don't emit warning when ending on a successful tool call (#12468)
This commit is contained in:
@@ -131,6 +131,24 @@ fn is_empty_response(message: &Message) -> bool {
|
||||
})
|
||||
}
|
||||
|
||||
pub fn ends_with_successful_tool_response(messages: &[Message]) -> bool {
|
||||
let Some(message) = messages.last() else {
|
||||
return false;
|
||||
};
|
||||
let mut responses = message
|
||||
.content
|
||||
.iter()
|
||||
.filter_map(MessageContent::as_tool_response)
|
||||
.peekable();
|
||||
responses.peek().is_some()
|
||||
&& responses.all(|response| {
|
||||
response
|
||||
.tool_result
|
||||
.as_ref()
|
||||
.is_ok_and(|result| !result.is_error.unwrap_or(false))
|
||||
})
|
||||
}
|
||||
|
||||
fn record_request_params(span: &tracing::Span, model_config: &ModelConfig) {
|
||||
if let Some(temperature) = model_config.temperature {
|
||||
span.record("gen_ai.request.temperature", temperature as f64);
|
||||
@@ -172,7 +190,11 @@ pub struct InferenceRunner<'a, S, E> {
|
||||
|
||||
/// The agent-visible conversation as the provider sees it: tool requests left
|
||||
/// unanswered by an earlier turn are dropped, since nothing will answer them now.
|
||||
fn messages_for_provider(conversation: &Conversation, turn: &[Message]) -> Vec<Message> {
|
||||
fn messages_for_provider(
|
||||
conversation: &Conversation,
|
||||
turn: &[Message],
|
||||
keep_empty_messages: bool,
|
||||
) -> Vec<Message> {
|
||||
let answered: std::collections::HashSet<&str> = conversation
|
||||
.messages()
|
||||
.iter()
|
||||
@@ -194,7 +216,7 @@ fn messages_for_provider(conversation: &Conversation, turn: &[Message]) -> Vec<M
|
||||
}
|
||||
message
|
||||
})
|
||||
.filter(|message| !message.content.is_empty())
|
||||
.filter(|message| keep_empty_messages || !message.content.is_empty())
|
||||
.collect()
|
||||
}
|
||||
|
||||
@@ -220,6 +242,17 @@ fn ends_with_provider_turn(messages: &[Message]) -> bool {
|
||||
})
|
||||
}
|
||||
|
||||
fn should_infer(conversation: &Conversation, turn: &[Message]) -> bool {
|
||||
let projected = messages_for_provider(conversation, turn, true);
|
||||
if projected
|
||||
.last()
|
||||
.is_some_and(|message| message.content.is_empty())
|
||||
{
|
||||
return false;
|
||||
}
|
||||
ends_with_provider_turn(&messages_for_provider(conversation, turn, false))
|
||||
}
|
||||
|
||||
fn cancellation_response(persisted: &[Message], pending: &[Message]) -> Option<Message> {
|
||||
let mut answered = persisted
|
||||
.iter()
|
||||
@@ -325,8 +358,7 @@ impl<S: Sync, E: InferenceEffect> Inference<S, E> for InferenceRunner<'_, S, E>
|
||||
let Ok(turn) = messages_since_kickoff(conversation) else {
|
||||
return false;
|
||||
};
|
||||
trailing_error(conversation).is_none()
|
||||
&& ends_with_provider_turn(&messages_for_provider(conversation, turn))
|
||||
trailing_error(conversation).is_none() && should_infer(conversation, turn)
|
||||
}
|
||||
|
||||
async fn infer(
|
||||
@@ -341,10 +373,10 @@ impl<S: Sync, E: InferenceEffect> Inference<S, E> for InferenceRunner<'_, S, E>
|
||||
return not_applicable();
|
||||
}
|
||||
|
||||
let mut messages_for_provider = messages_for_provider(conversation, messages);
|
||||
if !ends_with_provider_turn(&messages_for_provider) {
|
||||
if !should_infer(conversation, messages) {
|
||||
return not_applicable();
|
||||
}
|
||||
let mut messages_for_provider = messages_for_provider(conversation, messages, false);
|
||||
|
||||
let span = inference_span(self.provider.as_ref(), &self.model_config);
|
||||
|
||||
@@ -479,6 +511,7 @@ impl<S: Sync, E: InferenceEffect> Inference<S, E> for InferenceRunner<'_, S, E>
|
||||
}
|
||||
|
||||
let empty_response = !cancelled
|
||||
&& !ends_with_successful_tool_response(conversation.messages())
|
||||
&& !accumulator
|
||||
.iter()
|
||||
.any(|message| message.metadata.output_token_limit_reached)
|
||||
@@ -490,7 +523,24 @@ impl<S: Sync, E: InferenceEffect> Inference<S, E> for InferenceRunner<'_, S, E>
|
||||
return yielded_with(usage_effects);
|
||||
}
|
||||
|
||||
usage_effects.extend(accumulator.into_iter().map(|message| E::from(message)));
|
||||
if ends_with_successful_tool_response(conversation.messages())
|
||||
&& !accumulator
|
||||
.iter()
|
||||
.any(|message| message.metadata.output_token_limit_reached)
|
||||
&& accumulator.iter().all(is_empty_response)
|
||||
{
|
||||
let mut message = accumulator
|
||||
.into_iter()
|
||||
.last()
|
||||
.unwrap_or_else(Message::assistant);
|
||||
message.content.clear();
|
||||
message.metadata.user_visible = false;
|
||||
message.metadata.agent_visible = true;
|
||||
let message = emit.message(message).await;
|
||||
usage_effects.push(E::from(message));
|
||||
} else {
|
||||
usage_effects.extend(accumulator.into_iter().map(|message| E::from(message)));
|
||||
}
|
||||
applied(usage_effects)
|
||||
}
|
||||
.instrument(span)
|
||||
@@ -584,4 +634,9 @@ mod tests {
|
||||
&Message::assistant().with_content(MessageContent::thinking("", "sig-omitted"))
|
||||
));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn whitespace_only_text_is_an_empty_response() {
|
||||
assert!(is_empty_response(&Message::assistant().with_text(" \n\t ")));
|
||||
}
|
||||
}
|
||||
|
||||
@@ -6,6 +6,7 @@ use std::sync::Arc;
|
||||
use anyhow::{anyhow, Context, Result};
|
||||
use futures::stream::BoxStream;
|
||||
use futures::{stream, FutureExt, StreamExt, TryStreamExt};
|
||||
use goose_agent::inference::ends_with_successful_tool_response;
|
||||
use tracing_futures::Instrument;
|
||||
|
||||
use super::container::Container;
|
||||
@@ -3379,7 +3380,9 @@ impl Agent {
|
||||
session_manager.replace_conversation(&session_config.id, &conversation).await?;
|
||||
yield AgentEvent::HistoryReplaced(conversation.clone());
|
||||
}
|
||||
Ok(RetryResult::Skipped) if empty_response => {
|
||||
Ok(RetryResult::Skipped)
|
||||
if empty_response
|
||||
&& !ends_with_successful_tool_response(conversation.messages()) => {
|
||||
// No recipe retry configured, and this empty
|
||||
// turn would otherwise fall through to a
|
||||
// silent exit. Retry a bounded number of
|
||||
|
||||
Reference in New Issue
Block a user