alex/g3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

attempted fix of autocontinue

Browse Source
This commit is contained in:
Dhanji R. Prasanna
2025-12-22 15:01:27 +11:00
parent 38fcaaf449
commit 0e4febc3fb
3 changed files with 344 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

163
crates/g3-core/src/lib.rs
View File

@@ -531,6 +531,113 @@ impl StreamingToolParser {
self.message_stopped
}
/// Check if the text buffer contains an incomplete JSON tool call
/// This detects cases where the LLM started emitting a tool call but the stream ended
/// before the JSON was complete (truncated output)
pub fn has_incomplete_tool_call(&self) -> bool {
let patterns = [
r#"{"tool":"#,
r#"{ "tool":"#,
r#"{"tool" :"#,
r#"{ "tool" :"#,
];
// Find the last occurrence of a tool call pattern
let mut best_start: Option<usize> = None;
for pattern in &patterns {
if let Some(pos) = self.text_buffer.rfind(pattern) {
if best_start.map_or(true, |best| pos > best) {
best_start = Some(pos);
}
}
}
if let Some(start_pos) = best_start {
// Check if we can parse a complete JSON object from this position
// If NOT complete, it's an incomplete tool call
let json_text = &self.text_buffer[start_pos..];
!Self::is_complete_json_object(json_text)
} else {
false
}
}
/// Check if the text buffer contains an unexecuted tool call
/// This detects cases where the LLM emitted a complete tool call JSON
/// but it wasn't parsed/executed (e.g., due to parsing issues)
pub fn has_unexecuted_tool_call(&self) -> bool {
let patterns = [
r#"{"tool":"#,
r#"{ "tool":"#,
r#"{"tool" :"#,
r#"{ "tool" :"#,
];
// Find the last occurrence of a tool call pattern
let mut best_start: Option<usize> = None;
for pattern in &patterns {
if let Some(pos) = self.text_buffer.rfind(pattern) {
if best_start.map_or(true, |best| pos > best) {
best_start = Some(pos);
}
}
}
if let Some(start_pos) = best_start {
// Check if we can parse a complete JSON object from this position
let json_text = &self.text_buffer[start_pos..];
// If the JSON IS complete, it means there's an unexecuted tool call
if let Some(json_end) = Self::find_complete_json_object_end(json_text) {
// Extract just the JSON object (not any trailing text)
let json_only = &json_text[..=json_end];
// Try to parse it as a tool call to confirm it's valid JSON
return serde_json::from_str::<serde_json::Value>(json_only).is_ok();
}
false
} else {
false
}
}
/// Check if a string contains a complete JSON object
fn is_complete_json_object(text: &str) -> bool {
Self::find_complete_json_object_end(text).is_some()
}
/// Find the end position (byte index) of a complete JSON object in the text
/// Returns None if no complete JSON object is found
fn find_complete_json_object_end(text: &str) -> Option<usize> {
let mut brace_count = 0;
let mut in_string = false;
let mut escape_next = false;
let mut found_start = false;
for (i, ch) in text.char_indices() {
if escape_next {
escape_next = false;
continue;
}
match ch {
'\\' => escape_next = true,
'"' if !escape_next => in_string = !in_string,
'{' if !in_string => {
brace_count += 1;
found_start = true;
}
'}' if !in_string => {
brace_count -= 1;
if brace_count == 0 && found_start {
return Some(i); // Return the byte index of the closing brace
}
}
_ => {}
}
}
None // No complete JSON object found
}
/// Reset the parser state for a new message
pub fn reset(&mut self) {
self.text_buffer.clear();
@@ -4941,18 +5048,41 @@ impl<W: UiWriter> Agent<W> {
let has_response = !current_response.is_empty() || !full_response.is_empty();
// Check if there's an incomplete tool call in the buffer
let has_incomplete_tool_call = parser.has_incomplete_tool_call();
// Check if there's a complete but unexecuted tool call in the buffer
let has_unexecuted_tool_call = parser.has_unexecuted_tool_call();
// Auto-continue if tools were executed but final_output was never called
// This is the simple rule: LLM must call final_output before returning control
if any_tool_executed && !final_output_called {
// OR if the LLM emitted an incomplete tool call (truncated JSON)
// OR if the LLM emitted a complete tool call that wasn't executed
// This ensures we don't return control when the LLM clearly intended to call a tool
if (any_tool_executed && !final_output_called) || has_incomplete_tool_call || has_unexecuted_tool_call {
if auto_summary_attempts < MAX_AUTO_SUMMARY_ATTEMPTS {
auto_summary_attempts += 1;
warn!(
"LLM stopped without calling final_output after executing tools ({} iterations, auto-continue attempt {})",
iteration_count, auto_summary_attempts
);
self.ui_writer.print_context_status(
"\n🔄 Model stopped without calling final_output. Auto-continuing...\n"
);
if has_incomplete_tool_call || has_unexecuted_tool_call {
warn!(
"LLM emitted {} tool call ({} iterations, auto-continue attempt {})",
if has_incomplete_tool_call { "incomplete" } else { "unexecuted" },
iteration_count, auto_summary_attempts
);
self.ui_writer.print_context_status(
if has_incomplete_tool_call {
"\n🔄 Model emitted incomplete tool call. Auto-continuing...\n"
} else {
"\n🔄 Model emitted tool call that wasn't executed. Auto-continuing...\n"
}
);
} else {
warn!(
"LLM stopped without calling final_output after executing tools ({} iterations, auto-continue attempt {})",
iteration_count, auto_summary_attempts
);
self.ui_writer.print_context_status(
"\n🔄 Model stopped without calling final_output. Auto-continuing...\n"
);
}
// Add any text response to context before prompting for continuation
if has_response {
@@ -4971,10 +5101,17 @@ impl<W: UiWriter> Agent<W> {
}
// Add a follow-up message asking for continuation
let continue_prompt = Message::new(
MessageRole::User,
"Please continue until you are done. You **MUST** call `final_output` with a summary when done.".to_string(),
);
let continue_prompt = if has_incomplete_tool_call {
Message::new(
MessageRole::User,
"Your previous response was cut off mid-tool-call. Please complete the tool call and continue.".to_string(),
)
} else {
Message::new(
MessageRole::User,
"Please continue until you are done. You **MUST** call `final_output` with a summary when done.".to_string(),
)
};
self.context_window.add_message(continue_prompt);
request.messages = self.context_window.conversation_history.clone();