tool-call: fix passing of tools to template + allow agent to finish

common/tool-call.h

@@ -21,7 +21,6 @@ struct llama_tool_call_handler {
     std::string grammar;
     std::vector<std::string> grammar_trigger_words;
     std::vector<std::string> additional_stop_words;
-    nlohmann::ordered_json updated_tools;
 };
 
 llama_tool_calls parse_tool_calls(llama_tool_call_style style, const nlohmann::ordered_json & tools, const std::string& input);

examples/agent/README.md

@@ -15,8 +15,7 @@
 - Run some tools inside a docker container (check http://localhost:8088/docs once running):
 
   ```bash
-  docker run -p 8088:8088 -w /src \
-    -v $PWD/examples/agent:/src \
+  docker run -p 8088:8088 -w /src -v $PWD/examples/agent:/src \
     --rm -it ghcr.io/astral-sh/uv:python3.12-alpine \
     uv run fastify.py --port 8088 tools.py
   ```

examples/agent/run.py

@@ -182,6 +182,7 @@ def main(
         else:
             assert content
             print(content)
+            return
 
         i += 1
 

examples/server/server.cpp

@@ -986,6 +986,7 @@ struct server_context {
 
         {
             slot.antiprompts.clear();
+            slot.sparams.grammar_trigger_words.clear();
 
             auto copy_string_array = [&](const json & data, const std::string & key, std::vector<std::string> & vec) {
                 const auto & arr = data.find(key);
@@ -1004,7 +1005,7 @@ struct server_context {
             copy_string_array(data, "grammar_trigger_words", slot.sparams.grammar_trigger_words);
 
             slot.antiprompts.build(ctx, stop_words, slot.sparams.grammar_trigger_words);
-            
+
         }
 
         {

examples/server/tests/features/tool_call.feature

@@ -31,11 +31,11 @@ Feature: llama.cpp server
     Examples: Prompts
       | template_name                         | n_predict | tool_name | tool_arguments         | tools |
       | meetkai-functionary-medium-v3.1       | 128       | test      | {}                     | [{"type":"function", "function": {"name": "test", "description": "", "parameters": {"type": "object", "properties": {}}}}]                                                                       |
-      | meetkai-functionary-medium-v3.1       | 128       | ipython   | {"code": "I'm sorry,"} | [{"type":"function", "function": {"name": "ipython", "description": "", "parameters": {"type": "object", "properties": {"code": {"type": "string", "description": ""}}, "required": ["code"]}}}] |
+      | meetkai-functionary-medium-v3.1       | 128       | ipython   | {"code": "Yes, you can."} | [{"type":"function", "function": {"name": "ipython", "description": "", "parameters": {"type": "object", "properties": {"code": {"type": "string", "description": ""}}, "required": ["code"]}}}] |
       | meetkai-functionary-medium-v3.2       | 128       | test      | {}                     | [{"type":"function", "function": {"name": "test", "description": "", "parameters": {"type": "object", "properties": {}}}}]                                                                       |
       | meetkai-functionary-medium-v3.2       | 128       | ipython   | {"code": "Yes,"}       | [{"type":"function", "function": {"name": "ipython", "description": "", "parameters": {"type": "object", "properties": {"code": {"type": "string", "description": ""}}, "required": ["code"]}}}] |
       | meta-llama-Meta-Llama-3.1-8B-Instruct | 64        | test      | {}                     | [{"type":"function", "function": {"name": "test", "description": "", "parameters": {"type": "object", "properties": {}}}}]                                                                       |
-      | meta-llama-Meta-Llama-3.1-8B-Instruct | 16        | ipython   | {"code": ". A"}    | [{"type":"function", "function": {"name": "ipython", "description": "", "parameters": {"type": "object", "properties": {"code": {"type": "string", "description": ""}}, "required": ["code"]}}}] |
+      | meta-llama-Meta-Llama-3.1-8B-Instruct | 16        | ipython   | {"code": "it and "}    | [{"type":"function", "function": {"name": "ipython", "description": "", "parameters": {"type": "object", "properties": {"code": {"type": "string", "description": ""}}, "required": ["code"]}}}] |
 
 
   Scenario Outline: OAI Compatibility w/ tools and auto tool_choice

examples/server/utils.hpp

@@ -384,9 +384,6 @@ static json oaicompat_completion_params_parse(
                 }
                 llama_params["grammar_trigger_words"] = triggers;
             }
-            if (handler.updated_tools.is_null()) {
-                tools = handler.updated_tools;
-            }
             if (!handler.grammar.empty()) {
                 if (llama_params.contains("grammar")) {
                     throw std::runtime_error("Cannot use custom grammar constraints with tools.");

src/llama-sampling.cpp

@@ -1371,7 +1371,7 @@ static struct llama_sampler_i llama_sampler_grammar_i = {
     /* .clone  = */ llama_sampler_grammar_clone,
     /* .free   = */ llama_sampler_grammar_free,
 };
- 
+
 bool llama_sampler_is_grammar_empty_impl(struct llama_sampler * gsmpl) {
     struct llama_sampler_grammar * ctx = (struct llama_sampler_grammar *) gsmpl->ctx;
     return ctx->grammar == nullptr;

src/llama.cpp

@@ -21312,7 +21312,7 @@ int32_t llama_chat_apply_template(
 struct llama_sampler * llama_sampler_init_grammar(const struct llama_model * model, const char * grammar_str, const char * grammar_root) {
     return llama_sampler_init_grammar_impl(model->vocab, grammar_str, grammar_root);
 }
- 
+
 bool llama_sampler_is_grammar_empty(struct llama_sampler * gsmpl) {
     return llama_sampler_is_grammar_empty_impl(gsmpl);
 }