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

hive-claude/src/classify.rs

@@ -75,18 +75,27 @@ impl Sentinels {
         if matches!(ty, Some("assistant" | "user")) {
             return;
         }
-        // The terminal `result` event carries the model's final answer in its
-        // `result` field — model-authored text, same trust level as an
-        // assistant message. A marker quoted there (e.g. 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 `result`
-        // event, scan the control fields (`subtype` / `error` / `is_error`)
-        // but scrub the model-authored `result` field first; a genuine
-        // prompt-too-long/auth/session signal lives in those control fields
-        // (it's emitted *instead of* a successful model answer), never in the
-        // `result` text. Other control events (`error` / `system`) are
-        // claude-authored end to end, so they scan raw.
-        if ty == Some("result") {
+        // 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.
+        //
+        // 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);
+        if ty == Some("result") && !error_result {
             let mut scrubbed = event.clone();
             if let Some(obj) = scrubbed.as_object_mut() {
                 obj.remove("result");
@@ -157,11 +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)));
+
+        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]