vinzenz/hyperhive
vinzenz/hyperhive
Watch
0
0
Fork
You've already forked hyperhive
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix(#2351): key error-result scan on is_error to match the real event shape

Browse source
This commit is contained in:
damocles 2026-07-10 17:02:06 +02:00
commit c48d6319f7
1 changed files with 29 additions and 41 deletions
Download patch file Download diff file Expand all files Collapse all files

70
hive-claude/src/classify.rs
View file

@ -75,28 +75,26 @@ impl Sentinels {
if matches!(ty, Some("assistant" | "user")) {
return;
}
// A `result` event's `result` field is the model's final answer on a
// SUCCESSFUL turn — model-authored text, same trust level as an
// assistant message. A marker quoted there (the model explaining *this*
// code, or echoing an error string back) would falsely trip a sentinel
// and needlessly kill/downgrade the turn, so for a *success* result we
// scrub `result` before scanning and rely on the control fields.
// A `result` event's `result` field is model-authored answer text on a
// SUCCESSFUL turn — same trust level as an assistant message. A marker
// quoted there (the model explaining *this* code, or echoing an error
// string back) would falsely trip a sentinel and needlessly kill /
// downgrade the turn, so for a successful result we scrub `result`
// before scanning and rely on the other (claude-authored) fields.
//
// But on an ERROR result the same `result` field carries claude-code's
// OWN failure text — a genuine "Prompt is too long" (an API 400) lands
// there, emitted *instead of* a model answer. The model can't forge
// `is_error` / a non-`success` subtype (the CLI sets them from the real
// outcome), so an error result is claude-authored end to end and is
// safe to scan raw. Scrubbing it unconditionally (the original
// result-scrub fix) blinded prompt-too-long / auth detection — hence
// keying the scrub on success only. Other control events are
// claude-authored throughout, so they also scan raw.
// On a FAILED result the same `result` field instead carries
// claude-code's OWN failure text, emitted *instead of* a model answer.
// The real event (verified against captured stream-json) is:
// {"type":"result","is_error":true,"subtype":"success",
// "result":"Prompt is too long","terminal_reason":"blocking_limit"}
// Note `subtype` is "success" even on a hard failure — so `is_error` is
// the only reliable discriminator, and the model can't forge it (the CLI
// sets it from the real outcome). Scan those raw. Scrubbing `result`
// unconditionally (the original fix) blinded prompt-too-long / auth
// detection. Other control events are claude-authored throughout, so
// they also scan raw.
let error_result = ty == Some("result")
&& (event.get("is_error").and_then(serde_json::Value::as_bool) == Some(true)
|| event
.get("subtype")
.and_then(serde_json::Value::as_str)
.is_some_and(|s| s != "success"));
&& event.get("is_error").and_then(serde_json::Value::as_bool) == Some(true);
if ty == Some("result") && !error_result {
let mut scrubbed = event.clone();
if let Some(obj) = scrubbed.as_object_mut() {
@ -168,31 +166,21 @@ mod tests {
}
#[test]
fn control_event_marker_is_detected() {
fn genuine_failure_result_is_detected() {
// Both fixtures are the real claude-code failure shape, verified against
// captured stream-json: the marker lives in the `result` field with
// `is_error: true` and — counterintuitively — `subtype: "success"`. The
// earlier unconditional `result` scrub blinded this; `is_error` is the
// discriminator that restores it (and the model can't forge it).
let s = Sentinels::default();
let raw = r#"{"type":"result","subtype":"error","error":"Prompt is too long"}"#;
let raw = r#"{"type":"result","is_error":true,"subtype":"success","result":"Prompt is too long","terminal_reason":"blocking_limit","stop_reason":"stop_sequence"}"#;
s.scan_stdout_json(&json(raw), raw);
assert!(matches!(s.soft_error(), Some(Error::PromptTooLong)));
}
#[test]
fn prompt_too_long_in_error_result_field_is_detected() {
// Regression guard: claude-code reports a genuine "Prompt is too long"
// (API 400) in the `result` field of an ERROR result event. The earlier
// result-scrub blinded this by stripping `result` unconditionally. An
// error result is claude-authored (the model can't forge `is_error`),
// so it must be scanned raw.
for raw in [
r#"{"type":"result","subtype":"error_during_execution","is_error":true,"result":"API Error: 400 Prompt is too long"}"#,
r#"{"type":"result","subtype":"success","is_error":true,"result":"Prompt is too long"}"#,
] {
let s = Sentinels::default();
s.scan_stdout_json(&json(raw), raw);
assert!(
matches!(s.soft_error(), Some(Error::PromptTooLong)),
"prompt-too-long in an error result's `result` field must be detected: {raw}"
);
}
let s = Sentinels::default();
let raw = r#"{"type":"result","is_error":true,"subtype":"success","result":"Failed to authenticate. API Error: 401 Invalid authentication credentials","api_error_status":401}"#;
s.scan_stdout_json(&json(raw), raw);
assert!(matches!(s.soft_error(), Some(Error::AuthFailed)));
}
#[test]