End the classified section at an end-of-generation token in the prompt

llama-cpp-bindings-tests/src/classify_sample_loop.rs

@@ -123,7 +123,7 @@ mod tests {
 
         assert_eq!(outcome.observed_tool_call, 1);
         assert_eq!(outcome.generated_raw, "{");
-        assert!(outcome.content_stream.is_empty());
+        assert_eq!(outcome.content_stream, "");
     }
 
     #[test]
@@ -149,8 +149,8 @@ mod tests {
         });
 
         assert_eq!(outcome.observed_undeterminable, 1);
-        assert!(outcome.content_stream.is_empty());
-        assert!(outcome.reasoning_stream.is_empty());
+        assert_eq!(outcome.content_stream, "");
+        assert_eq!(outcome.reasoning_stream, "");
     }
 
     #[test]

llama-cpp-bindings-tests/tests/chat_protocol.rs

@@ -6,6 +6,7 @@ use llama_cpp_bindings::ChatTemplateError;
 use llama_cpp_bindings::ParseChatMessageError;
 use llama_cpp_bindings::model::LlamaChatMessage;
 use llama_cpp_bindings_tests::build_user_prompt_with_media_marker::build_user_prompt_with_media_marker;
+use llama_cpp_bindings_types::ParsedToolCall;
 use llama_cpp_test_harness::LlamaFixture;
 use llama_cpp_test_harness::llama_test;
 
@@ -254,7 +255,7 @@ fn parses_pure_content_response(fixture: &LlamaFixture<'_>) -> Result<()> {
     let ChatMessageParseOutcome::Recognized(parsed) = outcome else {
         bail!("expected Recognized for plain content; got Unrecognized");
     };
-    assert!(parsed.tool_calls.is_empty());
+    assert_eq!(parsed.tool_calls, Vec::<ParsedToolCall>::new());
     assert!(!parsed.is_empty());
     assert!(parsed.content.contains("hello world"));
 
@@ -348,7 +349,7 @@ fn parses_empty_input_yields_empty_message(fixture: &LlamaFixture<'_>) -> Result
     let ChatMessageParseOutcome::Recognized(parsed) = outcome else {
         bail!("expected Recognized for empty input; got Unrecognized");
     };
-    assert!(parsed.tool_calls.is_empty());
+    assert_eq!(parsed.tool_calls, Vec::<ParsedToolCall>::new());
 
     Ok(())
 }

llama-cpp-bindings-tests/tests/context_state.rs

@@ -232,7 +232,7 @@ fn decode_and_get_logits(fixture: &LlamaFixture<'_>) -> Result<()> {
     assert!(decode_result.is_ok());
 
     let logits = context.get_logits()?;
-    assert!(!logits.is_empty());
+    assert_ne!(logits, Vec::<f32>::new());
 
     Ok(())
 }
@@ -282,7 +282,7 @@ fn token_data_array_has_entries_after_decode(fixture: &LlamaFixture<'_>) -> Resu
 
     let token_data_array = context.token_data_array()?;
 
-    assert!(!token_data_array.data.is_empty());
+    assert_ne!(token_data_array.data, Vec::<LlamaTokenData>::new());
 
     Ok(())
 }

llama-cpp-bindings-tests/tests/embedding_models.rs

@@ -597,8 +597,8 @@ fn embedding_model_exposes_tool_call_markers(fixture: &LlamaFixture<'_>) -> Resu
         .tool_call_markers()?
         .expect("Qwen3 embedding model must expose tool-call markers");
 
-    assert!(!markers.open.is_empty());
-    assert!(!markers.close.is_empty());
+    assert_ne!(markers.open, "");
+    assert_ne!(markers.close, "");
 
     Ok(())
 }

llama-cpp-bindings-tests/tests/model_introspection.rs

@@ -464,7 +464,7 @@ fn meta_count_returns_positive(fixture: &LlamaFixture<'_>) -> Result<()> {
 )]
 fn meta_key_by_index_returns_valid_key(fixture: &LlamaFixture<'_>) -> Result<()> {
     let key = fixture.model.meta_key_by_index(0)?;
-    assert!(!key.is_empty());
+    assert_ne!(key, "");
     Ok(())
 }
 
@@ -502,7 +502,7 @@ fn meta_key_by_index_returns_valid_key(fixture: &LlamaFixture<'_>) -> Result<()>
 )]
 fn meta_val_str_by_index_returns_valid_value(fixture: &LlamaFixture<'_>) -> Result<()> {
     let value = fixture.model.meta_val_str_by_index(0)?;
-    assert!(!value.is_empty());
+    assert_ne!(value, "");
     Ok(())
 }
 
@@ -618,7 +618,7 @@ fn meta_val_str_returns_value_for_known_key(fixture: &LlamaFixture<'_>) -> Resul
     let model = fixture.model;
     let first_key = model.meta_key_by_index(0)?;
     let value = model.meta_val_str(&first_key)?;
-    assert!(!value.is_empty());
+    assert_ne!(value, "");
     Ok(())
 }
 
@@ -1184,12 +1184,12 @@ fn token_attr_returns_attrs_for_bos(fixture: &LlamaFixture<'_>) -> Result<()> {
 fn str_to_token_roundtrip(fixture: &LlamaFixture<'_>) -> Result<()> {
     let model = fixture.model;
     let tokens = model.str_to_token("hello world", AddBos::Never, ParseSpecialTokens::Always)?;
-    assert!(!tokens.is_empty());
+    assert_ne!(tokens, Vec::<LlamaToken>::new());
     let mut decoder = encoding_rs::UTF_8.new_decoder();
     let piece =
         model.token_to_piece(&SampledToken::Content(tokens[0]), &mut decoder, false, None)?;
 
-    assert!(!piece.is_empty());
+    assert_ne!(piece, "");
 
     Ok(())
 }
@@ -1375,7 +1375,7 @@ fn token_to_piece_bytes_returns_bytes_for_known_token(fixture: &LlamaFixture<'_>
     let tokens = model.str_to_token("hello", AddBos::Never, ParseSpecialTokens::Always)?;
     let bytes = model.token_to_piece_bytes(tokens[0], 32, false, None)?;
 
-    assert!(!bytes.is_empty());
+    assert_ne!(bytes, Vec::<u8>::new());
     Ok(())
 }
 
@@ -1563,7 +1563,7 @@ fn token_to_piece_decodes_reasoning_variant(fixture: &LlamaFixture<'_>) -> Resul
         None,
     )?;
 
-    assert!(!piece.is_empty());
+    assert_ne!(piece, "");
     Ok(())
 }
 
@@ -1607,7 +1607,7 @@ fn token_to_piece_decodes_tool_call_variant(fixture: &LlamaFixture<'_>) -> Resul
     let piece =
         model.token_to_piece(&SampledToken::ToolCall(tokens[0]), &mut decoder, true, None)?;
 
-    assert!(!piece.is_empty());
+    assert_ne!(piece, "");
     Ok(())
 }
 
@@ -1655,7 +1655,7 @@ fn token_to_piece_decodes_undeterminable_variant(fixture: &LlamaFixture<'_>) ->
         None,
     )?;
 
-    assert!(!piece.is_empty());
+    assert_ne!(piece, "");
     Ok(())
 }
 

llama-cpp-bindings-tests/tests/multimodal_vision.rs

@@ -18,6 +18,7 @@ use llama_cpp_bindings::mtmd::MtmdInputText;
 use llama_cpp_bindings::mtmd::NonCausalChunkMicroBatchMismatch;
 use llama_cpp_bindings::mtmd::mtmd_default_marker;
 use llama_cpp_bindings::sampling::LlamaSampler;
+use llama_cpp_bindings::token::LlamaToken;
 use llama_cpp_bindings_tests::build_user_prompt_with_media_marker::build_user_prompt_with_media_marker;
 use llama_cpp_bindings_tests::chunk_token_breakdown::ChunkTokenBreakdown;
 use llama_cpp_bindings_tests::classify_sample_loop::ClassifySampleLoop;
@@ -290,7 +291,7 @@ fn text_chunk_returns_text_tokens(fixture: &LlamaFixture<'_>) -> Result<()> {
         .text_tokens_of(first_chunk.chunk_type()?)
         .ok_or_else(|| anyhow::anyhow!("the first chunk of a text prompt must carry tokens"))?;
 
-    assert!(!tokens.is_empty());
+    assert_ne!(tokens, Vec::<LlamaToken>::new());
 
     Ok(())
 }

llama-cpp-bindings-tests/tests/structured_chat_output.rs

@@ -996,7 +996,7 @@ What is 2 + 2?
 
     let usage = classifier.usage();
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert_eq!(outcome.observed_reasoning, 0);
     assert_eq!(outcome.observed_undeterminable, 0);
     assert_eq!(usage.reasoning_tokens, 0);
@@ -1080,7 +1080,7 @@ What is 2 + 2?
     let usage = classifier.usage();
     let parsed = parse_partial_reasoning_response(model, &outcome.generated_raw)?;
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert!(outcome.observed_reasoning > 0);
     assert!(usage.reasoning_tokens > 0);
     assert_eq!(outcome.observed_undeterminable, 0);
@@ -1253,7 +1253,7 @@ fn mistral3_classifier_does_not_emit_reasoning_for_thinking_disabled_prompt(
 
     let usage = classifier.usage();
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert_eq!(outcome.observed_reasoning, 0);
     assert_eq!(outcome.observed_undeterminable, 0);
     assert_eq!(usage.reasoning_tokens, 0);
@@ -1334,7 +1334,7 @@ to the user.[/THINK]Here, provide a self-contained response.[/SYSTEM_PROMPT]\
     let usage = classifier.usage();
     let parsed = parse_partial_reasoning_response(model, &outcome.generated_raw)?;
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert!(outcome.observed_reasoning > 0);
     assert!(usage.reasoning_tokens > 0);
     assert_eq!(outcome.observed_undeterminable, 0);
@@ -1343,7 +1343,7 @@ to the user.[/THINK]Here, provide a self-contained response.[/SYSTEM_PROMPT]\
         usage.completion_tokens(),
         outcome.observed_content + outcome.observed_reasoning,
     );
-    assert!(!parsed.reasoning_content.is_empty());
+    assert_ne!(parsed.reasoning_content, "");
     assert_eq!(outcome.reasoning_stream, parsed.reasoning_content);
     assert_eq!(outcome.content_stream, parsed.content);
 
@@ -1459,7 +1459,7 @@ fn qwen35_chat_inference_emits_reasoning_when_template_auto_opens(
     }
     .run()?;
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert!(outcome.observed_reasoning > 0);
     assert_eq!(outcome.observed_undeterminable, 0);
     assert_eq!(outcome.observed_tool_call, 0);
@@ -1528,6 +1528,35 @@ fn qwen35_shared_reasoning_close_and_tool_call_open_is_one_transition(
     Ok(())
 }
 
+#[llama_test(
+    model_source = HuggingFace("unsloth/Ministral-3-3B-Reasoning-2512-GGUF", "Ministral-3-3B-Reasoning-2512-Q4_K_M.gguf"),
+    n_gpu_layers = 999,
+    load_mode = Mmap,
+    n_ctx = 2048,
+    n_batch = 512,
+    n_ubatch = 128,
+)]
+fn mistral3_tool_call_turn_finished_in_the_prompt_leaves_the_classifier_pending(
+    fixture: &LlamaFixture<'_>,
+) -> Result<()> {
+    let model = fixture.model;
+    let mut finished_tool_call_turn = model.str_to_token(
+        "[TOOL_CALLS]read[ARGS]{\"path\":\"marker.txt\"}",
+        AddBos::Never,
+        ParseSpecialTokens::Always,
+    )?;
+
+    finished_tool_call_turn.push(model.token_eos().context("model declares an EOS token")?);
+
+    let mut classifier = model.sampled_token_classifier(BareJsonToolCalls::Detect)?;
+
+    classifier.ingest_prompt_tokens(&finished_tool_call_turn);
+
+    assert_eq!(classifier.current_section(), SampledTokenSection::Pending);
+
+    Ok(())
+}
+
 #[llama_test(
     model_source = HuggingFace("unsloth/Qwen3.5-0.8B-GGUF", "Qwen3.5-0.8B-Q4_K_M.gguf"),
     n_gpu_layers = 999,
@@ -1599,7 +1628,7 @@ What is 2 + 2?<|im_end|>
 
     let usage = classifier.usage();
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert_eq!(outcome.observed_reasoning, 0);
     assert_eq!(outcome.observed_undeterminable, 0);
     assert_eq!(usage.reasoning_tokens, 0);
@@ -1683,7 +1712,7 @@ What is 2 + 2?<|im_end|>
     let usage = classifier.usage();
     let parsed = parse_partial_reasoning_response(model, &outcome.generated_raw)?;
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert!(outcome.observed_reasoning > 0);
     assert!(usage.reasoning_tokens > 0);
     assert_eq!(outcome.observed_undeterminable, 0);
@@ -2048,14 +2077,14 @@ fn qwen36_chat_inference_emits_reasoning_when_template_auto_opens(
     }
     .run()?;
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert!(outcome.observed_reasoning > 0);
     assert!(outcome.observed_content > 0);
     assert_eq!(outcome.observed_undeterminable, 0);
     assert_eq!(outcome.observed_tool_call, 0);
 
     let parsed = parse_partial_reasoning_response(model, &outcome.generated_raw)?;
-    assert!(!parsed.content.is_empty());
+    assert_ne!(parsed.content, "");
 
     let usage = classifier.into_usage();
     assert_eq!(usage.prompt_tokens, prompt_token_count);
@@ -2136,7 +2165,7 @@ What is 2 + 2?<|im_end|>
 
     let usage = classifier.usage();
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert_eq!(outcome.observed_reasoning, 0);
     assert_eq!(outcome.observed_undeterminable, 0);
     assert_eq!(usage.reasoning_tokens, 0);
@@ -2220,7 +2249,7 @@ What is 2 + 2?<|im_end|>
     let usage = classifier.usage();
     let parsed = parse_partial_reasoning_response(model, &outcome.generated_raw)?;
 
-    assert!(!outcome.generated_raw.is_empty());
+    assert_ne!(outcome.generated_raw, "");
     assert!(outcome.observed_reasoning > 0);
     assert!(usage.reasoning_tokens > 0);
     assert_eq!(outcome.observed_undeterminable, 0);
@@ -2230,7 +2259,7 @@ What is 2 + 2?<|im_end|>
         outcome.observed_content + outcome.observed_reasoning,
     );
 
-    assert!(!parsed.reasoning_content.is_empty());
+    assert_ne!(parsed.reasoning_content, "");
     assert_eq!(outcome.reasoning_stream, parsed.reasoning_content);
     assert_eq!(outcome.content_stream, parsed.content);
 

llama-cpp-bindings-types/src/parsed_tool_call.rs

@@ -46,8 +46,8 @@ mod tests {
     fn default_is_empty_strings_and_invalid_arguments() {
         let parsed = ParsedToolCall::default();
 
-        assert!(parsed.id.is_empty());
-        assert!(parsed.name.is_empty());
+        assert_eq!(parsed.id, "");
+        assert_eq!(parsed.name, "");
         assert_eq!(
             parsed.arguments,
             ToolCallArguments::InvalidJson(String::new())

llama-cpp-bindings/src/chat_message_parse_outcome.rs

@@ -11,6 +11,7 @@ pub enum ChatMessageParseOutcome {
 #[cfg(test)]
 mod tests {
     use llama_cpp_bindings_types::ParsedChatMessage;
+    use llama_cpp_bindings_types::ParsedToolCall;
 
     use super::ChatMessageParseOutcome;
     use crate::raw_chat_message::RawChatMessage;
@@ -37,7 +38,7 @@ mod tests {
                 ChatMessageParseOutcome::Recognized(parsed) => {
                     assert_eq!(parsed.content, "content");
                     assert_eq!(parsed.reasoning_content, "reasoning");
-                    assert!(parsed.tool_calls.is_empty());
+                    assert_eq!(parsed.tool_calls, Vec::<ParsedToolCall>::new());
                     saw_recognized = true;
                 }
                 ChatMessageParseOutcome::Unrecognized(raw) => {