness’s context JSONL.

• Some(path) — persistent mode: harness loads prior messages from this file at turn start (creating it on first use), appends new messages incrementally, and rewrites it on compaction. prior_messages is ignored. TurnEnd.final_messages is empty.

• None — in-memory mode: prior_messages is used as the seed; harness never touches the filesystem. TurnEnd.final_messages carries the full history snapshot for the caller to persist. Suitable for runtime-driver (manages history in RAM) and tests.

turn_end_grace: Option<Duration>

How long the loop waits for in-flight tool futures to resolve after the cancel token fires before publishing TurnEnd{interrupt}.

None falls back to the harness-level grace configured via crate::agent_loop::AgentLoopHarness::with_turn_end_grace (which itself defaults to crate::agent_loop::DEFAULT_TURN_END_GRACE, 1s — the historical hard-coded value). Hosts whose tool runtimes need longer to shut down remote work (e.g. sandbox exec with a TERM-grace + SIGKILL tail) should raise this so TurnEnd is only published once writers actually stopped.

Regardless of the window’s outcome, the event carries TurnEnd.pending_tools: the ids still unresolved when it fired.

Trait Implementations

impl Clone for NativeTurnInput

fn clone(&self) -> NativeTurnInput

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for NativeTurnInput

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

Formats the value using the given formatter.

impl PartialEq for NativeTurnInput

fn eq(&self, other: &Self) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.