From c0230b4109e1eccd74996d9880ef3dd25201f5f7 Mon Sep 17 00:00:00 2001
From: Rudy Celekli <rudy@gradiahq.com>
Date: Sat, 17 Oct 2026 20:58:24 -0400
Subject: [PATCH] fix: cut a long conversation opening on a character boundary

`standing::openings` shortens an opening longer than 300 bytes to its
last word in the first 300 bytes, and took those bytes as
`text[a..a + 300]`. When byte 300 falls inside a multi-byte character
(any accented, Cyrillic or CJK text) the slice panics before the
boundary fix that follows it runs, so the observer's conversation
outline (`observer::outline`, which calls `openings` on each
conversation document it outlines) panics on such a document.

The 300-byte cut now steps back to a character boundary before the
slice; the result is unchanged whenever the old code did not panic.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Rudy Celekli <rudy@gradiahq.com>
---
 crates/graphicle-memory/src/standing.rs | 26 ++++++++++++++++++++++---
 1 file changed, 23 insertions(+), 3 deletions(-)

diff --git a/crates/graphicle-memory/src/standing.rs b/crates/graphicle-memory/src/standing.rs
index 35a4eec..6b8bfbd 100644
--- a/crates/graphicle-memory/src/standing.rs
+++ b/crates/graphicle-memory/src/standing.rs
@@ -151,10 +151,13 @@ pub fn openings(text: &str) -> Vec<(usize, usize)> {
         }
         let mut b = b;
         if b - a > 300 {
-            b = a + text[a..a + 300].rfind(' ').unwrap_or(300);
-            while !text.is_char_boundary(b) {
-                b -= 1;
+            // The first 300 bytes, back to a character boundary: byte 300
+            // can fall inside a multi-byte character.
+            let mut cut = a + 300;
+            while !text.is_char_boundary(cut) {
+                cut -= 1;
             }
+            b = text[a..cut].rfind(' ').map_or(cut, |space| a + space);
         }
         found.push((a, b));
     }
@@ -387,6 +390,23 @@ mod tests {
         );
     }
 
+    #[test]
+    fn a_long_opening_ends_on_a_character_boundary() {
+        // Byte 300 of the opening falls inside an "é" (two bytes).
+        let text = format!("User: I need help with {}\n", "\u{e9}".repeat(200));
+        let spans = openings(&text);
+        assert_eq!(spans.len(), 1, "{spans:?}");
+        let (a, b) = spans[0];
+        assert_eq!(&text[a..b], "I need help with");
+        // Without a space in the first 300 bytes the cut is at the last
+        // whole character.
+        let text = format!("User: x{} b c d\n", "\u{4f60}".repeat(120));
+        let spans = openings(&text);
+        assert_eq!(spans.len(), 1, "{spans:?}");
+        let (a, b) = spans[0];
+        assert_eq!(text[a..b], format!("x{}", "\u{4f60}".repeat(99)));
+    }
+
     #[test]
     fn conflicts_pair_a_denial_with_the_users_own_statements_on_its_subject() {
         let early = "Conversation date: 2024/03/10 (Sun) 00:00\n\n\