rig_agent/agent/streaming.rs
1//! Streamed agent events, bounded event feeds, and terminal response shaping.
2//!
3//! ```no_run
4//! # fn example(agent: &rig_agent::Agent) {
5//! let stream = agent.prompt("Explain ownership.").stream();
6//! # }
7//! ```
8
9use rig_core::error::ProviderError;
10use rig_core::{
11 message::AssistantContent,
12 wasm_compat::{WasmBoxedStream, WasmCompatSend},
13};
14
15use crate::{
16 agent::engine::{DriveItem, StreamingTurnSource, Surfaced, drive_agent},
17 agent::runner::AgentRunner,
18 streaming::{Item, StreamEvent},
19};
20use futures::{SinkExt, Stream, StreamExt, channel::mpsc, stream::FusedStream};
21use serde::{Deserialize, Serialize};
22use std::io::Write;
23use std::pin::Pin;
24use tracing_futures::Instrument;
25
26use crate::completion::PromptError;
27use crate::run::response::{CompletionCall, PromptResponse};
28
29/// The stream a streamed run yields: its items, then its ending.
30pub type StreamingResult = WasmBoxedStream<'static, Result<MultiTurnStreamItem, PromptError>>;
31
32#[derive(Serialize, Deserialize, Debug, Clone)]
33#[serde(tag = "type", rename_all = "camelCase")]
34/// One item of a streamed run: a provider stream event, a committed tool
35/// call, a lifecycle marker, or the run's final response. An item reads back
36/// from its serialized form on its own, without checking its order against
37/// the items around it; [`StreamRecord::parse`](rig_core::streaming::StreamRecord::parse)
38/// checks a run's provider stream items.
39#[non_exhaustive]
40pub enum MultiTurnStreamItem {
41 /// A provider stream item containing model-emitted content: part
42 /// starts and ends, text and reasoning fragments, a tool call's
43 /// argument fragments as they stream (a call to a tool the turn does not
44 /// allow is held for its resolution), and unmodeled passthrough
45 /// payloads. A call's end carries the validated call; the same call id
46 /// is reported again as [`ToolCall`](Self::ToolCall) once history holds it.
47 StreamAssistantItem(Item<StreamEvent>),
48 /// A tool call **committed to the run's history**, in commit order.
49 ///
50 /// The stream's tool items are a projection of committed history
51 /// ([`project`](crate::run::project)): every call history holds is
52 /// reported, and later answered by a [`ToolResult`](Self::ToolResult),
53 /// whether it ran, a hook skipped it, or invalid-call recovery or an
54 /// output reprompt answered it; a run resumed mid-batch reports the batch's
55 /// calls again ([`projection_start`](crate::run::AgentRun::projection_start)).
56 /// It is **not** an execution-lifecycle event
57 /// (see [`ToolExecutionCommitted`](Self::ToolExecutionCommitted)). A
58 /// structured-output Tool-mode output-tool call that finalizes the run is
59 /// committed as text, so it is reported in the
60 /// [`FinalResponse`](Self::FinalResponse) instead.
61 ToolCall {
62 /// The call as the model emitted it. Its id is equal on its
63 /// execution commit and its result.
64 tool_call: rig_core::message::ToolCall,
65 },
66 /// Confirmation that Rig **executed and committed** a tool call. This is not
67 /// a real-time start notification: it is surfaced together with its
68 /// `ToolResult` only after the whole batch settles successfully. Use tool
69 /// hooks for live host-side start/result observation.
70 ///
71 /// This item is emitted only for a tool whose body actually ran (it passed
72 /// its `ToolCall` hook checks), never for a call dropped by a sibling's
73 /// termination, skipped by a hook, or resolved by invalid-call recovery.
74 /// Correlate it with the model call and result through the call's id.
75 ToolExecutionCommitted {
76 /// The tool call as **executed**: the model's call with any
77 /// [`DispatchAction::Patch`](crate::agent::DispatchAction::Patch) hook rewrite
78 /// applied (so a redaction rewrite is reflected here, not leaked). The
79 /// model's *original* call is reported via
80 /// [`StreamAssistantItem`](Self::StreamAssistantItem).
81 tool_call: rig_core::message::ToolCall,
82 },
83 /// A tool result **committed to the run's history**, answering a reported
84 /// [`ToolCall`](Self::ToolCall). The tool
85 /// batch commits and surfaces atomically at every `tool_concurrency`
86 /// (including the sequential default): results are surfaced in call order
87 /// only after the whole batch settles successfully; a run that terminates
88 /// mid-batch surfaces no successful tool results.
89 ToolResult {
90 /// The result; `tool_result.call` is the id of the call it answers.
91 tool_result: rig_core::message::ToolResult,
92 },
93 /// Details for one successfully completed completion request made by this agent stream.
94 ///
95 /// This is emitted when a provider call finishes. Usage is the provider's
96 /// final usage for that completion request when available; it is not
97 /// incremental per streamed token.
98 ///
99 /// ```
100 /// use rig_agent::agent::MultiTurnStreamItem;
101 /// fn input_tokens(item: &MultiTurnStreamItem) -> Option<u64> {
102 /// match item {
103 /// MultiTurnStreamItem::CompletionCall(call) => call.usage.input_tokens,
104 /// _ => None,
105 /// }
106 /// }
107 /// ```
108 CompletionCall(CompletionCall),
109 /// The completed model turn was rejected by a hook for retry.
110 ///
111 /// Text and reasoning deltas emitted for this turn were provisional. A
112 /// consumer should discard or visually reset output associated with `turn`.
113 /// A subsequent attempt is made only if the run's total model-call budget
114 /// permits it.
115 ModelTurnRetried {
116 /// One-based model-call index of the rejected turn.
117 turn: usize,
118 },
119 /// The final result from the stream: the unified [`PromptResponse`] shared
120 /// with the blocking surface.
121 ///
122 /// Terminal for the run: no retry, further turn, or tool execution follows.
123 /// This item is the stream-side
124 /// counterpart of the `on_run_settled` hook's success outcome. Error
125 /// termination surfaces as the stream's `Err` item instead, which is
126 /// equally terminal.
127 FinalResponse(PromptResponse),
128}
129
130/// The one mapping from what the engine surfaces directly; tool items are
131/// built only from committed history.
132impl From<Surfaced> for MultiTurnStreamItem {
133 fn from(surfaced: Surfaced) -> Self {
134 match surfaced {
135 Surfaced::Provider(item) => Self::StreamAssistantItem(item),
136 Surfaced::CompletionCall(call) => Self::CompletionCall(call),
137 Surfaced::ModelTurnRetried { turn } => Self::ModelTurnRetried { turn },
138 Surfaced::Final(response) => Self::FinalResponse(response),
139 }
140 }
141}
142
143impl MultiTurnStreamItem {
144 /// Build a final response from structured content and aggregate usage.
145 /// Concatenates text for output; completion details and history remain empty.
146 pub fn final_response(
147 content: Vec<AssistantContent>,
148 aggregated_usage: crate::completion::Usage,
149 ) -> Self {
150 Self::FinalResponse(PromptResponse::from_content(content, aggregated_usage))
151 }
152}
153
154impl AgentRunner {
155 /// Drive the agent loop as a stream of assistant content, tool activity
156 /// and, last, the [`FinalResponse`](MultiTurnStreamItem::FinalResponse).
157 /// Hooks fire at every observable point, including streamed text and
158 /// tool-call deltas.
159 ///
160 /// Like [`run`](AgentRunner::run), this is lazy: agent span creation
161 /// begins only when the stream is first polled, and a stream that is
162 /// dropped unpolled has done nothing. The stream is `Send` on
163 /// native targets, so it can be built in synchronous code and handed to
164 /// whatever polls it; it runs under the span it was built in (a caller's
165 /// enabled span is adopted, otherwise a root `invoke_agent` is created),
166 /// not under whichever span first polls it.
167 ///
168 /// ```rust,no_run
169 /// # use rig_agent::{Agent, agent::StreamingResult};
170 /// fn start(agent: &Agent, prompt: &str) -> StreamingResult {
171 /// // Nothing runs until whoever holds this polls it.
172 /// agent.prompt(prompt).stream()
173 /// }
174 /// ```
175 ///
176 /// Shares the drive loop, run construction, tool execution and fail-closed
177 /// hook handling with the blocking [`run`](AgentRunner::run) via
178 /// `drive_agent`, so the two behave identically apart from the streamed
179 /// delta events.
180 #[must_use = "a stream does nothing until polled"]
181 pub fn stream(self) -> StreamingResult {
182 // The span the stream is built under is the one it runs under, not
183 // whichever span first polls it: a host may build the stream in a
184 // request span and hand it to a task of its own.
185 self.stream_under(tracing::Span::current())
186 }
187
188 /// Build a lazy stream under an explicitly captured ambient span.
189 fn stream_under(self, ambient: tracing::Span) -> StreamingResult {
190 let run_under = ambient.clone();
191 let stream = async_stream::stream! {
192 let mut inner = self.start_stream(run_under);
193 while let Some(item) = inner.next().await {
194 yield item;
195 }
196 };
197 Box::pin(stream.instrument(ambient))
198 }
199
200 /// The eager half of [`stream`](Self::stream): build the run and return
201 /// the driver as a stream. Called on the first poll, under `ambient`, the
202 /// span the stream was built in.
203 fn start_stream(self, ambient: tracing::Span) -> StreamingResult {
204 let (agent_span, created_agent_span) = self.open_agent_span(ambient);
205
206 let hook_ctx = self.hook_context(true);
207 let run = self.build_run();
208 let source = StreamingTurnSource::new(
209 &self.config.hooks,
210 self.agent_name_or_default().to_string(),
211 created_agent_span,
212 self.config.record_telemetry_content,
213 );
214
215 let driver = drive_agent(
216 self,
217 source,
218 run,
219 agent_span.clone(),
220 created_agent_span,
221 hook_ctx,
222 )
223 .flat_map(|item| {
224 let (head, tools) = match item {
225 Ok(DriveItem::Surfaced(item)) => (Some(Ok(item.into())), None),
226 Ok(DriveItem::Projected(items)) => (None, Some(items)),
227 Ok(DriveItem::Done(_)) => (None, None),
228 Err(err) => (Some(Err(err)), None),
229 };
230 futures::stream::iter(head.into_iter().chain(tools.into_iter().flatten().map(Ok)))
231 });
232 Box::pin(driver.instrument(agent_span))
233 }
234}
235
236/// Capacity of the event queue behind [`AgentRunner::run_channel`]: the number
237/// of [`MultiTurnStreamItem`]s the run may buffer ahead of the consumer before
238/// it parks on back-pressure.
239pub const RUN_EVENTS_CAPACITY: usize = 32;
240
241/// Event feed of an agent run started with [`AgentRunner::run_channel`].
242///
243/// Every [`MultiTurnStreamItem`] the run would have streamed is delivered here
244/// in order, ending with [`MultiTurnStreamItem::FinalResponse`]. The feed is a
245/// bounded queue ([`RUN_EVENTS_CAPACITY`]): a slow consumer applies
246/// back-pressure to the run instead of losing events. Poll it as a
247/// [`Stream`] from async code, or drain it with the non-blocking
248/// [`try_next`](RunEvents::try_next) from a synchronous tick, such as a game loop,
249/// UI frame, an ECS system.
250///
251/// Dropping the feed does not cancel the run; it simply stops receiving events
252/// and the run future still resolves with the final
253/// [`PromptResponse`].
254#[derive(Debug)]
255pub struct RunEvents {
256 receiver: mpsc::Receiver<MultiTurnStreamItem>,
257}
258
259impl RunEvents {
260 /// Take the next buffered event without waiting.
261 ///
262 /// Returns `None` both when no event is queued yet and once the run has
263 /// finished and the feed is drained; use [`is_done`](RunEvents::is_done)
264 /// to tell the two apart.
265 pub fn try_next(&mut self) -> Option<MultiTurnStreamItem> {
266 self.receiver.try_recv().ok()
267 }
268
269 /// Whether the run has finished and every event has been taken.
270 pub fn is_done(&self) -> bool {
271 self.receiver.is_terminated()
272 }
273}
274
275impl Stream for RunEvents {
276 type Item = MultiTurnStreamItem;
277
278 fn poll_next(
279 mut self: Pin<&mut Self>,
280 cx: &mut std::task::Context<'_>,
281 ) -> std::task::Poll<Option<Self::Item>> {
282 Pin::new(&mut self.receiver).poll_next(cx)
283 }
284
285 fn size_hint(&self) -> (usize, Option<usize>) {
286 self.receiver.size_hint()
287 }
288}
289
290impl FusedStream for RunEvents {
291 fn is_terminated(&self) -> bool {
292 self.receiver.is_terminated()
293 }
294}
295
296impl AgentRunner {
297 /// Split the run into a driving future and a [`RunEvents`] feed.
298 ///
299 /// The future performs the same agent loop as
300 /// [`run`](AgentRunner::run) and [`stream`](AgentRunner::stream), and
301 /// resolves with the final [`PromptResponse`]; the
302 /// feed receives each intermediate [`MultiTurnStreamItem`] as it happens.
303 /// Spawn the future on any executor and poll the feed from wherever the
304 /// events are consumed; neither side assumes a runtime.
305 ///
306 /// The feed is bounded ([`RUN_EVENTS_CAPACITY`]); when it is full the run
307 /// waits for the consumer rather than dropping events. Dropping the feed
308 /// lets the run continue to completion unobserved. Like
309 /// [`stream`](Self::stream), the run belongs to the span this method was
310 /// called in, wherever the future is later polled.
311 #[must_use = "the run does nothing until the future is driven"]
312 pub fn run_channel(
313 self,
314 ) -> (
315 impl Future<Output = Result<PromptResponse, PromptError>> + WasmCompatSend,
316 RunEvents,
317 ) {
318 let (mut sender, receiver) = mpsc::channel(RUN_EVENTS_CAPACITY);
319 // Captured here, not when the future is first polled: the doc above
320 // says to spawn the future, and the run must still belong to the span
321 // that split it.
322 let ambient = tracing::Span::current();
323 let future = async move {
324 let mut stream = self.stream_under(ambient);
325 let mut response = None;
326 let mut forward = true;
327 while let Some(item) = stream.next().await {
328 let item = match item {
329 Ok(item) => item,
330 Err(err) => {
331 // Drain after the terminal error so span and lineage teardown
332 // completes instead of being dropped at the yield.
333 while stream.next().await.is_some() {}
334 return Err(err);
335 }
336 };
337 match item {
338 MultiTurnStreamItem::FinalResponse(done) => {
339 if forward {
340 let _ = sender
341 .send(MultiTurnStreamItem::FinalResponse(done.clone()))
342 .await;
343 }
344 response = Some(done);
345 }
346 item => {
347 if forward && sender.send(item).await.is_err() {
348 forward = false;
349 }
350 }
351 }
352 }
353 response.ok_or_else(|| {
354 PromptError::Provider(ProviderError::Response(
355 "agent run ended without producing a final response".to_string(),
356 ))
357 })
358 };
359 (future, RunEvents { receiver })
360 }
361}
362
363/// Why [`stream_to_stdout`] returned no response.
364#[derive(Debug, thiserror::Error)]
365pub enum StreamToStdoutError {
366 /// The run failed.
367 #[error(transparent)]
368 Run(#[from] PromptError),
369 /// Writing to stdout failed.
370 #[error(transparent)]
371 Io(#[from] std::io::Error),
372 /// The stream ended without a final response.
373 #[error("the stream ended without a final response")]
374 Incomplete,
375}
376
377/// Print a streamed run's assistant text and reasoning to stdout and return
378/// its final response.
379///
380/// Streaming metadata events, such as `MultiTurnStreamItem::CompletionCall`,
381/// are not printed; metadata is returned on the [`PromptResponse`] via
382/// accessors such as [`PromptResponse::completion_calls`]. A model-turn retry
383/// prints a visible boundary because text already written to stdout cannot be
384/// retracted. Fails with the run's error as soon as the stream yields it, a
385/// stdout write failure, or [`StreamToStdoutError::Incomplete`] when the
386/// stream ends without a final response.
387pub async fn stream_to_stdout(
388 stream: &mut StreamingResult,
389) -> Result<PromptResponse, StreamToStdoutError> {
390 let mut stdout = std::io::stdout();
391 write!(stdout, "Response: ")?;
392 while let Some(content) = stream.next().await {
393 match content? {
394 MultiTurnStreamItem::StreamAssistantItem(Item::Event(StreamEvent::Text {
395 text,
396 ..
397 })) => {
398 write!(stdout, "{text}")?;
399 stdout.flush()?;
400 }
401 MultiTurnStreamItem::StreamAssistantItem(Item::Event(StreamEvent::End {
402 content: AssistantContent::Reasoning(reasoning),
403 ..
404 })) => {
405 write!(stdout, "{}", reasoning.text)?;
406 stdout.flush()?;
407 }
408 MultiTurnStreamItem::FinalResponse(response) => return Ok(response),
409 MultiTurnStreamItem::ModelTurnRetried { turn } => {
410 write!(
411 stdout,
412 "\n[model turn {turn} rejected; retry requested]\nResponse: "
413 )?;
414 stdout.flush()?;
415 }
416 _ => {}
417 }
418 }
419 Err(StreamToStdoutError::Incomplete)
420}
421
422#[cfg(test)]
423mod malformed_tool_args_tests;
424#[cfg(test)]
425#[allow(irrefutable_let_patterns, unreachable_patterns)]
426mod tests;