surma
c728429d04
- The serve loop ends the process after EOF or exit_after_response, also with background tasks or listeners. - One schema generator per direction for the whole document, so types with the same name get distinct names. - The agent gives the complete error (code, message, data) to the model, writes SKIP only after a successful request, caps tool output at 64 KiB, writes each history line in one write, and skips a damaged line. - Endpoints skip blank and non-RPC lines. client_rs does not panic. - #[must_use] on builders and the lock guard; #[schemars] attributes and duplicate hooks are compile errors; raw identifiers; -32000 for local SDK errors; x-sausage-params only for methods with params.
89 lines
3.0 KiB
Rust
89 lines
3.0 KiB
Rust
//! History window selection and context-size error detection.
|
|
|
|
use crate::llm::Entry;
|
|
|
|
/// Recognize provider errors that permit a retry with fewer complete turns.
|
|
pub fn is_too_large(error: &str) -> bool {
|
|
let Some((status, body)) = error
|
|
.strip_prefix("HTTP ")
|
|
.and_then(|error| error.split_once(':'))
|
|
else {
|
|
return false;
|
|
};
|
|
if status.split_whitespace().next() == Some("413") {
|
|
return true;
|
|
}
|
|
if status.split_whitespace().next() != Some("400") {
|
|
return false;
|
|
}
|
|
let body = body.to_ascii_lowercase();
|
|
body.contains("context_length_exceeded")
|
|
|| body.contains("maximum context length")
|
|
|| body.contains("context length exceeded")
|
|
|| body.contains("too large")
|
|
|| (body.contains("too long")
|
|
&& ["context", "input", "prompt"]
|
|
.iter()
|
|
.any(|term| body.contains(term)))
|
|
}
|
|
|
|
/// Find a user-turn boundary without discarding any of the current turn.
|
|
pub fn next_turn_start(history: &[Entry], from: usize, current: usize) -> Option<usize> {
|
|
let end = current.min(history.len().checked_sub(1)?);
|
|
(from..=end).find(|&index| matches!(&history[index], Entry::User { .. }))
|
|
}
|
|
|
|
#[cfg(test)]
|
|
mod tests {
|
|
use super::*;
|
|
|
|
#[test]
|
|
fn detects_context_size_errors_only_for_http_400_or_413() {
|
|
assert!(is_too_large("HTTP 400: context_length_exceeded"));
|
|
assert!(is_too_large("HTTP 413: prompt is too long"));
|
|
assert!(is_too_large("HTTP 413: {\"type\":\"request_too_large\"}"));
|
|
assert!(is_too_large("HTTP 413: {}"));
|
|
assert!(is_too_large("HTTP 400: maximum context length"));
|
|
assert!(is_too_large("HTTP 400: request is too large"));
|
|
assert!(!is_too_large("HTTP 500: prompt is too long"));
|
|
assert!(!is_too_large("HTTP 400: invalid token"));
|
|
}
|
|
|
|
#[test]
|
|
fn history_skip_advances_by_whole_user_turns() {
|
|
let history = vec![
|
|
Entry::User {
|
|
text: "first".into(),
|
|
},
|
|
Entry::ToolCall {
|
|
id: "a".into(),
|
|
name: "tool".into(),
|
|
arguments: "{}".into(),
|
|
},
|
|
Entry::ToolResult {
|
|
id: "a".into(),
|
|
output: "done".into(),
|
|
},
|
|
Entry::User {
|
|
text: "second".into(),
|
|
},
|
|
Entry::ToolCall {
|
|
id: "b".into(),
|
|
name: "tool".into(),
|
|
arguments: "{}".into(),
|
|
},
|
|
Entry::ToolResult {
|
|
id: "b".into(),
|
|
output: "done".into(),
|
|
},
|
|
Entry::User {
|
|
text: "current".into(),
|
|
},
|
|
];
|
|
assert_eq!(next_turn_start(&history, 0, 6), Some(0));
|
|
assert_eq!(next_turn_start(&history, 1, 6), Some(3));
|
|
assert_eq!(next_turn_start(&history, 4, 6), Some(6));
|
|
assert_eq!(next_turn_start(&history, usize::MAX, 6).unwrap_or(6), 6);
|
|
}
|
|
}
|