fix: keep active request after prompt-guided tool results

crates/tinyinference-llm/src/prompt_tools/mod.rs

@@ -229,6 +229,41 @@ pub fn ensure_resolvable_user_turn(messages: &[Message]) -> Vec<Message> {
     out
 }
 
+/// Keep the active request at the end of a prompt-guided tool continuation.
+/// Some chat templates treat the synthetic tool-result user turn as transport
+/// data and resolve an earlier user message as the query. Repeating the latest
+/// real text request after a terminal result gives those templates the correct
+/// query without changing the durable transcript.
+#[must_use]
+pub fn anchor_user_request_after_tool_result(messages: &[Message]) -> Vec<Message> {
+    if !messages.last().is_some_and(|message| {
+        matches!(message, Message::User(_))
+            && message
+                .text()
+                .trim_start()
+                .starts_with(TOOL_RESULTS_PREFIX.trim_end())
+    }) {
+        return messages.to_vec();
+    }
+    let Some(request) = messages
+        .iter()
+        .rev()
+        .skip(1)
+        .find(|message| is_resolvable_user_query(message))
+    else {
+        return messages.to_vec();
+    };
+    let request = request.text();
+    if request.trim().is_empty() {
+        return messages.to_vec();
+    }
+    let mut out = messages.to_vec();
+    out.push(Message::user(format!(
+        "Continue the latest user request using the tool result above. Latest user request:\n{request}"
+    )));
+    out
+}
+
 /// Converts a recovered call into this crate's [`ToolCall`], minting a
 /// process-unique id.
 ///

crates/tinyinference-llm/src/prompt_tools/test.rs

@@ -166,6 +166,44 @@ fn user_turn_normalization_does_not_count_folded_tool_results() {
     assert_eq!(out[1].text(), CONTINUATION_USER_TURN);
 }
 
+#[test]
+fn tool_continuation_anchors_the_latest_request_after_search_results() {
+    let messages = coalesce_tool_results(&[
+        Message::system("system"),
+        Message::user("hey"),
+        Message::assistant("Hey! What's up?"),
+        Message::user("fetch my latest email"),
+        Message::assistant(""),
+        Message::tool("search-1", "GMAIL_FETCH_EMAILS schema"),
+    ]);
+    let anchored = anchor_user_request_after_tool_result(&messages);
+    assert_eq!(anchored.len(), messages.len() + 1);
+    assert!(
+        anchored[anchored.len() - 2]
+            .text()
+            .contains("GMAIL_FETCH_EMAILS")
+    );
+    assert!(
+        anchored
+            .last()
+            .unwrap()
+            .text()
+            .contains("fetch my latest email")
+    );
+    assert!(!anchored.last().unwrap().text().contains("Hey! What's up?"));
+    assert_eq!(anchor_user_request_after_tool_result(&anchored), anchored);
+}
+
+#[test]
+fn tool_continuation_without_a_real_user_request_stays_unchanged() {
+    let messages = coalesce_tool_results(&[
+        Message::system("system"),
+        Message::assistant("calling"),
+        Message::tool("call-1", "result"),
+    ]);
+    assert_eq!(anchor_user_request_after_tool_result(&messages), messages);
+}
+
 #[test]
 fn user_turn_normalization_ignores_blank_and_accepts_non_text_turns() {
     let out = ensure_resolvable_user_turn(&[Message::system("system"), Message::user("   ")]);

crates/tinyinference-llm/src/providers/openai/transport.rs

@@ -1141,8 +1141,9 @@ impl OpenAiModel {
             // into the text protocol before the protocol block is added.
             let coalesced = crate::prompt_tools::coalesce_tool_results(&request.messages);
             let resolvable = crate::prompt_tools::ensure_resolvable_user_turn(&coalesced);
+            let anchored = crate::prompt_tools::anchor_user_request_after_tool_result(&resolvable);
             instructed_messages = crate::prompt_tools::with_tool_instructions(
-                &resolvable,
+                &anchored,
                 &prompt_tool_schemas,
                 &request.tool_choice,
             );